diff --git a/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/f5964ab7-5331-420f-bb50-825216bb8b90_content_list.json b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/f5964ab7-5331-420f-bb50-825216bb8b90_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c12b6f22afb8ddef6230719a11c6ffc5460aab73 --- /dev/null +++ b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/f5964ab7-5331-420f-bb50-825216bb8b90_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45e547aec2bf075294722342c28bbb89438b0591bb9d8bc97c3f8c966906cc73 +size 91380 diff --git a/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/f5964ab7-5331-420f-bb50-825216bb8b90_model.json b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/f5964ab7-5331-420f-bb50-825216bb8b90_model.json new file mode 100644 index 0000000000000000000000000000000000000000..336dc6ffac7793a1c19091cdffc47df261076479 --- /dev/null +++ b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/f5964ab7-5331-420f-bb50-825216bb8b90_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:519bc4c3fb1695c35af18c78daf25100ced05b4050e6fc561ba9d9f0e574dba1 +size 118369 diff --git a/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/f5964ab7-5331-420f-bb50-825216bb8b90_origin.pdf b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/f5964ab7-5331-420f-bb50-825216bb8b90_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..817cd977213d8bab011990bea056366874880479 --- /dev/null +++ b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/f5964ab7-5331-420f-bb50-825216bb8b90_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90bcc5cb30dcedc99cb8f404735873373256ddffb782247e1a1e32129fe0e046 +size 1334321 diff --git a/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/full.md b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5ef823aa3ef833c10d2ffb95532965814d5c9414 --- /dev/null +++ b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/full.md @@ -0,0 +1,346 @@ +# egoPPG: Heart Rate Estimation from Eye-Tracking Cameras in Egocentric Systems to Benefit Downstream Vision Tasks + +Björn Braun, Rayan Armani, Manuel Meier, Max Moebus, and Christian Holz +Department of Computer Science, ETH Zürich, Switzerland + +https://siplab.org/projects/egoPPG + +![](images/1724e17bb27953ade15bcb51b2ff24ea5f2693970c30834cc7acc3dd41321650.jpg) +Figure 1. We propose egoPPG as a novel computer vision task: tracking a person's heart rate (HR) on unmodified egocentric vision headsets. Taking eye tracking videos as input, our method PulseFormer estimates the photoplethysmogram (PPG) from areas around the eyes to derive HR values. For training and validation, we introduce egoPPG-DB, a dataset of eye tracking videos while participants performed everyday activities with synchronized ground-truth PPG (via nose-based contact sensor) and HR values (via ECG chest strap). + +![](images/c824c6fec23be4a1bff6b2fc2a8a5c585ead81e207f2b5445d4e78e02dbd6c81.jpg) + +# Abstract + +Egocentric vision systems aim to understand the spatial surroundings and the wearer's behavior inside it, including motions, activities, and interactions. We argue that egocentric systems must additionally detect physiological states to capture a person's attention and situational responses, which are critical for context-aware behavior modeling. In this paper, we propose egoPPG, a novel vision task for egocentric systems to recover a person's cardiac activity to aid downstream vision tasks. We introduce PulseFormer, a method to extract heart rate as a key indicator of physiological state from the eye tracking cameras on unmodified egocentric vision systems. PulseFormer continuously estimates the photoplethysmogram (PPG) from areas around the eyes and fuses motion cues from the headset's inertial measurement unit to track HR values. We demonstrate egoPPG's downstream benefit for a key task on EgoExo4D, an existing egocentric dataset for which we find PulseFormer's estimates of HR to improve proficiency estimation by $14\%$ . To train and validate PulseFormer, we + +collected a dataset of $13+$ hours of eye tracking videos from Project Aria and contact-based PPG signals as well as an electrocardiogram (ECG) for ground-truth HR values. Similar to EgoExo4D, 25 participants performed diverse everyday activities such as office work, cooking, dancing, and exercising, which induced significant natural motion and HR variation (44-164 bpm). Our model robustly estimates HR $(\mathrm{MAE} = 7.67\mathrm{bpm})$ and captures patterns $(r = 0.85)$ . Our results show how egocentric systems may unify environmental and physiological tracking to better understand users and that egoPPG as a complementary task provides meaningful augmentations for existing datasets and tasks. We release our code, dataset, and HR augmentations for EgoExo4D to inspire research on physiology-aware egocentric tasks. + +# 1. Introduction + +Egocentric vision systems, such as Mixed Reality (MR) glasses by Meta [56], Magic Leap [43], and others have emerged as powerful devices for capturing and analyzing a person's behavior and their environment from a first-person + +perspective. The wider availability of promising wearable capture platforms has sparked a large amount of research on egocentric vision tasks for environment understanding and navigation [18], including localization [39, 71, 74], and simultaneous localization and mapping [15, 33, 68]. Since egocentric systems simultaneously capture parts of the wearer's behavior in addition to their environment, prior work has investigated egocentric action recognition [51, 87, 89, 96] and hand-object interaction [26, 73, 97] to understand user behavior. Several large-scale datasets now accelerate data-driven research in this domain with multimodal data for training and evaluation (e.g., Ego4D [26], Nymeria [50], EgoExo4D [27]). + +Most recently, Meta's Project Aria 2 introduced a contact-based heart rate (HR) sensor, with which egocentric systems can gauge the wearer's cognitive performance, attention, and situational responses [13, 20, 52, 76, 83]. Numerous additional conditions manifest in a person's HR, such as emotions, stress and fatigue [1, 9, 59, 65, 80]—capturing these dynamics can thus benefit models of human behavior to enable a richer understanding of user behavior. + +In this paper, we introduce a method to make such HR estimates available to many existing egocentric systems and already recorded large datasets, such as EgoExo4D [27] or Nymeria [50]. Our method PulseFormer accurately recovers a person's HR from the eye tracking videos in egocentric headsets. PulseFormer first estimates the person's photoplethysmogram (PPG) from the subtle fluctuations in skin intensity due to pulsatile artery expansion beneath the surface following a blood volume pulse (BVP), in particular deriving it from regions around the wearer's eye for robust tracking. Our spatial attention module ensures that PPG is estimated from robust regions around the eye, while our cross-attention fusion with the system's inertial measurement unit (IMU) learns a motion-informed temporal attention to optimally weight the eye tracking images for more accurate PPG estimates in scenarios with heavy motion. + +We validate our method's efficacy on a novel dataset that we collected to capture some of the activities included in large-scale egocentric datasets alongside physiological reference recordings. Our dataset egoPPG-DB contains 13 hours of recordings from 25 participants, who wore Project Aria glasses and performed six real-world tasks with varying motion and intensity, causing their HR values to reach levels between 44–164 bpm. + +# Downstream benefits for egocentric vision tasks + +A key contribution of our paper is that we demonstrate that knowing a person's continuous HR values benefits egocentric vision tasks downstream. We augment an existing architecture with PulseFormer's HR estimates and show its impact on EgoExo4D's proficiency estimation benchmark, which improves accuracy on this task by $14.1\%$ . + +# Contributions + +We summarize our key contributions as follows: + +1. egoPPG as a novel task and PulseFormer as an HR estimation method for egocentric systems that operates on eye tracking videos. Our method robustly predicts continuous HR across a series of activities and interactions (MAE=7.67 bpm), with a $23.8\%$ lower error than current state-of-the-art rPPG models [10, 45, 90, 92, 93]. + +2. egoPPG-DB, a dataset of eye tracking videos and synchronized BVP (contact-based) and ECG recordings (chest strap-based) to verify all physiological signals. We captured these across diverse everyday activities that were inspired by those included in existing large-scale egocentric datasets, such as EgoExo4D [26, 27, 50]. + +3. a validation of egoPPG's downstream benefits for egocentric vision tasks. We demonstrate the implications of our method PulseFormer on the proficiency estimation benchmark of the EgoExo4D dataset, which increases the accuracy by $14.1\%$ when augmenting EgoExo4D with our continuously predicted HR values. + +# 2. Related work + +Egocentric vision. In recent years, research in egocentric vision has surged, driven by advances in AR/VR glasses [4, 18, 31, 43, 56, 57], which provide new ways for understanding user interaction from a first-person perspective. Much of this work has focused on tasks such as action recognition [41, 51, 87, 89, 96] and anticipation [14, 25, 87], full-body pose estimation [36, 37, 75], responding to user needs [67, 69, 91], and social behavior analysis [21, 26, 35]. Additionally, tracking vital signs in AR/VR settings and for affective computing applications [1, 9, 59, 65, 80] has become an important tool for understanding users' physiological states [53], their behavior, attention, and intent [3, 58, 88]. + +Physiological measurements. Wearable sensors have had a tremendous impact on health monitoring in recent years, enabling continuous measurement of key physiological metrics, such as heart rate (HR), oxygen saturation, and activity levels [11, 17, 55, 61, 62]. HR, in particular, is a key measure for assessing an individual's health and performance [19, 23, 40, 48, 72]. In parallel to wearable sensors such as smartwatches, recent research has extensively explored using cameras as an unobtrusive, non-contact alternative for measuring HR, generally called remote photoplethysmography (rPPG) [34, 66, 84]. rPPG measures HR based on the BVP via subtle color changes of the skin. Generally, rPPG methods can be broadly divided into traditional signal processing techniques [7, 16, 34, 66, 85] and deep learning-based approaches [8, 10, 45, 78, 90, 92]. So far, rPPG has been mostly applied to facial videos with the camera and user being stationary, such as while sitting in front of a laptop, as it requires a continuous video feed + +of the same skin region. This limitation is shown in current rPPG datasets, which primarily capture individuals in seated positions with either a stationary camera directed at their face [6, 29, 70, 79] or requiring users to hold a smartphone steadily in front of their face [82]. As a result, rPPG is not feasible to be deployed in more dynamic settings. + +Eye tracking cameras. Eye tracking in egocentric vision systems is mostly done using inward-facing cameras directed at the eyes [2]. Even during motion, eye tracking in VR devices demonstrated accurate performance showcasing that the cameras remain almost stationary relative to the user's eyes [12]. Furthermore, IR illumination makes them robust to lighting variations and low-light conditions [49]. To the best of our knowledge, videos from eye tracking cameras have not yet been explored for HR estimation. + +# 3. Overview + +Our aim is to enable egocentric vision systems i) to model a person's physiological state via continuously estimated HR and ii) to integrate these HR estimates into downstream tasks that benefit from knowledge of the user's state. Sec. 4 first describes our dataset of synchronized eye tracking videos and ground-truth HR measurements. Sec. 5 introduces our method PulseFormer for recovering continuous HR from eye tracking videos. Sec. 6 outlines the downstream benefits of our novel task, using HR as input for modeling user proficiency on the EgoExo4D dataset. + +# 4. egoPPG-DB + +The egoPPG-DB dataset was developed to support HR estimation from eye tracking videos under real-world conditions and contains significant motion and HR fluctuations. By including diverse everyday activities, we provide a challenging benchmark for egocentric HR estimation models. + +# 4.1. Recruiting and recording + +We recruited $N = 25$ participants (12 female, 13 male, ages 19-32, $\mu = 25.1$ and $\sigma = 3.3$ ) on a voluntary basis, resulting in over 13 hours of video recordings. Based on the Fitzpatrick scale [22], 9 participants had skin type II, 8 had skin type III, 3 had skin type IV, and 5 had skin type V. All participants signed a consent form before the data collection, agreeing with using and sharing their data for academic and non-commercial purposes. The data collection was approved by the ETH Zurich Ethics Commission (no. 2023-N-08). In terms of duration, egoPPG-DB is among the longest rPPG datasets as listed in Tab. 10 (Supplementary). Participants of egoPPG-DB are not included in EgoExo4D. + +# 4.2. Apparatus + +Fig. 2 illustrates our experimental setup. We used Project Aria glasses [18] with Profile 21 to record eye tracking + +![](images/7ac2c332063a6493be3c1bd908f83cd100fe34196b696642ccc219b219b5d8c1.jpg) +Figure 2. Apparatus used to record the egoPPG-DB dataset. + +videos at 30 fps with a resolution of $320 \times 240$ pixels per eye. To capture ground truth PPG measurements, with which we train our model, we developed a custom sensor that records PPG data offline at $128\mathrm{Hz}$ . The sensor consists of a main board, mounted on the left side of the frame, featuring a DA14695 system-on-chip interfacing with a MAX86141ENP+ PPG sensor. The LEDs and photodiodes used by the PPG sensor are embedded in the left nose pad and connected to the main board using a flat flexible cable. For each participant, we individually adjusted the nose pad position to ensure the sensor aligned with their left angular artery [30]. To validate our custom PPG sensor, we also recorded gold-standard ECG data using a movisens ECGMove 4 chest belt sampling at $1024\mathrm{Hz}$ . We synchronized all devices at the start and end of each recording with a synchronization pattern, using their built-in IMUs. + +# 4.3. Capture protocol + +The average recording lasted 32 minutes. The capture protocol comprised 5 activities (Tab. 1): watching a video, office work, kitchen work, dancing, and exercising on an indoor bike (Fig. 1). We included these activities for three purposes: (1) Incorporate everyday activities including the corresponding HR changes and motion artifacts; (2) cover a wide range of HR values (low HR when watching a video vs. high HR when exercising), and (3) resemble activities that were captured in large-scale egocentric vision datasets, such as EgoExo4D [27]. In Tab. 11 (Supplementary), we give a detailed description of all activities. Tab. 3 shows mean HR values for each activity. Exercising on the bike produced the highest mean HR values (113 bpm), whereas watching the video resulted in the lowest (71 bpm). + +# 4.4. Dataset and signal quality verification + +To ensure the contact PPG sensor, whose signal we later use as the target for model training, produces accurate HR + +
ActivityActionsMinutes
Watch videoWatch a documentary5
Work on a computer4
Office workWrite on a paper2
Talk to the experimenter2
WalkingWalk to the kitchen1
Cut vegetables
Kitchen workPrepare a sandwich5
Wash the dishes
WalkingWalk to the dancing room1.5
DancingFollow random dance video5
Exercise bikeRide an exercise bike5
WalkingWalk back to the start1.5
+ +Table 1. Capture protocol for recording the egoPPG-DB dataset. + +values, we evaluate it against the gold-standard ECG. We calculated the MAE and Pearson correlation between HR estimates from the ECG and PPG signals for each participant using a 30-second sliding window. For activity labeling, we manually annotated the start and end times of each task (see Tab. 1) for each participant using the Point of View (POV) RGB videos recorded by the Project Aria glasses. To ensure that the signal quality of the contact PPG is sufficient for model training, we excluded all tasks with an MAE over $3.0\mathrm{~bpm}$ between the PPG and ECG (e.g. when the PPG sensor moved). This applied to 20 of 150 tasks ( $13\%$ , see Tab. 6 in Supplementary). During the remaining tasks, our custom-built PPG nose sensor achieved very high accuracy, with an MAE of $1.3\mathrm{~bpm}$ and a correlation of 0.94 compared to the ECG signal, showing its suitability as ground truth. + +# 5. PulseFormer method + +# 5.1. Problem definition + +Our objective is to estimate BVP and HR from periodic changes in pixel intensity in eye tracking video frames $\pmb{F} \in \mathbb{R}^{w \times h}$ . Physically, this means extracting physiological signals from the information in the light reflected by the arteries and arterioles that carry blood beneath the skin. This light reflection can be modeled as a combination of diffuse and specular reflections. Wang et al. [85] model the reflected light intensity $C(t)$ as: + +$$ +\boldsymbol {C} (\boldsymbol {t}) = I (t) \left(\boldsymbol {v} _ {\boldsymbol {s}} (t) + \boldsymbol {v} _ {\boldsymbol {d}} (t)\right) + \boldsymbol {v} _ {\boldsymbol {n}} (t) \tag {1} +$$ + +where $I(t)$ is the luminance intensity, $\pmb{v}_{s}(t)$ the specular reflection, $\pmb{v}_{d}(t)$ the diffuse reflection, and $\pmb{v}_{n}(t)$ the sensor noise. While the specular reflection $\pmb{v}_{s}(t)$ lacks pulsatile information, the diffuse reflection $\pmb{v}_{d}(t)$ contains information about the absorption and scattering of the light in skin + +tissue [85]. Thus, $\pmb{v}_d(t)$ can be further decomposed as: + +$$ +\boldsymbol {v} _ {\boldsymbol {d}} (t) = \boldsymbol {u} _ {\boldsymbol {d}} d _ {0} + \boldsymbol {u} _ {\boldsymbol {p}} p (t) \tag {2} +$$ + +where $\pmb{u}_d$ is the unit color vector of the skin, $d_0$ the stationary reflection strength, $\pmb{u_p}$ the relative absorption, and $p(t)$ the signals of interest. $p(t)$ is in our case the BVP, which our model aims to learn from the camera recordings. + +# 5.2. Deep learning model + +Our architecture is built upon a 3D CNN backbone (PhysNet) [92] with a temporal input length of $T = 128$ frames (corresponding to 4.3 seconds) downsampled to $(h = 48) \times (w = 128)$ pixels, resulting in an input of dimensions $(T, C, w, h)$ . The channel is $C = 1$ in our case, as our input is from monochrome videos. The input in our network is the consecutive standardized frame differences (per participant frame-wise differences divided by the STD of the frames) of the eye tracking videos to help the network focus on the changes between frames [10]. As labels, we use the standardized consecutive differences of the PPG signals. Since eye tracking videos offer additional challenges compared to facial videos, usually used for rPPG tasks, we have designed our model to address these challenges (see Fig. 3). + +Motion-informed temporal attention (MITA). Egocentric glasses are body-worn and subject to considerable motion artifacts when the user moves. Therefore, we propose to leverage the IMU within the glasses to obtain a motion-informed temporal attention. We employ a cross-attention module to integrate the IMU data with the video input, allowing our model to weigh each frame differently along the temporal dimension based on the motion intensity encoded by the IMU. This allows the model, e.g., to give less emphasis to frames heavily affected by motion. Given the input feature map $\mathbf{F}_{\mathrm{in}} \in \mathbb{R}^{T \times 1 \times w \times h}$ , we use ResNet18 [28] and a linear layer to obtain the image embeddings $\mathbf{F}_e \in \mathbb{R}^{T \times D}$ where $D = 128$ is the embedding dimension. Given IMU measurements $\mathbf{I}_{\mathrm{in}} \in \mathbb{R}^{T \times 1}$ , we use two 1D convolutional layers to obtain the IMU embeddings $\mathbf{I}_e \in \mathbb{R}^{T \times D}$ . We then calculate the cross-attention $\mathbf{A} \in \mathbb{R}^{T \times D}$ as: + +$$ +\mathbf {A} = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q K} ^ {\top}}{\sqrt {D}}\right) \mathbf {V} \tag {3} +$$ + +where $\mathbf{I}_{\mathrm{e}}$ serve as queries $Q$ and $\mathbf{F}_{\mathrm{e}}$ as keys $K$ and values $V$ . Using a linear layer, we obtain the motion-informed temporal attention $\mathbf{T} \in \mathbb{R}^{T \times 1 \times 1 \times 1}$ , which we multiply with $F_{in}$ . Spatial attention (SA). While the bulbar conjunctiva (white of the eyes) contains many blood vessels from which the BVP could theoretically be estimated, eyes typically move strongly during everyday situations and are closed while blinking (see participant 2 in Fig. 4). Consequently, extracting the BVP from the eye regions would introduce substantial motion artifacts and reduce the signal-to-noise + +![](images/214d747cf48fc2a92a922764a114d30dd88a84b7d31c0f750054f112ec60bcda.jpg) +Figure 3. Architecture of our model for continuous BVP estimation from eye tracking videos and consecutive HR computation. + +ratio (SNR). In contrast, when qualitatively analyzing eye tracking images, we see that the skin around the eyes exhibits considerably less motion than the eyes themselves and could thus provide a more stable source of BVP information. To address this, we introduce spatial attention modules [32, 64, 86] before each pooling (see Fig. 3) to allow our network to focus on high-SNR regions, such as the skin, and reduce the influence of low-SNR regions with frequent motion, like the eyes. Given some feature map $\pmb{F} \in \mathbb{R}^{T \times C \times w \times h}$ , the spatial attention modules infer a spatial attention map $M_{s} \in \mathbb{R}^{T \times 1 \times w \times h}$ as: + +$$ +M _ {s} (\boldsymbol {F}) = \sigma * (f ^ {7 \times 7} ([ \boldsymbol {F} _ {a v g}; \boldsymbol {F} _ {m a x} ])) \qquad (4) +$$ + +where $\sigma$ is the sigmoid function, $f^{7\times 7}$ a $7\times 7$ convolution operation and $F_{avg}\in \mathbb{R}^{T\times 1\times w\times h}$ and $F_{max}\in \mathbb{R}^{T\times 1\times w\times h}$ are the average-pooled and max-pooled feature maps respectively. The final output $F_{out}\in \mathbb{R}^{T\times C\times w\times h}$ of each attention process is then the product of $M_{s}$ and $F$ . + +Data augmentation. Furthermore, individual variations in the fit of the glasses result in different parts of the skin around the eyes being visible. For some individuals, the eye tracking cameras capture only the areas above the eyes, for others, only below, and in some cases, the glasses sit at an incline (see Fig. 4). To account for such variations, we apply three targeted data augmentations during training that reflect these specific differences in camera angles and coverage: (1) random rotation between -20 and +20 degrees to account for slight inclinations; (2) random horizontal cropping to help the network distinguish between high and low SNR regions across various skin areas and camera positions; and (3) horizontal and vertical flipping to further increase robustness to differences in skin region visibility. Our model requires approximately 399 GFLOPS per batch and has about 12M parameters. The frame rate is 2.9k fps on an RTX 4090 and 180 fps on an AMD EPYC CPU. + +# 5.3. Experiments setup + +Training. We trained all models using five-fold cross-validation split by participants to ensure a strict separation + +![](images/f6d0a910c278ba737a6985df2b7b4e1a99484842a4fbb3b64b5c57fbd2575ff1.jpg) + +![](images/bfc2ee38dde9f796edbb9b2b4694fdf8cef874b1432f78d7367b8187a063b113.jpg) + +![](images/5f67eeb752fc6c48a4757fdce41297ac772e221d5681760856c2cc0fde0a378a.jpg) +Figure 4. Left: Head geometry determines the regions that the eye tracker captures. Right: Learned spatial attention maps show that eye regions are excluded and PulseFormer instead extracts BVP from the surrounding skin regions, which moves less than the eyes. + +![](images/0f284d013faa6627b596f3cf454a5f32b804828e9413da1f886aa443973aee72.jpg) + +![](images/5def554dd65694728f438df85033d819b23949efb22d263584d4c3dc9062be10.jpg) + +![](images/48418df0a743d2836e0283a0c440a0921d88df1887088e537c2c36bafdc2819c.jpg) + +between training, validation, and test sets. We iteratively held out the data from five participants (20%) as the test set, two as validation, and trained on the remaining with a batch size of 4 for 100 epochs, a learning rate of 0.0009, and mean squared error (MSE) as loss. In addition to our model, we used ten state-of-the-art rPPG baseline networks to compare the performance of our proposed model to these established models. Our model was trained on a GeForce RTX 4090, with a total runtime of about 20 hours for all folds. + +Evaluation. To calculate the HR, we filter the predicted BVP with a Butterworth filter $(0.5 - 2.8\mathrm{Hz})$ and then detect peaks. To assess model accuracy, we use the mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and Pearson correlation (r) using a non-overlapping 60-second window [16, 45, 46]. + +Video sampling rate. While we recorded the eye tracking videos with 30 fps, large-scale datasets such as EgoExo4D [27] or Nymeria [50] used only 10 fps. To assess the impact of reduced fps, we evaluated model performance when (1) downsampling our videos to 10 fps by retaining only every third frame and (2) downsampling to 10 fps, then + +linearly interpolating between frames to upsample to 30 fps. + +# 6. Downstream use for proficiency estimation + +To demonstrate the utility of predicting a user's physiological state for egocentric vision applications, we use the user proficiency estimation benchmark from the EgoExo4D dataset, which contains over 5000 videos from 740 participants performing skilled human activities [27]. This benchmark aims to classify the proficiency of a user (novice, early expert, intermediate expert, late expert) using only egocentric videos $(Ego)$ , only exocentric videos $(Exo)$ , or all videos together $(Ego + Exo)$ . Our goal was to assess if we can improve the performance of the current baseline model (TimeSFformer [5]) when integrating our predicted HR data into the network. This results in three additional configurations: using egocentric/exocentric videos and HR $(Ego + HR/Exo + HR)$ and using all videos and HR $(Ego + Exo + HR)$ . To predict the continuous HRs for all EgoExo4D videos, we use PulseFormer, pre-trained on egoPPG-DB. + +We implement the TimeSFormer model in exactly the same configuration as for the benchmark results [27] with a clip size of 16 frames and a sampling rate of 16, trained for 15 epochs. We use all videos of the EgoExo4D dataset, for which the proficiency estimation labels are available (using the official benchmark training/validation sets) and which have at least 16 frames at a sampling rate of 16, resulting in 2044 videos. From the official training set, we use $10\%$ as validation, and the held-out official validation set for testing. We summarize our predicted HR data by calculating five features (mean, STD, minimum and maximum HR, and mean HR change) for the corresponding videos. We integrate these features via normalization and a 50-parameter linear layer whose output we concatenate with the output of TimeSformer's backbone before feeding it into the classification head. We train all models from random initialization and evaluate using top-1 accuracy per EgoExo4D protocol. + +# 7. Experiments + +# 7.1. Heart rate estimation + +# 7.1.1. Signal-processing baseline + +We employed signal processing to verify that the BVP signal is present in the eye tracking videos, to determine in which regions the SNR is highest, and to establish a baseline (see Tab. 2). Since the glasses remain mostly stable throughout the recording, we manually define two spatial cropping regions per participant. One region that includes mostly skin, and one region that includes mainly eyes (see Fig. 4). We calculate the mean pixel intensity for both regions, remove motion artifacts by discarding any changes outside the interquartile range and finally filter the signal with a $4^{\text{th}}$ order Butterworth bandpass filter $(0.5$ to $2.8\mathrm{Hz})$ to obtain the BVP (see Fig. 5 in Supplementary). + +# 7.1.2. PulseFormer method + +Using our proposed network PulseFormer, we obtain an MAE of 7.67 bpm and a correlation of 0.85 between our predicted HR and the ground truth HR (see Tab. 2). This is an improvement of 2.40 bpm $(23.8\%)$ of the MAE and 0.13 for the correlation compared to the current SOTA FactorizePhys [38]. Split by activity, we obtain the lowest MAE while the participants are watching a video (MAE=5.52 bpm) and the highest MAE during exercising on a bike (MAE=12.91 bpm), which is the task with the highest mean HR (113.1 bpm) and the second highest motion magnitude. In addition, the MAE decreases for all activities when adding MITA, with the greatest performance improvement for dancing. We define the motion magnitude as the root-mean-squared sum of the absolute differences across the 3-axis IMU recorded by the Aria glasses and normalize it between zero and one across all activities to get a measure of motion of each activity. See Fig. 6 (Supplementary) for a boxplot of the MAEs of PulseFormer's predictions. Using signal processing, we obtain an MAE of 12.40 bpm when using the skin region around the eyes and an MAE of 14.60 bpm using the eye regions as input. This is also reflected in the spatial attention maps that our model implicitly learns (see Fig. 4), which exclude the eyes to predict the HR. To qualitatively cross-check these results, Fig. 5 (Supplementary) shows an example plot of the raw mean intensity values (before filtering) of the skin region compared to the eye region, with the BVP clearly visible for the skin region. Tab. 4 shows the results when downsampling our videos to 10 fps. MAE increases to 11.13 bpm and the correlation decreases to 0.7 when training and testing using 10 fps. When upsampling the videos again to 30 fps, the MAE decreases to 10.20 bpm and the correlation increases to 0.77. In Sec. 16 (Supplementary), we show that PulseFormer also outperforms the baselines in a cross-dataset evaluation. + +# 7.2. Downstream task: proficiency estimation + +Tab. 5 summarizes the results of our experiments to evaluate the value of HR estimation for the proficiency estimation benchmark on EgoExo4D. We see that integrating our predicted HRs into the TimeSFformer model [5] improved accuracy for all scenarios but one (Soccer). We also achieved the highest accuracy for each of these individual scenarios with our HR integration. When combining the egocentric videos with our predicted HRs, we achieved an overall accuracy of $45.29\%$ , a $14.1\%$ increase compared to using egocentric videos alone. The largest gains appeared in the cooking and dancing tasks, where accuracy rose from $20.00\%$ to $40.00\%$ and from $43.44\%$ to $53.27\%$ . Also, when using the egocentric and exocentric videos, and our predicted HRs together, the accuracy increased by $12.67\%$ (4.94 percentage points) from $39.00\%$ to $43.94\%$ compared + +
ModelMAERMSEMAPEr
Yue et al. [94]29.6332.9937.860.1
DeepPhys [10]28.2631.9736.680.08
TS-CAN [45]26.3232.3929.130.11
ContrastPhys+ [81]19.1224.1322.570.21
RhythmMamba [99]15.0519.7817.46-0.16
Baseline eyes14.6018.1818.370.20
PhysMamba [90]13.9416.8617.760.61
RhythmFormer [98]13.1317.4314.730.51
Baseline skin12.4015.5415.290.50
PhysNet [92]12.0915.4315.140.66
PhysFormer [93]10.7113.9712.690.72
PulseFormer w/o SA10.4913.6212.830.73
FactorizePhys [38]10.0713.4312.360.67
PulseFormer w/o MITA8.8212.0310.820.81
PulseFormer (ours)7.6710.699.450.85
Improvement over second-best method-2.40-2.74-2.91+0.13
+ +Table 2. Results for HR prediction from eye tracking videos using different models (PulseFormer, PulseFormer without SA, PulseFormer without MITA and established rPPG baselines). + +
Activityμ HRMotion magnitudePulseFormerPulseFormer w/o MITA
Video71.505.525.97
Office75.70.457.508.22
Kitchen85.30.547.228.89
Dancing89.11.007.8510.54
Bike113.10.7712.9114.62
Walking93.70.308.238.29
+ +Table 3. Results for HR prediction (MAE) split by activity using PulseFormer and PulseFormer without MITA. + +
Input videoMAERMSEMAPEr
10 fps (other datasets)11.1315.1812.280.70
Upsampled to 30 fps10.1813.0712.480.77
+ +Table 4. Results for HR prediction with different frame rates. In the first row, we downsample our videos to a frame rate of 10 fps, commonly used by large-scale datasets such as EgoExo4D [27]. In the second row, we first downsample our videos to 10 fps and then upsample them to 30 fps by linearly interpolating between frames. + +to using only the egocentric and exocentric videos. + +# 8. Discussion + +# 8.1. Heart rate estimation + +Evaluating PulseFormer on egoPPG-DB, we showed that HR can be reliably predicted from eye tracking videos and IMU signals from unmodified egocentric vision headsets. While SOTA rPPG models (e.g., PhysFormer) achieve MAEs as low as 0.50 bpm [93, 98] on datasets such as UBFC-RPPG [6] or OBF [44], these datasets captured participants while calmly sitting at a table looking at the camera (with very little motion or HR changes). However, during even just light motion (e.g., on MMPD [82] or VIPLHR [63]), their MAE increases to 5.0–12.0 bpm [93, 98]. + +On egoPPG-DB, which contains much stronger motion (dancing, exercise bike) and HR fluctuations (between 44-164 bpm), PulseFormer's MAE is 7.67 bpm and outperforms the rPPG SOTA FactorizePhys (MAE=10.07 bpm). Given the strong motion and diverse everyday activities in egoPPG-DB, we believe that our results demonstrate the robustness of PulseFormer in dynamic, everyday conditions. For context, even HR estimates from contact sensors tightened to the body (e.g., Apple Watch) yield an MAE of 3.0 bpm during rest and an MAE of 4.6 bpm on a bike [24]. + +# 8.1.1. Performance depending on method + +We introduced MITA and leverage SA modules to improve the performance of our model. The performance of PulseFormer decreases from 7.67 bpm to 8.82 bpm when removing the MITA and to 10.49 bpm when removing the SA modules (see Tab. 2). When qualitatively analyzing the learned SA maps, we see that our model implicitly learned to exclude the eyes for estimating BVP from the eye tracking videos (see Fig. 4). This aligns with our results using signal processing, obtaining better performance for the skin region compared to the eyes (see Tab. 2). + +# 8.1.2. Performance depending on activity + +Analyzing our results split by activity (see Tab. 3), we obtain the highest MAE when exercising on a bike (MAE=12.91 bpm) and the lowest MAE when watching a video (MAE=5.52 bpm). While watching a video yields the lowest MAE, it is higher than MAEs typically reported for rPPG datasets, such as UBFC-RPPG [6], despite similar levels of motion. We attribute this to two factors: first, the higher variability in HR across egoPPG-DB, requiring improved generalization, and second, the inherent motion artifacts in eye tracking videos from blinking and natural eye movements, even during static tasks like watching videos. Such inherent motion artifacts and, e.g., slipping glasses can make capturing rPPG more difficult in this manner. Furthermore, although dancing has the highest motion magnitude, its MAE (7.85 bpm) is comparable to that of lower-motion tasks such as office and kitchen activities. When comparing performance with and without our MITA module, we + +
ScenarioMajorityEgoEgo + HR (ours)ExoExo + HR (ours)Ego + ExoEgo + Exo + HR (ours)
Basketball38.0045.4547.4748.4848.4849.4950.50
Cooking0.0020.0040.0035.0040.0025.0040.00
Dancing24.5943.4453.2742.6248.3650.8259.84
Music57.8978.9481.5857.8957.8957.8960.53
Bouldering15.2924.5027.818.6112.5815.8921.19
Soccer62.5050.0056.2581.2575.0075.0062.50
Overall27.8039.6945.2934.7537.6739.0043.94
+ +Table 5. Results for proficiency estimation benchmark on EgoExo4D dataset. Note that for all scenarios except Soccer, the accuracy increases when integrating PulseFormer's heart rate estimate into the existing and otherwise unmodified baseline model. + +observe an improvement of $2.6\mathrm{~bpm}$ for dancing, indicating that MITA effectively addresses motion-induced artifacts. + +# 8.1.3. Performance depending on camera fps + +Using eye tracking videos recorded at only 10 fps considerably decreases performance (see Tab. 4). However, up sampling the frame rate to 30 fps through linear interpolation between frames substantially improves the performance again. This is especially important as many large-scale datasets, such as EgoExo4D [27] or Nymeria [50], for which predicting a user's physiological state could help for further downstream tasks, are recorded at only 10 fps. + +# 8.2. Benefits for proficiency estimation downstream + +We found that incorporating HR data into the baseline model of the proficiency estimation task substantially improved accuracy across all three configurations. The egocentric videos combined with the HR achieved the highest overall accuracy at $45.29\%$ , marking a $14.1\%$ increase over using only egocentric videos $(39.69\%)$ . Adding HR especially improved accuracy for cooking (from $20\%$ to $40\%$ ) and dancing (from $43.44\%$ to $53.27\%$ ), which had the lowest accuracies besides bouldering when using only egocentric videos, demonstrating the value of HR in enhancing model performance. Combining egocentric videos, exocentric videos, and HR provided further accuracy gains for some scenarios, achieving the best results for basketball, cooking, and dancing. Results using exocentric views alone were lower overall, which is consistent with benchmark results [27]. Soccer was the only scenario, for which the performance decreased for $Exo+HR$ and $Ego+Exo+HR$ . We see two reasons for that. 1) Of EgoExo4D's 2044 official train/test videos, only 77 are soccer, making it the scenario with the least training/test data by far. 2) Our HR estimates may be less accurate for "Stop-and-Go" sports, which are not captured in egoPPG-DB right now. In Tab. 9 (Supplementary), we show that we obtain the best downstream performance when using the HR features calculated with PulseFormer compared to the baselines. For training and testing, we used the available subset of EgoExo4D videos + +for which proficiency labels are available, following the official training and test splits. While our used data shows slight variations from the official release in majority class distributions and accuracy scores, the observed trends align well with the established benchmark results. + +# 8.3. Broader impacts + +Beyond health applications, such as predicting stress and fatigue [9, 60, 65], cardiac measurements could also help models better understand user behavior to, e.g., improve personalized assistance [42]. Furthermore, we believe that our approach requires the user's explicit consent, regardless of application. Mechanisms must make users aware of measurements and require consent, e.g., on the Aria platform. + +# 9. Conclusion + +egoPPG is a novel task for egocentric vision systems to extract the wearer's heart rate for integrating their physiological state into egocentric vision tasks downstream. We have introduced PulseFormer, a method that processes input from the eye tracking cameras on unmodified egocentric vision systems and fuses them with motion cues from the headset's IMU to robustly estimate the person's HR in various everyday scenarios. We validate PulseFormer's robustness on our dataset egoPPG-DB and demonstrate significant improvements over existing rPPG models. With HR estimations from PulseFormer, we also significantly improve the proficiency estimation benchmark on the largescale EgoExo4D dataset. Our results emphasize the potential of physiological insights obtained via egoPPG methods for further egocentric vision applications. By making our dataset available to the community, we aim to support physiological state estimation via HR in future research and new downstream tasks for egocentric vision systems. Given our promising results, we believe that future work could now focus on collecting more participants with a broader demographic background across different age groups, skin types, and ethnicities, and also extend data collection to outdoor settings to assess the impact of varying lighting conditions. + +# References + +[1] Mojtaba Khomami Abadi, Ramanathan Subramanian, Seyed Mostafa Kia, Paolo Avesani, Ioannis Patras, and Nicu Sebe. Decaf: Meg-based multimodal database for decoding affective physiological responses. IEEE Transactions on Affective Computing, 6(3):209-222, 2015. 2 +[2] Isayas Berhe Adhanom, Paul MacNeilage, and Eelke Folmer. Eye tracking in virtual reality: a broad review of applications and challenges. Virtual Reality, 27(2):1481-1505, 2023. 3 +[3] Henny Admoni and Siddhartha Srinivasa. Predicting user intent through eye gaze for shared autonomy. In 2016 AAAI fall symposium series, 2016. 2 +[4] Apple. Apple vision pro. https://www.apple.com/apple-vision-pro/, 2024. Accessed: 2024.11.13.2 +[5] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 6 +[6] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 3, 7, 2 +[7] Björn Braun, Daniel McDuff, Tadas Baltrusaitis, and Christian Holz. Video-based sympathetic arousal assessment via peripheral blood flow estimation. Biomedical Optics Express, 14(12):6607-6628, 2023. 2 +[8] Björn Braun, Daniel McDuff, and Christian Holz. How suboptimal is training rpgg models with videos and targets from different body sites? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 410-418, 2024. 2 +[9] Rafael A Calvo, Sidney D'Mello, Jonathan Matthew Gratch, and Arvid Kappas. The Oxford handbook of affective computing. Oxford University Press, USA, 2015. 2, 8 +[10] Weixuan Chen and Daniel McDuff. Deepphys: Video-based physiological measurement using convolutional attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 349-365, 2018. 2, 4, 7 +[11] Hsueh-Wen Chow, Chao-Ching Yang, et al. Accuracy of optical heart rate sensing technology in wearable fitness trackers for young and older adults: validation and comparison study. JMIR mHealth and uHealth, 8(4):e14707, 2020. 2 +[12] Viviane Clay, Peter König, and Sabine Koenig. Eye tracking in virtual reality. Journal of eye movement research, 12(1), 2019. 3 +[13] Stephen A Coombes, Torrie Higgins, Kelly M Gamble, James H Cauraugh, and Christopher M Janelle. Attentional control theory: Anxiety, emotion, and motor planning. Journal of anxiety disorders, 23(8):1072-1079, 2009. 2 +[14] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1-23, 2022. 2 +[15] Davison. Real-time simultaneous localisation and mapping with a single camera. In Proceedings Ninth IEEE International Conference on Computer Vision, pages 1403-1410. IEEE, 2003. 2 + +[16] Gerard De Haan and Vincent Jeanne. Robust pulse rate from chrominance-based rppg. IEEE transactions on biomedical engineering, 60(10):2878-2886, 2013. 2, 5 +[17] Jessilyn Dunn, Ryan Runge, and Michael Snyder. Wearables and the medical revolution. Personalized medicine, 15(5): 429-448, 2018. 2 +[18] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, et al. Project aria: A new tool for egocentric multi-modal ai research. arXiv preprint arXiv:2308.13561, 2023. 2, 3 +[19] Harun Evrengul, Halil Tanriverdi, Sedat Kose, Basri Amasyali, Ayhan Kilic, Turgay Celik, and Hasan Turhan. The relationship between heart rate recovery and heart rate variability in coronary artery disease. Annals of Noninvasive Electrocardiology, 11(2):154-162, 2006. 2 +[20] Michael W Eysenck, Nazanin Derakshan, Rita Santos, and Manuel G Calvo. Anxiety and cognitive performance: attentional control theory. Emotion, 7(2):336, 2007. 2 +[21] Alircza Fathi, Jessica K Hodgins, and James M Rehg. Social interactions: A first-person perspective. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1226-1233. IEEE, 2012. 2 +[22] Thomas B Fitzpatrick. The validity and practicality of sun-reactive skin types i through vi. Archives of dermatology, 124(6):869-871, 1988. 3 +[23] Kim Fox, Jeffrey S Borer, A John Camm, Nicolas Danchin, Roberto Ferrari, Jose L Lopez Sendon, Philippe Gabriel Steg, Jean-Claude Tardif, Luigi Tavazzi, Michal Tendera, et al. Resting heart rate in cardiovascular disease. Journal of the American College of Cardiology, 50(9):823-830, 2007. 2 +[24] Stephen Gillinov, Muhammad Etiwy, Robert Wang, Gordon Blackburn, Dermot Phelan, A Marc Gillinov, Penny Houghtaling, Hoda Javadikasgari, and Milind Y Desai. Variable accuracy of wearable heart rate monitors during aerobic exercise. Medicine & Science in Sports & Exercise, 2017. 7 +[25] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13505-13515, 2021. 2 +[26] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2 +[27] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2, 3, 5, 6, 7, 8 +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 + +[29] Guillaume Heusch, André Anjos, and Sébastien Marcel. A reproducible study on remote heart rate measurement. arXiv preprint arXiv:1709.00962, 2017. 3 +[30] Christian Holz and Edward J Wang. Glabella: Continuously sensing blood pressure behavior using an unobtrusive wearable device. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 1(3):1-23, 2017. 3 +[31] HTC. Htc vivo. https://vive.com/, 2024. Accessed: 2024.11.13. 2 +[32] Min Hu, Fei Qian, Xiaohua Wang, Lei He, Dong Guo, and Fuji Ren. Robust heart rate estimation with spatial-temporal attention network from facial videos. IEEE Transactions on Cognitive and Developmental Systems, 14(2):639-647, 2021. 5 +[33] Patrick Hübner, Kate Clintworth, Qingyi Liu, Martin Weinmann, and Sven Wursthorn. Evaluation of hololens tracking and depth sensing for indoor mapping applications. Sensors, 20(4):1021, 2020. 2 +[34] Markus Huelsbusch and Vladimir Blazek. Contactless mapping of rhythmical phenomena in tissue perfusion using ppgi. In Medical Imaging 2002: Physiology and Function from Multidimensional Images, pages 110-117. International Society for Optics and Photonics, 2002. 2 +[35] Hao Jiang, Calvin Murdock, and Vamsi Krishna Ithapu. Ego-centric deep multi-channel audio-visual active speaker localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544–10552, 2022. 2 +[36] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 2 +[37] Jiaxi Jiang, Paul Streli, Manuel Meier, Andreas Fender, and Christian Holz. Egoposer: Robust real-time ego-body pose estimation in large scenes. arXiv preprint arXiv:2308.06493, 3(7), 2023. 2 +[38] Jitesh Joshi, Sos S Agaian, and Youngjun Cho. Factorizephys: Matrix factorization for multidimensional attention in remote physiological sensing. arXiv preprint arXiv:2411.01542, 2024. 6, 7 +[39] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 2 +[40] Robert E Kleiger, Phyllis K Stein, and J Thomas Bigger Jr. Heart rate variability: measurement and clinical utility. Annals of Noninvasive Electrocardiology, 10(1):88-101, 2005. 2 +[41] Dan Kondratyuk, Liangzhe Yuan, Yandong Li, Li Zhang, Mingxing Tan, Matthew Brown, and Boqing Gong. Movinets: Mobile video networks for efficient video recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16020-16030, 2021. 2 +[42] Nikola Kovacevic, Christian Holz, Markus Gross, and Rafael Wampfler. On multimodal emotion recognition for humanchatbot interaction in the wild. In Proceedings of the 26th + +International Conference on Multimodal Interaction, pages 12-21, 2024. 8 +[43] Magic Leap. Magic leap 2. https://www.magicleap.com/magic-leap-2, 2024. Accessed: 2024.11.13. 1, 2 +[44] Xiaobai Li, Iman Alikhani, Jingang Shi, Tapio Seppanen, Juhani Junttila, Kirsi Majamaa-Voltri, Mikko Tulppo, and Guoying Zhao. The obf database: A large face video database for remote physiological signal measurement and atrial fibrillation detection. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 242-249. IEEE, 2018. 7, 3 +[45] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems, 33:19400-19411, 2020. 2, 5, 7 +[46] Xin Liu, Girish Narayanswamy, Akshay Paruchuri, Xiaoyu Zhang, Jiankai Tang, Yuzhe Zhang, Roni Sengupta, Shwetak Patel, Yuntao Wang, and Daniel McDuff. rppg-toolbox: Deep remote ppg toolbox. Advances in Neural Information Processing Systems, 36, 2024. 5 +[47] Camillo Lugasesi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Yong, Juhyun Lee, et al. Mediapipe: A framework for perceiving and processing reality. In Third workshop on computer vision for AR/VR at IEEE computer vision and pattern recognition (CVPR), 2019. 2 +[48] Tiffany Luong and Christian Holz. Characterizing physiological responses to fear, frustration, and insight in virtual reality. IEEE Transactions on Visualization and Computer Graphics, 28(11):3917-3927, 2022. 2 +[49] Zhaoyang Lv, Nicholas Charron, Pierre Moulon, Alexander Gamino, Cheng Peng, Chris Sweeney, Edward Miller, Huixuan Tang, Jeff Meissner, Jing Dong, et al. Aria everyday activities dataset. arXiv preprint arXiv:2402.13349, 2024. 3 +[50] Lingni Ma, Yuting Ye, Fangzhou Hong, Vladimir Guzov, Yifeng Jiang, Rowan Postyeni, Luis Pesqueira, Alexander Gamino, Vijay Baiyya, Hyo Jin Kim, et al. Nymeria: A massive collection of multimodal egocentric daily motion in the wild. arXiv preprint arXiv:2406.09905, 2024. 2, 5, 8 +[51] Minghuang Ma, Haoqi Fan, and Kris M Kitani. Going deeper into first-person activity recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1894-1903, 2016. 2 +[52] Alan C MacPherson, Dave Collins, and Sukhvinder S Obhi. The importance of temporal structure and rhythm for the optimum performance of motor skills: A new focus for practitioners of sport psychology. Journal of Applied Sport Psychology, 21(S1):S48-S61, 2009. 2 +[53] Javier Marín-Morales, Carmen Llinares, Jaime Guixeres, and Mariano Alcaniz. Emotion recognition in immersive virtual reality: From statistics to affective computing. Sensors, 20(18):5163, 2020. 2 +[54] Daniel McDuff, Miah Wander, Xin Liu, Brian Hill, Javier Hernandez, Jonathan Lester, and Tadas Baltrusaitis. Scamps: Synthetics for camera measurement of physiological signals. Advances in Neural Information Processing Systems, 35: 3744-3757, 2022. 3 + +[55] Manuel Meier, Berken Utku Demirel, and Christian Holz. Wildppg: a real-world ppg dataset of long continuous recordings. In Proceedings of the 38th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2025. Curran Associates Inc. 2 +[56] Meta. Meta quest. https://www.meta.com/quest/, 2024. Accessed: 2024.11.13. 1, 2 +[57] Microsoft. Microsoft hololens. https://learn.microsoft.com/en-us/hololens/, 2024. Accessed: 2024.11.13. 2 +[58] Kyle Min and Jason J Corso. Integrating human gaze into attention for egocentric activity recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1069-1078, 2021. 2 +[59] Juan Abdon Miranda-Correa, Mojtaba Khomami Abadi, Nicu Sebe, and Ioannis Patras. Amigos: A dataset for affect, personality and mood research on individuals and groups. IEEE transactions on affective computing, 12(2):479-493, 2018. 2 +[60] Max Moebus, Shkurta Gashi, Marc Hilty, Pietro Oldrati, and Christian Holz. Meaningful digital biomarkers derived from wearable sensors to predict daily fatigue in multiple sclerosis patients and healthy controls. Iscience, 27(2), 2024. 8 +[61] Max Moebus, Lars Hauptmann, Nicolas Kopp, Berken Demirel, Björn Braun, and Christian Holz. Nightbeat: Heart rate estimation from a wrist-worn accelerometer during sleep. IEEE Journal of Biomedical and Health Informatics, 2024. 2 +[62] Subhas Chandra Mukhopadhyay. Wearable sensors for human activity monitoring: A review. IEEE sensors journal, 15(3):1321-1330, 2014. 2 +[63] Xuesong Niu, Hu Han, Shiguang Shan, and Xilin Chen. Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part V 14, pages 562–576. Springer, 2019. 7, 3 +[64] Xuesong Niu, Xingyuan Zhao, Hu Han, Abhijit Das, Antitza Dantcheva, Shiguang Shan, and Xin Chen. Robust remote heart rate estimation from face utilizing spatial-temporal attention. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1-8. IEEE, 2019. 5 +[65] Rosalind W Picard. Affective computing. MIT press, 2000. 2, 8 +[66] Ming-Zher Poh, Daniel McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics express, 18(10):10762-10774, 2010. 2 +[67] Ivan Rodin, Antonino Furnari, Dimitrios Mavroeidis, and Giovanni Maria Farinella. Predicting the future from first person (egocentric) vision: A survey. Computer Vision and Image Understanding, 211:103252, 2021. 2 +[68] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerfslam: Real-time dense monocular slam with neural radiance fields. In 2023 IEEE/RSJ International Conference on Intel- + +ligent Robots and Systems (IROS), pages 3437-3444. IEEE, 2023. 2 +[69] Michael S Ryoo, Thomas J Fuchs, Lu Xia, Jake K Aggarwal, and Larry Matthies. Robot-centric activity prediction from first-person videos: What will they do to me? In Proceedings of the tenth annual ACM/IEEE international conference on human-robot interaction, pages 295–302, 2015. 2 +[70] Rita Meziati Sabour, Yannick Benezeth, Pierre De Oliveira, Julien Chappe, and Fan Yang. Ubfc-phys: A multimodal database for psychophysiological studies of social stress. IEEE Transactions on Affective Computing, 2021. 3 +[71] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Fast image-based localization using direct 2d-to-3d matching. In 2011 International Conference on Computer Vision, pages 667–674. IEEE, 2011. 2 +[72] Fred Shaffer and Jay P Ginsberg. An overview of heart rate variability metrics and norms. Frontiers in public health, 5: 258, 2017. 2 +[73] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 2 +[74] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2733-2742, 2021. 2 +[75] Takaaki Shiratori, Hyun Soo Park, Leonid Sigal, Yaser Sheikh, and Jessica K Hodgins. Motion capture from body-mounted cameras. In ACM SIGGRAPH 2011 papers, pages 1-10, 2011. 2 +[76] Isabelle M Shuggi, Hyuk Oh, Helena Wu, Maria J Ayoub, Arianna Moreno, Emma P Shaw, Patricia A Shewokis, and Rodolphe J Gentili. Motor performance, mental workload and self-efficacy dynamics during learning of reaching movements throughout multiple practice sessions. Neuroscience, 423:232-248, 2019. 2 +[77] Mohammad Soleymani, Jeroen Lichtenauer, Thierry Pun, and Maja Pantic. A multimodal database for affect recognition and implicit tagging. IEEE transactions on affective computing, 3(1):42-55, 2011. 3 +[78] Jeremy Speth, Nathan Vance, Patrick Flynn, and Adam Czajka. Non-contrastive unsupervised learning of physiological signals from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14464–14474, 2023. 2 +[79] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In The 23rd IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062, 2014. 3, 2 +[80] Ramanathan Subramanian, Julia Wache, Mojtaba Khomami Abadi, Radu L Vieriu, Stefan Winkler, and Nicu Sebe. Ascertain: Emotion and personality recognition using commercial sensors. IEEE Transactions on Affective Computing, 9 (2):147-160, 2016. 2 + +[81] Zhaodong Sun and Xiaobai Li. Contrast-phys+: Unsupervised and weakly-supervised video-based remote physiological measurement via spatiotemporal contrast. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 7 +[82] Jiankai Tang, Kequan Chen, Yuntao Wang, Yuanchun Shi, Shwetak Patel, Daniel McDuff, and Xin Liu. Mmpd: multidomain mobile video physiology dataset. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1-5. IEEE, 2023. 3, 7, 2 +[83] Chai M Tyng, Hafeez U Amin, Mohammad NM Saad, and Aamir S Malik. The influences of emotion on learning and memory. Frontiers in psychology, 8:235933, 2017. 2 +[84] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Optics express, 16(26):21434-21445, 2008. 2 +[85] Wenjin Wang, Albertus C Den Brinker, Sander Stuijk, and Gerard De Haan. Algorithmic principles of remote ppg. IEEE Transactions on Biomedical Engineering, 64(7):1479-1491, 2016. 2, 4 +[86] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5 +[87] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 2 +[88] Kentaro Yamada, Yusuke Sugano, Takahiro Okabe, Yoichi Sato, Akihiro Sugimoto, and Kazuo Hiraki. Attention prediction in egocentric video using motion and visual saliency. In Advances in Image and Video Technology: 5th Pacific Rim Symposium, PSIVT 2011, Gwangju, South Korea, November 20-23, 2011, Proceedings, Part I 5, pages 277-288. Springer, 2012. 2 +[89] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3333-3343, 2022. 2 +[90] Zhixin Yan, Yan Zhong, Wenjun Zhang, Lin Shu, Hongbin Xu, and Wenxiong Kang. Physmamba: Leveraging dual + +stream cross-attention sdd for remote physiological measurement. arXiv preprint arXiv:2408.01077, 2024. 2, 7 +[91] Yu Yao, Mingze Xu, Chiho Choi, David J Crandall, Ella M Atkins, and Behzad Dariush. Egocentric vision-based future vehicle localization for intelligent driving assistance systems. In 2019 International Conference on Robotics and Automation (ICRA), pages 9711-9717. IEEE, 2019. 2 +[92] Zitong Yu, Xiaobai Li, and Guoying Zhao. Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks. arXiv preprint arXiv:1905.02419, 2019. 2, 4, 7 +[93] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip HS Torr, and Guoying Zhao. Physformer: Facial video-based physiological measurement with temporal difference transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4186-4196, 2022. 2, 7 +[94] Zijie Yue, Miaojing Shi, and Shuai Ding. Facial video-based remote physiological measurement via self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13844-13859, 2023. 7 +[95] Zheng Zhang, Jeff M Girard, Yue Wu, Xing Zhang, Peng Liu, Umur Ciftci, Shaun Canavan, Michael Reale, Andy Horowitz, Huiyuan Yang, et al. Multimodal spontaneous emotion corpus for human behavior analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3438-3446, 2016. 3 +[96] Yue Zhao, Ishan Misra, Philipp Krahenbihl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 2 +[97] Chenchen Zhu, Fanyi Xiao, Andrés Alvarado, Yasmine Babaei, Jiabo Hu, Hichem El-Mohri, Sean Culatana, Roshan Sumbaly, and Zhicheng Yan. Egoobjects: A large-scale egocentric dataset for fine-grained object understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20110-20120, 2023. 2 +[98] Bochao Zou, Zizheng Guo, Jiansheng Chen, and Huimin Ma. Rhythmformer: Extracting rppg signals based on hierarchical temporal periodic transformer. arXiv preprint arXiv:2402.12788, 2024. 7 +[99] Bochao Zou, Zizheng Guo, Xiaocheng Hu, and Huimin Ma. Rhythmmamba: Fast remote physiological measurement with arbitrary length videos. arXiv preprint arXiv:2404.06483, 2024. 7 \ No newline at end of file diff --git a/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/images.zip b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b323888ac41ba93e66f5b97a066f3b1dc8b47eaf --- /dev/null +++ b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d20770baefc1bc9ab98a259eb0c4f2768af8d82ca357857d8aaea22232117a9 +size 506385 diff --git a/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/layout.json b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d916438a8405b5444926ac8980a56b0afe741fe5 --- /dev/null +++ b/egoppgheartrateestimationfromeyetrackingcamerasinegocentricsystemstobenefitdownstreamvisiontasks/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5319420ce0ce6582e0f5e78a751160bf20da824d8ebed67f956c35d274d5ec2a +size 446661 diff --git a/imanipskillincrementallearningforroboticmanipulation/47162734-cb83-448a-b459-5b0fc2ff10ae_content_list.json b/imanipskillincrementallearningforroboticmanipulation/47162734-cb83-448a-b459-5b0fc2ff10ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..42913f45f3809d14cfae467543acf5327768f046 --- /dev/null +++ b/imanipskillincrementallearningforroboticmanipulation/47162734-cb83-448a-b459-5b0fc2ff10ae_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5576728101cc4ddc62cdd07a03358a849f9ecc4ffa634261865adc3ecb8ac343 +size 81990 diff --git a/imanipskillincrementallearningforroboticmanipulation/47162734-cb83-448a-b459-5b0fc2ff10ae_model.json b/imanipskillincrementallearningforroboticmanipulation/47162734-cb83-448a-b459-5b0fc2ff10ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..add34696bf07431d738123caafaf44c91082707f --- /dev/null +++ b/imanipskillincrementallearningforroboticmanipulation/47162734-cb83-448a-b459-5b0fc2ff10ae_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a450f681fda251b49b222adbb058dc78feff261ec035f6f39432b99bb176186a +size 104397 diff --git a/imanipskillincrementallearningforroboticmanipulation/47162734-cb83-448a-b459-5b0fc2ff10ae_origin.pdf b/imanipskillincrementallearningforroboticmanipulation/47162734-cb83-448a-b459-5b0fc2ff10ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9843533b9ad37ffdfcb2603b2343f61d2c47226d --- /dev/null +++ b/imanipskillincrementallearningforroboticmanipulation/47162734-cb83-448a-b459-5b0fc2ff10ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:821f2d04e636e87feeb9d8fd866ed3064f18c0abbe2f9c66adb9e591f6e6a001 +size 1876323 diff --git a/imanipskillincrementallearningforroboticmanipulation/full.md b/imanipskillincrementallearningforroboticmanipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e9985da3831337323ad018ffd6615c2fe2e15539 --- /dev/null +++ b/imanipskillincrementallearningforroboticmanipulation/full.md @@ -0,0 +1,313 @@ +# iManip: Skill-Incremental Learning for Robotic Manipulation + +Zexin Zheng $^{1*}$ , Jia-Feng Cai $^{1*}$ , Xiao-Ming Wu $^{1}$ , Yi-Lin Wei $^{1}$ + +Yu-Ming Tang $^{1}$ , Ancong Wu $^{1,2\dagger}$ , Wei-Shi Zheng $^{1,2\dagger}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China $^{2}$ Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +{zhengzx25, caijf23}@mail2.sysu.edu.cn wuanc@mail.sysu.edu.cn wszheng@ieee.org + +# Abstract + +The development of a generalist agent with adaptive multiple manipulation skills has been a long-standing goal in the robotics community. In this paper, we explore a crucial task, skill-incremental learning, in robotic manipulation, which is to endow the robots with the ability to learn new manipulation skills based on the previous learned knowledge without re-training. First, we build a skill-incremental environment based on the RLBench benchmark, and explore how traditional incremental methods perform in this setting. We find that they suffer from severe catastrophic forgetting due to the previous methods on classification overlooking the characteristics of temporality and action complexity in robotic manipulation tasks. Towards this end, we propose an incremental Manipulation framework, termed iManip, to mitigate the above issues. We firstly design a temporal replay strategy to maintain the integrity of old skills when learning new skill. Moreover, we propose the Extendable PerceiverIO, consisting of an action prompt with extendable weight to adapt to new action primitives in new skill. Extensive experiments show that our framework performs well in Skill-Incremental Learning. + +# 1. Introduction + +Imagine that we are in a household setting with a robot assistant that already has basic functions like folding clothes and fetching items. Now, as the owners, we want it to learn new skills. For instance, today we've purchased a dishwasher, and we'd like the robot to learn how to load dishes into it. It could quickly acquire these new skills by observing and mimicking our demonstrations. This is an interesting and challenging requirement for robotics manipulation, which needs the robot to learn new skills based on previously learned knowledge without retraining. However, in the area of robotic manipulation, previous research mainly + +![](images/6b85a78518db05dc9b1a8f9d5b967e11ffa78553939d675199133c1cb795ff56.jpg) +Figure 1. (a) An overview of our skill-incremental learning for robotic manipulation that requires the agent to learn skill sequences over time. (b) A comparison of model performance between the traditional incremental baseline (TIB) and our iManip. + +focuses on how to acquire better manipulation performance [6, 38, 47, 57, 69] or how to transfer the knowledge from the pretrained large language or vision models [4, 10, 21, 30] to learn robotics manipulations, rarely works explore how to incrementally learn new skills. LIBERO [35] is a preliminary benchmark to learn lifelong robotics control, but it explores the incremental abilities in new objects or new spatial positions, still limited in the characteristics of the same skills, where different tasks share the same skill. + +In this paper, we thoroughly explore this crucial task, skill-incremental learning, in robotic manipulation, which is to leverage the previous learned knowledge to enable the robots to learn new manipulation skills without training from scratch. To begin with, we construct a skill- + +incremental environment based on the RLBench benchmark. It requires the agent to continuously acquire a sequence of 10 challenging language-conditioned manipulation skills. Each skill consists of at least two variations encompassing several types, such as variations in shape and color, totaling 166 variations. Then, we explore how the traditional incremental learning methods work on this skill-incremental environment. As seen in Figure 1, we discover that after learning new manipulation skills, the agent's performance on prior skills significantly deteriorates. Thus, we can conclude that traditional incremental learning methods still suffer from catastrophic forgetting in this new setting, which, we regard, is due to their neglect of the temporality and action complexity in robotic manipulation tasks. The temporal complexity arises from the dynamic changes in the environment and in the robot states over time, resulting in each action having an impact on subsequent actions. And the complexity of actions requires the agent to learn new action primitives for action planning in novel environments and interactions, which is representative in 3D dimensions with rotation and shift, and highly complex. + +To mitigate the above problems, we propose a new skill incremental Manipulation framework, termed iManip, for this new setting. The key idea of our framework is to modify the traditional incremental learning methods to fully consider the characteristics of temporality and action complexity in robotic manipulation tasks. Two designs make our framework nontrivial. First, to address temporal complexity, we design the temporal replay strategy to maintain the integrity of the temporal data and propose to replay a fixed number of keyframe samples at different time points for each manipulation skill, using the farthest-distance entropy sampling strategy. Second, to address the complexity of actions, we propose the Extendable PerceiverIO, consisting of an action prompt with extendable weight to adapt to new action primitives. When learning new skills, we freeze the learned parameters of PerceiverIO while learning a new small set of skill-specific action prompts and weight matrices for new action primitives learning. + +Extensive experiments show that our iManip framework maintains several excellent capabilities: (1) Effectiveness: it performs well in the skill-incremental learning setting, outperforming the traditional incremental baseline with an increase of 9.4 points. (2) Robustness: it demonstrates robust performance in several different incremental settings. (3) Lightweight: it only needs lightweight finetuning of the policy decoder with fewer training steps, comparable with full weights finetuning. (4) Extendability: it also has extraordinary performance in real-world experiments. + +# 2. Related Work + +Robotic Manipulation. Learning robot manipulation conditioned on both vision and language has gained increas + +ing attention [11, 16-18, 49, 50, 53, 58-60, 63, 65], with robot imitation learning using scripted trajectories [23, 39] or tele-operation data [12, 42, 62] gradually becoming a mainstream approach. Previous work [5, 13, 24] has focused on using 2D images to predict actions, while more recent studies have leveraged the rich spatial information of 3D point clouds for motion planning. For instance, PerAct [51] feeds voxel tokens into a PerceiverIO [22]-based transformer policy, achieving impressive results across various tasks. GNFactor [67] optimizes a generalizable neural field for semantic extraction, while ManiGaussian [37] introduces a dynamic Gaussian Splatting [27] framework for semantic propagation. 3DDA [26] proposes a 3D denoising transformer to predict noise in noised 3D robot pose trajectories. However, these methods suffer from catastrophic forgetting in skill-incremental learning and we propose our iManip framework to continuously learn new skills and mitigate forgetting of learned knowledge. + +Conventional Lifelong Learning Approaches. One of the most commonly used methods is the rehearsal-based method [2, 8, 9, 19, 20, 25, 33, 46, 54, 56]. iCaRL [46] proposes a herding-based step for prioritized exemplar selection to store old exemplars. RWalk [8] proposes a hard-exemplar sampling strategy for replay. Distillation-based methods [7, 14, 19, 31, 34, 52, 55, 68] propose distilling old knowledge from the old network to the current network or maintaining the old feature space during new tasks. ABD [52] proposes distilling synthetic data for incremental learning and EEIL [7] proposes to distill the knowledge from the classification layers of the old classes. Moreover, there are dynamic-architecture-based methods [1, 32, 36, 43, 44, 66] that dynamically adjust the model's representation ability to fit the evolving data stream. In this work, we use the traditional rehearsal-based method [46] and the distillation-based methods [7] for robotic skill-incremental learning. We find that they also suffer from catastrophic forgetting due to overlooking the temporal and action complexities of robotic manipulation. + +A Preliminary Lifelong Robot Learning Benchmark LIBERO. Previous works [15, 28, 29, 40, 41, 61] explore different strategies for incremental robotic learning based on different testing environments. Recently, to promote community development, LIBERO [35] proposed a benchmark, which explores incremental abilities with new objects, goals, or spatial positions, as shown in Figure 2. The limitation of LIBERO lies in the fact that most tasks are constrained by similar skill characteristics. For example, it regards "Put the bowl on the plate" and "Put the bowl on stove" as different tasks. In this paper, we explore a more realistic and challenging task, skill-incremental learning, where the agent learns a sequence of skills over time, each involving multiple poses and object variations in placement, color, shape, size, and category. + +![](images/9e9a8f717d27a1ead0e89bee60f2db6f54dfa0c183e497bed7b66e01ea6cedcc.jpg) +Figure 2. Overview of robotic incremental learning. Previous works focus on incremental abilities in new objects, goals, or spatial positions, where different tasks may share the same skill. The iManip focuses on skill-incremental learning which better captures the true adaptability and flexibility required for real-world robotic learning. + +# 3. Skill-incremental Learning for Robotic Manipulation + +# 3.1. Challenges + +While robotic manipulation has received increasing attention, few previous studies explore how to incrementally learn new skills. In this paper, we focus on skill-incremental learning for robotic manipulation, a challenging setting that requires agents to continuously acquire new skills without training from scratch. + +In this new setting, we find that applying traditional incremental learning methods [7, 8, 37, 46, 51] still suffers from catastrophic forgetting of previously learned skills, as seen in Figure 1 (b). There are two key challenges when applying previous methods of visual classification to robotic skill-incremental learning: 1) Previous methods overlook the temporal complexity inherent in robotic manipulation tasks, where dynamic changes in the environment and the robot states over time cause actions to impact subsequent ones. For example, classical replay algorithms focus on sampling the most representative samples per class, directly storing representative samples from demonstrations may result in temporal imbalance of the trajectory, leading to instability during task execution. 2) Previous methods focus primarily on general visual features while neglecting the actions complexity in robotic manipulation. Robotic manipulation involves action planning through interactions with the physical environment, such as visual and language input. When a new manipulation skill arises, the agent learns new visual-language interactions and quickly acquires new action primitives based on prior knowledge. + +# 3.2. Overall Pipeline + +To tackle the two challenges, we propose the temporal replay strategy and the extendable PerceiverIO architecture in our iManip framework. Specifically, as shown in Figure 3, we present the overall framework for robotic skill-incremental learning, which can sequentially learn robotic manipulation skills while mitigating catastrophic forgetting of learned skills. Specifically, the agent learns a sequence of manipulation skills with a stream of training data denoted as $\mathcal{D} = \{\mathcal{D}_i\}_{i=1}^T$ , where $\mathcal{D}_i = \{(o_i^{(1)},a_i^{(1)}),(o_i^{(2)},a_i^{(2)}),\ldots\}$ represents the demonstration trajectories of skill $i$ . The visual input $o_i^{(t)} = (I_i^{(t)},D_i^{(t)},P_i^{(t)})$ consists of the $t$ -th single-view images $I_i^{(t)}$ , depth images $D_i^{(t)}$ , and proprioception matrix $P_i^{(t)} \in \mathbb{R}^4$ that includes the openness, end-effector position, and the current timestep. After learning skill $i$ , a compact memory $\mathcal{M}$ stores a fixed number of demonstration replays for skills up to $i-1$ . Following [37, 51, 67], the agent combines the visual input $o_i^{(t)}$ and language instructions $l_i$ to generate the optimal action $a_i^{(t)} = (a_{i,\mathrm{trans}}^{(t)},a_{i,\mathrm{rot}}^{(t)},a_{i,\mathrm{open}}^{(t)},a_{i,\mathrm{col}}^{(t)})$ , which respectively demonstrates the target translation in voxel $a_{i,\mathrm{trans}}^{(t)} \in \mathbb{R}^{100^3}$ , rotation $a_{i,\mathrm{rot}}^{(t)} \in \mathbb{R}^{(360/5)\times 3}$ , openness $a_{i,\mathrm{open}}^{(t)} \in [0,1]$ and collision avoidance $a_{i,\mathrm{col}}^{(t)} \in [0,1]$ . + +Our framework consists of a voxel encoder for learning 3D scene features, a latent transformer (the extendable PerceiverIO), and a policy decoder to predict optimal robot actions. Specifically, our approach employs a temporal replay strategy, maximizing the information entropy of replay demos to effectively address the first challenges caused by + +![](images/2cd7c2f50fc3c6e20ad9a895a5c0541d1b8ca0100ab12bb44e9bd9826b5b073b.jpg) +Figure 3. The overall framework of iManip, which primarily consists of a temporal replay strategy to store the samples with the farthest distance entropy for each keyframe of old demonstrations and an extendable PerceiverIO consisting of action prompts with extendable weights to adapt to new action primitives. + +classic replay methods. Additionally, we introduce the extendable PerceiverIO, consisting of action prompts with extendable weights to adapt to new action primitives, while preserving knowledge of previous skills to prevent catastrophic forgetting. + +# 3.3. The iManip Framework + +Temporal Replay Strategy. Rehearsal-based methods [8, 19, 46] are classic algorithms in incremental vision classification tasks while overlooking the temporal complexity inherent in robotic manipulation. Apart from random sampling, popular methods such as herding sampling [46] and hard-exemplar sampling [3, 8] effectively select the most representative samples from each class. However, robotic manipulation data consists of temporal samples from entire expert episodes. Directly storing representative samples can lead to temporal imbalance of episodes, resulting in instability during task execution. + +Therefore, we propose a temporal replay strategy to balance the sampling of different keyframes of episodes for each skill. Keyframes [51] are the samples from episodes when the end-effector changes state (e.g., gripper closing) or when its velocity approaches zero, representing critical temporal landmarks within the trajectory. + +Furthermore, to sample a greater variety of variants, we + +propose the farthest-distance entropy sampling to store an equal number of each keyframe. It requires the buffer to contain the replays that exhibit the largest entropy divergence as + +$$ +S = \underset {S \in E, | S | = K} {\arg \max } \sum_ {i \in S} \sum_ {j \in S} A [ i ] [ j ], \tag {1} +$$ + +where $S$ is the sampled set with size $K$ , $E$ is the demo set of a specific keyframe with size $N$ , and $A$ is the distance array of the demos' action prediction entropy $\mathcal{L}_{\mathrm{act}}$ . Specifically, we propose to store the demo $j$ to the replay buffer that has the farthest distance of sampled demos in $S$ as + +$$ +j = \underset {j \in E} {\arg \max } \sum_ {k \in S} A [ j ] [ k ]. \tag {2} +$$ + +This ensures the storage of temporally balanced, information-rich samples from previous episodes, helping to mitigate catastrophic forgetting of learned skills. + +Based on the above analysis, the pseudocode for the temporal replay strategy is shown in Algorithm 1. The algorithm can get the optimal solution for the objective 1, with time complexity of $O(N^2)$ , which is the size of a specific keyframe and not relevant to the size of the previous data, suitable for incremental skill learning. + +# Algorithm 1 Farthest-distance Entropy Sampling + +Require: Entropy set corresponding to the keyframe samples $E = \{e_1, e_2, \ldots, e_N\}$ , sampling size $K$ + +Ensure: Sampled set $S = \{s_1, s_2, \ldots, s_K\}$ + +1: Calculate the distance array $A$ , where $A[i][j] = \text{distance}(e_i, e_j)$ +2: Select the sample $i$ and add to $S$ , where $i = \arg \max_{1 \leq i \leq N} \sum_{j=1}^{N} A[i][j]$ +3: for $k = 2$ to $K$ do +4: Find the value in $E$ that has the largest difference with the values in $S$ , + +$$ +j = \operatorname *{arg max}_{j\in E}\sum_{k\in S}A[j][k] +$$ + +5: Add sample $j$ to $S$ as the $k$ -th sample + +6: end for +7: return Sampled set $S$ + +Extendable PerceiverIO. Unlike traditional vision classification tasks, robotic manipulation requires the integration of multiple modalities to enable complex decision-making and long-term action planning through interactions with the physical environment. Classic methods [7, 8, 19, 46] for lifelong classification overlook the complexities of action in robotic manipulation tasks that require the agent to learn various action primitives for different skills. In our iManip framework, we propose an extendable PerceiverIO, which learns skill-specific action prompts with extendable weights to adapt to new action primitives. + +Specifically, as illustrated in Figure 3, the input to the extendable PerceiverIO consists of multimodal patches, $X = [X_{\mathrm{voxel}}, X_{\mathrm{language}}, X_{\mathrm{action}}]$ , where $X_{\mathrm{voxel}}$ and $X_{\mathrm{language}}$ represent the input sequences of voxel and language encodings, respectively, and $X_{\mathrm{action}} = [X_{\mathrm{action}}^{\mathrm{old}}, X_{\mathrm{action}}^{\mathrm{new}}]$ is the skill-specific action prompt that concatenates both the old and new action prompts. The notation $[\cdot, \cdot]$ denotes the concatenation operation along the token dimension. Subsequently, $X$ undergoes a cross-attention computation between the input and a much smaller set of latent vectors through the latent encoder, producing $X_{\mathrm{latent}}$ . Then $X_{\mathrm{latent}}$ is encoded with weight-extendable self-attention layers as + +$$ +\begin{array}{l} Q = X _ {\text {l a t e n t}} \cdot W _ {Q} ^ {\text {s c a l e}}, \quad K = X _ {\text {l a t e n t}} \cdot W _ {K} ^ {\text {s c a l e}}, \tag {3} \\ V = X _ {\text {l a t e n t}} \cdot W _ {V}, \\ \end{array} +$$ + +$$ +X _ {\text {a t t}} = \operatorname {s o f t m a x} \left[ \frac {Q \cdot K ^ {\top}}{\sqrt {d}} \right] \cdot V, \tag {4} +$$ + +where $X_{\mathrm{latent}}, X_{\mathrm{att}} \in \mathbb{R}^{T \times d}$ are respectively a set of $T$ input and output tokens with channel dimension $d$ , $W_{Q}^{\mathrm{scale}}, W_{K}^{\mathrm{scale}} \in \mathbb{R}^{d \times d'}$ , $W_{V} \in \mathbb{R}^{d \times d}$ are learnable weight matrices. Notably, $W_{Q}^{\mathrm{scale}}, W_{K}^{\mathrm{scale}}$ are extendable by appending newly weight matrices $W_{Q}^{\mathrm{new}}$ , $W_{K}^{\mathrm{new}} \in \mathbb{R}^{d \times d_{\mathrm{new}}}$ as + +$$ +W _ {Q} ^ {\text {s c a l e}} = \left[ W _ {Q} ^ {\text {o l d}}, W _ {Q} ^ {\text {n e w}} \right], \quad W _ {K} ^ {\text {s c a l e}} = \left[ W _ {K} ^ {\text {o l d}}, W _ {K} ^ {\text {n e w}} \right], \tag {5} +$$ + +where $W_{Q}^{\mathrm{old}}, W_{K}^{\mathrm{old}} \in \mathbb{R}^{d \times d_{\mathrm{old}}}$ are the old weight matrices and the expanded dimension $d' = d_{\mathrm{old}} + d_{\mathrm{new}}$ . Finally, these encoded latents are cross-attended with the input once again through the latent decoder to ensure alignment with the input size. In our iManip framework, we freeze the old PerceiverIO while learning the action prompts $X_{\mathrm{action}}^{\mathrm{new}}$ and a small set of newly weight matrices $W_{Q}^{\mathrm{new}}, W_{K}^{\mathrm{new}}$ of new skills. This enables the agent to quickly adapt to new action primitives while preventing the forgetting of previous skills. + +Knowledge distillation between the old and new agents. To better preserve the knowledge of previous skills while learning new ones, we employ knowledge distillation [7, 64], where the output probability distribution of the old model is used to train the new model. This enables the transfer of knowledge from the old to the new agent, as defined by the following objective: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d i s}} = \mathcal {L} _ {2} \left(\mathcal {Q} _ {\text {t r a n s}} ^ {\text {o l d}}, \mathcal {Q} _ {\text {t r a n s}} ^ {\text {n e w}}\right) + \mathcal {L} _ {2} \left(\mathcal {Q} _ {\text {r o t}} ^ {\text {o l d}}, \mathcal {Q} _ {\text {r o t}} ^ {\text {n e w}}\right) + \\ \left| \mathcal {Q} _ {\text {o p e n}} ^ {\text {o l d}} - \mathcal {Q} _ {\text {o p e n}} ^ {\text {n e w}} \right| + \left| \mathcal {Q} _ {\text {c o l l i d e}} ^ {\text {o l d}} - \mathcal {Q} _ {\text {c o l l i d e}} ^ {\text {n e w}} \right|, \\ \end{array} +$$ + +where $\mathcal{L}_2$ is the MSE loss, $[\mathcal{Q}_{\mathrm{trans}}^{\mathrm{old}},\mathcal{Q}_{\mathrm{rot}}^{\mathrm{old}},\mathcal{Q}_{\mathrm{open}}^{\mathrm{old}},\mathcal{Q}_{\mathrm{collide}}^{\mathrm{old}}]$ and $[\mathcal{Q}_{\mathrm{trans}}^{\mathrm{new}},\mathcal{Q}_{\mathrm{rot}}^{\mathrm{new}},\mathcal{Q}_{\mathrm{open}}^{\mathrm{new}},\mathcal{Q}_{\mathrm{collide}}^{\mathrm{new}}]$ denote the probabilities of the ground truth actions in expert demonstrations for translation, rotation, gripper openness, and collision avoidance for the old and new robots, respectively. + +# 3.4. Learning Objectives + +Our approach is to address the problem of skill-incremental learning for robotic manipulation from multiple aspects. First, for each manipulation skill, there is an action loss to facilitate robot imitation learning. Following [37, 51, 67], we employ cross-entropy loss (CE) to ensure accurate action prediction: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {a c t}} = - \mathbb {E} _ {Y _ {\text {t r a n s}}} [ \log \mathcal {V} _ {\text {t r a n s}} ] - \mathbb {E} _ {Y _ {\text {r o t}}} [ \log \mathcal {V} _ {\text {r o t}} ] \tag {7} \\ - \mathbb {E} _ {\mathcal {Y} _ {\text {o p e n}}} \left[ \log \mathcal {V} _ {\text {o p e n}} \right] - \mathbb {E} _ {\mathcal {Y} _ {\text {c o l l i d e}}} \left[ \log \mathcal {V} _ {\text {c o l l i d e}} \right], \\ \end{array} +$$ + +where $\mathcal{V}_i = \mathrm{softmax}(\mathcal{Q}_i)$ for $\mathcal{Q}_i\in [\mathcal{Q}_{\mathrm{trans}},\mathcal{Q}_{\mathrm{open}},\mathcal{Q}_{\mathrm{rot}},$ $\mathcal{Q}_{\mathrm{collide}}]$ and $Y_{i}\in [Y_{\mathrm{trans}},Y_{\mathrm{rot}},Y_{\mathrm{open}},Y_{\mathrm{collide}}]$ is the ground truth one-hot encoding. + +Furthermore, when learning new skills, we propose the temporal replay strategy to preserve a fixed number of representative samples from old demonstrations. The cached memory $\mathcal{M}$ will be used in conjunction with the new skill demos $\mathcal{D}_{\mathrm{new}}$ for learning new skills. Additionally, our extendable PerceiverIO will dynamically expand new learnable weights for the new skill. We find that training only the skill-specific action prompts $X_{\mathrm{action}}^{\mathrm{new}}$ with newly appended weights $W_{Q}^{\mathrm{new}}$ , $W_{K}^{\mathrm{new}}$ and the policy decoder effectively prevents catastrophic forgetting, more analysis can be seen in the 3rd experiments in Section 4.2. Finally, we employ knowledge distillation loss $\mathcal{L}_{\mathrm{dis}}$ to help the agent retain the knowledge of previous skills. Overall, in skill-incremental + +
MethodsBaseStep 1Step 2Step 3Step 4Step 5Average
OldAllOldAllOldAllOldAllOldAllOldAll
multi-task methods
PerAct [51]44.04.07.32.75.11.19.02.56.71.31.62.35.9
ManiGaussian [37]55.212.020.76.712.05.715.53.09.35.35.26.512.5
skill-incremental methods
P-TIB [7, 46, 51]44.033.634.726.025.122.326.017.016.411.610.422.122.5
M-TIB [7, 37, 46]55.242.445.336.737.134.339.531.031.629.826.834.836.1
Ours (iManip)56.057.656.750.748.045.147.542.039.138.736.046.845.5
+ +learning, our training loss is formulated as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {a c t}} + \lambda_ {\text {d i s}} \mathcal {L} _ {\text {d i s}}, \tag {8} +$$ + +where $\lambda_{\mathrm{dis}}$ is a hyperparameter that controls the importance of the knowledge distillation loss $\mathcal{L}_{\mathrm{dis}}$ during training. + +# 4. Experiments + +# 4.1. Experimental setup and details + +Experimental setup. Following [37, 67], we select 10 representative manipulation skills in RLBench[23] and 5 daily manipulation skills in the real world for our experiments. Each skill has at least two variations and 20 demonstrations during training that cover multiple types, such as position, shape, and color. To achieve a high success rate for these skills, the manipulation policy needs to learn generalizable knowledge rather than overfitting the limited given demonstrations. For visual observation, we only use the front RGB-D image with $128 \times 128$ resolutions. To demonstrate the performance of our method under different incremental settings, we define several configurations, represented as Bn-kNm. This notation indicates that the policy is initially trained on n base skills, followed by the addition of m new skills in each step, with a total of k steps. + +Evaluation Metric. We report the performance of the agent on each learned skill by the average success rate. At each incremental step, we present the average success rate for old, new, and all (combined old and new) skills. In the simulation, we evaluate the agent with 25 episodes per skill, whereas in the real world, we use 10 episodes per skill. During evaluation, the agent continues to take actions until an oracle signals task completion or the agent reaches a maximum of 25 steps. + +Implementation Details. For model design, we use different encoders to transform corresponding modality data into tokens, which serve as the input for the Extendable PerceiverIO. The RGB-D images are projected and transformed into voxels, which are then encoded by a 3D convolutional encoder with a UNet architecture, while text instructions are encoded using CLIP RN50 [45], and the proprioception data is encoded by a single-layer MLP. After + +Table 1. Performance comparison of different methods of B5-5N1 in Rlbench. We show the average success rate of old and all learned skills, and the average performance of all new steps. The traditional incremental methods [7, 46] on baseline [37, 51] is termed TIB. + +
TRSEPIODISB5-1N1B5-5N1
R120.75.2
R2✓49.327.6
R3✓✓54.032.4
Ours✓✓✓56.736.0
+ +Table 2. Ablation Study on two experiment setup. We report the average success rate of all learned skills. + +encoding, the tokens from all modalities have the same dimension of 512. The hyperparameter $\lambda_{\mathrm{dis}}$ is set as 0.01 and the action prompt length is 16. We store 2 keyframe replays of the total 20 demonstrations of learned skills and train the agent on two NVIDIA RTX 4090 GPUs with a batch size of 1, a learning rate of 0.002, and 100k iterations. More studies about the hyperparameters are shown in the Appendix. + +# 4.2. Simulation results + +Performance comparison with different methods. We conduct the skill-incremental learning experiment in the B5-5N1 setting, where we first train the policy on five base skills and then gradually learn a new skill at each subsequent step with a total of five steps. + +To demonstrate the performance of our method in robotic skill-incremental learning, we compare with two standard multi-task manipulation policies, PerAct [51] and Mani-Gaussian [37], by retraining the agent to learn new manipulation skills. Furthermore, we apply two Traditional Incremental Baselines (TIB) for visual classification to the above policies for comparison, termed P-TIB and M-TIB respectively. As shown in Table 1, the results demonstrate that the performance of our method significantly outperforms the others at each subsequent step. This demonstrates that our method better facilitates the learning of new skills while mitigating the forgetting of previous skills. More detailed results of each learning skill at every step are shown in the Appendix. + +Ablation Study. We conduct the ablation study to validate the effectiveness of each policy, as shown in Table 2. R1 is the control group where the agent does not have any incremental policy. When we add the Temporal Replay Strategy + +
Frozen layerConvergence stepsTrained paramSlide blockPut in drawerDrag stickPush buttonsStack blocks
OldNewOldNewOldNewOldNewOldNew
Non-frozen10000047M43.260.044.816.040.892.044.028.045.612.0
Encoder7500037M50.454.051.212.045.688.048.424.052..08.0
EPIO7000018M52.052.052.816.046.484.050.424.049.68.0
Decoder7500039M45.624.047.20.042.440.044.84.046.40.0
Encoder+EPIO600008M57.652.056.812.050.484.055.220.056.08.0
+ +Table 3. Performance of five sets B5-1N1 experiments with the same base skills and different new skills, while freezing different network layers. For each new skill, we train a total of 100k iterations and report the average success rate and the average model convergence steps. + +
MethodsB5-1N5B2-4N2B3-2N3
ManiGaussian [37]25.610.417.3
M-TIB [7, 37, 46]30.828.433.3
Ours (iManip)37.236.841.3
+ +Table 4. Average success rate of all learned skills on different skill-incremental setup. + +(TRS) to the agent, the setup of B5-1N1 and B5-5N1 improve by $28.6\%$ and $22.4\%$ in the success rate, respectively (see R2). The significant performance improvement stems from the success of our temporal replay strategy that maintains the integrity of the temporal data. + +When we add Extendable PerceiverIO (EPIO) to the agent, the success rates of B5-1N1 and B5-5N1 further improve by $4.7\%$ and $4.8\%$ , respectively (see R3). The EPIO design works because the skill-specific action prompts help the agents incrementally learn action primitives for new skills, and the extendable weights designed in transformer blocks allow the model to preserve the old knowledge while adapting to new skills. Our complete policy achieves the best success rate, where the Distillation mechanism (DIS) improves the performance by $2.7\%$ and $3.6\%$ , respectively. + +Through the ablation study, we find that the replay policy has the greatest impact on overall performance. Without old data for retraining, the agent is more likely to forget previously learned knowledge. This occurs because data plays a crucial role in robotic manipulation, and without the support of previous data, the agent is prone to overfitting the data of new skills. + +Effect of parameter freezing on skill-incremental learning. The agent consists of three main components: the encoders, the extendable PerceiverIO, and the policy decoder. we freeze each component individually to evaluate its effect. As shown in Table 3, five sets B5-1N1 experiments on different new skills demonstrate the following: (1) Freezing the encoders or the extendable PerceiverIO helps retain knowledge from the old skills without significantly hindering the learning of the new skill (Lines 1,2,3). (2) The policy decoder is crucial for learning new skills (Lines 1,4). (3) Freezing the above components helps decrease the number of parameters and accelerate convergence. + +![](images/8be5d9e6aa45629464b497996141ab6ef5c57c585f78c107da5806e844e8548a.jpg) +Figure 4. Average success rate on different replay methods. + +Based on these findings, we freeze both the encoder and the extendable PerceiverIO, leaving only the decoder with newly appended action prompts and weights for new skill training, achieving faster convergence, fewer parameters, and better performance (Line 5)! + +Experiments on different skill-incremental setup. We implement different skill-incremental settings to validate the generalizability of our method and report the average success rate across all previously learned skills after the last incremental step. As shown in Table 4, compared to Mani-Gaussian and M-TIB, our method achieves higher success rates in all incremental experimental settings. This shows that our approach has stronger performance and better generalization capabilities. + +Exploring data replay methods. We compare our temporal replay strategy with the classical rehearsal-based method on the setup of B5-1N1, as shown in Figure 4. Herding [46] and Hardsample [8] are two methods for selecting the most representative samples from the data. Episode refers to replaying a complete trajectory. Random refers to random sampling. The results show that classic herding sampling and hard-exemplar sampling perform poorly on old skills due to neglecting temporal integrity in robotic demonstrations. In contrast, replaying complete trajectories or random sampling better preserves the temporal integrity of the samples, leading to better performance. Our temporal replay strategy, leveraging the farthest-distance entropy sampling for each keyframe can sample more different variants and achieves the best success rate. + +
Manipulation skillsBaseStep 1Step 2Step 3Step 4
BLOursBLOursBLOursBLOursBLOurs
Slide toy90.090.010.080.0080.0060.0060.0
Open drawer--70.060.0060.0050.0040.0
Pick and place----60.060.0060.0050.0
Pour water------40.040.0010.0
Close jar--------50.040.0
Old manipulation skills90.090.010.080.0 +70.0070.0 +70.0056.7 +56.7040.0 +40.0
All manipulation skills90.090.040.070.0 +30.020.066.7 +46.710.052.5 +42.510.040.0 +30.0
+ +Table 5. Real world experiments. The table reports the success rate of BaseLine (BL) and Ours. $^{+}\mathrm{num}$ is the improvement of our method compared to the baseline. + +![](images/72510ab72b7257097afdda261ff23b985de09f8e1e96f409afc630d98e291b26.jpg) +Figure 5. Visualization of skill-specific action prompts by Grad-CAM as the agent executes two manipulation skills: close jar (the first row) and slide block (the second row). + +![](images/3addfa99c55557951d287571c934840b0c234fcab7499cac625b3345223c2967.jpg) + +Visualization of the skill-specific action prompts. We visualize our skill-specific action prompts by Grad-CAM [48], as shown in Figure 5. Our experiments train the agent in the B5-5N1 setting with totaling 10 action prompts. We first compute the sum of the parameter gradients for each action prompt and then normalize these values to calculate Grad-CAM weights, displayed in different colors. We show the results across two different skills. When the agent executes the same action, e.g., "move down to the target", the third action prompt weight is maximized (see the first column). Furthermore, when performing different actions, different weights of skill-specific action prompts are maximized (see the second column). It demonstrates that action prompts can learn skill-specific action primitives. Freezing old action prompts while learning new ones helps prevent forgetting and adapt to new action primitives. + +# 4.3. Real world experiments + +We conduct five manipulation skills in the real-world environment to further validate the effectiveness of our method. We use a Franka Panda robotic arm to execute the action and a Realsense D455 camera to capture RGB-D images as + +observations. For each training step, we collect 20 demonstrations. The training setup is B1-4N1 where we first train on a base skill, then incrementally add one new skill at a time for training, with a total of four new skills added. During testing, we perform 10 test runs for each learned skill and report the success rate of task execution. More details about the real world experiments and videos are shown in the supplementary material. + +We compare our method with the baseline [37] without any incremental policy. As shown in Table 5, it is evident that without the incremental strategy, the knowledge of previous skills is rapidly forgotten when training on new skills. After incorporating our incremental strategy, the success rate on new skills is lower than the baseline. This occurs because the baseline has overfitted to the new skill, while our model, which is designed to retain knowledge from previous skills, experiences a slight decrease in its ability to learn new skills. This trade-off is an inherent challenge in incremental learning. + +# 5. Conclusion + +In this work, we focus on a new and challenging setting, skill-incremental learning in robotic manipulation, which is to continually learn new skills while maintaining the previously learned skills. We conduct experiments on the RL-Bench benchmark and find that traditional methods suffer from catastrophic forgetting because they overlook the temporal and action complexities of robotic manipulation. Our approach proposes a temporal replay strategy to address the temporal complexities and an extendable PerceiverIO model with adaptive action prompts to address the action complexities. Extensive experiments demonstrate that our iManip framework excels in effectiveness, robustness, lightweight design, and extendability. + +# Acknowledgements + +This work was supported partially by NSFC (92470202, U21A20471), Guangdong NSF Project (No.2023B1515040025), Guangdong Key Research and Development Program (No.2024B0101040004). + +# References + +[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2 +[2] Arjun Ashok, KJ Joseph, and Vineeth N Balasubramanian. Class-incremental learning with cross-space clustering and controlled transfer. In Proceedings of the European Conference on Computer Vision, 2022. 2 +[3] Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 4 +[4] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. $pi\_0$ : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 1 +[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2 +[6] Jia-Feng Cai, Zibo Chen, Xiao-Ming Wu, Jian-Jian Jiang, Yi-Lin Wei, and Wei-Shi Zheng. Real-to-sim grasp: Rethinking the gap between simulation and real world in grasp detection. 2024. 1 +[7] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision, 2018. 2, 3, 5, 6, 7 +[8] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision, 2018. 2, 3, 4, 5, 7 +[9] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 2 +[10] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 1 +[11] Shizhe Chen, Ricardo Garcia, Cordelia Schmid, and Ivan Laptev. Polarnet: 3d point clouds for language-guided robotic manipulation. arXiv preprint arXiv:2309.15596, 2023.2 +[12] Xuxin Cheng, Jialong Li, Shiqi Yang, Ge Yang, and Xiaolong Wang. Open-television: Teleoperation with immersive active visual feedback. arXiv preprint arXiv:2407.01512, 2024. 2 +[13] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. + +Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, 2023. 2 +[14] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the European Conference on Computer Vision, 2020. 2 +[15] Chongkai Gao, Haichuan Gao, Shangqi Guo, Tianren Zhang, and Feng Chen. Cril: Continual robot imitation learning via generative and prediction model. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2021. 2 +[16] Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: 3d feature field transformers for multi-task robotic manipulation. arXiv preprint arXiv:2306.17817, 2023. 2 +[17] Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. In Conference on Robot Learning, 2023. +[18] Pierre-Louis Guhur, Shizhe Chen, Ricardo Garcia Pinel, Makarand Tapaswi, Ivan Laptev, and Cordelia Schmid. Instruction-driven history-aware policies for robotic manipulations. In Conference on Robot Learning, 2023. 2 +[19] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2, 4, 5 +[20] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[21] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 1 +[22] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, 2021. 2 +[23] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 2020. 2, 6 +[24] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In Conference on Robot Learning, 2022. 2 +[25] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[26] Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations. arXiv preprint arXiv:2402.10885, 2024. 2 + +[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2 +[28] Khimya Khetarpal, Shagun Sodhani, Sarath Chandar, and Doina Precup. Environments for lifelong reinforcement learning. arXiv preprint arXiv:1811.10732, 2018. 2 +[29] Byeonghwi Kim, Minhyuk Seo, and Jonghyun Choi. Online continual learning for interactive instruction following agents. arXiv preprint arXiv:2403.07548, 2024. 2 +[30] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. 2024. 1 +[31] Kibok Lee, Kimin Lee, Jinwoo Shin, and Honglak Lee. Overcoming catastrophic forgetting with unlabeled data in the wild. In Proceedings of the IEEE/CVF international conference on computer vision, 2019. 2 +[32] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In International conference on machine learning, 2019. 2 +[33] Yuan-Ming Li, Ling-An Zeng, Jing-Ke Meng, and Wei-Shi Zheng. Continual action assessment via task-consistent score-discriminative feature distribution modeling. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2 +[34] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 2017. 2 +[35] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 2024. 1, 2 +[36] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 2021. 2 +[37] Guanxing Lu, Shiyi Zhang, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation. In Proceedings of the European Conference on Computer Vision, 2024. 2, 3, 5, 6, 7, 8 +[38] Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multitask robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1 +[39] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 2022. 2 +[40] Jorge Mendez, Shashank Shivkumar, and Eric Eaton. Life-long inverse reinforcement learning. Advances in neural information processing systems, 2018. 2 +[41] Yuan Meng, Zhenshan Bing, Xiangtong Yao, Kejia Chen, Kai Huang, Yang Gao, Fuchun Sun, and Alois Knoll. Pre + +serving and combining knowledge in robotic lifelong reinforcement learning. Nature Machine Intelligence, 2025. 2 +[42] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2 +[43] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2 +[44] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. Advances in Neural Information Processing Systems, 2021. 2 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 6 +[46] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 4, 5, 6, 7 +[47] Hyunwoo Ryu, Jiwoo Kim, Hyunseok An, Junwoo Chang, Joohwan Seo, Taehan Kim, Yubin Kim, Chaewon Hwang, Jongeun Choi, and Roberto Horowitz. Diffusion-eds: Bi-equivariant denoising generative modeling on se (3) for visual robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1 +[48] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF international conference on computer vision, 2017. 8 +[49] Lin Shao, Toki Migimatsu, Qiang Zhang, Karen Yang, and Jeannette Bohg. Concept2robot: Learning manipulation concepts from instructions and human demonstrations. The International Journal of Robotics Research, 2021. 2 +[50] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on robot learning, 2022. 2 +[51] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, 2023. 2, 3, 4, 5, 6 +[52] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2 +[53] Simon Stepputtis, Joseph Campbell, Mariano Phiplipp, Stefan Lee, Chitta Baral, and Henri Ben Amor. Language-conditioned imitation learning for robot manipulation tasks. Advances in Neural Information Processing Systems, 2020. 2 + +[54] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. Learning to imagine: Diversify memory for incremental learning using unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[55] Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In European Conference on Computer Vision, 2024. 2 +[56] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[57] Yoshihisa Tsurumine and Takamitsu Matsubara. Goal-aware generative adversarial imitation learning from imperfect demonstration for robotic cloth manipulation. Robotics and Autonomous Systems, 2022. 1 +[58] An-Lan Wang, Nuo Chen, Kun-Yu Lin, Li Yuan-Ming, and Wei-Shi Zheng. Task-oriented 6-dof grasp pose detection in clutters. arXiv preprint arXiv:2502.16976, 2025. 2 +[59] Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xian-Tuo Tan, Xiao-Ming Wu, Hao Li, Mark Cutkosky, and Wei-Shi Zheng. Grasp as you say: Language-guided dexterous grasp generation. arXiv preprint arXiv:2405.19291, 2024. +[60] Yi-Lin Wei, Mu Lin, Yuhao Lin, Jian-Jian Jiang, Xiaoming Wu, Ling-An Zeng, and Wei-Shi Zheng. Afforddexgrasp: Open-set language-guided dexterous grasp with generalizable-instructive affordance. arXiv preprint arXiv:2503.07360, 2025. 2 +[61] Maciej Wolczyk, Michal Zajac, Razvan Pascanu, Łukasz Kuciński, and Piotr Miłos. Continual world: A robotic benchmark for continual reinforcement learning. Advances in Neural Information Processing Systems, 2021. 2 +[62] Philipp Wu, Yide Shentu, Zhongke Yi, Xingyu Lin, and Pieter Abbeel. Gello: A general, low-cost, and intuitive teleoperation framework for robot manipulators. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2024. 2 +[63] Xiao-Ming Wu, Jia-Feng Cai, Jian-Jian Jiang, Dian Zheng, Yi-Lin Wei, and Wei-Shi Zheng. An economic framework for 6-dof grasp detection. In European Conference on Computer Vision, 2024. 2 +[64] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 5 +[65] Zhou Xian and Nikolaos Gkanatsios. Chaineddiffuser: Unifying trajectory diffusion and keypose prediction for robotic manipulation. In Conference on Robot Learning, 2023. 2 +[66] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 2 +[67] Yanjie Ze, Ge Yan, Yueh-Hua Wu, Annabella Macaluso, Yuying Ge, Jianglong Ye, Nicklas Hansen, Li Erran Li, and Xiaolong Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields. In Conference on Robot Learning, 2023. 2, 3, 5, 6 + +[68] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Co-transport for class-incremental learning. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2 +[69] Jiaming Zhou, Teli Ma, Kun-Yu Lin, Zifan Wang, Ronghe Qiu, and Junwei Liang. Mitigating the human-robot domain discrepancy in visual pre-training for robotic manipulation. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 1 \ No newline at end of file diff --git a/imanipskillincrementallearningforroboticmanipulation/images.zip b/imanipskillincrementallearningforroboticmanipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ecda9b0ab2464a6131d48d8c8ee30966c9e0d97a --- /dev/null +++ b/imanipskillincrementallearningforroboticmanipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5701c1c488b71adda6634664e6af68fa627e95a3263908be222fe32f0d9fad31 +size 712637 diff --git a/imanipskillincrementallearningforroboticmanipulation/layout.json b/imanipskillincrementallearningforroboticmanipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..809d70a39bc24cf3edbd4318a5204cb7d220ed98 --- /dev/null +++ b/imanipskillincrementallearningforroboticmanipulation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b34b2a89dbc8be731323f75dea43d1a4a7e64decc2483a75f953e0711d9c6dfc +size 401384 diff --git a/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/83403940-00e7-485d-b6f5-2066a9a36d93_content_list.json b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/83403940-00e7-485d-b6f5-2066a9a36d93_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fb37aa70b35b1af730854f4ea619f1836e71c5e2 --- /dev/null +++ b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/83403940-00e7-485d-b6f5-2066a9a36d93_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d849d2a2f4c2974284d024bfe3cebc75876d4b6a9fcbb0fcbf77dfd2cc39b1c +size 82060 diff --git a/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/83403940-00e7-485d-b6f5-2066a9a36d93_model.json b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/83403940-00e7-485d-b6f5-2066a9a36d93_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99239c1a06897275a777bc611f7645bef8726384 --- /dev/null +++ b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/83403940-00e7-485d-b6f5-2066a9a36d93_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9e46caf4d76fd0ca2f9e8896959e5fb0c05cc9b5496df0cf5cef127c4412b79 +size 107320 diff --git a/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/83403940-00e7-485d-b6f5-2066a9a36d93_origin.pdf b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/83403940-00e7-485d-b6f5-2066a9a36d93_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..31b797ee9ce1e4c0549c65c229d938da60371e84 --- /dev/null +++ b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/83403940-00e7-485d-b6f5-2066a9a36d93_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8b75e21c9200cd9a9367d4b05973dd9368dec8dfb9a1868645e8830a09561a4 +size 11511121 diff --git a/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/full.md b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1bce31fc178ef1ddd61bb90bfe882bc18da49fe3 --- /dev/null +++ b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/full.md @@ -0,0 +1,377 @@ +# $\chi$ : Symmetry Understanding of 3D Shapes via Chirality Disentanglement + +Weikang Wang* + +Tobias Wei $\text{堡}$ rg\* University of Bonn + +Nafie El Amrani +Lamarr Institute + +Florian Bernard + +![](images/85d6d33df26ff04c521ef04252a1613c1b5ee1c9ec2465b67bd912871c32dd8f.jpg) +Figure 1. Left: Our chirality features can disentangle the left and right parts of 3D shapes. By leveraging the generalisation abilities of foundation image models, our method remains effective even on partial shapes. Right: Our chirality features resolve left-right ambiguities in 3D shape matching by augmenting state-of-the-art vertex features like Diff3F [16] with structural information. + +# Abstract + +Chirality information (i.e. information that allows distinguishing left from right) is ubiquitous for various data modes in computer vision, including images, videos, point clouds, and meshes. While chirality has been extensively studied in the image domain, its exploration in shape analysis (such as point clouds and meshes) remains underdeveloped. Although many shape vertex descriptors have shown appealing properties (e.g. robustness to rigid-body transformations), they are often not able to disambiguate between left and right symmetric parts. Considering the ubiquity of chirality information in different shape analysis problems and the lack of chirality-aware features within current shape descriptors, developing a chirality feature extractor becomes necessary and urgent. Based on the recent Diff3F framework [16], we propose an unsupervised chirality feature extraction pipeline to decorate shape vertices with chirality-aware information, extracted from 2D foundation models. We evaluated the extracted chirality features through quantitative and qualitative exper + +iments across diverse datasets. Results from downstream tasks including left-right disentanglement, shape matching, and part segmentation demonstrate their effectiveness and practical utility. Project page: https://wei-kang-wang.github.io/chirality/ + +# 1. Introduction + +Symmetry and chirality are two sides of the same coin: Symmetry highlights the similarities between two parts, whereas chirality focuses on the differences between them. As a broadly applicable and intuitive assumption for many objects in the physical world, symmetry has been extensively studied in various computer vision domains for a long time, including 3D reconstruction [57, 58], pose estimation [10, 37, 43, 68], and generative models [1, 60]. However, chirality, despite its close relationship to symmetry, remains less explored and has just re-attracted researchers' attention in recent years [32, 34, 52, 62, 70]. + +In the field of shape analysis, symmetry and chirality also play an important role in many problems, including matching [11, 33, 64, 67], deformation [69], symmetry plane detection [44], etc. For many shape analysis prob + +![](images/5e2be48211a2e38ae5125f0f2bc063358f8bf7b46a45f7b7eba861624b258632.jpg) +SD+DINO +(a) Left/Right + +![](images/f7e598aba66d8633ca3cec163a63edca590de082381e2337d373632a57ea701c.jpg) +Source +(b) Shape matching +Figure 2. Left: The 2-center clustering of vertex features aggregated from StableDiffusion and DINO [40, 49] features does not lead to the desired chirality indicator. Right: Shapes matched using Diff3F [16] display a lack of chirality awareness. + +![](images/59e5a5bd70a29b96e6174f10022a6fab566dd28328e52e9e819d5635abc6ff0b.jpg) +Target + +![](images/36c7b9b0c9fbe1db01aa51cfdc704c5b462204411c4b7eb71215d1c11eb6de2f.jpg) +Diff3F + +lems, vertex feature descriptors play a central role, and various methods to generate feature descriptors have been proposed [4, 8, 50]. Recently, Diff3F [16] was introduced as a novel framework for generating vertex feature descriptors. It aggregates 2D foundation model features from rendered multi-view images of a shape to create vertex feature descriptors. While recent features have shown semantic and geometric robustness, none of them are able to disambiguate between left-and-right symmetric shape parts, which could cause severe problems, as shown in Fig. 2. + +In order to fill this gap, we propose an unsupervised method based on Diff3F [16] that disentangles left-and-right (chirality) information to decorate shapes with chirality-aware vertex descriptors using features aggregated from 2D foundation models (such as DINO-V2 [40] and StableDiffusion [49]). We conduct experiments on various datasets and show that our extracted chirality feature provides left-and-right information of a shape. We show that left-and-right ambiguities, which are a common issue of state-of-the-art shape matching methods [8, 16], can be effectively mitigated by combining our chirality feature with other vertex feature descriptors. The generalization experiments of our pre-trained model on partial shapes and anisotropic shapes additionally prove the robustness of our chirality feature extractor. + +To summarize, our contributions are as follows: + +- We propose an unsupervised method to extract left-and-right (chirality) information for decorating shape vertex descriptors. Left-and-right disentanglement experiments on various datasets show the validity of our extracted chirality features quantitatively and qualitatively. +- For shape matching and part segmentation, we demonstrate that augmenting standard vertex descriptors with our chirality features effectively resolves left-right ambiguities across multiple datasets. +- Finally, cross-dataset and cross-category generalisation experiments, particularly those conducted on partial and anisotropic shapes, demonstrate the robustness of our trained chirality feature extractor. + +# 2. Related work + +We briefly review related work on chirality (Sec. 2.1), feature extraction (Sec. 2.2), and shape matching (Sec. 2.3). + +# 2.1. Chirality in visual computing + +In the context of studies on images, Lin et al. [32] explored how the statistics of visual data change under reflection and coined the term visual chirality. Yeh et al. [62] proposed a network whose layers can be mirror-flipped, using it for human pose estimation by resolving left/right ambiguities in the human body. Zheng et al. [70] investigated the vertical flipping associated with visual chirality in freehand sketches. Moreover, Tan et al. [52] leveraged the property of image-level visual chirality and reformulated it as a learnable pixel-level cue for mirror detection. In facial analysis, Lo et al. [34] employed chirality information from the left and right halves of faces to learn robust facial embeddings for expression recognition. Recently, to obtain geometry-aware features, Zhang et al. [65] fine-tuned a model on aggregated DINO-V2 $[40] + \mathrm{SD}$ [49] features with labelled keypoint correspondences between image pairs, inherently capturing chirality information. + +In the field of shape analysis, chirality extraction is closely related to the detection of intrinsic or extrinsic shape symmetries. Tevs et al. [53] showed that finding correspondences between shapes of widely varying geometry could benefit from extrinsic symmetry. Je et al. [26] applied Langevin dynamics within a redefined symmetry space to enhance robustness of extrinsic symmetry detection against noise. E3Sym [31] established robust point correspondences using E(3)-invariant features extracted from a lightweight neural network, enabling dense symmetry predictions. Ovsjanikov et al. [41] utilized Euclidean symmetries in the signature space defined by the eigenfunctions of the Laplace-Beltrami operator to identify intrinsic symmetries. Kim et al. [28] modelled intrinsic symmetry as Möbius transformations derived from critical points of the Average Geodesic Distance (AGD) function. Liu et al. [33] detected intrinsic symmetry on genus-zero mesh surfaces by extracting closed curves through conformal maps based on triplets of extrema points identified via the AGD function. Additionally, Nagar and Raman [38] hypothesized that if a shape was intrinsically symmetric, the shortest geodesic between two symmetric points was also intrinsically symmetric, facilitating the extraction of intrinsic correspondences. + +However, to our best knowledge, there exist no methods in the shape analysis field to extract chirality-aware vertex feature descriptors. Based on the recent paper Diff3F [16], which aggregated features from 2D foundation models (DINO-V2 [40] and StableDiffusion [49]) to obtain vertex descriptors, we propose an unsupervised pipeline to distil chirality information from features extracted by 2D foundation models to decorate vertex feature descriptors. + +# 2.2. Feature descriptors from 2D foundation models + +In recent years, various 2D foundation models have been proposed, including DINO [9], DINO-V2 [40], CLIP [45], StableDiffusion (SD) [49], etc. They have been used in various 2D/3D areas as feature descriptors and have surpassed both handcrafted features and other deep features, due to their rich semantic and geometric information obtained from large-scale and/or multi-modal datasets. Hedlin et al. [22] treated SD features as pixel-level feature descriptors to do semantic correspondence, while Hedlin et al. [23] leveraged SD features as an unsupervised 2D keypoint detector. FeatureNeRF [61] learned generalizable NeRFs by distilling pre-trained vision foundation models to perform on other downstream tasks beyond synthesis, such as semantic understanding and parsing. In this paper, similar to the setting in Diff3F [16] (see Sec. 3.1 for more details), we also aggregate per vertex features from 2D foundation model features of multiple views. However, we additionally derive chirality information embedded in the 2D foundation model features to augment vertex descriptors, enabling them to be left-right aware and consistent across shapes. + +# 2.3. Shape matching + +Shape matching is a well-studied problem in computer vision and graphics [51, 54], aimed at finding correspondences between pairs of shapes. Classical approaches solve this problem by establishing correspondences based on geometric relations [24, 48], while other methods rely on non-rigid shape registration [5, 19, 25]. One prominent mesh-based approach, the functional map framework [42], encodes shape correspondences into a compact matrix using truncated basis functions, usually the first $k$ Laplacian eigenfunctions [30]. Although it has been extended to handle non-isometries [39, 46], partial shapes [3, 59] or multishape matching [8, 20], its reliance on spectral basis calculation involves high computational costs, pre-processing, mesh connectivity, and lacks semantic consideration. + +An alternative line of works focuses on spatial approaches and operates on point clouds directly, such as 3D-Coded [21] and Deprelle et al. [13], which deform a template shape to find correspondences. Other methods use supervised learning for point-to-point correspondences [14, 55, 63], while recent unsupervised approaches for point clouds, such as SE-ORNet [27], first align point clouds and then use a teacher-student network with DGCNN backbone [56] to learn point-wise embeddings. DPC [29] leverages self- and cross-attention to learn discriminative per-point features for smooth mappings. While both mesh- and point cloud-based methods achieve good matching performance, they primarily focus on geometric features and overlook semantic information. To address this, Diff3F [16] leverages foundation models to extract semantic vertex features. Our method builds on Diff3F [16] by incorporating chirality in + +formation, which makes features left/right aware and improves the quality of shape matching. + +# 3. Chirality feature optimization + +In this section, we introduce our unsupervised pipeline to decorate an textured shape with chirality-aware features. Our method is based on Diff3F [16], a recently proposed method that utilizes in-painting diffusion [49] along with DINO-V2 [40] features to decorate textured shapes with semantical aware features. We briefly introduce Diff3F [16] in Sec. 3.1. In Sec. 3.2, we introduce the generation of chiral image pairs and their features, and also the aggregation from image features to per-vertex chiral feature pairs, which are the core idea of our method. Finally, in Sec. 3.3, we introduce our model architecture and our unsupervised losses. + +# 3.1. Background: Diff3F + +In this subsection, we provide a summary of the Diff3F [16] method. We consider a 3D shape $\mathcal{X}$ represented by a vertex set $V\in \mathbb{R}^{|V|\times 3}$ and an edge set $E\in \mathbb{N}^{|E|\times 2}$ . Diff3F [16] generates vertex-wise features as follows: + +1. Rendering from $N$ different camera poses of shape $\mathcal{X}$ results in $N$ images $I_{j}^{S} \in \mathbb{R}^{H \times W \times 3}$ , corresponding depth map $D_{j} \in \mathbb{R}^{H \times W \times 1}$ and normal map $N_{j} \in \mathbb{R}^{H \times W \times 3}$ , where $j \in \{0, \dots, N-1\}$ and $H$ and $W$ denote the height and width, respectively. +2. Using a pre-trained StableDiffusion model [49] with ControlNet [66], the images $I_{j}^{S}$ are textured with the guidance of depth and normal maps and a category prompt of the shape to produce $I_{j}^{\mathrm{tex}} \in \mathbb{R}^{H \times W \times 3}$ . +3. During the texturing process, diffusion features from multiple layers are extracted and aggregated over multiple time steps of the diffusion process to form a set of features for each image, $F_{j}^{\mathrm{SD}} \in \mathbb{R}^{H \times W \times 1280}$ . +4. The textured images $I_{j}^{\mathrm{tex}}$ are fed through DINO-V2 [40] to extract features $F_{j}^{\mathrm{DINO}} \in \mathbb{R}^{H \times W \times 768}$ . +5. Diffusion $(F_{j}^{\mathrm{SD}})$ and DINO-V2 $(F_{j}^{\mathrm{DINO}})$ features are then normalised and concatenated to form the final image features $F_{j}^{\mathrm{img}}\in \mathbb{R}^{H\times W\times 2048}$ . +6. Using the camera poses, each $F_{j}^{\mathrm{img}}$ is projected back onto the vertices of $\mathcal{X}$ and the final feature $\mathcal{F}_v$ for vertex $v\in V$ is the average of aggregated features from all $N$ camera poses. + +For convenience, we refer to StableDiffusion as SD throughout the remainder of this paper. Similarly, we abbreviate DINO-V2 as DINO, as it is the only version used in this work. + +# 3.2. Generating chiral pairs + +In our method, the input is an untextured mesh $\mathcal{X}$ , defined by a vertex set $V\in \mathbb{R}^{|V|\times 3}$ and an edge set $E\in \mathbb{N}^{|E|\times 2}$ + +![](images/e9f900fce05e5dbd7c88074508fabf57c265534bfc2be869a94db957207118a2.jpg) +Figure 3. Overview of the pipeline of our method. We consider a single 3D mesh for which we generate $N$ textured images following Diff3F [16]. We flip the images horizontally to receive $N$ pairs of textured images. The images are then independently processed by a frozen feature extractor, supplying us with features $F_{img}$ and $\bar{F}_{img}$ . After reprojecting onto the shape, we aggregate the features to receive $\mathcal{F}_v$ and $\bar{\mathcal{F}}_v$ . Finally, we train a chirality network $\tilde{g}_{\Phi}$ on $\mathcal{F}_v$ and $\bar{\mathcal{F}}_v$ to learn chirality features $\chi$ and $\bar{\chi}$ that tells left from right. + +without any feature descriptors but accompanied by a category label used for diffusion guidance. + +Following Diff3F [16], we get the textured images $\{I_j\}_{j=1}^N$ . Then, we flip these textured images $\{I_j\}_{j=1}^N$ horizontally to get a set of flipped textured images $\{\bar{I}_j\}_{j=1}^N$ . Note that flipping the image vertically is equivalent to flipping the image horizontally together with a 180 degree in-plane rotation, and that in-plane rotations will not change the chirality of objects within an image. + +After that, we feed both the textured image set $\{I_j\}_{j=1}^N$ and the flipped textured image set $\{\bar{I}_j\}_{j=1}^N$ into DINO [40] and StableDiffusion [49] to get the feature sets $\{(F_j^{\mathrm{SD}}, F_j^{\mathrm{DINO}})\}_{j=1}^N, \{(\bar{F}_j^{\mathrm{SD}}, \bar{F}_j^{\mathrm{DINO}})\}_{j=1}^N$ . We adopt a similar approach to [16, 65] for concatenating features from each DINO and SD feature pair: each feature is first normalized individually, then concatenated along the feature dimension, followed by normalization of the combined feature vector. This process yields $\{F_j^{\mathrm{img}}\}_{j=1}^N$ and $\{\bar{F}_j^{\mathrm{img}}\}_{j=1}^N$ corresponding to the textured images and their flipped versions, respectively. Next, we flip $\{\bar{F}_j^{\mathrm{img}}\}_{j=1}^N$ horizontally to get $\{\hat{F}_j^{\mathrm{img}}\}_{j=1}^N$ , which is spatially aligned with the original features $\{F_j^{\mathrm{img}}\}_{j=1}^N$ . + +Finally, for each vertex $v \in V$ of shape $\mathcal{X}$ , we utilize the camera information to locate its corresponding pixels + +in each rendered view. The vertex feature $\mathcal{F}_v$ is given by averaging the features of these pixels. Similarly, per vertex feature $\bar{\mathcal{F}}_v$ are obtained from $\hat{F}^{\mathrm{img}}$ . This process results in the chirality vertex feature pair $(\mathcal{F}_v,\bar{\mathcal{F}}_v)$ + +Using this construction, $\mathcal{F}_v$ aggregates semantic and geometric information from the image foundation model features, while its counterpart $\bar{\mathcal{F}}_v$ aggregates information differing only in left-and-right (chirality) information. For example, if $v$ denotes the keypoint at the left eye in shape $\mathcal{X}$ , then $\mathcal{F}_v$ will average left eye image foundation model features from different views, while $\bar{\mathcal{F}}_v$ gathers right eye features from different views (note that $v$ still denotes the left eye keypoint of $\mathcal{X}$ , and these right eye features are virtually constructed and gathered by our flipping process). + +# 3.3. Chirality feature extraction + +In this section, we describe the extraction of our chirality features $\chi, \bar{\chi}$ from the generated pairs of vertex features $\{(\mathcal{F}_v, \bar{\mathcal{F}}_v)\}_{v \in V}$ , where $\mathcal{F}_v, \bar{\mathcal{F}}_v \in \mathbb{R}^D$ . We denote the stacked vertex features of $\{\mathcal{F}_v\}_{v \in V}$ and $\{\bar{\mathcal{F}}_v\}_{v \in V}$ as $\mathcal{F} \in \mathbb{R}^{|V| \times D}$ and $\bar{\mathcal{F}} \in \mathbb{R}^{|V| \times D}$ , respectively. + +To this end, we use an encoder $g_{\Phi}:\mathbb{R}^{D}\to \mathbb{R}^{D}$ together with a linear projection layer $A\in \mathbb{R}^{D\times D}$ as our chirality feature extractor denoted as $\tilde{g} (\cdot) = Ag_{\Phi}(\cdot)$ , with $\Phi$ and $A$ being the optimised parameters. A decoder $h_{\Psi}:\mathbb{R}^{D}\to \mathbb{R}^{D}$ is included to avoid trivial solutions. + +The chirality feature vector $\chi := (\chi_{v_1},\dots ,\chi_{v_{|V|}})^\top \in \mathbb{R}^{|V|\times 1}$ of all vertices of shape $\mathcal{X}$ is given by + +$$ +\chi_ {v} := \frac {[ \tilde {g} (\mathcal {F} _ {v}) ] _ {1}}{\| \tilde {g} (\mathcal {F} _ {v}) \| _ {2}} = \frac {[ A g (\mathcal {F} _ {v}) ] _ {1}}{\| A g (\mathcal {F} _ {v}) \| _ {2}}, \tag {1} +$$ + +where $[\cdot ]_1$ denotes taking the first entry of the feature vector. + +This indexing and normalisation will ensure that chirality features $\chi_v$ are in the range $[-1, 1]$ . Similarly, we obtain $\bar{\chi}$ from $\bar{\mathcal{F}}$ using the model $\tilde{g}$ . More details about the architecture of our feature extractor are included in Sec. B in the supplementary material. To train our model, we employ the following losses: + +Dissimilarity loss $\mathcal{L}_{\mathrm{dis}}$ . The dissimilarity loss is defined as + +$$ +\mathcal {L} _ {\mathrm {d i s}} = - \frac {1}{\sqrt {| V |}} \| \chi - \bar {\chi} \| _ {2}. \tag {2} +$$ + +It serves as the main loss, since it maximises the difference between the chirality feature obtained from $\mathcal{F}_v$ and $\bar{\mathcal{F}}_v$ , which should only differ in left-and-right information. + +Invertibility loss $\mathcal{L}_{\mathrm{inv}}$ . To keep our encoder $g$ from learning trivial solutions, we introduce a decoder $h_{\Psi}$ , where $\Psi$ is the learnable parameter. The invertibility loss + +$$ +\mathcal {L} _ {\mathrm {i n v}} = \frac {1}{\sqrt {| V |}} \| \left[ \begin{array}{l l} \mathcal {F} ^ {\top} & \bar {\mathcal {F}} ^ {\top} \end{array} \right] ^ {\top} - h \left(g \left(\left[ \begin{array}{l l} \mathcal {F} ^ {\top} & \bar {\mathcal {F}} ^ {\top} \end{array} \right] ^ {\top}\right)\right) \| _ {F}, \tag {3} +$$ + +ensures that $h$ is able to reconstruct the inputs $\mathcal{F}$ and $\bar{\mathcal{F}}$ from the outputs $g(\mathcal{F})$ and $g(\bar{\mathcal{F}})$ , respectively. Here, $[\mathcal{F}^\top \bar{\mathcal{F}}^\top ]^\top$ corresponds to stacking the rows of $\mathcal{F}$ and $\bar{\mathcal{F}}$ , and $g$ and $h$ are applied row-wise. + +Total variation loss $\mathcal{L}_{\mathrm{var}}$ . To achieve spatial smoothness, we add a total variation loss + +$$ +\mathcal {L} _ {\text {v a r}} = \frac {1}{| E |} \sum_ {(u, v) \in E} \| \chi_ {u} - \chi_ {v} \| _ {1} + \| \bar {\chi} _ {u} - \bar {\chi} _ {v} \| _ {1}, \tag {4} +$$ + +where $|E|$ is the number of edges inside the mesh. + +Fifty-fifty loss $\mathcal{L}_{\mathrm{ff}}$ . The total variation loss $\mathcal{L}_{\mathrm{var}}$ introduces a bias towards solutions with small boundary length. We counteract this by introducing + +$$ +\mathcal {L} _ {\mathrm {f i f}} = \frac {1}{| V |} \left(\frac {| \chi^ {\top} \mathbf {1} _ {| V |} |}{\| \chi \| _ {\infty}} + \frac {| \bar {\chi} ^ {\top} \mathbf {1} _ {| V |} |}{\| \bar {\chi} \| _ {\infty}}\right), \tag {5} +$$ + +which penalises solutions which assign more vertices to one of the two halves of each shape. + +The overall training loss is the linear combination of the above losses, i.e. + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d i s}} + \lambda_ {1} \mathcal {L} _ {\text {i n v}} + \lambda_ {2} \mathcal {L} _ {\text {v a r}} + \lambda_ {3} \mathcal {L} _ {\text {f i f}}. \tag {6} +$$ + +# 4. Experiments + +We conduct various experiments to show the effectiveness of our chirality feature. In Sec. 4.1, we compare the left- and right disentanglement performance using our chirality feature compared with other shape feature descriptors and methods. Then, in Sec. 4.2 and Sec. 4.3, we show the effectiveness of our chirality feature in various shape analysis tasks, including shape matching and shape segmentation. To evaluate the generalisation ability of our method, we test on partial and anisotropic shapes in Sec. 4.4 and Sec. 4.5, respectively. Finally, the ablation study of losses and 2D foundation features are conducted in Sec. 4.6. In each experiment section, we present the datasets, baselines and evaluation metrics used. For more details about the implementation, see Sec. E in the supplementary material. + +# 4.1. Left-right disentanglement + +Datasets. To evaluate our approach's ability to distinguish between left and right of 3D shapes, we use datasets where left/right annotations are available. To this end, we use Be-CoS [18], a recently proposed framework to generate shape matching datasets with dense correspondences and left/right annotations. This framework combines shapes from multiple shape matching datasets and connects them with cross-category and cross-class dense correspondences. Using this framework, we generate multiple versions to evaluate the performance and generalisation of our approach. + +- BeCoS consists of humanoid and four-legged animals with 1980/284/274 train/test/validation split, generated from 7 different shape datasets, namely TOSCA [7], FAUST [6], SCAPE [2], KIDS [47], DT4D [35], SMAL [72] and SHREC'20 [17]. See [18] for more details. +- BeCoS $_{h}$ consists of only humanoid shapes from TOSCA [7], FAUST [6], SCAPE [2], KIDS [47], DT4D [35] with 366/64/58 train/test/validation split. +- BeCoS-a consists of only four-legged animals from TOSCA [7], DT4D [35], SMAL [72] and SHREC'20 [17] with 1614/220/216 train/test/validation split. + +For more details about the data generation please refer to Sec. A in the supplementary material. Additionally, we use well-established shape matching datasets to evaluate our method, namely FAUST [6], SCAPE [2], SMAL [69], and TOSCA [7]. We use the same train/test/Validation splits given by BeCoS [18]. Since these datasets do not provide left/right annotations, we use the BeCoS [18] framework to get these annotations. + +Baselines. We compare the performance of our chirality features against other feature descriptors, including Diff3F [16], DINO+SD [40, 49], and DINO+SD fine-tuned with Zhang et al. [65]. We also include an axiomatic method Liu et al. [33] that extracts closed intrinsic/extrinsic symmetric + +
TrainBeCoSBeCoS-hBeCoS-aFAUSTSMAL
TestBeCoSBeCoS-hBeCoS-aBeCoS-hBeCoS-aFAUSTSCAPESMALTOSCA
Diff3F [16]50.8754.4350.2353.2850.7751.2152.5350.9151.48
DINO+SD [40, 49]51.1654.3150.3052.6850.9651.0552.5550.8051.42
Zhang et al. [65]51.1854.1650.7852.8351.0250.9051.4750.4150.97
Liu et al. [33]79.9879.8380.4679.8380.4690.4580.8475.7172.88
\( \chi_{\text{DINO+SD}} \)91.8494.0984.1990.3691.1094.7695.5196.5994.09
+ +Table 1. Left and right distinguishment accuracy $(acc_{\chi}\uparrow)$ on BeCoS [18] datasets for various methods. Our method outperforms all others across all datasets. + +![](images/e08e4ec9c463123fa8a5fc8fa9f5ffaee50d806fa4e269d02ba55e376ba0872f.jpg) +Figure 4. Qualitative results on left-right disentanglement of Zhang et al. [67], Liu et al. [33], DINO + SD [40, 49], Diff3F [16] and our features. Our method provides the only features that consistently and accurately distinguish between left and right of the object. + +curves on surfaces of a genus-0 mesh that splits the mesh into left and right parts. + +Evaluation metrics. Given a set of shapes $X = \{\mathcal{X}_1, \dots, \mathcal{X}_N\}$ , where each shape $\mathcal{X}_n$ has a vertex set $V_{\mathcal{X}_n}$ , the chirality accuracy $\mathrm{acc}_{\chi}$ is defined as + +$$ +\operatorname {a c c} _ {\chi} = \max \left\{\operatorname {a c c}, 1 - \operatorname {a c c} \right\}, \tag {7} +$$ + +where + +$$ +\operatorname {a c c} = \frac {1}{N} \sum_ {n = 1} ^ {N} \frac {1}{| V _ {\chi_ {n}} |} \sum_ {v \in V _ {\chi_ {n}}} \mathbb {1} (\operatorname {s i g n} \left(\chi_ {v}\right) = \chi_ {v} ^ {g t}). \tag {8} +$$ + +$\chi_v$ is our chirality features for vertex $v \in V_{\mathcal{X}_n}$ of each shape $\mathcal{X}_n$ , $\mathbb{1}$ is the indicator function, and $\chi_v^{gt}$ is the ground truth of left/right annotation of vertex $v \in V_{\mathcal{X}_n}$ . + +Note that in Eq. 7, we use the maximum of acc and $1 - \mathrm{acc}$ from Eq. 8, because there is no inherent assignment of whether $\chi_v > 0$ or $\chi_v < 0$ corresponds to left or right. + +Experiment results. The results are summarised in Tab. 1. Our method substantially outperforms all baselines in the context of distinguishing left/right of 3D shapes. Additionally, our method achieves good cross-dataset and cross-category generalisation ability. Fig. 4, provides qualitative results of our method compared to baselines. + +# 4.2. Shape matching + +In this section, we evaluate the performance of our chirality features on shape matching. To ensure a fair comparison with prior work, we adopt a similar experimental setup as in DPC [29], SE-ORNet [12] and Diff3F [16]. + +
TOSCA-aSHREC'19
acc (↑)err (↓)acc (↑)err (↓)
DPC [29]30.793.7417.406.26
SE-ORNet [12]33.254.3221.414.56
3D-CODED [21]--2.108.10
Diff3F [16]20.255.4426.401.69
Diff3F [16] + χDINO+SD22.734.7227.321.02
+ +Table 2. Shape matching results of TOSCA-a, SHREC'19. (The results of [12, 21, 29] are reported from [16]). + +![](images/87b73485d46b295b3abc280821ef1e5d696d7c254d52b2450c61e73f4d708b68.jpg) +Figure 5. Quantitative comparison (with (AUC x 100) shown in brackets) on TOSCA-A and SHREC'19 datasets. With our chirality features, the vertex feature descriptor archives better matching performance on both datasets. + +![](images/b1c8024e4874b3f3d7a7a10344d688fc5eb3354dca8ca1b36cafdde71dbf62f3.jpg) + +![](images/5c070798d5374502925945f597d5c95dcf9c443af30187dde7dbbede0d4ae27a.jpg) +Figure 6. Qualitative results of shape correspondence using Diff3F and our chirality features. Our approach is able to correct for the left/right ambiguity as shown by the blue ovals. + +Datasets. We test our features on both human and animal shapes. For human shapes, we evaluate our method on SHREC'19 [36]. This dataset comprises of 44 human scans with a test set of 430 shape pairs. We use the more challenging re-meshed version from [15]. For the non-human shapes, we use only animal the 41 shapes from TOSCA [7] (noted as TOSCA-a) and pair shapes from the same category to create a testing set of 286 pairs. + +Baselines. We compare our approach to recent state-of-the-art shape matching methods. One supervised method (3D-CODED [21]) that we only have access to its model trained on humans. And two unsupervised methods, namely DPC [29] and SR-ORNET [12], that were trained on hu + +man and animal datasets separately. We also compare our method to Diff3F [16] that extracts vertex features of each shape and requires no training. + +Evaluation metrics. Following Diff3F [16], we use the average matching error and the matching accuracy as our evaluation metrics. For a source shape $\mathcal{X}$ and a target shape $\mathcal{Y}$ , represented with a vertex sets $V_{\mathcal{X}} \in \mathbb{R}^{|V_{\mathcal{X}}| \times 3}$ and $V_{\mathcal{Y}} \in \mathbb{R}^{|V_{\mathcal{Y}}| \times 3}$ , respectively, the matching error is + +$$ +\operatorname {e r r} = \frac {1}{| V _ {\mathcal {X}} |} \sum_ {v \in V _ {\mathcal {X}}} \| f (v) - y _ {g t} \| _ {2}, \tag {9} +$$ + +where $f(v)$ is the predicted matching point of $v \in V_{\mathcal{X}}$ in $\mathcal{Y}$ and $y_{gt} \in V_{\mathcal{Y}}$ is the ground truth corresponding point of $v$ . Furthermore, the matching accuracy is defined as + +$$ +a c c (\epsilon) = \frac {1}{| V _ {\mathcal {X}} |} \sum_ {v \in V _ {\mathcal {X}}} \mathbb {1} \left(\left\| f (v) - y _ {g t} \right\| _ {2} < \epsilon d\right), \tag {10} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function, $d$ is the maximal Euclidean distance between points in $V_{\mathcal{Y}}$ , and $\epsilon \in [0,1]$ is the error tolerance. + +Experiment results. The results are summarised in Tab. 2. Our chirality features enhance Diff3F [16] performance in shape matching task since it resolves the left/right ambiguity. In Fig. 5, we summarise the PCK curves of our method compared to the state-of-the-art method on both human and animal dataset. Fig. 6 provides qualitative results of our method compared to baselines. + +# 4.3. Part segmentation + +We apply $k$ -means clustering to our chirality features combined with Diff3F [16] features to segment shapes into $k$ parts. Our features enhance Diff3F [16] features to be able to differentiate between left and right parts of shapes as seen in Fig. 7 (left and right legs are not clustered together; the part segmentation is still consistent across shapes). + +![](images/cb07e6d4b71919b4635ca2c66d1d5d0d16e2f63c31b9a3257542d4319446cc6a.jpg) +Figure 7. Visualisations of part segmentation using Diff3F [16] and our features. Our chirality features can differentiate between the left and right parts of the body (e.g. hands and legs), whereas Diff3F clusters the left and right hands and legs together. + +# 4.4. Partial shapes + +To showcase the potential of our method, we conduct a proof-of-concept experiment on partially observed meshes, which is a common real-world scenario for 3D shapes collected by scanning devices. To this end, we extract our chirality features from both human and animal shapes from CP2P dataset [3] and use a model trained on complete shapes from BeCoS [18]. In Fig. 8, we observe that our chirality features can distinguish between left and right parts of partial shapes even when most of the mesh is missing. + +![](images/98ba9f89af5c711555613b69cba2292b074991359089729f930e2e50c540fb97.jpg) +Figure 8. Our chirality features can distinguish left and right parts of partial shapes on both human and animal shapes. + +![](images/95e1694d665dd7b4b5ad8c2af01da47d5b65a59f1263d7c49519b5f212223a06.jpg) + +![](images/f01a9ad844ce40d5f9a591932b1631e83ee66ef0b8ac4dc0f331b7b7df941a16.jpg) + +# 4.5. Anisotropic shapes + +To further demonstrate the robustness of our method to different discretisation of shapes, we show its performance on anisotropic meshes. We use the anisotropic version of FAUST and SCAPE from [14]. These non-uniform meshes are often encountered in adaptive refinement and characterised by having a non-consistent discretisation granularity. Qualitative results of the model trained on isotropic BeCoS [18] on anisotropic shapes are shown in Fig. 9. + +![](images/b9128b5413fd5e986df62848d2b8d51e8e820bc13d51df76466185824d67d4b9.jpg) +Figure 9. Our approach is robust to anisotropic meshes and can handle meshes with different discretisation. + +![](images/5e4b52431ff3c35ccc3531e0beb229325f8cd22b052bf86af71c16c6644636b2.jpg) + +![](images/8be674e1b20f37d4ff4471ac1caaf86422b1d2356aa7546b11f9c63e747127fd.jpg) + +![](images/b4f1727dfd3c1098c39197d380bd5548ad7ec6ef070e52dafb48c7d5c25c0268.jpg) + +# 4.6. Ablation study + +We conduct ablative experiments to verify our design choices, using use the FAUST [6] and SMAL [69] datasets. Tab. 3 summarises our findings. By comparing the first four columns, we conclude all losses are crucial for obtaining accurate chirality features. By comparing the last three columns, we observe combining SD and DINO yields the best results across both human and animal datasets. + +# 5. Limitations & Future works + +Although our proposed method shows superiority compared to other methods for left-and-right disentanglement, and + +
w/o\( {\mathcal{L}}_{\text{dis }} \)\( {\mathcal{L}}_{\text{var }} \)\( {\mathcal{L}}_{\text{fif }} \)\( {\mathcal{L}}_{\text{inv }} \)SDDINOfull
FAUST51.34 ± 0.9677.02 ± 8.8790.95 ± 7.8696.26 ± 0.4989.21 ± 2.4881.47 ± 24.9695.79 ± 0.50
SMAL51.33 ± 0.9174.37 ± 0.9796.41 ± 0.2776.72 ± 3.8071.65 ± 1.1794.21 ± 3.8994.71 ± 2.59
+ +Table 3. Ablation study on the FAUST and SMAL datasets. We use the chirality accuracy $acc_{\chi}$ ( $\uparrow$ ) as evaluation metric. The best/second best results in each column are highlighted/underlined, respectively. + +the extracted chirality feature performs well on various downstream tasks, further improvements are still possible. Firstly, methods using 3D descriptors aggregated from 2D images (e.g. Diff3F [16], DenseMatcher [71]) struggle with occluded vertices and fail to learn reliable features. Our method simply inherits this shortcoming from Diff3F [16]. Deforming shapes, to reduce occlusions, is an interesting direction for future work. Secondly, our model takes meshes as input, since $\mathcal{L}_{\mathrm{var}}$ relies on the edges of the mesh, which limits the direct applicability of our method to point clouds. However, we show in Sec. C of the supplementary material that our method (with a simple $k$ -nearest neighbour approximation of the connectivity of the input point cloud) can give reasonable results. Additionally, despite incorporating the total variation loss, we occasionally observe patches with incorrect features, as seen in Fig. 1 (left side, top right). Another type of inaccuracy might arise when two body parts with different chirality are in close proximity or touching each other (Fig. 9, middle left). Future works combining geometric constraints might alleviate these problems. + +# 6. Conclusion + +Chirality information plays an important role in visual computing and has been severely under-explored in the shape analysis field. In this paper, we propose an unsupervised method to disentangle per vertex chirality features from semantic and geometric features aggregated from 2D foundation models (DINO-V2 [40] and StableDiffusion [49]). The chirality features disentangled by our proposed pipeline show superiority compared to various other features/methods on left-right distinguishing tasks both quantitatively and qualitatively. Furthermore, combined with our chirality features, other vertex feature descriptors show better performance on both shape matching and part segmentation tasks on various datasets. Additionally, the generalisation tests on partial and anisotropic shapes confirm the robustness of our method and also enlarge the application scenarios of our model due to more realistic properties. To conclude, we believe that our proposed pipeline and extracted chirality features will benefit future research in shape analysis and other visual computing areas. + +# Acknowledgments + +We thank Paul Roetzer for the valuable feedback on earlier drafts of this manuscript. This work is supported by the ERC starting grant no. 101160648 (Harmony). + +# References + +[1] James Allingham, Bruno Mlodozeniec, Shreyas Padhy, Javier Antorán, David Krueger, Richard Turner, Eric Nalisnick, and José Miguel Hernández-Lobato. A generative model of symmetry transformations. In NeurIPS, 2024. 1 +[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. ACM SIGGRAPH, 2005. 5 +[3] Souhaib Attaiki, Gautam Pai, and Maks Ovsjanikov. Dpfm: Deep partial functional maps. In 3DV, 2021. 3, 8 +[4] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The wave kernel signature: A quantum mechanical approach to shape analysis. In ICCV workshops, 2011. 2 +[5] Florian Bernard, Zeeshan Khan Suri, and Christian Theobalt. Mina: Convex mixed-integer programming for non-rigid shape alignment. In CVPR, 2020. 3 +[6] Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. Faust: Dataset and evaluation for 3d mesh registration. In CVPR, 2014. 5, 8 +[7] Alexander Bronstein, Michael Bronstein, and Ron Kimmel. Numerical Geometry of Non-Rigid Shapes. Springer Publishing Company, Incorporated, 1 edition, 2008. 5, 7 +[8] Dongliang Cao and Florian Bernard. Unsupervised deep multi-shape matching. In ECCV, 2022. 2, 3 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3 +[10] Enric Corona, Kaustav Kundu, and Sanja Fidler. Pose estimation for objects with rotational symmetry. In IROS, 2018. 1 +[11] Luca Cosmo, Emanuele Rodola, Andrea Albarelli, Facundo Memoli, and Daniel Cremers. Consistent partial matching of shape collections via sparse modeling. In Computer Graphics Forum, 2017. 1 +[12] Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In CVPR, 2023. 6, 7 +[13] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 3 +[14] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In CVPR, 2020. 3, 8 +[15] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric maps: Robust feature learning for shape correspondence. In CVPR, 2020. 7 + +[16] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[17] Roberto M Dyke, Yu-Kun Lai, Paul L Rosin, Stefano Zappala, Seana Dykes, Daoliang Guo, Kun Li, Riccardo Marin, Simone Melzi, and Jingyu Yang. Shrec'20: Shape correspondence with non-isometric deformations. Computers & Graphics, 2020. 5 +[18] Viktoria Ehm, Nafie El Amrani, Yizheng Xie, Lennart Bastian, Maolin Gao, Weikang Wang, Lu Sang, Dongliang Cao, Tobias Weißberg, Zorah Lähner, Daniel Cremers, and Florian Bernard. Beyond Complete Shapes: A Benchmark for Quantitative Evaluation of 3D Shape Matching Algorithms. Computer Graphics Forum, 2025. 5, 6, 8 +[19] Marvin Eisenberger, Zorah Lähner, and Daniel Cremers. Divergence-free shape correspondence by deformation. In Computer Graphics Forum, 2019. 3 +[20] M. Gao, Z. Lähner, J. Thunberg, D. Cremers, and F. Bernard. Isometric multi-shape matching. In CVPR, 2021. 3 +[21] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 3, 7 +[22] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. NeurIPS, 2023. 3 +[23] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Xingzhe He, Hossam Isack, Abhishek Kar, Helge Rhodin, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised keypoints from pretrained diffusion models. In CVPR, 2024. 3 +[24] Benjamin Holzschuh, Zorah Lähner, and Daniel Cremers. Simulated annealing for 3d shape correspondence. In 3DV, 2020. 3 +[25] Qi-Xing Huang, Bart Adams, Martin Wicke, and Leonidas J Guibas. Non-rigid registration under isometric deformations. In Computer Graphics Forum, 2008. 3 +[26] Jihyeon Je, Jiayi Liu, Guandao Yang, Boyang Deng, Shengqu Cai, Gordon Wetzstein, Or Litany, and Leonidas Guibas. Robust symmetry detection via riemannian Langevin dynamics. In ACM SIGGRAPH Asia, 2024. 2 +[27] Deng Jiacheng, Wang ChuXin, Lu Jiahao, He Jianfeng, Zhang Tianzhu, Yu Jiyang, and Zhang Zhe. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In CVPR, 2023. 3 +[28] Vladimir G Kim, Yaron Lipman, Xiaobai Chen, and Thomas Funkhouser. Möbius transformations for global intrinsic symmetry analysis. In Computer Graphics Forum, 2010. 2 +[29] Itai Lang, Dvir Ginzburg, Shai Avidan, and Dan Raviv. Dpc: Unsupervised deep point correspondence via cross and self construction. In 3DV, 2021. 3, 6, 7 +[30] B. Levy. Laplace-beltrami eigenfunctions towards an algorithm that "understands" geometry. In SMI, 2006. 3 +[31] Ren-Wu Li, Ling-Xiao Zhang, Chunpeng Li, Yu-Kun Lai, and Lin Gao. E3sym: Leveraging e (3) invariance for unsupervised 3d planar reflective symmetry detection. In ICCV, 2023. 2 + +[32] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, 2020. 1, 2 +[33] Tianqiang Liu, Vladimir G Kim, and Thomas Funkhouser. Finding surface correspondences using symmetry axis curves. In Computer Graphics Forum, 2012. 1, 2, 5, 6 +[34] Ling Lo, Hongxia Xie, Hong-Han Shuai, and Wen-Huang Cheng. Facial chirality: From visual self-reflection to robust facial feature learning. IEEE Transactions on Multimedia, 2022. 1, 2 +[35] Robin Magnet, Jing Ren, Olga Sorkine-Hornung, and Maks Ovsjanikov. Smooth non-rigid shape matching via effective dirichlet energy optimization. In 3DV, 2022. 5 +[36] S. Melzi, R. Marin, E. Rodola, U. Castellani, J. Ren, A. Poulenard, P. Wonka, and M. Ovsjanikov. Matching humans with different connectivity. In Eurographics Workshop on 3D Object Retrieval, 2019. 7 +[37] Nathaniel Merrill, Yuliang Guo, Xingxing Zuo, Xinyu Huang, Stefan Leutenegger, Xi Peng, Liu Ren, and Guoquan Huang. Symmetry and uncertainty-aware object slam for 6 dof object pose estimation. In CVPR, 2022. 1 +[38] Rajendra Nagar and Shanmuganathan Raman. Fast and accurate intrinsic symmetry detection. In ECCV, 2018. 2 +[39] Dorian Nogneng and Maks Ovsjanikov. Informative descriptor preservation via commutativity for shape matching. In Computer Graphics Forum, 2017. 3 +[40] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 4, 5, 6, 8 +[41] Maks Ovsjanikov, Jian Sun, and Leonidas Guibas. Global intrinsic symmetries of shapes. In Computer Graphics Forum, 2008. 2 +[42] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. ACM ToG, 2012. 3 +[43] Giorgia Pitteri, Michael Ramamonjisoa, Slobodan Ilic, and Vincent Lepetit. On object symmetries and 6d pose estimation from images. In 3DV, 2019. 1 +[44] Joshua Podolak, Philip Shilane, Aleksey Golovinskiy, Szymon Rusinkiewicz, and Thomas Funkhouser. A planar-reflective symmetry transform for 3d shapes. In ACM SIGGRAPH, 2006. 1 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[46] Jing Ren, Mikhail Panine, Peter Wonka, and Maks Ovsjanikov. Structured regularization of functional map computations. In Computer Graphics Forum, 2019. 3 +[47] Emanuele Rodola, Samuel Rota Bulo, Thomas Windheuser, Matthias Vestner, and Daniel Cremers. Dense non-rigid shape correspondence using random forests. In CVPR, 2014. 5 +[48] Paul Roetzer, Paul Swoboda, Daniel Cremers, and Florian Bernard. A scalable combinatorial solver for elastic geometrically consistent 3d shape matching. In CVPR, 2022. 3 + +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 6, 8 +[50] Jian Sun, Maks Ovsjanikov, and Leonidas Guibas. A concise and provably informative multi-scale signature based on heat diffusion. In Computer Graphics Forum, 2009. 2 +[51] Gary K.L. Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C. Langbein, Yonghuai Liu, David Marshall, Ralph R. Martin, Xian-Fang Sun, and Paul L. Rosin. Registration of 3d point clouds and meshes: A survey from rigid to nonrigid. IEEE TVCG, 2013. 3 +[52] Xin Tan, Jiaying Lin, Ke Xu, Pan Chen, Lizhuang Ma, and Rynson WH Lau. Mirror detection with the visual chirality cue. IEEE TPAMI, 2022. 1, 2 +[53] Art Tevs, Qixing Huang, Michael Wand, Hans-Peter Seidel, and Leonidas Guibas. Relating shapes via geometric symmetries and regularities. ACM ToG, 2014. 2 +[54] Oliver van Kaick, Hao Zhang, Ghassan Hamarneh, and Daniel Cohen-Or. A survey on shape correspondence. Computer Graphics Forum, 2011. 3 +[55] Yue Wang and Justin M. Solomon. Deep closest point: Learning representations for point cloud registration. In ICCV, 2019. 3 +[56] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM ToG, 2019. 3 +[57] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In CVPR, 2020. 1 +[58] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Magicpony: Learning articulated 3d animals in the wild. In CVPR, 2023. 1 +[59] Yizheng Xie, Viktoria Ehm, Paul Roetzer, Nafie El Amrani, Maolin Gao, Florian Bernard, and Daniel Cremers. Echomatch: Partial-to-partial shape matching via correspondence reflection. In CVPR, 2025. 3 +[60] Jianke Yang, Robin Walters, Nima Dehmamy, and Rose Yu. Generative adversarial symmetry discovery. In ICML, 2023. 1 +[61] Jianglong Ye, Naiyan Wang, and Xiaolong Wang. Featureerf: Learning generalizable nerfs by distilling foundation models. In ICCV, 2023. 3 +[62] Raymond Yeh, Yuan-Ting Hu, and Alexander Schwing. Chirality nets for human pose regression. In NeurIPS, 2019. 1, 2 +[63] Zi Jian Yew and Gim Hee Lee. RPM-net: Robust point matching using learned features. In CVPR, 2020. 3 +[64] Yusuke Yoshiyasu, Eiichi Yoshida, Kazuhito Yokoi, and Ryusuke Sagawa. Symmetry-aware nonrigid matching of incomplete 3d surfaces. In CVPR, 2014. 1 +[65] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In CVPR, 2024. 2, 4, 5, 6 +[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 4 + +[67] Zhiyuan Zhang, KangKang Yin, and Kelvin WC Foong. Symmetry robust descriptor for non-rigid surface matching. In Computer Graphics Forum, 2013. 1, 6 +[68] Heng Zhao, Shenxing Wei, Dahu Shi, Wenming Tan, Zheyang Li, Ye Ren, Xing Wei, Yi Yang, and Shiliang Pu. Learning symmetry-aware geometry correspondences for 6d object pose estimation. In ICCV, 2023. 1 +[69] Qian Zheng, Zhuming Hao, Hui Huang, Kai Xu, Hao Zhang, Daniel Cohen-Or, and Baoquan Chen. Skeleton-intrinsic symmetrization of shapes. In Computer Graphics Forum, 2015. 1, 5, 8 +[70] Ying Zheng, Yiyi Zhang, Xiaogang Xu, Jun Wang, and Hongxun Yao. Visual chirality meets freehand sketches. In ICIP, 2021. 1, 2 +[71] Junzhe Zhu, Yuanchen Ju, Junyi Zhang, Muhan Wang, Zhecheng Yuan, Kaizhe Hu, and Huazhe Xu. Densemarketer: Learning 3d semantic correspondence for category-level manipulation from a single demo. In ICLR, 2025. 8 +[72] Silvia Zuffi, Angjoo Kanazawa, David Jacobs, and Michael J. Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 5 \ No newline at end of file diff --git a/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/images.zip b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9e4efaeb4e677d264281b2ae243848be2a3c7b7c --- /dev/null +++ b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5483350d139f3f03fa77f64afafd87320b7d3fe6e7eb8755fcd3df4e1ae001bb +size 516432 diff --git a/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/layout.json b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2b3ec396405d193b5d2f827aecaf57d0e6391bba --- /dev/null +++ b/khsymmetryunderstandingof3dshapesviachiralitydisentanglement/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b21738cb2ce67c4f872826efe2e34f124e4608a33b340d27f2781e765124208a +size 497504 diff --git a/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/3ae20890-4db9-45fb-acac-a9261ca916aa_content_list.json b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/3ae20890-4db9-45fb-acac-a9261ca916aa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1946ac4a29a0bcc75370b14e05b4d12147ee168e --- /dev/null +++ b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/3ae20890-4db9-45fb-acac-a9261ca916aa_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f652c02fa2ea90b2737eec46d015441327c5d80331e10054c8aa829cb39114b +size 80910 diff --git a/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/3ae20890-4db9-45fb-acac-a9261ca916aa_model.json b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/3ae20890-4db9-45fb-acac-a9261ca916aa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f81916f635618497f49c0c57a0b6898e4105d0c --- /dev/null +++ b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/3ae20890-4db9-45fb-acac-a9261ca916aa_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a733aa02dab8a5083cb405bce05cae6a2dd9932574652c04ccc1ffe590784063 +size 101414 diff --git a/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/3ae20890-4db9-45fb-acac-a9261ca916aa_origin.pdf b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/3ae20890-4db9-45fb-acac-a9261ca916aa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f13642c1f13a70b9870d6146b4557351bcdfe4a2 --- /dev/null +++ b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/3ae20890-4db9-45fb-acac-a9261ca916aa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bbe386d7961df758f038989aa4f98c867f594d4093e998b088941ed5a3ab003 +size 2282671 diff --git a/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/full.md b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8bf17f18e8b955ab6a189335c625c42d3ca46a2 --- /dev/null +++ b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/full.md @@ -0,0 +1,293 @@ +# mmCooper: A Multi-agent Multi-stage Communication-efficient and Collaboration-robust Cooperative Perception Framework + +Bingyi Liu $^{1}$ , Jian Teng $^{1}$ , Hongfei Xue $^{2*}$ , Enshu Wang $^{3*}$ , Chuanhui Zhu $^{1}$ , Pu Wang $^{2}$ , Libing Wu $^{3}$ $^{1}$ Wuhan University Of Technology, $^{2}$ University of North Carolina at Charlotte, $^{3}$ Wuhan University +{byliu, tengjian, zhuchuanhui}@whut.edu.cn, +{hongfei.xue, Pu.Wang}@charlotte.edu, {wanges17, wu}@whu.edu.cn + +# Abstract + +Collaborative perception significantly enhances individual vehicle perception performance through the exchange of sensory information among agents. However, real-world deployment faces challenges due to bandwidth constraints and inevitable calibration errors during information exchange. To address these issues, we propose mm-Cooper, a novel multi-agent, multi-stage, communication-efficient, and collaboration-robust cooperative perception framework. Our framework leverages a multi-stage collaboration strategy that dynamically and adaptively balances intermediate- and late-stage information to share among agents, enhancing perceptual performance while maintaining communication efficiency. To support robust collaboration despite potential misalignments and calibration errors, our framework prevents misleading low-confidence sensing information from transmission and refines the received detection results from collaborators to improve accuracy. The extensive evaluation results on both real-world and simulated datasets demonstrate the effectiveness of the mm-Cooper framework and its components. + +# 1. Introduction + +With the advancement of autonomous driving systems [19, 57], perception devices such as cameras and LiDARs have been widely deployed. Although perception technologies have witnessed rapid advancements driven by deep learning-based algorithms, the traditional single-vehicle perception paradigms [23, 30, 52] cannot meet the safety and reliability requirements of autonomous vehicles due to unavoidable factors such as object occlusion and limitations in detection range [1, 6, 14, 27, 61]. + +Benefiting from infrastructure improvements and the advancement of Internet of Vehicles (IoVs) technologies like V2X [25, 26], autonomous vehicles can achieve multi + +vehicle collaborative perception by sharing perception information [5, 47], which significantly enhances the performance of perception systems. Although cooperative perception has made progress recently, its practical deployment still faces challenges such as communication resource constraints [8, 13], localization errors [31, 34], and low information fusion efficiency [46, 55]. + +In cooperative perception systems, the effectiveness of fusion strategies significantly affects both perception performance and communication efficiency. The fusion strategies can be divided into three categories: early, intermediate, and late fusion methods. Specifically, early fusion methods [7, 61] aggregate raw sensor observations from all agents, providing abundant information but requiring immense bandwidth, making this approach costly for communication. Intermediate fusion methods [8, 13, 53-55] partially mitigate this issue by sharing encoded features rather than raw data to reduce communication overhead. However, transmitting the feature map of the entire sensing scenario still results in substantial bandwidth consumption [13]. Late fusion methods [38, 40] on the other hand, minimize communication demands by sharing lightweight perception results among agents. However, this approach is highly sensitive to environmental noise, as even minor disturbances can degrade collaborative quality [31]. Overall, these above single-stage cooperative perception approaches face challenges in managing the trade-off between perception accuracy and communication efficiency, raising a compelling question: Can we improve the perception performance while preserving communication efficiency by strategically sharing information across multiple stages to harness the strengths of each fusion stage? + +In addition, the data exchanged by agents inevitably experiences misalignment and contains calibration errors during the synchronization process due to communication delays and pose noise [15, 41]. These calibration errors can produce misleading features and proposals in subsequent steps, thereby affecting the performance of the perception system. While existing intermediate fusion methods at + +tempt to address the calibration errors through position-wise feature fusion [20, 22, 29], they often overlook the critical role of neighboring region information for enhancing collaborative robustness. Meanwhile, late fusion methods [3, 36, 42] simply merge bounding boxes from collaborative agents, failing to leverage information-rich intermediate features from the ego agent, which could help refine and calibrate these bounding boxes for higher accuracy. + +To this end, we propose mmCooper, a novel multi-agent, multi-stage, communication-efficient, and collaboration-robust cooperative perception framework. The proposed mmCooper framework facilitates effective information aggregation across fusion stages and enhances communication efficiency through carefully designed core components. Specifically, we design an adaptive multistage data fusion mechanism that facilitates multi-agent cooperation by dynamically fusing information at both intermediate and late stages. Guided by a confidence-based filtering strategy, this mechanism automatically assesses and determines the data volume to be processed at intermediate and late stages. High-confidence bounding boxes are shared directly, while intermediate features are selectively shared instead of bounding boxes for regions requiring further contextual reference, and low-confidence areas are excluded to prevent propagating misleading information. This dual-stage fusion approach enhances perception performance while maintaining communication efficiency. + +Besides, to further address potential misalignment and calibration noise among agents, we incorporate specific designs in the proposed mmCooper framework for both intermediate and late stages. In the intermediate stage, we introduce a Multi-scale Offset-aware Fusion Module that not only fuses data from target locations related to cooperative agents' views but also from nearby regions, adding redundancy to mitigate calibration noise. For the late stage, we design a Bounding Box Filtering & Calibration Module that uses the information-rich intermediate-stage features from the ego agent to filter inaccurate bounding boxes received from other agents and refine the remaining noisy ones. By integrating these corrected and ego-generated bounding boxes, our mmCooper framework can achieve high perception accuracy and robustness. To summarize, our main contributions are summarized as follows: + +- We propose mmCooper, a multi-agent multi-stage communication-efficient collaboration-robust cooperative perception framework. To our knowledge, this is the first framework to address the tradeoff between limited communication bandwidth and desired perception performance by sharing both intermediate and late-stage information for multi-agent collaborative perception. +- To address the potential misalignment and calibration error among agents, we design a Multi-scale Offset-aware Fusion Module to integrate spatially adjacent contextual + +information at different feature scales in the intermediate stage and a Bounding Box Filtering & Calibration Module to filter and improve the bounding boxes from other agents in the late stage, guided by the information-rich intermediate features. + +- Extensive experiments show excellent performance on both real-world and simulated datasets, with a $7.29\% / 1.31\% / 2.09\%$ improvement in AP@0.7 on OPV2V [50], DAIR-V2X [56] and V2XSet [49] datasets over the second-best SOTA methods. Meanwhile, our approach requires only 1/9153, 1/156, and 1/18305 of the communication volume used by comparable methods. Furthermore, our method consistently outperforms previous approaches under various configurations. + +# 2. Related Work + +# 2.1. Multi-Agent Communication + +Communication strategies in multi-agent systems have been widely studied [39]. Early works [21, 37, 44] often used predefined protocols or heuristic methods to determine how agents communicate with each other. However, these methods are difficult to generalize to complex tasks. Recently, several learning-based communication strategies have been proposed to adapt to diverse scenario applications. For instance, Vain [12] utilizes attention neural structures to specify the information that needs to be shared in agent interactions. ATOC [17] introduces recurrent units to decide who agents communicate with by receiving local observations and action intentions from other agents. TarMAC [9] designs an architecture oriented towards reinforcement learning, learning to communicate from task-specific rewards. CommNet [43] learns continuous communication in multi-agent systems. Most of these previous works have considered decision-making tasks and have employed reinforcement learning due to the lack of explicit supervision. Different from these works, our paper focuses on LiDAR-based 3D perception tasks in complex autonomous driving scenarios and achieves efficient communication between agents using a multi-stage communication strategy. + +# 2.2. Collaborative Perception + +Collaborative perception addresses the limitations of single-agent perception by aggregating perception information from surrounding agents to adapt to complex and dynamic autonomous driving environments. Current work is categorized into early, intermediate, and late fusion based on collaboration stages. Early fusion [7, 59, 61] shares raw point cloud data among agents, achieving high perception performance but consuming significant bandwidth resources. EMP [61] achieved scalability and adaptability on highly fluctuating open roads by dynamically partitioning the point cloud range shared by agents. Late fusion [38, 40, 51] shares lightweight proposals between agents but performs poorly and is sensitive to noise. VIPS [38] implemented + +![](images/059fbaa0bf57d05c170a6d9b333bf75773184ad7642b025d23f8a39f9ab09be3.jpg) +Figure 1. Comparison between our proposed multi-stage mmCooper framework and existing methods. The BEV (Bird's Eye View) communication map illustrates the information shared by agents at each scene location. (a)(b)(c) depict existing methods, which transmit the entire scene's data in a single stage without considering sensing confidence, leading to excessive communication overhead and degraded performance. (d) demonstrates how mmCooper selectively transmits non-overlapping information across multiple stages, reducing communication costs and enhancing model performance. + +efficient graph-structured matching to enable object-level fusion. Intermediate fusion [13, 22, 29, 49, 58] shared features among agents, yet complete Intermediate features still incur high bandwidth consumption. Recent work focused on novel communication mechanisms to filter Intermediate features and reduce communication volume. for instance, Where2comm [13] employed a confidence-based scheme to guide agents in focusing on sharing spatially critical information. ERMVP [58] adopted a hierarchical feature sampling strategy to reduce communication overhead while leveraging sparse consensus features to mitigate localization errors. Different from these works, our framework conducts multi-stage fusion instead of single-stage which balances the perception performance and communication overhead. As far as we know, only a few works have considered multi-stage settings. Zhang et al. [60] proposes a defense method that combines raw data and intermediate features to fuse redundant information, thereby reducing the impact of spurious data. However, this work focuses on security issues related to collaborative perception instead of improving sensing performance. Xie et al. [48] applies reinforcement learning for dynamic partitioning of early, intermediate, and late-stage information. However, this work merely divides the data without integrating information across stages. In contrast to these works, this paper introduces a collaborative perception framework that aims to improve sensing performance by conducting multi-stage data fusion. + +# 3. Methodology + +# 3.1. Overview + +As illustrated in Fig. 1, our key idea is to conduct selective multi-stage agent collaboration, rather than the single + +stage fusion used in existing methods. Conventional approaches indiscriminately transmit data encompassing the entire scene through a single stage, introducing undesired low-quality sensing information during communication. This not only increases communication overhead but also complicates fusion, as the system must first evaluate the quality of received data to ensure reliable results. In contrast, our approach adopts a selective multi-stage collaboration strategy, where the transmission decision, whether to send data and at which stage, is determined based on the sensing confidence level of each location. For high-confidence regions, we transmit late-stage bounding boxes, minimizing communication costs. For moderate-confidence regions, we transmit intermediate-stage features for richer information. For low-confidence regions, transmission is suppressed to prevent misleading other agents. Additionally, our framework incorporates alignment and calibration corrections during both intermediate and late-stage fusion, further enhancing the reliability of the fused results. + +The details of our proposed mmCooper framework are illustrated in Fig. 2, which comprises three main components: (1) the Information Broadcasting, (2) the Intermediate-stage Fusion, and (3) the Late-stage Fusion. The encoded point cloud features will be sent to two branches. In one branch, the Information Broadcasting (Sec. 3.3) employs a Confidence-based Filter Generation Module to dynamically determine whether data should be suppressed, transmitted at the intermediate stage, or transmitted at the late stage. In the other branch, the Intermediate-stage Fusion (Sec. 3.4) contains the Multi-scale Offset-aware Fusion Module, which uses cross-attention to integrate features from both target and neighboring regions. This module mitigates potential misalignment between features from the ego and other agents, ensuring robust feature fusion. In Late-stage Fusion (Sec. 3.5), a BBox Filtering & Calibration Module removes inaccurate bounding boxes from other agents and refines the remaining ones using the ego agent's information-rich intermediate features. The calibrated bounding boxes from all agents will be merged with those generated by the ego agent to produce the final outputs. + +# 3.2. Observation Encoding + +In a multi-agent cooperative perception scenario, consider the collaboration of $n$ agents, represented by the agent set $N = \{1,\dots,n\}$ . The $i$ -th agent serves as the ego vehicle and the remaining $n - 1$ agents act as collaborators. All the agents first broadcast basic positional information to their collaborators, including coordinates and heading angle. In this way, the agents obtain the necessary information to project the information from the collaborators' systems to the ego vehicle's coordinate system. The ego agent processes its point cloud through a shared encoder to extract Bird's Eye View (BEV) features, represented as + +![](images/d6269f5e5b6546661ddf04054790d05e3e55c69b6cc399b8d5e537fcee6076a5.jpg) +Figure 2. Overview of mmCooper framework: a cooperative perception system with adaptive multi-stage fusion. It consists of three main components: (1) Information Broadcasting (Sec. 3.3) filters features and bounding boxes to achieve bandwidth efficiency; (2) Intermediate-stage Fusion (Sec. 3.4) captures surrounding information from the received feature maps for robust feature fusion; (3) Late-stage Fusion (Sec. 3.5) utilizes the information-rich fused features for Filtering & Calibration of the received bounding boxes. + +![](images/a5e8ec273546d64c7a27a2c839a6b787abab5324c70ef45d2e421c46fc4d3a15.jpg) +Figure 3. The design of Filter Generator in CFG Module, generating confidence scores to guide information transmission at each location. + +$F^i = \psi_E(x^i) \in \mathbb{R}^{C \times H \times W}$ , where $\psi_E$ denotes the shared PointPillar [18] encoder, $x^i$ represents the projected local observation, and $C$ , $H$ , and $W$ correspond to the channel, height, and width of the feature map, respectively. + +# 3.3. Information Broadcasting + +As shown in the top-left corner of Fig. 2, when conducting information broadcasting, the Confidence-based Filter Generation (CFG) Module dynamically decides whether data at each location should be suppressed, transmitted as intermediate-stage features, or transmitted as late-stage bounding boxes. Notably, no additional model or data compression techniques [32] are applied in this framework, only raw features or bounding boxes are transmitted. The bounding boxes are generated by the Detection Decoder from the features and the transmitting decision is guided by the confidence filter generated by the Filter Generator (Fig. 3) based on the sensing confidence, as detailed below. + +Filter Generator. The encoded BEV features (i.e., $F^i$ ) are fed into this Confidence Generator to produce intermediate-stage confidence map $C_f^i$ and late-stage confidence map $C_b^i$ for all the locations of the scene using CNN. A Gaussian Filter is then applied to the generated confidence maps to smooth out noise and reduce anomalies, improving the selection of critical regions. To suppress the information of the locations with low confidence, only the top $p\%$ highest- + +confidence positions are selected from each confidence map to generate the spatial filters $M_{f / b,t}^{i}$ (i.e., $M_{f,t}^{i}$ or $M_{b,t}^{i}$ ). Then, for each location, a classification is required to decide which stage to transmit the data. However, using the traditional Softmax will prevent the back-propagation of the gradient due to its non-differentiability as a discrete operation. Instead, we utilize the Gumbel Softmax [16] to obtain approximate samples from a discrete distribution ensuring the propagation of gradient through the straight-through estimator [2] while preserving standard forward propagation. The Gumbel filter $M_{f / b,g}^{i}$ obtained in this process is given by the following equation: + +$$ +M _ {f / b, g} ^ {i} = f _ {m a x} \left(\frac {\exp \left((\log C _ {f / b} ^ {i} + g _ {f / b}) / \tau\right)}{\sum_ {s} \exp \left((\log C _ {s} ^ {i} + g _ {s}) / \tau\right)}\right), +$$ + +where $g_{f / b}$ is Gumbel noise [33], $\tau$ denotes the temperature parameter of the distribution $(\tau >0)$ , and $s\in \{f,b\}$ . Based on the above process, our final filter $\mathcal{M}_{f / b}^{i}$ is represented as: + +$$ +\mathcal {M} _ {f / b} ^ {i} = M _ {f / b, g} ^ {i} \odot M _ {f / b, t} ^ {i}. +$$ + +Based on the aforementioned filter, the intermediate-stage feature $F^i$ and the coarse bounding box $\mathcal{B}^i$ are filtered to obtain the filtered feature $\hat{F}^i$ and coarse bounding box $\hat{\mathcal{B}}^i$ . + +# 3.4. Intermediate-stage Fusion + +The data shared between agents inevitably contains calibration errors, causing potential misalignment in feature maps. To address this, we propose a Multi-scale Offset-aware Fusion (MOF) Module in the intermediate-stage fusion, which aggregates information from neighboring regions, allowing the fusion process at each location to account for potential misalignment. By incorporating the surrounding context, it effectively mitigates the impact of calibration errors, ensuring more accurate and robust agent collaboration. + +Multi-scale Offset-aware Fusion (MOF). The overall pipeline of this module is shown in Fig. 4(a). First, the ego features and received collaborator features are encoded at + +![](images/05a67b23b20765ce2fb208830b51cad3fd57eef5b486d3232e9c55e3e9d2b295.jpg) +Figure 4. (a) The Multi-scale Offset-aware Fusion Module. (b) The Multi-scale Offset-aware Attention Module. + +![](images/256b77720cf1f6ad3e4516fae48de1908056bee20ee335cccbbe30d1cb134912.jpg) +Figure 5. (a) The BBox Filtering & Calibration (BFC) Module. (b) The Deformable BBox Attention (DBA) Module. + +three different scales to capture target information of varying sizes. Next, Multi-scale Offset-aware Attention (MOA) is applied for feature fusion. The features are then upsampled to a uniform size and concatenated across scales to generate the final fused feature $F^i$ . + +Multi-scale Offset-aware Attention (MOA). As shown in Fig. 4(b), we encode the ego feature at each location into queries using MLP. Meanwhile, features from a small neighborhood around the corresponding positions in each collaborator's feature map are extracted as keys and values for cross-attention. Note that low-confidence features are suppressed by collaborators and not transmitted, making them irrelevant during cross-attention. Specifically, we denote the focused location as $(r,c)$ on the feature map, with a neighborhood size of $s\times s$ . The output representation of feature fusion at location $(r,c)$ is expressed as: + +$$ +\mathcal {F} _ {r, c} ^ {i} = \operatorname {C r o s s A t t n} (M L P (F _ {r, c} ^ {i}), \hat {F} _ {r, c, s \times s} ^ {j}, \hat {F} _ {r, c, s \times s} ^ {j}), +$$ + +Where $\hat{F}_{r,c,s\times s}^{j}$ represents the transmitted collaborator features within an $s\times s$ neighborhood centered at position $(r,c)$ . By selectively aggregating the features at this posi + +tion and its neighborhood, we achieve collaboratively robust feature fusion. + +# 3.5. Late-stage Fusion + +The bounding boxes received from collaborators are often redundant, imprecise, misaligned, or even incorrect. To address this, the BBox Filtering & Calibration (BFC) Module in the Late-stage Fusion leverages the information-abundant fused intermediate features to refine the received bounding boxes, either filtering out erroneous ones or correcting their positions for improved accuracy. + +BBox Filtering & Calibration (BFC). As depicted in Fig. 5(a), first, the received bounding boxes and the fused intermediate features will be put into the Deformable BBox Attention (DBA) Module as inputs. It then extracts relevant information for each received bounding box from collaborators, which is processed by a vanilla transformer [45] to generate quality scores and positional offsets. Note that, to prevent unreasonable offsets and scores, we bound the output range using the Tanh activation function. Finally, bounding boxes with low quality scores are discarded, while the remaining ones are refined by applying the predicted positional offsets. The refined bounding boxes are then merged with those detected by the ego agent, followed by Non-Maximum Suppression (NMS) to generate the final results. To train the BFC module, we compute the offset loss $\mathcal{L}_{off}$ using the smooth absolute error loss [11] and the score loss $\mathcal{L}_{score}$ using focal loss [24]. + +Deformable BBox Attention (DBA). As shown in Fig. 5(b), we first encode the received bounding boxes $\hat{\mathcal{B}}^j$ by applying am MLP to their coordinates, size, and orientation information to generate feature representations. These features are then mapped onto the BEV space based on the bounding box locations, forming the bounding box feature map $F_{b}(q)$ . To enhance these representations, we employ the deformable cross-attention mechanism to selectively attend to and integrate relevant information from the fused intermediate features. Specifically, the initial bounding box features $F_{b}(q)$ serve as the query embedding to calculate the offsets of reference points using MLP. Using these offsets, we retrieve the corresponding reference features from the fused intermediate feature map as $F_{r}^{i} = \mathcal{F}^{i}(q + \triangle q_{m})$ . In the meanwhile, an MLP and Softmax are applied on $F_{b}(q)$ to generate the aggregation weights for $F_{r}^{i}$ , which are then used to weight and integrate the reference features. Finally, after another MLP layer, we obtain the enhanced bounding box feature map: + +$$ +D B A (q) = \sum_ {\alpha = 1} ^ {A} W _ {\alpha} \left[ \sum_ {n = 1} ^ {N} \sum_ {m = 1} ^ {M} \omega \left(W _ {\beta} F _ {b} (q)\right) F _ {r} ^ {i} \right] + F _ {b} (q), +$$ + +where $A$ is the number of attention heads, $W_{\alpha}$ and $W_{\beta}$ are learnable parameters, $m$ represents the number of reference locations, and $\omega$ denotes the softmax operation. + +# 3.6. Loss Functions + +We use the regression head and classification head of the Detection Decoder [4] to generate the bounding box detection results. The regression results represent the position, size, and yaw angle of each predefined box, expressed as $O_{reg} = f_{dec}^{r}(\mathcal{F}^{i}) \in \mathbb{R}^{7 \times H \times W}$ . The classification results indicate the confidence score for each predefined box, represented as $O_{cls} = f_{dec}^{c}(\mathcal{F}^{i}) \in \mathbb{R}^{2 \times H \times W}$ . To optimize the proposed system, we use smooth absolute error loss to supervise regression and focal loss for classification, denoted as $\mathcal{L}_{reg}$ and $\mathcal{L}_{cls}$ , respectively. Combining the previous losses, our total loss is represented as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {r e g}} + \mathcal {L} _ {\text {c l s}} + \mathcal {L} _ {\text {o f f}} + \mathcal {L} _ {\text {s c o r e}}. +$$ + +# 4. Experiments + +# 4.1. Datasets and experimental settings + +Datasets. We conduct an extensive evaluation on three benchmark datasets, namely OPV2V [50], DAIR-V2X [56], and V2XSet [49]. OPV2V is a large-scale public V2V cooperative perception simulation dataset, which comprises 73 different scenes, 11,464 frames of point clouds and RGB images, and over 230,000 annotated 3D detection boxes. The training, validation, and test sets are divided into 6,374, 1,980, and 2,170 frames, respectively. DAIR-V2X is a large-scale real-world 3D object detection dataset, including 71,254 frames of point cloud and image data, with the training, validation, and test sets split in a 5:2:3 ratio. The V2XSet dataset is a large-scale synthetic dataset designed for V2X perception. It contains a total of 11,447 frames, which is divided into training, validation, and test sets, consisting of 6,694, 1,920, and 2,833 frames, respectively. + +Evaluation metrics. To evaluate the 3D object detection performance of the baseline and the proposed framework, we use the average precision (AP) [10] at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 as the evaluation metric. The communication volume between agents is expressed by the following equation [13]: + +$$ +\mathbf {V} = \log_ {2} \left(\left(| M | \times H \times W \times C + N _ {b} \times 7\right) \times 3 2 / 8\right), +$$ + +where $|M|$ represents the feature retention ratio, $N_{b}$ denotes the number of bounding boxes per agent, and 7 is used to describe the coordinates, dimensions, and angle information of each bounding box. Each value is represented using float32, with a division by 8 for bytes. + +Implementation details. We implement our model on the PyTorch toolbox [35] and train it on NVIDIA GeForce RTX 4090 GPUs. We use the Adam optimizer and adopt batch sizes of 3, 5, and 3 for the OPV2V, DAIR-V2X, and V2XSet datasets, respectively, with 60 epochs for each. All models use the PointPillars [18] backbone to extract features from raw point clouds. The confidence ratio $p$ selected for the CFG module is $70\%$ for OPV2V, $60\%$ for DAIR-V2X, and + +Table 1. Collaborative perception performance on the OPV2V, DAIR-V2X and V2XSet dataset with a time delay of $100~ms$ , localization errors of $0.2m$ , and heading errors of $0.2^{\circ}$ using Average Percision(AP) $@0.7/0.5$ as metrics. Bold numbers indicate the best results, while underlined numbers represent the second-best results. * represents a variation of PointPillar [18]. + +
ModelsOPV2VDAIR-V2XV2XSet
AP@0.7/0.5AP@0.7/0.5AP@0.7/0.5
No Fusion* [18]48.66/68.7143.57/50.0340.20/60.60
Late Fusion* [18]59.48/79.6234.47/51.1430.75/54.92
Intermediate Fusion* [18]70.82/ 88.4139.38/56.2259.38/83.18
When2com [29]57.55/74.1133.68/48.2041.85/67.41
DiscoNet [22]68.64/84.7240.69/52.6754.11/79.82
Where2comm [13]69.73/85.1643.71/59.5263.77/83.17
V2X-ViT [49]70.06/84.6540.43/53.0861.49/83.63
Select2col [28]62.46/82.3034.82/51.9651.22/76.88
ERMVP [58]69.71/86.6346.96/64.2158.44/81.54
Ours78.11/ 88.9348.27/ 65.1265.86/84.40
+ +$70\%$ for V2XSet, and the surrounding grid size chosen for the MOF is $3 \times 3$ . To closely simulate real traffic conditions, we introduce localization and heading errors with standard deviations of $0.2 \, \text{m}$ and $0.2^{\circ}$ , respectively, sampled from a Gaussian distribution. The communication range is limited to $70 \, \text{m}$ , excluding agents beyond this range from collaborative communication. + +# 4.2.Quantitative Evaluation + +Detection Performance. Tab. 1 presents the comparison of the perception performance of our proposed mmCooper model and various baseline models on OPV2V, DAIR-V2X, and V2XSet datasets with a time delay of $100~ms$ , localization errors of $0.2m$ , and heading errors of $0.2^{\circ}$ . The No Fusion relies solely on ego agent observations. Late Fusion shares only the bounding box results among agents. Intermediate Fusion allows agents to share complete features extracted from raw point clouds. The above three models are all variants based on the PointPillar [18] point cloud detector. Additionally, we consider SOTA models, including When2com [29], DisoNet [22], Where2comm [13], V2X-ViT [49], Select2col [28] and ERMVP [58]. As shown in the table, our proposed mmCooper outperforms all the baselines on both simulated and real-world datasets. Compared to existing SOTA models, mmCooper outperforms the second-best baseline on the OPV2V, DAIR-V2X, and V2XSet datasets by $7.29\% / 0.59\%$ , $1.31\% / 0.91\%$ and $2.09\% / 0.77\%$ in AP@0.7/0.5, respectively. This demonstrates the effectiveness and superiority of our multi-stage fusion framework. + +Communication Volume. Fig. 6 illustrates the model performance under different bandwidth conditions on the OPV2V, DAIR-V2X, and V2XSet datasets. Experimental results show that mmCooper significantly reduces communication costs, achieving performance levels comparable to late fusion. In comparison to other SOTA models, the communication volume required to achieve optimal performance is 9153/156/18305 times more than our method. Notably, on the OPV2V and V2XSet datasets, our + +Figure 6. Collaborative perception performance and communication volumes of all the models on the OPV2V, DAIR-V2X, and V2XSet datasets. +![](images/29ca86b54613944ac6a0ae3caceebc3d9701256e8fb86d85acc02638b9d979e8.jpg) +$\star$ mmCooper(Ours)---ERMVP Where2comm V2X-ViT When2com Select2Col Intermediate Fusion Late Fusion + +![](images/a32c12834aeb4d187f5947c8d06a36b4817758fac8d826c323302e2b345a9a26.jpg) + +![](images/427d19ddd9190e1051b8f28da4d9fb7d715a44fdcb0c0cd1c84df7b6f074901d.jpg) + +method even achieves lower communication volume than late fusion. This is because our CFG module suppresses low-confidence information from the transmission and selectively transmits complementary and exclusive information at intermediate and late stages rather than transmitting through only one stage encompassing all information of the entire scenario. In addition, since some Late Fusion methods (e.g., OpenCOOD [50]) transmit non-processed (e.g., NMS) BBoxes, our mmCooper even achieves a lower communication volume compared to the late fusion model on certain datasets. + +Robustness to Localization Errors and Transmission Delay. We verified the robustness of the agents during collaborative processes in the presence of localization errors and transmission delay on the OPV2V and DAIR-V2X datasets. Following the noise settings in [49], localization noise was sampled from a Gaussian distribution with a mean of $0\,m$ and a standard deviation $\sigma \in \{0, 0.2, 0.4\}m$ . As shown in Tab. 2, as the standard deviation of the localization noise gradually increases, the detection performance of the models decreases. On the DAIR-V2X dataset, when2com even performs worse than the No Fusion method when the localization error exceeds $0.2\,m$ . Our method outperforms all the baselines under all localization noise settings, demonstrating its robustness to varying levels of localization error. Transmission delay can lead to misalignment of features and proposals, thereby impacting detection performance. We evaluated the models on the OPV2V and DAIR-V2X datasets with time delays $\tau \in \{0, 200, 400\}ms$ set to $0\,ms$ , $200\,ms$ , and $400\,ms$ . With increasing time delay, the detection performance of all methods gradually degrades. However, mmCooper achieves the highest Average Precision (AP) over all the models on both datasets. This robustness to localization errors and transmission delay demonstrates the effectiveness of the MOF and BFC modules in addressing the potential misalignment and errors during communication among agents. For more results regarding errors during transmission, please refer to the supplementary. + +![](images/1ed8a4e237d15a7ad46ee7c8230d487ec8ff961ebb94ae4854006a71b7d40fc1.jpg) +Figure 7. Visualization of two-stage fusion on DAIR-V2X. We show the results from the Confidence-based Filtering Generation Module, including discarded information (white background), BBoxes for transmission (red dots), and features for transmission (yellow background). We also show the BFC module results, including uncalibrated BBoxes (blue boxes), calibrated BBoxes (red boxes), and ground truth BBoxes (green boxes). + +# 4.3. Ablation Study + +We conducted comprehensive ablation studies on the OPV2V and DAIR-V2X datasets to demonstrate the importance of different components, as shown in Tab. 3. + +Importance of Core Components. All the core components contribute to performance improvement. Removing the CFG module results in the transmission of information for both the intermediate and late stages across all locations without suppression or selection. While this variant increases the amount of transmitted information compared to mmCooper, it leads to a performance drop. This highlights the importance of filtering out low-confidence information and selectively transmitting data at different stages to optimize performance. The absence of the MOF module replaces the MOA module with a self-attention mechanism, while removing the BFC module eliminates the filtering and refinement of received bounding boxes. Both modifications result in a performance drop, highlighting the importance of incorporating spatial neighborhood information to mitigate misalignment and applying bounding box filtering and refinement to address redundancy, imprecision, or potential errors in received detections. + +Superiority of Multi-stage Fusion. We entirely removed the Late Fusion (LF) or the Intermediate Fusion (IF) to degrade our model to single-stage fusion. The experimental results indicate that removing either module leads to a performance decline, underscoring the effectiveness of the proposed multi-stage cooperation framework. + +Table 2. Results for different localization errors (m) and transmission delay (ms) on the OPV2V and DAIR-V2X dataset evaluated using AP@0.7. Bold numbers indicate the best results, while underlined numbers represent the second-best results. + +
Noise TypeLocalization Errors(m)Transmission Delay(ms)
DatasetsOPV2VDAIR-V2XOPV2VDAIR-V2X
Noise Level0.00.20.40.00.20.402004000200400
No Fusion48.6648.6648.6643.5743.5743.5748.6648.6648.6643.5743.5743.57
When2com [29]70.8264.3559.9839.5036.6235.0370.8243.9636.5539.5036.8333.72
DiscoNet [22]76.0573.5965.8546.9446.0344.9676.0560.0850.3346.9444.0341.57
Where2comm [13]78.4775.4569.7752.3449.3446.9678.4765.2353.3352.3447.7645.25
V2X-ViT [49]77.8375.2166.7246.1243.8142.5377.8366.0850.3346.1245.4943.99
ERMVP [58]80.5576.1972.7853.2748.6645.9780.5568.8968.3153.2749.6848.60
Ours86.4182.5376.8056.0651.5247.6686.4177.4475.2656.0650.9650.81
+ +Table 3. Ablation study results of different designs in mmCooper on the OPV2V and DAIR-V2X datasets. CFG: Confidence-based Filter Generation Module; MOF: Multi-scale Offset-aware Fusion; BFC: BBox Filtering & Calibration Module; LF: Late-stage Fusion; IF: Intermediate-stage Fusion. + +
CFGMOFBFCLFIFAP@0.7/0.5(↑)
OPV2VDAIR-V2X
✓✓✓✓✓78.11/88.9348.27/65.12
Importance of Core Components
×✓✓--73.46/88.8447.66/63.63
✓×✓--72.60/88.5147.27/63.49
✓✓×--76.77/88.5644.31/59.68
Results of Single-stage Fusion
---×✓76.83/88.6546.65/62.84
---✓×66.80/78.6245.78/56.88
+ +# 4.4. Qualitative Evaluation + +Visualization of Two-stage Fusion. As shown in the Fig. 7, we visualize the transmission decisions and bounding boxes refinement in the two-stage fusion on the DAIR-V2X dataset. The yellow background and red dots show the selective transmission of data through two stages. The white background shows the locations discarded by the CFG module, showing a substantial amount of low-confidence or irrelevant information in the scenario. The removal and correction of the uncalibrated blue boxes demonstrates the BFC module's effectiveness in filtering and refining received bounding boxes by eliminating redundancies and improving alignment with the ground truth. + +Visualization of Detection Results. Fig. 8 shows visualization results for a scenario in the DAIR-V2X dataset. Compared to baseline methods, mmCooper demonstrates high-precision 3D object detection, accurately predicting nearly all ground truth objects. In contrast, baseline methods either fail to detect certain vehicles or produce inaccurate bounding boxes. †Refer to the supplementary materials for additional visualized detection results. + +# 5. Conclusion + +In this paper, we have proposed mmCooper, a novel multi-agent, multi-stage, communication-efficient, and + +![](images/2cce017427ea8143b0e824f643696330eb4a7ed0bd22ee14de751e05029b7d46.jpg) + +![](images/9f78615aa87cdb346ef96226927ff946f71c91cfb5838c867aa6950d9e86cb8f.jpg) + +![](images/faafa67d9cccb3cc0d64642f79990f0029143a867411e0ce43e0e47a888f5444.jpg) +(a) Where2comm +(c) ERMVP + +![](images/42380c077f7a2fee54647d54e03ae4c3cce81b2e73a802a4b5529598d8a8e969.jpg) +(b) V2X-ViT +(d) Ours +Figure 8. Visualization comparison of detection results on the DAIR-V2X dataset. Green and red boxes represent the ground truth and the model-predicted bounding boxes, respectively. + +collaboration-robust cooperative perception framework. mmCooper is the first framework to achieve adaptive fusion of complementary data across stages for cooperative perception. We introduce the Confidence-based Filter Generation Module to enable dynamic partitioning of intermediate features and late-stage bounding boxes, balancing communication load and perception performance. Additionally, the Multi-scale Offset-aware Fusion module and BBox Filtering & Calibration module are incorporated to address potential misalignment and calibration noise among agents. mmCooper outperforms the second-best state-of-the-art models on the OPV2V, DAIR-V2X, and V2Xset datasets, achieving improvements of $7.29\%$ , $1.31\%$ , and $2.09\%$ in AP@0.7, respectively, while reducing communication volume by factors of 9153, 156 and 18305. + +Acknowledgements. This work is supported by the National Natural Science Foundation of China under Grant 62272357 and 62302326, Wuhan Science and Technology Project for Key Research and Development under Grant 2024050702030090 and Wuhan Science and Technology Joint Project for Building a Strong Transportation Country under Grant 2024-2-7. + +# References + +[1] Christoph Allig and Gerd Wanielik. Alignment of perception information for cooperative perception. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1849-1854. IEEE, 2019. 1 +[2] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 4 +[3] Kunyang Cai, Ting Qu, Bingzhao Gao, and Hong Chen. Consensus-based distributed cooperative perception for connected and automated vehicles. IEEE Transactions on Intelligent Transportation Systems, 24(8):8188-8208, 2023. 2 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 6 +[5] Cheng Chang, Jiawei Zhang, Kunpeng Zhang, Wenqin Zhong, Xinyu Peng, Shen Li, and Li Li. Bev-v2x: Cooperative birds-eye-view fusion and grid occupancy prediction via v2x-based data sharing. IEEE Transactions on Intelligent Vehicles, 2023. 1 +[6] Guoying Chen, Xuanming Zhao, Zhenhai Gao, and Min Hua. Dynamic drifting control for general path tracking of autonomous vehicles. IEEE Transactions on Intelligent Vehicles, 8(3):2527-2537, 2023. 1 +[7] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 1, 2 +[8] Ziming Chen, Yifeng Shi, and Jinrang Jia. Transiff: An instance-level feature fusion framework for vehicleinfrastructure cooperative 3d detection with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18205-18214, 2023. 1 +[9] Abhishek Das, Theophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. In International Conference on machine learning, pages 1538-1546. PMLR, 2019. 2 +[10] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6 +[11] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 5 +[12] Yedid Hoshen. Vain: Attentional multi-agent predictive modeling. Advances in neural information processing systems, 30, 2017. 2 +[13] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 1, 3, 6, 8 + +[14] Min Hua, Guoying Chen, Buyang Zhang, and Yanjun Huang. A hierarchical energy efficiency optimization control strategy for distributed drive electric vehicles. Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering, 233(3):605-621, 2019. 1 +[15] Tao Huang, Jianan Liu, Xi Zhou, Dinh C Nguyen, Mostafa Rahimi Azghadi, Yuxuan Xia, Qing-Long Han, and Sumei Sun. V2x cooperative perception for autonomous driving: Recent advances and challenges. arXiv preprint arXiv:2310.03525, 2023. 1 +[16] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4 +[17] Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation. Advances in neural information processing systems, 31, 2018. 2 +[18] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 4, 6 +[19] Jesse Levinson, Jake Askeland, Jan Becker, Jennifer Dolson, David Held, Soeren Kammel, J Zico Kolter, Dirk Langer, Oliver Pink, Vaughan Pratt, et al. Towards fully autonomous driving: Systems and algorithms. In 2011 IEEE intelligent vehicles symposium (IV), pages 163-168. IEEE, 2011. 1 +[20] Jinlong Li, Runsheng Xu, Xinyu Liu, Jin Ma, Zicheng Chi, Jiaqi Ma, and Hongkai Yu. Learning for vehicle-to-vehicle cooperative perception under lossy communication. IEEE Transactions on Intelligent Vehicles, 8(4):2650-2660, 2023. 2 +[21] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International Conference on Image Processing, pages 4325-4328. IEEE, 2010. 2 +[22] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 2, 3, 6, 8 +[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1 +[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[25] Bingyi Liu, Weizhen Han, Enshu Wang, Shengwu Xiong, Chunming Qiao, and Jianping Wang. An efficient message dissemination scheme for cooperative drivings via cooperative hierarchical attention reinforcement learning. IEEE Transactions on Mobile Computing, 2023. 1 +[26] Bingyi Liu, Jingxiang Hao, Enshu Wang, Dongyao Jia, Weizhen Han, Libing Wu, and Shengwu Xiong. Multiagent reinforcement learning based resource allocation for + +efficient message dissemination in c-v2x networks. In 2024 IEEE/ACM 32nd International Symposium on Quality of Service (IWQoS), pages 1-10. IEEE, 2024. 1 +[27] Wei Liu, Lu Xiong, Xin Xia, Yishi Lu, Letian Gao, and Shunhui Song. Vision-aided intelligent vehicle sideslip angle estimation based on a dynamic model. IET Intelligent Transport Systems, 14(10):1183-1189, 2020. 1 +[28] Yuntao Liu, Qian Huang, Rongpeng Li, Xianfu Chen, Zhifeng Zhao, Shuyuan Zhao, Yongdong Zhu, and Honggang Zhang. Select2col: Leveraging spatial-temporal importance of semantic information for efficient collaborative perception. IEEE Transactions on Vehicular Technology, 2024. 6 +[29] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2, 3, 6, 8 +[30] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1 +[31] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1 +[32] Yunsheng Ma, Juanwu Lu, Can Cui, Sicheng Zhao, Xu Cao, Wenqian Ye, and Ziran Wang. Macp: Efficient model adaptation for cooperative perception. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3373-3382, 2024. 4 +[33] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. 4 +[34] Zhenyang Ni, Zixing Lei, Yifan Lu, Dingju Wang, Chen Feng, Yanfeng Wang, and Siheng Chen. Self-localized collaborative perception. arXiv preprint arXiv:2406.12712, 2024. 1 +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[36] Andrea Piazzoni, Jim Cherian, Roshan Vijay, Lap-Pui Chau, and Justin Dauwels. Copem: Cooperative perception error models for autonomous driving. In 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), pages 3934-3939. IEEE, 2022. 2 +[37] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the IEEE, 96(10): 1640-1656, 2008. 2 +[38] Shuyao Shi, Jiahe Cui, Zhehao Jiang, Zhenyu Yan, Guoliang Xing, Jianwei Niu, and Zhenchao Ouyang. Vips: + +Real-time perception fusion for infrastructure-assisted autonomous driving. In Proceedings of the 28th annual international conference on mobile computing and networking, pages 133-146, 2022. 1, 2 +[39] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multiagent cooperative and competitive tasks. arXiv preprint arXiv:1812.09755, 2018. 2 +[40] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 1, 2 +[41] Hao Su, Shin'Ichi Arakawa, and Masayuki Murata. 3d multi-object tracking based on two-stage data association for collaborative perception scenarios. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-7. IEEE, 2023. 1 +[42] Sanbao Su, Songyang Han, Yiming Li, Zhili Zhang, Chen Feng, Caiwen Ding, and Fei Miao. Collaborative multi-object tracking with conformal uncertainty propagation. IEEE Robotics and Automation Letters, 2024. 2 +[43] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation. Advances in neural information processing systems, 29, 2016. 2 +[44] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pages 330-337, 1993. 2 +[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5 +[46] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1 +[47] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. Advances in Neural Information Processing Systems, 36, 2024. 1 +[48] Qi Xie, Xiaobo Zhou, Tie Qiu, Qingyu Zhang, and Wenyu Qu. Soft actor-critic-based multilevel cooperative perception for connected autonomous vehicles. IEEE Internet of Things Journal, 9(21):21370-21381, 2022. 3 +[49] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 2, 3, 6, 7, 8 +[50] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 6, 7 +[51] Runsheng Xu, Weizhe Chen, Hao Xiang, Xin Xia, Lantao Liu, and Jiaqi Ma. Model-agnostic multi-agent perception framework. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1471–1478. IEEE, 2023. 2 + +[52] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 1 +[53] Dingkang Yang, Kun Yang, Yuzheng Wang, Jing Liu, Zhi Xu, Rongbin Yin, Peng Zhai, and Lihua Zhang. How2comm: Communication-efficient and collaboration-pragmatic multi-agent perception. Advances in Neural Information Processing Systems, 36, 2024. 1 +[54] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23383-23392, 2023. +[55] Kun Yang, Dingkang Yang, Jingyu Zhang, Hanqi Wang, Peng Sun, and Liang Song. What2comm: Towards communication-efficient collaborative perception via feature decoupling. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7686-7695, 2023. 1 +[56] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 2, 6 +[57] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1 +[58] Jingyu Zhang, Kun Yang, Yilei Wang, Hanqi Wang, Peng Sun, and Liang Song. Ermvp: Communication-efficient and collaboration-robust multi-vehicle perception in challenging environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12575–12584, 2024. 3, 6, 8 +[59] Qingzhao Zhang, Xumiao Zhang, Ruiyang Zhu, Fan Bai, Mohammad Naserian, and Z Morley Mao. Robust real-time multi-vehicle collaboration on asynchronous sensors. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2023. 2 +[60] Qingzhao Zhang, Shuowei Jin, Ruiyang Zhu, Jiachen Sun, Xumiao Zhang, Qi Alfred Chen, and Z Morley Mao. On data fabrication in collaborative vehicular perception: Attacks and countermeasures. In 33rd USENIX Security Symposium (USENIX Security 24), pages 6309-6326, 2024. 3 +[61] Xumiao Zhang, Anlan Zhang, Jiachen Sun, Xiao Zhu, Y Ethan Guo, Feng Qian, and Z Morley Mao. Emp: Edge-assisted multi-vehicle perception. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 545–558, 2021. 1, 2 \ No newline at end of file diff --git a/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/images.zip b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f1e6c3b9db8cd14198b917c361977c6896c49a43 --- /dev/null +++ b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc11aba38c8124f454dc501e318c36df2971249efe99048eb3674877e79b4279 +size 643877 diff --git a/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/layout.json b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1e69b61dacdcda3c4bed0ce595e7a425299b326 --- /dev/null +++ b/mmcooperamultiagentmultistagecommunicationefficientandcollaborationrobustcooperativeperceptionframework/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:838da75b061c80000d0becadc47b535c0f41b93603ceb498d9bb40090f856937 +size 389216 diff --git a/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/22fa6baf-2e02-4481-a20e-d89061e5dbe8_content_list.json b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/22fa6baf-2e02-4481-a20e-d89061e5dbe8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..915bf0fb013bc3ffc05eeb7d7fc5752fbfd9858b --- /dev/null +++ b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/22fa6baf-2e02-4481-a20e-d89061e5dbe8_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39f929bed77a9bb9434ec9840595623686c350d24a0699b595e256bbfc955bcb +size 82093 diff --git a/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/22fa6baf-2e02-4481-a20e-d89061e5dbe8_model.json b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/22fa6baf-2e02-4481-a20e-d89061e5dbe8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..db9b360b396ecc56cdf3681abe2229eff6923d53 --- /dev/null +++ b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/22fa6baf-2e02-4481-a20e-d89061e5dbe8_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21c2b3ded0a90a964abd9a8878d9a4b462487450d591af9945f509b3605d489e +size 102135 diff --git a/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/22fa6baf-2e02-4481-a20e-d89061e5dbe8_origin.pdf b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/22fa6baf-2e02-4481-a20e-d89061e5dbe8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4d425e26806aed44a8ea0fbf108f420b6ebbbed7 --- /dev/null +++ b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/22fa6baf-2e02-4481-a20e-d89061e5dbe8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49726f61661237056678c66e3d5a8219f18221eb4e47cd007f258b022284822f +size 1490708 diff --git a/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/full.md b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..02e75f6904ae4ccc866ec123b7eb3e9df5c76309 --- /dev/null +++ b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/full.md @@ -0,0 +1,322 @@ +# monoVLN: Bridging the Observation Gap between Monocular and Panoramic Vision and Language Navigation + +Renjie Lu $^{1*}$ , Yu Zhou $^{1*}$ , Hao Cheng $^{2}$ , Jingke Meng $^{1\dagger}$ , Wei-Shi Zheng $^{1,3,4}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, $^{2}$ Hunan University, + +3Pengcheng Lab, 4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China {lurj3, zhouy635}@mail2.sysu.edu.cn, haocheng@hnu.edu.cn, mengjke@gmail.com, wszheng@ieee.org + +# Abstract + +Vision and Language Navigation(VLN) requires agents to navigate 3D environments by following natural language instructions. While existing methods predominantly assume access to panoramic observations, many practical robotics are equipped with monocular RGBD cameras, creating a significant configuration disparity. In this work, we address this critical gap by developing a novel 3DGS-based framework for monocular VLN agents, focusing on the intrinsic information incompleteness challenge. Our approach incorporates two key innovations: (1) implicit partial completion module for inferring representations of missing regions in incompletely rendered panoramic feature maps, and (2) an uncertainty-aware active perception strategy that enables the agent to actively acquire visual observation when uncertain about its decision. Extensive experiments on R2R-CE and RxR-CE datasets demonstrate that our monoVLN outperforms all existing monocular methods, significantly improve $8\%$ success rate on R2R-CE compared to previous monocular methods. We also validate our monoVLN in real-world environments, providing a practical solution for real-world VLN. + +# 1. Introduction + +Vision and Language Navigation (VLN) [5, 27] involves guiding an agent to navigate to a specified destination by interpreting natural language instructions. This process typically requires the agent to execute fine-grained actions (e.g., rotating 15 degrees left/right or advancing 0.25 meters) within navigable areas. VLN has attracted significant research interests[3, 10, 11, 13, 17, 20, 21, 34, 43, 46, 50, 52]. + +Current researches in VLN predominantly focus on navigation utilizing panoramic observations, which assumes that the 360-degree area surrounding the agent is fully ac + +![](images/9ee83174b772e30e0a0899655dae9ce76245b7bd2333e29552a883157204d214.jpg) +observed rendered +Panoramic + +![](images/4863e12ee5d2248084a07e9ca8126d13a04a65de710b29e82f4bdc98738b8ec9.jpg) +Panoramic views captured by cameras + +![](images/ebd56567ea52678938cf0f6ac4244e96cfdbe1559e07850dc70df623eb4cca3f.jpg) + +![](images/9786de176a1af84ddcea475abb4a3ead5faece71d9cd9bd764f4a26a367df920.jpg) +Monocular + +![](images/6663b0c854e2d97251129b1bf4aefaa8685456bb81c486efc54c594a6b36777d.jpg) +Rendered incomplete panoramic views +□front +Figure 1. An example of panoramic views captured by cameras and rendered panoramic views. The rendered views are incomplete due to unobserved areas during navigation, resulting in certain regions being rendered with black. + +![](images/0e0e40532f91ac1da27080538faa8efd869804bc3faf2e55ddabee81a2eab9ec.jpg) + +cessible. These methods achieve notable performance but facing significant limitations in real-world robotic deployment. The primary constraint stems from the prevalent use of monocular RGBD cameras in practical robotic applications, which restrict the observable area to a limited field of view around the agent at a step. Consequently, the development of models compatible with monocular camera is imperative for enhancing practicality and compatibility in real-world scenarios. Recent methods [46, 50] address this challenge by reconstructing 3D feature fields that integrate previously observed information and rendering panoramic views, enabling the direct transfer of panoramic VLN models to monocular settings and achieving impressive results. However, they ignore a fundamental challenge inherent to monocular VLN: the intrinsic information incompleteness resulting from the limited field of view. As illustrated in Figure 1, the rendered views by works[46, 50] are often incomplete, which prevents visual encoders from extracting comprehensive features and leads to suboptimal performance. In this work, we present the first attempt to solve this problem. + +An intuitive solution is to perform a 360-degree rotation at each step to mimic panoramic observations. While this method is conceptually simple, it introduces significant inefficiencies, leading to an increase of over $200\%$ in the num + +ber of navigation steps required. To alleviate this, drawing inspiration from human navigation, we first attempt to leverage the agent's subjective perspective, presenting a simple yet effective active perception strategy, enabling it to actively acquire visual observations when uncertain about its decision. Specifically, we focus on two key challenges: determining when to look, which involves identifying the moments for the agent to gather observations, and where to look, which involves selecting a informative direction for visual data collection. This strategy can potentially reduce severe incompleteness shown in Figure 1, where nearly all regions appear black in a given view while avoiding many extra steps. + +For the partial absence of the observation within a view, we further propose an implicit partial completion module. This module leverages contextual feature relationships to infer representations of missing regions rendered from feature field, enabling the extraction of comprehensive features for partially incomplete views. Specifically, to infer features for missing regions, we leverage the MAE framework, which operates on partial features. The view encoder takes rendered partial feature maps as input and generates corresponding latent representations. These latent features are then decoded to reconstruct complete feature maps, guided by the auxiliary loss inspired by MAE[18]. Through this reconstruction-based training paradigm, the decoder learns to predict full feature maps from partial inputs, enabling the view encoder to implicitly infer missing visual information. + +Based on the above, we propose monoVLN, a new 3DGS-based VLN framework designed for practical robotic applications using monocular RGB-D cameras. Unlike previous works [46, 50], which rely on NeRF for constructing 3D feature fields, we employ 3D Gaussian Splating(3DGS). This technique enables fast rendering and a training-free feature field, significantly improving efficiency. We conduct extensive experiments on the widely used R2R-CE[27] and RxR-CE[29] datasets. Our proposed model significantly improves $8\%$ success rate compared to previous monocular methods on R2R-CE, while only including $9\%$ extra steps. To validate practical applicability, we deploy monoVLN on real-world robotic platforms, where it achieves fast inference on RTX 4060 GPU, highlighting its potential for real-world application. + +In summary, our contributions are as follows: + +- We propose monoVLN, a 3DGS-based framework for monocular vision and language navigation, enabling efficiency and effectiveness. +- We present the first attempt to tackle the information incompleteness problem in monocular VLN by introducing implicit completion and active perception. +- Our model achieves a new state-of-the-art on R2R-CE and RxR-CE datasets, with an $8\%$ higher success rate on R2R-CE compared to previous monocular VLN methods. + +# 2. Related Works + +Vision and Language Navigation. The field of natural language-guided navigation in unseen environments[5, 27, 29] has seen substantial progress in recent years. A primary focus has been on architectural innovations, particularly in developing advanced decision-making backbones[10, 20] and enhancing vision modeling techniques[1, 2, 19, 32]. Building upon these foundations, various planning strategies have been proposed, including self-correction mechanisms[25, 35], graph-level planning[11, 34, 44] and hierarchical frameworks[14]. Complementing these advances, visual data augmentation methods[30, 31, 43] have emerged for improving visual generalization ability. Other line of work focuses on enhancing visual-language representation, which leverage back-translation-based data generation[13, 24, 45] and domain-specific pretraining techniques[16, 17, 36, 38, 39]. More recently, [47, 48] focus on scaling data generation, approaching near human performance on the well-recognized R2R benchmark. + +VLN in Continuous Environments. Despite the remarkable progress in VLN, these agents are developed on discrete predefined topological graphs, making it impractical to real-world robot applications. To bridge the gap, VLN-CE[27] pioneers continuous environment navigation by constructing the R2R-CE dataset. Early methods[8, 15, 27, 28, 41] on R2R-CE demonstrated significantly inferior performance compared to their discrete VLN counterparts. Subsequent studies[21, 26] analyze the difference between discrete and continuous VLN and developed frameworks that transfers discrete VLN agents to continuous domains, substantially reducing the gap between these two paradigms. Recently, ETPNav[3] enhances the framework through improved waypoint predictor and topological mapping. HNR[49] introduces a feature field representation combined with an expanded frontier-based action space. + +Real World VLN Models. Both discrete and continuous models are trained and evaluated in simulators. [6] first deploy VLN models in real world. VLMap [22] introduces a framework that constructs BEV maps and use LLMs to generate high-level navigation code. To enable agents to comprehend free-form instructions and avoid noisy depth in real-world scenarios, NaVid [52] utilizes a video-based LLM to process RGB frames and predict actions in text form. While VLN agents are predominantly developed with panoramic observations, recent work 3DFF[50] transfers panoramic VLN models to monocular observations via feature field, g3D-LF[46] further enhances their quality. However, these approaches overlook the inherent information incompleteness of monocular observations. Our work presents the first attempt to tackle this challenge through active perception and implicit inference of complete views, effectively mitigating the issue. + +![](images/bb6f49486227fb6bb321724121baa20ab862147b0ccd37b65f46b3f2dee967f4.jpg) +Figure 2. Typical panoramic VLN model pipeline. The blue node represents the current location, while the gray nodes represent other positions in the environment. The planner predicts a sub-goal on the graph through node classification, guiding the agent to move toward the predicted node. If the target node is a direct neighbor of the current node, the agent simply rotates and moves to it. Otherwise, the agent navigates to the target node by following the shortest path on the graph. + +# 3. Method + +# 3.1. VLN Background + +Problem Formulation. In VLN, an agent is provided with a natural language instruction $\mathcal{I}$ describing the navigation task. At time step $t$ , the agent observes an RGB image $r_t$ and a depth image $d_t$ captured by a monocular RGB-D camera, and its current 4D pose $p_t = (x_t, y_t, z_t, \theta_t)$ , where $(x_t, y_t, z_t)$ represents the agent's coordinates and $\theta_t$ denotes its heading direction. The agent selects an action $a_t$ from a discrete action space $\mathcal{A} = \{\text{turn left } 15^\circ, \text{turn right } 15^\circ, \text{move forward } 0.25\text{m}, \text{stop}\}$ . The navigation task is considered successful when the agent stops at the target location within 3 meters. + +VLN Models with Panoramic Views. In this part, we outline the general pipeline of panoramic VLN models. As illustrated in Figure 2, typical panoramic VLN framework facilitates navigation by constructing a topological map derived from panoramic views and planning paths on this map using visual observations and natural language instructions. Specifically, at step $t$ , the agent observes panoramic RGB images $\mathcal{R}_t = \{r_{t,i}\}_{i=1}^{12}$ and panoramic depths $\mathcal{D}_t = \{d_{t,i}\}_{i=1}^{12}$ towards 12 different orientations with 30 degrees between each. The RGB images $\mathcal{R}_t$ are encoded into features $\{f_{t,i}^v\}_{i=1}^{12}$ , which are stored in nodes[3] or edges[34] of the topological map $\mathcal{M}_t$ . The depth images $\mathcal{D}_t$ are used to predict the coordinates of nearby navigable waypoints[3, 21, 34], enabling the update of the topological map from $\mathcal{M}_{t-1}$ to $\mathcal{M}_t$ . A planner then encodes the topological map $\mathcal{M}_t$ along with the instruction $\mathcal{I}$ and selects the optimal waypoint as the sub-goal to move. With the predicted sub-goal, the agent navigates to it along the shortest path on $\mathcal{M}_t$ . If it fails to reach the sub-goal due to obstacles, we stop it immediately. + +# 3.2. Framework Overview + +We present the overall framework of our monoVLN in Figure 3, designed for practical VLN using monocular RGB-D cameras. Our framework primarily addresses the challenge of intrinsic information incompleteness caused by the limited field of view in monocular observations. Specif + +ically, with the monocular RGB-D observation, we incrementally construct a 3D feature field by leveraging the feature 3DGS [53]. This feature field is used for two functionalities: first, rendering an agent-centric top-down view feature map(BEV feature) input to a waypoint predictor to predict the coordinates of nearby navigable waypoints. Second, rendering panoramic feature maps processed by a implicit partial completion module to enable comprehensive panoramic understanding by inferring missing regions in rendered views. With the rendered panoramic features and predicted waypoints, we construct a topological map and perform decision planning, following the general pipeline of VLN models[3, 11, 21, 34]. When the agent encounters uncertainty in its decision-making process, it selects a informative directions and actively acquires visual observations to refine its decisions. Details are presented below. + +# 3.3. 3DGS-based Feature Field Construction + +In this section, we describe how we construct 3DGS feature field to effectively convert monocular observation to panoramic observation and enable transferring powerful panoramic VLN model to the monocular setting. By reconstructing the feature field during navigation and rendering panoramic views using it, the rendered panoramic views provide richer information compared to monocular observations since the reconstructed feature field integrates historical observations. We opt for a 3DGS-based feature field as it permits us to directly convert a point cloud into Gaussians without training. This choice allows us to concentrate on the core navigation task without diverting attention to scene reconstruction. + +Specifically, at time step $t$ , we utilize MobileSAM[51] to extract a feature map $f_{t}^{r} \in \mathbb{R}^{C \times H \times W}$ for the monocular RGB observation $r_t$ . Using downsampled depth image $d_t$ , the feature map $f_{t}^{r}$ is mapped to 3D world position through the camera pose and intrinsics, resulting in a point cloud where each point is associated with color and feature information. To transform the point cloud into a feature-based 3DGS [53], we transfer the attributes of the points to Gaussians. Each Gaussian has six attributes: coordinates, color, scale, rotation, opacity, and feature. We migrate the coordi + +![](images/2ca813ef62bb940fb9f64f950908d18cf3e51b84564938abfdc7ef3bfecc8fd5.jpg) +Figure 3. The overall framework of our method. We process observed RGB-D images to update the feature field for rendering panoramic feature maps and a BEV map, which are then processed by our Implicit Partial Completion(IPC) module and BEV-based waypoint predictor. Their predictions are fed into mapper and planner to predict a sub-goal. The robot moves toward this sub-goal or only rotates to achieve Uncertainty-aware Active Perception(UAP) based on its prediction uncertainty. + +nates, color, and feature directly from the point cloud while setting the remaining attributes to fixed values. This allows us to treat 3DGS as a point cloud rendering method rather than its conventional use for scene reconstruction. This simple method provides acceptable rendering quality without the need for training. + +# 3.4. BEV-based Waypiont Prediction + +Panoramic VLN models are predominantly designed using topological maps for global planning and backtracking[11, 14, 34]. These approaches typically rely on a waypoint predictor to estimate the coordinates of nearby waypoints for constructing such maps[3, 21]. While these predictors are built upon panoramic depth images, we present a waypoint predictor for the monocular setting that utilizes a top-down view feature map (BEV feature) to accomplish this, where the BEV can be directly derived from the feature field. + +As illustrated in Figure 3, the BEV feature is extracted from the feature field. To generate this map, we position the camera considerably higher above the agent and render an agent-centric BEV feature map $f_{t}^{b} \in \mathbb{R}^{C \times 192 \times 192}$ . Each pixel in the BEV map corresponds to a $5cm \times 5cm$ regions. To avoid occlusions caused by ceilings, we filter out Gaussians located more than 1 meter above the agent's height before rendering. The resulting BEV feature map is fed into the waypoint predictor composed of two U-Nets: the first U-Net completes the BEV feature map, while the second predicts a heatmap representing potential waypoints. Using Non-Maximum Suppression (NMS), discrete waypoints relative to the agent are extracted from the heatmap and converted to world coordinates via the agent's pose. These coordinates are then used to update the topological map from $\mathcal{M}_{t-1}$ to $\mathcal{M}_{t}$ shown in Figure 2. + +The waypoint predictor is trained independently using a heatmap loss [54], with the same dataset as previous works [21, 50]. Additionally, an inpainting loss is applied to ensure the completeness of the BEV feature map. For more implementation details, please refer to the Appendix. + +![](images/abf87501f2f1ac5e7857e4127843dae17d5e3381b1e614388452d65bc95e5cd9.jpg) +Figure 4. Training MAE on incomplete rendered feature map. The rendered incomplete feature map is randomly masked and input to the view encoder. Then the output is processed by the following decoder to reconstruct the full feature map. + +# 3.5. Implicit Partial Completion Module + +The feature fields rendered using 3DGS are often incomplete, particularly in unobserved areas (e.g., regions behind the agent), as demonstrated in Figure 4. For instance, unobserved structures such as stairs are rendered as black areas. To address this limitation, we propose an implicit partial completion module that leverages contextual relationships to infer representations of missing regions and extract comprehensive features from these partially incomplete views. + +To infer features for missing regions, we leverage the MAE framework[18], which operates on partial features. The MAE view encoder takes rendered partial feature maps as input and generates corresponding latent representations. These latent features are then decoded to reconstruct complete feature maps, guided by the auxiliary loss inspired by MAE. Through this reconstruction-based training paradigm, the decoder learns to predict full feature maps from partial inputs, enabling the view encoder to implicitly infer missing visual information. + +Formally, the 3DGS rendered partial feature maps denoted as $\{f_{t,i}^{m}\}_{i = 1}^{12}$ that corresponding to 12 orientations, with a 30-degree angle between adjacent directions. Given the rendered $i_{th}$ feature map $f_{t,i}^{m}$ , we convert it to tokens like ViT[12]. These tokens undergo random masking (set to zero) and, along with an additional [CLS] token, are processed by a transformer-based view encoder. The encoded + +tokens are subsequently decoded by another transformer to reconstruct the complete feature map. We train the completion encoder and decoder using the following reconstruction loss: + +$$ +\mathcal {L} _ {m a e} = \frac {1}{| \mathcal {S} |} \sum_ {x \in \mathcal {S}} \| x - x ^ {g t} \| ^ {2}, \tag {1} +$$ + +where $S$ denotes the set of predicted masked tokens, $x$ represents reconstructed tokens and $x^{gt}$ corresponds to the ground truth tokens. We follow [18] to exclude non-masked patches from the loss computation. Notably, the reconstruction process covers both randomly masked patches and rendered black areas. Upon training completion, the [CLS] token is used to serve as the global representation $f_{i}^{v}$ of the input feature map. For the rendered 12 views, we extract this global feature for each of them and obtain $\{f_{t,i}^{v}\}_{i = 1}^{12}$ . To maintain consistency with standard panoramic VLN training protocols that typically employ CLIP features, we also align $f_{t,i}^{v}$ to CLIP feature by maximizing feature similarity and minimizing distance: + +$$ +\mathcal {L} _ {c l i p} = 1 - \frac {f ^ {v} \cdot f ^ {c l i p}}{\| f ^ {v} \| \| f ^ {c l i p} \|} + \| f ^ {v} - f ^ {c l i p} \| _ {2} ^ {2} + \| f ^ {v} - f ^ {c l i p} \| _ {1}, \tag {2} +$$ + +$f^{clip}$ is the corresponding CLIP feature. We omit script $t, i$ here for clarity. + +# 3.6. Mapping and Planning + +With rendered panoramic features $\{f_{t,i}^{v}\}_{i = 1}^{12}$ and predicted waypoints at each step, we construct a topological map and perform decision planning, following the general VLN models [3, 11, 21, 34] (see Section 3.1). Specifically, topological map $\mathcal{M}_{t - 1}$ is updated by these predicted features and waypoints to produce $\mathcal{M}_t$ . Subsequently, the planner predicts an node $a_{t}$ guided by the instruction $\mathcal{I}$ : + +$$ +p _ {t} = \operatorname {P l a n n e r} (\mathcal {I}, \mathcal {M} _ {t}), \tag {3} +$$ + +$$ +a _ {t} = \arg \max p _ {t}. +$$ + +Here, $p_t$ is predicted probabilities and $a_t$ corresponds to a node on $\mathcal{M}_t$ and serves as a sub-goal for the navigation. The planner is trained using the ground-truth next node $a_t^*$ by minimizing the cross-entropy loss at each step. The overall loss for the planner is defined as follows: + +$$ +\mathcal {L} _ {\text {p l a n n e r}} = \frac {1}{T} \sum_ {t = 1} ^ {T} \mathrm {C E} \left(p _ {t}, a _ {t} ^ {*}\right). \tag {4} +$$ + +$T$ is the total planning step and CE is cross entropy loss. + +# 3.7. Uncertainty-aware Active Perception + +Predictions made by the planner may be compromised, as the implicit partial completion module cannot retrieve information from entirely unobserved directions. To alleviate + +this, we draw inspiration from human navigation and propose leveraging the agent's subjective perspective through a simple yet effective active perception strategy. This strategy enables the agent to actively acquire visual observations when uncertain about its decisions. We focus on two key challenges: determining when to look, which involves identifying the steps for the agent to gather observations, and where to look, which involves selecting a informative direction for more visual observations. + +(1) When to look? The agent should gather more information when it finds its prediction unreliable, as additional information have potential to improve prediction accuracy. We assess the reliability of a prediction by considering two conditions: directional novelty and low confidence. When both conditions are satisfied, the agent decides to gather more visual information. Directional novelty is defined as a direction pointing towards a previously unobserved sector, specifically any orientation outside the $\pm 30^{\circ}$ range of already viewed directions at current node. Low confidence is determined by thresholding the class-normalized entropy: $H_{norm} = H(p) / \log N > \tau$ , where $H(p)$ is the entropy of prediction $p_t$ and $N$ is the number of candidates. Threshold $\tau$ is a hyper-parameter. This normalization enables stable thresholding across varying class numbers. + +(2) Where to Look? When deciding to acquire additional visual information, we choose to have the agent look toward the orientation of the predicted node. Ideally, the optimal direction would be that of the ground truth(GT) next node. However, during inference, we cannot obtain this GT direction. Our strategy is to utilize the direction of the model-predicted node instead. This approach remains informative as the prediction model has been trained under the supervision of ground truth data. + +Our solution to these questions culminates in the simple Uncertainty-aware Active Perception(UAP) strategy: the agent rotates to where the model-predicted node is when it is uncertain about its prediction. This inserts a checking step in the traditional combined rotate-then-move action, as depicted in Fig.3. The checking step enables the agent to determine whether to move forward or plan again(i.e. replan) with acquired visual information in the rotation. Replan may occur several times until the agent makes a certain decision or every view is observed. + +# 4. Experiments + +# 4.1. Setups + +We evaluate our approach on the R2R-CE[27] and RxR-CE[29] datasets in simulated environments. + +R2R-CE is a dataset based on MP3D[7] using Habitat simulator[42]. It includes 5,611 trajectories divided into train, validation seen, validation unseen, and test unseen splits. Each trajectory has 3 English instructions, with an + +average length of 9.89 meters and average instruction length of 32 words. The camera's FOV is 90 degrees. + +RxR-CE is a larger multilingual VLN dataset containing 126K instructions in English, Hindi, and Telugu. It includes diverse trajectories in terms of length (the average is 15 meters), which is more challenging in continuous environments. The camera's FOV for RxR-CE is 63 degrees. + +Evaluation Metrics. We adopt standard metrics for evaluating the agent's performance in VLN, including Navigation Error(NE), Success Rate(SR), SR given the Oracle stop policy(OSR), Success rate weighted by normalized inverse Path Length(SPL[4]), the normalized DTW distance between the agent's trajectory and ground truth path, nDTW weighted by success rate(sDTW[23]). + +Training of each module. The proposed BEV-based waypoint predictor (in Section 3.4) and implicit partial completion module (in Section 3.5) are trained independently before the overall navigation framework pipeline. Specifically, for training the BEV-based waypoint predictor, we randomly navigate between nodes in the graph, collecting rendered BEV features and their corresponding ground truth(GT) waypoint labels, following the methodology of prior works[21, 50], to train the predictor. We train the implicit partial completion module using the samples collected from the MP3D Simulator[7, 42]. To prevent information leakage, we exclude data from the MP3D unseen and test splits. We also include HM3D dataset[40] data, following prior works [49, 50], to enhance generalization. + +Implementation Details. All models are trained with AdamW[33] optimizer. Scale of Gaussians is $2.5\mathrm{cm}$ , with opacity 1 and no rotation. MAE mask ratio is set to 0.5. For the R2R dataset, the planning model is trained on two NVIDIA 3090 GPUs for $10^{4}$ iterations, with batch size 4 per GPU, requiring approximately 2 days. The learning rate is initialized at $5 \times 10^{-6}$ and follows a cosine decay schedule. During training, the planning model is initialized with weights pre-trained on panoramic setting. For the RxR dataset, the planning model is similarly trained for $2 \times 10^{4}$ iterations with learning rate $5 \times 10^{-6}$ on 4 NVIDIA 3090 GPUs with batch size 2 per GPU, taking about 6 days. + +# 4.2. Ablation Study + +To evaluate the effectiveness of each module, we conduct ablation studies. Row 1 represents the baseline model, which is our proposed 3DGS-based navigation framework excluding the Implicit Partial Completion(IPC) module and the Uncertainty-Aware Active Perception(UAP) strategy. Row 2 incorporates the UAP strategy, while Row 3 adds the IPC module. Row 4 corresponds to the full model. + +Comparing Row 1 with Row 2, the $1.7\%$ success rate improvement confirms the effectiveness of our active perception strategy. Further analysis of Row 1 versus Row 3 + +
IPCUAPNE↓OSR↑SR↑SPL↑
4.8359.151.141.7
✓4.8759.852.843.2
✓4.7060.754.044.4
✓✓4.6162.454.844.4
+ +reveals the critical role of the IPC module: its $2.9\%$ performance gain demonstrates it extracts comprehensive representation from partially rendered feature maps. When integrating both UAP and IPC (Row 4), the full model achieves optimal performance across key metrics, notably $+3.7\%$ success rate. This synergy arises because the two components address information incompleteness in complementary ways: the UAP strategy proactively reduces significant incompleteness through active perception, while the IPC module implicitly enhances understanding of incomplete observations via implicit completion. + +Discussion. Our framework employs a different backbone (MobileSAM), 3DGS-based feature field, and BEV-based waypoint predictor, distinguishing it from 3DFF, which uses CLIP, NeRF-based feature field, and semantic-map-based waypoint predictor. Even without IPC and UAP modules, our baseline outperforms it. These may be due to better feature field and BEV-based waypoint predictor that don't rely on high-accuracy semantic segmentation. However, our main focus is on solving information incompleteness, so we skip detailed analysis of these components. + +# 4.3. Comparison With Other VLN Methods + +Table 1. Ablation study. + +
MethodsObsVal UnseenTest Unseen
NE↓OSR↑SR↑SPL↑NE↓OSR↑SR↑SPL↑
ETPNav[3]P4.716557495.12635548
BEVBert[2]P4.576759504.70675950
PRET[34]P4.586659494.74625647
CM2[15]M7.0241.534.327.67.74393224
WS-MGMap[9]M6.2847.638.934.37.11453528
NaVid[52]M5.4749.137.435.9----
3DFF[50]M5.9555.844.930.46.2454.443.728.9
g3D-LF[46]M5.7059.547.234.66.0057.546.332.2
OursM4.6162.454.844.44.9760.553.644.9
+ +Table 2. Evaluation on R2R dataset. Obs indicates the observation type. P means panoramic observation and M indicate monocular observation. + +Comparison on R2R-CE. Table 2 compares our approach with previous VLN methods. Our model, built upon the PRET panoramic planning model, outperforms all other monocular approaches. Specifically, our method achieves a success rate approximately $10\%$ higher than the 3DFF model. We also significantly outperform g3D-LF, suggesting that improving the feature field alone is insufficient for monocular observations, as information incompleteness remains a more critical issue. Importantly, our approach significantly narrows the performance gap between monocular + +
MethodsObsVal Unseen
NE↓OSR↑SR↑SPL↑sDTW↑
PRET[3]P5.8160.052.543.942.7
ETPNav[3]P5.64-54.844.845.3
HNR[49]P5.51-56.446.747.2
CM²[15]M8.9825.314.49.2-
WS-MGMap[9]M9.8329.815.012.1-
NaVid[52]M8.4134.523.812.2-
3DFF[50]M8.7936.725.518.1-
OursM8.2937.731.826.825.2
+ +and panoramic models, reducing it from $12\%$ to $4\%$ in terms of success rate on the val unseen split. + +Comparison on RxR-CE. Table 3 presents the performance comparison on the RxR-CE dataset. Our proposed monoVLN outperforms all previous monocular-based methods. Specifically, our method improves $6\%$ on the success rate compared to previous monocular-based SOTA methods, indicating that the information incompleteness is also a critical challenge on RxR-CE dataset. Results on R2R-CE and RxR-CE demonstrate the effectiveness of our method on different benchmarks. + +# 4.4. Design Choices of Each Module. + +Table 3. Evaluation on RxR-CE dataset. + +
NE↓OSR↑SR↑SPL↑
(1)5.9848.240.832.3
(2)4.6060.153.043.8
(3)4.7060.754.044.4
+ +How to incorporate the reconstruction auxiliary task in IPC module? The MAE view encoder is designed to achieve feature completion (Eq.1) and alignment with CLIP features (Eq.2). To this end, we explore three integration strategies for training: reconstruction-only, joint training, and pretrain-then-finetune. As demonstrated in Table 4, the exclusion of reconstruction during the training process results in substantially diminished performance metrics. The absence of CLIP alignment results in the lowest performance, primarily because the planning model is pretrained solely with CLIP features. Through the integration of both reconstruction and clip alignment during the view encoder training phase, we observe a marked improvement in model performance. Moreover, pretraining with reconstruction loss followed by finetuning with CLIP alignment produces the best results, achieving a $1\%$ improvement in success rate compared to joint training. Based on these empirical results, we adopt the pretrain-then-finetune paradigm in all other experiments. + +Comprehensive study of our active perception strategy. + +We evaluate the effectiveness of our Uncertainty-Aware Ac + +Table 4. Different ways to incorporate the reconstruction task. (1) Training with reconstruction loss (in Eq.1) only. (2) Training with joint CLIP aligning loss (in Eq.2) and reconstruction loss. (3) Pretrain with reconstruction then finetune with CLIP alignment. + +
Methodsinitial rotationNE↓OSR↑SR↑SPL↑num steps↓
w/o UAP5.7751.945.635.1108
ours5.4356.548.735.4153
rotate 360✓4.7660.954.245.4287
random view✓4.6462.154.244.2127
w/o UAP✓4.7060.754.044.4117
ours✓4.6162.454.844.4128
+ +Table 5. Effectiveness of our Active Perception Strategy. + +tive Perception(UAP) strategy, both with and without initial rotation, in Table 5. The initial rotation refers to the agent performing a 360-degree rotation before navigation, as implemented in 3DFF [50]. As shown, the first two rows (without initial rotation) demonstrate a significant improvement in success rate with our active perception strategy, achieving a $3.1\%$ absolute gain. To ensure a fair comparison, we also test the setup with initial rotation1. Rows 5 and 6 of the table reveal that the UAP strategy still enhances the success rate, albeit by a smaller margin of $0.8\%$ . This reduced gain can be attributed to the diminished incompleteness in the rendered views when the initial rotation is performed. Under this experimental setup, we further explore the performance of the approach that involves a 360-degree rotation at every step in Row 3 and acquires a random view observation within the UAP strategy in Row 4. Notably, our UAP strategy outperforms the 360-degree rotation approach in terms of the success rate. The reason behind this is that panoramic views obtained through continuous 360-degree rotations may introduce redundant information. In contrast, our method is designed to selectively collect data only when uncertainty is detected, enabling it to focus on capturing the most relevant details. Moreover, our approach exhibits remarkable efficiency in terms of navigation steps. It significantly reduces the number of extra steps, from 170 in the 360-degree rotation approach to only 11, which further demonstrates the superiority of our UAP strategy. + +Impact of entropy threshold in UAP. We investigate the optimal entropy threshold for the UAP strategy. As depicted in Figure 5, we train the model by utilizing a range of entropy thresholds, considering scenarios both with and without an initial rotation. A higher entropy threshold leads to a reduction in the frequency of replanning. This, in turn, results in a decrease in the number of navigation steps and model forward passes. When the threshold equals 1, no replanning occurs, equivalent to the baseline without UAP. In the scenario with an initial rotation, the optimal entropy threshold is 0.75, which is slightly larger than the optimal threshold of 0.5 in the scenario without an initial rotation. The underlying reason for this difference is that the 360 - degree rotation at the onset of navigation provides a more comprehensive panoramic understanding. + +![](images/a14cd0b75c63fae45b2ff27e9392d5076517326bb110f2d2e518bd772ae668ce.jpg) + +![](images/1b0c8794e0d497623ef8e838eb54fae4b3f297ca909d30a3d5813188fb6a45c4.jpg) +Figure 5. Success rate under different entropy threshold with or without initial rotation. + +Discussion. Our experiments imply that monocular VLN can match panoramic performance with improved methods, and current panoramic methods can be further enhanced. As illustrated in Table 5, performing a full 360-degree rotation at each step does not yield the highest success rate. Moreover, according to Figure 5, the best performance is not attained when the maximum amount of visual information is collected at a threshold of 0. This suggests that current panoramic agents can be enhanced by eliminating redundant observations, thus revealing a promising direction for improving VLN agents. + +# 4.5. Visualization + +![](images/719c0faae5b1d17fef21d5f146f38f8af3e4d8c115060bab37cfc04207ae78d7.jpg) +Figure 6. Visualization of completed feature maps in unseen environments. The feature map is visualized by PCA like DINOv2[37]. Baseline is the rendered feature map whereas ours is feature map completed by MAE decoder. + +We visualize the rendered feature maps in unseen environments in Figure 6. As shown, compared with baseline, our pretrained module can reconstruct small background regions(e.g., row 2, the ceiling) and completes large black areas such as stairs and floors, even though these areas were not observed. Meanwhile, our approach not only reconstructs missing feature maps but also refines those that are blurry or noisy. For instance, in row 1, the rendered feature maps exhibit significant noise, whereas the refined versions are cleaner and more closely aligned with the ground truth, thereby reducing the reliance on high-quality feature fields. + +Instruction: Walk out the room and turn right, move forward a bit and turn left, walk towards the corner and turn right, stop there. + +![](images/9a5192520f45f5a32cc9f877dcb0c3a510128badd54dfb4f1c13015b82159073.jpg) +Figure 7. Real world Example. + +
GPUbackbonewaypointviewplannertotal
406089ms140ms146ms8ms383ms
309031ms27ms99ms10ms167ms
+ +Table 6. Inference time on different GPUs. Backbone is MobileSAM's forward time. Waypoint includes BEV rendering and prediction time. View covers panoramic rendering and encoding time. Planner is VLN model forward time. + +# 4.6. Properties of Our Method. + +Fast Inference. Our model is efficient and achieves fast inference. We show the inference time of each module in Table 6. On 3090, our model's inference is significantly faster than g3D-LF[46], which is running at more than 330ms on 4090. On 4060, the total inference time is 383ms, comparable to g3D-LF on 4090. The planner inference time on 4060 is smaller than 3090, this accounts for 4060's higher clock speed, making it faster on small model inference. + +Deployable on Real-world Robot. We deploy our model on Turtlebot4. The inference is done on a laptop with RTX 4060 GPU. As depicted in Figure 7, with given instruction, our model can navigate following it with monocular observation. We find a dataset bias that substantially degrades performance upon deployment. We provide more details in supplementary materials. + +# 5. Conclusions + +We propose a novel 3DGS-based vision and language navigaiton framework with monocular observation. We tackle the information incompleteness challenge by implicitly completing partial feature maps. We also design a simple uncertainty-aware active perception strategy to gather more visual information to mitigate the challenge. Our experiments imply monocular VLN can match panoramic performance with improved methods. + +Limitations and future work. The generalization of our proposed monoVLN to real-world (e.g., our lab) deployment is limited due to the domain gap between the training data from the simulator and real-world scenarios. This limitation will be addressed in our future work. + +# Acknowledgments + +This work was supported partially by NSFC (No. 62206315, 92470202, U21A20471), Guangdong NSF Project (No. 2024A1515010101, No. 2023B1515040025), Guangdong Key Research and Development Program (No.2024B0101040004), Guangzhou Basic and Applied Basic Research Scheme (No. 2024A04J4067). + +# References + +[1] Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view enhanced model for vision and language navigation. In ACMMM, 2021. 2 +[2] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. Bevbert: Multimodal map pre-training for language-guided navigation. In ICCV, 2023. 2, 6 +[3] Dong An, Hanqing Wang, Wenguan Wang, Zun Wang, Yan Huang, Keji He, and Liang Wang. Etpnav: Evolving topological planning for vision-language navigation in continuous environments. PAMI, 2024. 1, 2, 3, 4, 5, 6, 7 +[4] Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir R. Zamir. On evaluation of embodied navigation agents. CoRR, 2018. 6 +[5] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, 2018. 1, 2 +[6] Peter Anderson, Ayush Shrivastava, Joanne Truong, Arjun Majumdar, Devi Parikh, Dhruv Batra, and Stefan Lee. Simto-real transfer for vision-and-language navigation. In CoRL, 2020. 2 +[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. 3DV, 2017. 5, 6 +[8] Kevin Chen, Junshen K. Chen, Jo Chuang, Marynel Vázquez, and Silvio Savarese. Topological planning with transformers for vision-and-language navigation. In CVPR, 2021. 2 +[9] Peihao Chen, Dongyu Ji, Kunyang Lin, Runhao Zeng, Thomas H. Li, Mingkui Tan, and Chuang Gan. Weakly-supervised multi-granularity map learning for vision-and-language navigation. In NeurIPS, 2022. 6, 7 +[10] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. NeurIPS, 2021. 1, 2 +[11] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022. 1, 2, 3, 4, 5 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4 +[13] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, 2018. 1, 2 +[14] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In CVPR, 2023. 2, 4 +[15] Georgios Georgakis, Karl Schmeckpeper, Karan Wanchoo, Soham Dan, Eleni Miltsakaki, Dan Roth, and Kostas Dani-ilidis. Cross-modal map learning for vision and language navigation. In CVPR, 2022. 2, 6, 7 +[16] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, 2021. 2 +[17] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In CVPR, 2020. 1, 2 +[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 4, 5 +[19] Keji He, Chenyang Si, Zhihe Lu, Yan Huang, Liang Wang, and Xinchao Wang. Frequency-enhanced data augmentation for vision-and-language navigation. In NeurIPS, 2023. 2 +[20] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent vision-and-language bert for navigation. In CVPR, 2021. 1, 2 +[21] Yicong Hong, Zun Wang, Qi Wu, and Stephen Gould. Bridging the gap between learning in discrete and continuous environments for vision-and-language navigation. In CVPR, 2022. 1, 2, 3, 4, 5, 6 +[22] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In ICRA, 2023. 2 +[23] Gabriel Ilharco, Vihan Jain, Alexander Ku, Eugene Ie, and Jason Baldridge. General evaluation for instruction conditioned navigation using dynamic time warping. In NeurIPS Workshop, 2019. 6 +[24] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning. In CVPR, 2023. 2 +[25] Liyiming Ke, Xiujun Li, Yonatan Bisk, Ari Holtzman, Zhe Gan, Jingjing Liu, Jianfeng Gao, Yejin Choi, and Siddhartha S. Srinivasa. Tactical rewind: Self-correction via backtracking in vision-and-language navigation. In CVPR, 2019. 2 +[26] Jacob Krantz and Stefan Lee. Sim-2-sim transfer for vision-and-language navigation in continuous environments. In ECCV, 2022. 2 + +[27] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, 2020. 1, 2, 5 +[28] Jacob Krantz, Aaron Gokaslan, Dhruv Batra, Stefan Lee, and Oleksandr Maksymets. Waypoint models for instruction-guided navigation in continuous environments. In ICCV, 2021. 2 +[29] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In EMNLP, 2020. 2, 5 +[30] Jialu Li, Hao Tan, and Mohit Bansal. Envedit: Environment editing for vision-and-language navigation. In CVPR, 2022. 2 +[31] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In ICCV, 2021. 2 +[32] Rui Liu, Wenguan Wang, and Yi Yang. Volumetric environment representation for vision-language navigation. In CVPR, 2024. 2 +[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 +[34] Renjie Lu, Jingke Meng, and Wei-Shi Zheng. PRET: planning with directed fidelity trajectory for vision and language navigation. In ECCV, 2024. 1, 2, 3, 4, 5, 6 +[35] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In ICLR, 2019. 2 +[36] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In ECCV, 2020. 2 +[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael G. Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. CoRR, 2023. 8 +[38] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: history-and-order aware pretraining for vision-and-language navigation. CoRR, 2022. 2 +[39] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP+: history-enhanced and order-aware pre-training for vision-and-language navigation. TPAMI, 2023. 2 +[40] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M. Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X. Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3D): 1000 large + +scale 3d environments for embodied AI. In NeurIPS, 2021. 6 +[41] Sonia Raychaudhuri, Saim Wani, Shivansh Patel, Unnat Jain, and Angel X. Chang. Language-aligned waypoint (LAW) supervision for vision-and-language navigation in continuous environments. In EMNLP, 2021. 2 +[42] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. In ICCV, 2019. 5, 6 +[43] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In NAACL, 2019. 1, 2 +[44] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In CVPR, 2021. 2 +[45] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks. In CVPR, 2022. 2 +[46] Zihan Wang and Gim Hee Lee. g3d-lf: Generalizable 3d-language feature fields for embodied tasks. CoRR, 2024. 1, 2, 6, 8 +[47] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV, 2023. 2 +[48] Zun Wang, Jialu Li, Yicong Hong, Songze Li, Kunchang Li, Shoubin Yu, Yi Wang, Yu Qiao, Yali Wang, Mohit Bansal, and Limin Wang. Bootstrapping language-guided navigation learning with self-refining data flywheel. CoRR, 2024. 2 +[49] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, Junjie Hu, Ming Jiang, and Shuqiang Jiang. Lookahead exploration with neural radiance representation for continuous vision-language navigation. In CVPR, pages 13753-13762, 2024. 2, 6, 7 +[50] Zihan Wang, Xiangyang Li, Jiahao Yang, Yeqi Liu, and Shuqiang Jiang. Sim-to-real transfer via 3d feature fields for vision-and-language navigation. In CoRL, 2024. 1, 2, 4, 6, 7 +[51] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 3 +[52] Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and He Wang. Nvid: Video-based vm plans the next step for vision-and-language navigation. RSS, 2024. 1, 2, 6, 7 +[53] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In CVPR, 2024. 3 +[54] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. CoRR, 2019. 4 \ No newline at end of file diff --git a/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/images.zip b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5da220ea6cdc8f0fe8da53bdad8b4148b7093e43 --- /dev/null +++ b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d521ac33d1391889cbb1d87a1a3cf67931f74b09dc9120e9a5c96f12f5a97c6f +size 348128 diff --git a/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/layout.json b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f5030fa9b28e4ea25093873b92c190a8c501b6cb --- /dev/null +++ b/monovlnbridgingtheobservationgapbetweenmonocularandpanoramicvisionandlanguagenavigation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d131a3ffc91259b4de4223a82117b58b3f07e7f3451af1e047ff23b97b5fd0dc +size 395298 diff --git a/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/87692093-3778-42e8-bdc6-19e39b1c5e1c_content_list.json b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/87692093-3778-42e8-bdc6-19e39b1c5e1c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..75f70e7f93b2259668236a680320dc36df0d3feb --- /dev/null +++ b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/87692093-3778-42e8-bdc6-19e39b1c5e1c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c21de20dc9c0b654568e56ae226c5b7479fc0161038150f843adf10d9f185d7 +size 79913 diff --git a/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/87692093-3778-42e8-bdc6-19e39b1c5e1c_model.json b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/87692093-3778-42e8-bdc6-19e39b1c5e1c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6d9e9176bc36c8dde91497ca78e068b14ca2b501 --- /dev/null +++ b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/87692093-3778-42e8-bdc6-19e39b1c5e1c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:886f274c939a1b585e2ece4dbfacffe3bc0501df1d2c09cd4f09cec4b862245f +size 100495 diff --git a/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/87692093-3778-42e8-bdc6-19e39b1c5e1c_origin.pdf b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/87692093-3778-42e8-bdc6-19e39b1c5e1c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..04474cc02b22fe01d9aeb77b84cbdd7db4a84d30 --- /dev/null +++ b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/87692093-3778-42e8-bdc6-19e39b1c5e1c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a859ca7a749fedc200799deb1b2caaa5b4d277229390e5299bba3e8afba1745d +size 2591427 diff --git a/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/full.md b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..982b72d6957d0e289f1320fc7e7d9660bc9871ff --- /dev/null +++ b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/full.md @@ -0,0 +1,322 @@ +# $\pi$ -AVAS: Can Physics-Integrated Audio-Visual Modeling Boost Neural Acoustic Synthesis? + +Susan Liang, Chao Huang, Yunlong Tang, Zeliang Zhang, Chenliang Xu +University of Rochester + +# Abstract + +The Audio-Visual Acoustic Synthesis (AVAS) task aims to model realistic audio propagation behavior within a specific visual scene. Prior works often rely on sparse image representations to guide acoustic synthesis. However, we argue that this approach is insufficient to capture the intricate physical properties of the environment and may struggle with generalization across diverse scenes. In this work, we review the limitations of existing pipelines and address the research question: Can we leverage physical audio-visual associations to enhance neural acoustic synthesis? We introduce Physics-Integrated Audio-Visual Acoustic Synthesis (PI-AVAS or $\pi$ -AVAS), a novel framework designed with two key objectives. i) Generalization: We develop a vision-guided audio simulation framework that leverages physics-based sound propagation. By explicitly modeling vision-grounded geometry and sound rays, our approach achieves robust performance across diverse visual environments. ii) Realism: While simulation-based approaches offer generalizability, they often compromise on realism. To mitigate this, we incorporate a second stage for data-centric refinement, where we propose a flow matching-based audio refinement model to narrow the gap between simulation and real-world audio-visual scenes. Extensive experiments demonstrate the effectiveness and robustness of our method. We achieve state-of-the-art performance on the RWAVS-Gen, RWAVS, and RAF datasets. Additionally, we show that our approach can be seamlessly integrated with existing methods to significantly improve their performance. + +# 1. Introduction + +The audio-visual acoustic synthesis task, as introduced by Chen et al. [6] and Liang et al. [21], aims to generate realistic binaural audio for new speaking and listening positions based on vision data. This task presents unique challenges, including synthesizing realistic binaural audio and achieving novel-view synthesis. Various approaches have been proposed to address this problem [6, 9, 21, 22, 26, 43]. NAF [26] simulates sound propagation within a scene us + +![](images/daf58c8a33a6bdb268f452ec066e2fa8d87a5c64ecd9b42a20ce2054efd080ce.jpg) +Geometry-Based Audio Simulation + +![](images/0840dc61c346dcaf56f71b816ff62ee5f445a7d085e478fceb37eccb2023ddaa.jpg) +Realism: +Generalization: +Figure 1. An intuitive comparison between audio simulation approaches, neural rendering approaches, and our $\pi$ -AVAS. Physics-based approaches generalize well but lack realism, while neural rendering produces high-quality results but struggles with novel sound sources. Our two-stage method achieves both realism and generalization using the physics-based sound simulation and audio refinement model. + +![](images/f7011286854e098e7b9124a2c220cf943d5d20103457b0fe34eb05009bd0311e.jpg) +Vision-Informed Neural Rendering + +![](images/ff3a52588a9b3383a3e82d5a8eec0e1e0c09bec481917488cb5890f706fe6f93.jpg) +Realism +Generalization +× + +![](images/6b5d10af15e4e0e852a282b5f7c4bbfe6d07f0ec83e40cda254f4dc1e8741939.jpg) +Ours (two stages) +Realism: + +![](images/bd33174e51765f0005e27aae565dbce962a8f65dec7551beb8df4874edfc30d7.jpg) +Generalization: +# + +ing a local feature grid and an implicit decoder. INRAS [43] divides sound modeling into a three-stage implicit neural field. Chen et al. [6] present a vision-conditioned audio transformation network to synthesize audio for new listening positions. Liang et al. [21] design a NeRF-like system to jointly render audio and visual content. + +While these methods produce plausible results for fixed sound sources, they struggle to generalize to novel sound sources due to inadequate modeling of sound propagation. Sound propagation is influenced by various environmental factors, including the positions of the sound emitter and receiver, the room geometry, and the material properties of surfaces. The complex interactions among these physical properties impact the propagation behavior in a given environment. To address these factors, existing methods typically learn from sparse images implicitly within neural networks. Although this implicit modeling of visual information can support audio synthesis, it falls short of accurately capturing the necessary physical properties that govern sound propagation. In other words, the limited image-audio pairs do not provide sufficient information to infer the + +complete room geometry and scene layout. Consequently, these models experience significant performance degradation when confronted with novel sound sources. + +Recognizing the limitations of implicit learning approaches in capturing physical audio-visual associations, we investigate the question: Can we leverage explicit physical priors in audio-visual modeling to bridge this gap? In this paper, we propose a novel two-stage method to address this challenge. In the first stage, we design a vision-guided sound simulation framework that improves generalization for novel emitter and receiver positions. We begin by reconstructing a 3D mesh environment from a set of images using NeRF [29] or Gaussian Splitting [18]. This scene mesh effectively captures the geometry and structure of the environment, which are critical factors influencing sound propagation. For a given sound source and receiver pair, we model sound propagation within the mesh scene by treating sound as a ray [36]. This physics-based simulation approach enables robust generalization to new positions, even in complex environments. + +While explicit modeling of scenes from visual input provides a significant advantage in generalization, it falls short of achieving realistic audio rendering [4, 46] for several reasons. (1) The material properties of each bounce surface, such as absorption, scattering, and transmission, are not fully considered. (2) The simulation traces only a limited number of sound rays due to computational restrictions, and low-frequency values are imprecise. (3) Background noise effects are not modeled. To address these issues, we propose our second stage to enhance the realism of the simulated sound. We utilize a conditional flow matching model [25, 47] to refine the coarsely simulated sound. With the powerful generative capability of flow-matching models, we can effectively correct errors from the first stage. An intuitive comparison between traditional audio simulation methods, neural rendering approaches, and our proposed method is presented in Fig. 1. + +We conduct extensive experiments on three real-world datasets: RWAVS-Gen, RWAVS [21], and RAF [9] datasets. RWAVS-Gen and RWAVS datasets measure the waveform sound generation quality, and the RAF dataset assesses the room impulse response rendering performance. The experimental results demonstrate that our method exhibits strong generalization and can be readily applied to novel sound sources. We also show that our method can be integrated into existing approaches to improve their generalization performance. In conclusion, our contributions are as follows: + +- We introduce a novel physics-integrated audio-visual acoustic synthesis framework to generate realistic audio content at novel positions based on visual information. +- We propose a vision-guided audio simulation method to enhance generalization for novel sources and listeners. +- We design a flow matching-based audio refinement model + +to bridge the gap between simulation sounds and realworld recordings. + +- Our experiments highlight the limitations of existing approaches, demonstrate the advantages of our method, and show the applicability of our model. + +# 2. Related Work + +Our work is closely related to areas such as vision-informed audio generation and flow matching models. We discuss each area and related work in the following section. + +# 2.1. Vision-Informed Audio Generation + +Vision-informed audio generation focuses on synthesizing audio based on visual inputs like images, videos, meshes, and poses. Many studies propose neural network-based audio generation pipelines [5, 6, 12, 15, 21, 23, 26, 27, 32, 33, 43, 52]. For instance, 2.5D Visual Sound [12] employs a U-Net [34] to synthesize binaural audio conditioned on an image, while Chen et al. [5] introduce a cross-modal transformer [49] to generate audio that matches room acoustics informed by an image. Some researchers have also developed video-guided audio generation methods, such as Diff Foley [27], and Movie Gen [32], which produce synchronized audio by considering temporal cues in videos. + +Another key area of vision-informed audio generation is pose-conditioned audio rendering, which is the focus of this paper. Inspired by Neural Radiance Field (NeRF) [29], several works [1, 2, 6, 7, 11, 14, 20, 21, 26, 33, 43] explore novel-pose audio synthesis by learning an audio field. Chen et al. [6] introduce a CNN-based network for transforming audio to synthesize sound at novel poses, while Liang et al. [21] design a NeRF-like system to jointly generate audio and visual content conditioned on poses. Although these pose-conditioned approaches generate plausible results for fixed sound sources, they face challenges with novel sound sources. In comparison, our method can easily render sound for new sources, thanks to our vision-guided audio simulation approach with physics integration. + +# 2.2. Flow Matching Models + +Deriving from Continuous Normalizing Flows (CNFs) [8], Lipman et al. [25] introduce Flow Matching to train CNFs in a simulation-free manner. Flow Matching, especially Optimal Transport Flow Matching [28], models the transformation between noise and data samples in a simpler and more efficient way than diffusion models [13, 41], leading to more stable training and better performance. Tong et al. [47] extend Flow Matching to arbitrary distribution transformations, including probability paths between different data distributions [24]. Inspired by this, we treat our simulated audio as one distribution and the target binaural audio as another distribution and use Flow Matching to refine + +![](images/d918661a929b644072d77b3eeeede2abdde240ac7ac37ceb5eb2e1a8b85d1344.jpg) +Figure 2. Overview of our approach. Our framework $\pi$ -AVAS consists of two stages: the vision-guided audio simulation and the audio refinement with flow matching. In the first stage, we conduct 3D scene reconstruction and simulate sound propagation between a speaker and a listener in the mesh scene. In the second stage, we refine the coarsely simulated sound with a flow matching model, enhancing the quality of the synthesized sound. + +the simulation results. The audio refinement network effectively corrects simulation errors from the first stage. + +# 3. Method + +Our method aims to boost the performance of audio-visual acoustic synthesis models for novel sound sources. We design a novel Physics-Integrated Audio-Visual Acoustic Synthesis model (PI-AVAS or $\pi$ -AVAS) with two stages. In the first stage (Sec. 3.1), we introduce a vision-guided audio simulation approach that integrates the physical properties of sound propagation. In the second stage (Sec. 3.2), we propose a conditional flow matching model that refines the prediction results from the first stage, improving the quality of audio generation. Additionally, we introduce a data augmentation strategy in Sec. 3.3 to facilitate model training. The complete pipeline is illustrated in Fig. 2. + +# 3.1. Vision-Guided Audio Simulation + +To enhance the generalization for audio sources and receivers in novel positions, we design a vision-guided audio simulation framework. + +3D Scene Mesh Extraction. We aim to reconstruct meshes of an audio-visual scene from input images in the first step (see the second subfigure in Fig. 2). Given a set of images and their corresponding camera poses, we utilize Neural Radiance Field (NeRF) [29, 45] or 3D Gaussian Splatting [18] to learn a neural representation of the given environment. Then, we convert the NeRF model weights or Gaussian points to 3D point clouds, which are a more compatible data format. Once the point clouds are reconstructed, we use Poisson surface reconstruction [17] to generate meshes of the audio-visual scene. Empirically, we do not observe a noticeable difference between the reconstructed 3D meshes of NeRF and Gaussian Splattings. We leave a detailed comparison between NeRF, Gaussian Splatting, and traditional Structure-from-Motion approaches [38] for future work, as + +this is not the main focus of our paper. + +Physics-Integrated Sound Simulation. After we generate the meshes of an audio-visual scene, we can simulate the audio propagation between arbitrary sound emitters and sound receivers in this scene (see the third subfigure in Fig. 2). + +Specifically, for a pair of sound source tx and sound receiver rx, we treat the sound emitted by the source $x_{\mathrm{tx}} \in \mathbb{R}^n$ ( $n$ is the audio length) as a collection of rays, tracing each ray's interaction with the room's meshes [3, 40]. We denote the energy received at the receiver from the transmitter as $E(\mathrm{tx} \to \mathrm{rx})$ , from which we omit the time delay variable for simplicity. We model both direct propagation and indirect reflection to calculate the received energy: + +$$ +E (\mathrm {t x} \rightarrow \mathrm {r x}) = \underbrace {E _ {\mathrm {d}} (\mathrm {t x} \rightarrow \mathrm {r x})} _ {\text {D i r e c t p r o p a g a t i o n}} + \underbrace {E _ {\mathrm {i d}} (\mathrm {t x} \rightarrow \mathrm {r x})} _ {\text {I n d i r e c t r e f l e c t i o n}}. \tag {1} +$$ + +$E_{\mathrm{d}}(\mathrm{tx}\rightarrow \mathrm{rx})$ is the energy of direct propagation and $E_{\mathrm{id}}(\mathrm{tx}\rightarrow \mathrm{rx})$ is the reflected energy defined as + +$$ +\begin{array}{l} E _ {\mathrm {i d}} (\mathrm {t x} \rightarrow \mathrm {r x}) = \int_ {\Omega} \underbrace {E (\omega \rightarrow \mathrm {t x})} _ {\text {E n e r g y}} \underbrace {G (\omega \leftrightarrow \mathrm {t x})} _ {\text {G e o m e t r y}} \tag {2} \\ \underbrace {\rho (\omega \to \mathrm {t x} \to \mathrm {r x})} _ {\text {M a t e r i a l p r o p e r t y}} * \underbrace {M (\omega \leftrightarrow \mathrm {t x})} _ {\text {S c e n e s i z e}} d \omega , \\ \end{array} +$$ + +where $\Omega$ is the entire mesh space, $\omega$ is an area of $\Omega$ , $M(\omega \leftrightarrow \mathrm{tx})$ measures energy absorption and time delay, $G(\omega \leftrightarrow \mathrm{tx})$ means the energy dispersion and occlusion during sound propagation, $\rho (\omega \rightarrow \mathrm{tx}\rightarrow \mathrm{rx})$ represents the acoustic property of each bounce surface, and the asterisk $*$ is the convolution operation. + +Eq. (1) can be converted to an infinite sum of integrals and solved with the Neumann series expansion [19]. We then calculate the impulse response based on the accumulated energy $E$ and convolve that with the sound source to generate the simulated sound. To improve the realism of the + +![](images/66b43decdf3557187bcec57ae8d343862841938cbf62358d3575a251e8631ca1.jpg) +Figure 3. Visualization of the audio refinement process with the flow matching model. The flow matching model learns a vector field that gradually transforms a simulated sound $x_{\mathrm{sim}}$ to a ground-truth sound $x_{\mathrm{rx}}$ . + +simulated sound, we also apply the Head-Related Transfer Function (HRTF) to generate binaural audio $x_{\mathrm{sim}} \in \mathbb{R}^{2 \times n}$ based on the listener's head direction. Because it is challenging to estimate material property solely based on visual information [10, 39], we assign default material coefficients to all surfaces. + +So far, we simulate sound propagation between the sound source and the receiver using visual information obtained from a set of images. Since this approach integrates the physical properties of sound propagation, it generalizes well to novel sound sources and listeners. + +# 3.2. Audio Refinement With Flow Matching + +Although the vision-guided audio simulation approach offers robustness for novel poses, there is a noticeable fidelity gap between simulated and recorded sounds [4, 46] as mentioned in the introduction section, e.g., the material property is not correctly modeled. Therefore, we introduce our second stage to further enhance the quality of the generated sound $x_{\mathrm{sim}}$ . We propose a flow matching model to refine and improve the simulated audio. We illustrate the audio refinement process with the flow matching model in Fig. 3. Flow Matching Formula. Specifically, we treat the simulated sound $x_{\mathrm{sim}} \in \mathbb{R}^{2 \times n}$ as one data distribution $\mathcal{N}(x|x_{\mathrm{sim}}, \sigma^2 I)$ and the recorded sound (ground-truth sound) $x_{\mathrm{rx}} \in \mathbb{R}^{2 \times n}$ as another data distribution $\mathcal{N}(x|x_{\mathrm{rx}}, \sigma^2 I)$ , where $\sigma$ is a predefined standard deviation. We model the sound refinement process as a transfer between these two distributions. We design the following time-dependent probability path $p_t: [0,1] \times \mathbb{R}^n \to \mathbb{R}_{>0}$ : + +$$ +p _ {t} (x) = \mathcal {N} (x | t x _ {\mathrm {r x}} + (1 - t) x _ {\mathrm {s i m}}, \sigma^ {2} I), \tag {3} +$$ + +where $p_0(x) = \mathcal{N}(x|x_{\mathrm{sim}},\sigma^2 I)$ , $p_1(x) = \mathcal{N}(x|x_{\mathrm{rx}},\sigma^2 I)$ , and time step $t\in [0,1]$ . We define the time-dependent flow $\psi_t(x):[0,1]\times \mathbb{R}^n\to \mathbb{R}^n$ as follows: + +$$ +\psi_ {t} (x) = t x _ {\mathrm {r x}} + (1 - t) x _ {\mathrm {s i m}} + \sigma \epsilon , \tag {4} +$$ + +where $\epsilon$ is sampled from a standard Gaussian distribution $\mathcal{N}(0,I)$ . According to the definition of vector fields, we can derive the vector field $v_{t}:[0,1]\times \mathbb{R}^{n}\to \mathbb{R}^{n}$ using the flow defined in Eq. (4): + +$$ +v _ {t} \left(\psi_ {t} (x)\right) = \frac {d}{d t} \psi_ {t} (x) = x _ {\mathrm {r x}} - x _ {\mathrm {s i m}}. \tag {5} +$$ + +Then we train a deep neural network $u_{t}(\psi_{t}(x), x_{\mathrm{tx}}, p; \theta)$ to fit the vector field defined in Eq. (5), where $\theta$ is the trainable parameters of a neural network, $x_{\mathrm{tx}}$ is the source sound, and $p$ is the pose information of sound source tx and listener rx. The training objective is + +$$ +\mathcal {L} _ {\mathrm {F M}} (\theta) = \mathbb {E} _ {q _ {0} (x), q _ {1} (x), t} \| u _ {t} \left(\psi_ {t} (x), x _ {\mathrm {t x}}, p; \theta\right) - \left(x _ {\mathrm {r x}} - x _ {\mathrm {s i m}}\right) \| ^ {2}. \tag {6} +$$ + +We provide pseudo-code for training the audio refinement model in the appendix. In our experiments, we use Short-time Fourier transform (STFT) to transform both the target $x_{\mathrm{rx}} - x_{\mathrm{sim}}$ and predicted $u_{t}(\psi_{t}(x),x_{\mathrm{tx}},p;\theta)$ vector fields from a time space to a time-frequency space and calculate the L2 distance as the training loss. + +After training, we utilize the network for audio refinement. Given a simulated sound $x_{\mathrm{sim}}$ , we generate an enhanced sound $x_{\mathrm{rx}}$ by solving the ordinary differential equation: + +$$ +\frac {d}{d t} \psi_ {t} (x) = u _ {t} \left(\psi_ {t} (x), x _ {\mathrm {t x}}, p; \theta\right), \tag {7} +$$ + +$$ +\psi_ {0} (x) = x _ {\mathrm {s i m}}. +$$ + +The solution $\psi_{1}(x)$ is the synthesized binaural audio. In this paper, we study one first-order solver (Euler solver) and two second-order solvers (Midpoint solver and Heun solver). We compare different solvers and provide the pseudo-code of inference in the appendix. + +Audio Refinement Network. In Fig. 4 (a), we show the architecture of our network that is designed to approximate the vector field $v_{t}(x)$ . Given an intermediate sound $\psi_t(x)$ sampled from Eq. (4), we concatenate it with the source sound $x_{\mathrm{tx}}$ emitted by the loudspeaker. The concatenated sound is first fed to a two-layer convolutional encoder to enrich the channel dimension and then passed to a stack of multi-scale gated convolution blocks. To condition the vector field prediction on the source and listener's poses $p$ and time step $t$ , we first project them into a high-frequency space using Random Gaussian Fourier Embedding (RGFE) [44], followed by MLPs. The resulting embeddings form the flow matching conditions $c$ , which are then passed to the multi-scale gated convolution blocks. Each multi-scale gated convolution block uses condition $c$ to adjust the sound features $f_{i-1}$ and predicts the next sound features $f_{i}$ and a skip feature $s_i$ . Finally, we combine all + +![](images/600c5c581329fc367796877f3351b7f5e56e5e68e4d997fe633778bb1aa50e13.jpg) +(a) Audio Refinement Network + +![](images/368b291e12eeb875aefdab13bbe5e2c2f131d7451660f0f36084ab1698ffd467.jpg) +(b) Multi-Scale Gated Convolution Block + +skip features $\{s_0,s_1,\ldots ,s_{L - 1}\}$ together, where $L$ is the number of blocks, and use a two-layer convolutional decoder to predict the vector field $u_{t}$ . + +We present our multi-scale gated convolution block in Fig. 4 (b). Inspired by WaveNet [48] and ViGAS [6], we design a gate operator to filter and control intermediate features. Given a feature $f_{i-1}$ from the previous block, we utilize dilated convolution layers (D-Conv) to learn meaningful features $D_1(f_{i-1})$ and $D_2(f_{i-1})$ , where $D_1$ and $D_2$ are dilated convolution layers. We also use pointwise convolution layers (P-Conv) to extract condition features $P_1(c)$ and $P_2(c)$ , where $P_1$ and $P_2$ are pointwise convolution layers. We apply the gated operator using the following equation: + +$$ +y = \tanh \left(D _ {1} \left(f _ {i - 1}\right) + P _ {1} (c)\right) \otimes \sigma \left(D _ {2} \left(f _ {i - 1}\right) + P _ {2} (c)\right), \tag {8} +$$ + +where $\otimes$ means Hadamard product. + +Then we feed $y$ through a pointwise convolution layer to form the residual and add it to the input feature $f_{i - 1}$ to generate the new feature $f_{i}$ . We feed $y$ through another pointwise convolution layer to generate a skip feature $s_i$ . Since we gradually increase the dilation size of each block, we name it the multi-scale gated convolution block. + +# 3.3. Data Augmentation + +Data scarcity presents a challenge when we train audiovisual acoustic synthesis models, as each scene is typically recorded in a 10 to 20-minute video. This yields only 600 to 1200 samples (at 1 fps) per scene for training. To facilitate training, we propose a data augmentation strategy (see Fig. 5). Given that the video captures continuous camera movement, we approximate the camera's entire trajectory by interpolating between discrete training poses (shown as orange poses in the figure). We then shift each training pose randomly along this trajectory by up to one second forward or backward. These shifted positions serve as augmented + +![](images/c6fb163afc1466f0313466e4373ab6985a0ee8d893f7a3d86237119796df763d.jpg) +Figure 4. Model architecture. (a) shows the audio refinement network. Given an intermediate sound $\psi_t(x)$ and a source sound $x_{\mathrm{tx}}$ , we concatenate them and pass them through an encoder followed by a series of multi-scale gated convolution blocks. We use timestep $t$ and pose $p$ as conditions by encoding them with RGFE and MLPs. All skip features $s_i$ are gathered and used to predict the vector field $u_t$ . (b) illustrates the multi-scale gated convolution block. We design a gate operator to filter and control the intermediate features $f_i$ . +Figure 5. Data augmentation. We randomly shift each training pose, represented in orange, along the interpolated camera trajectory by up to one second forward or backward. The shifted poses, shown in blue, serve as augmented poses to aid in model training. + +poses (blue poses), and we pair them with temporally corresponding audio clips as augmented sound samples. + +# 4. Experiments + +# 4.1. Generalization Evaluation + +We first evaluate the generalization ability of various methods to novel sound sources. + +Experiment Setup. We use the Real-World Audio-Visual Scene (RWAVS) dataset [21] to benchmark our method. The RWAVS dataset captures multimodal data, including the source position, camera poses, mono source sounds, and binaural received sounds for each scene. The dataset captures data from diverse environments, such as offices, multi-room houses, apartments, and outdoor spaces. For each scene, the original RWAVS dataset contains multiple videos with varying sound source locations. To examine generalization issues in existing approaches, we design a + +Table 1. Quantitative comparison with state-of-the-art methods on the RWAVS-Gen dataset using the generalization evaluation setup. We also show the inference speed and the model size of all methods. We highlight the best result in bold. + +
MethodsOfficeHouseApartmentOutdoorsOverallSpeed (ms)Size (MB)
MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓
INRAS [43]2.1260.1823.6050.2204.5350.2322.0580.1573.0810.1982.90.79
NAF [26]2.2750.1812.8730.1864.8780.2311.5750.1352.9000.1834.60.74
ViGAS [6]2.1370.1833.8780.2133.9460.2211.9670.1542.9820.19312.89.72
AV-NeRF [21]2.0860.1803.7590.2214.5200.2302.3080.1653.1680.1992.13.04
π-AVAS (Ours)1.8560.1631.9460.1403.8980.2091.3260.1252.2570.15910.45.62
+ +![](images/0986ecf767b2d770c8dd08fff549ec6ac79ba92cbdea7f96cee2d65a8cb4fdf3.jpg) +Figure 6. Visualization of synthesized binaural sounds for novel sound sources and listeners. + +new evaluation setup: in each environment, we select one video as training data, with the remaining videos used for evaluation. This setup allows us to effectively assess how well existing methods generalize to new sound source locations within the same environment. We present an example in the appendix. We denote this new benchmark as RWAVS-Gen to distinguish it from the original RWAVS benchmark. + +Following AV-NeRF [21], we select NAF [26], INRAS [43], ViGAS [6], and AV-NeRF as our baselines. NAF models sound propagation within a scene by using a local feature grid and an implicit decoder. INRAS disentangles sound modeling through a three-stage implicit neural field. ViGAS predicts sound at a new location by leveraging audio-visual features from source viewpoints. AV-NeRF constructs a multi-modal neural field to condition audio modeling on 3D visual scene context. + +We choose magnitude distance (MAG) [51] and envelope distance (ENV) [30] metrics to evaluate audio quality following RWAVS [21]. + +Quantitative Results. We compare our model with existing approaches on the RWAVS-Gen dataset and report their generalization performance in Tab. 1. All methods show better audio rendering performance in the office and the outdoor scenes while causing worse audio generation in the house and apartment scenes because the office is a classic shoebox environment with four parallel walls and the outdoor scene is an open space without occlusion. Our model achieves the best performance across all four environments, achieving the lowest metric losses, with 2.257 MAG and 0.159 ENV. This demonstrates the strong robustness and + +generalization capability of our physics-integrated model to novel sound sources. + +Inference Speed. We test the inference speed of different models to render one second of binaural audio with one RTX 4090 GPU and report the results in the rightmost column in Tab. 1. Since our approach involves audio simulation (2.1ms) and requires 4 steps to complete flow matching estimation (8.3ms), it is slower than some methods that directly output sound. However, (1) our approach still meets the real-time requirement (16.7ms for 60 FPS and 33.3ms for 30 FPS), meaning it causes no noticeable delay in real-world audio applications. (2) There are many successful works we can use to reduce inference time without performance degradation, such as consistency models [42] and adversarial diffusion distillation [35]. If we reduce the number of steps to 1, our method needs only 4.2 ms for inference. + +Qualitative Comparison. We visualize the generated sounds of different approaches in Fig. 6 for an intuitive comparison. As depicted, existing methods encounter challenges when applied to novel sound sources, resulting in inaccurate sound volume. In contrast, our physics-based approach effectively considers the changes in sound source locations and accurately produces binaural audio. + +# 4.2. Standard Evaluation + +We then compare the rendering performance of our $\pi$ -AVAS model with other methods using standard benchmarks. + +Experiment Setup. We use the original RWAVS and Real Acoustic Field (RAF) datasets to benchmark our approach. For the RWAVS dataset, we make no modification to the benchmark and use the original setup to test our model. The RAF dataset is a real-world room impulse response dataset that densely captures real room acoustic data. It provides impulse response signals of an office environment in two conditions: empty and furnished. It allows to study the difference in acoustic fields introduced by furniture. + +Besides the baselines used in the RWAVS-Gen experiment, we include state-of-the-art approaches on RWAVS and RAF datasets to augment our experiment. AV-GS [1] introduces an audio-visual Gaussian Splitting method that explicitly represents a scene for acoustic synthesis. SOAF [11] designs an occlusion-aware acoustic field. AVR [20] utilizes the volume rendering technique to generate acoustic impulse responses. + +Table 2. Quantitative comparison with state-of-the-art methods on the original RWAVS dataset. We highlight the best-performing result in bold and underline the second-best result. + +
MethodsOfficeHouseApartmentOutdoorsOverall
MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓
Mono-Mono9.2690.41111.8890.42415.1200.47413.9570.47012.5590.445
Mono-Energy1.5360.1424.3070.1803.9110.1921.6340.1272.8470.160
Stereo-Energy1.5110.1394.3010.1803.8950.1911.6120.1242.8300.159
INRAS [43]1.4050.1413.5110.1823.4210.2011.5020.1302.4600.164
NAF [26]1.2440.1373.2590.1783.3450.1931.2840.1212.2830.157
ViGAS [6]1.0490.1322.5020.1612.6000.1871.1690.1211.8300.150
AV-NeRF [21]0.9300.1292.0090.1552.2300.1840.8450.1111.5040.145
AV-GS [1]0.8610.1241.9700.1522.0310.1770.7910.1071.4170.140
SOAF [11]0.8280.1261.9510.1532.0970.1820.7700.1091.4110.142
π-AVAS (Ours)0.6740.1091.9920.1492.0410.1730.7850.1061.3730.134
+ +Table 3. Quantitative comparison with state-of-the-art methods on the RAF dataset. We highlight the best-performing result in bold and underline the second-best result. + +
MethodsRAF-FurnishedRAF-Empty
T60↓C50↓EDT↓Amp.↓Phase↓Env.↓T60↓C50↓EDT↓Amp.↓Phase↓Env.↓
AAC-nearest13.03.4173.51.091.604.8313.03.4173.31.091.604.83
AAC-linear12.43.6590.20.991.603.8113.13.2571.51.101.595.22
Opus-nearest14.43.7880.31.191.605.3513.34.25100.61.161.594.58
Opus-linear13.13.5577.81.471.605.7412.73.9495.50.951.594.26
NAF [26]7.10.9820.60.931.625.348.01.2226.30.851.624.67
INRAS [43]6.91.0821.40.961.626.437.61.2125.80.881.624.72
AVR [20]5.00.9517.90.751.584.525.51.0423.30.671.583.96
π-AVAS (Ours)4.80.8116.30.241.584.955.00.9119.30.261.564.42
+ +We use MAG and ENV metrics to measure performance on the RWAVS dataset. Following AVR [20], we choose T60, C50, EDT, Amplitude (Amp.), Phase, and Envelope (Env.) to assess performance on the RAF dataset. T60, C50, and EDT are the most important metrics for measuring impulse response quality by analyzing energy decay. Amplitude and Phase metrics evaluate the impulse response in the time-frequency domain. The Envelope metric evaluates the impulse response in the time domain. + +RWAVS Results. As shown in Tab. 2, we compare our method with existing baselines on the original RWAVS dataset. Mono-Mono, Mono-Energy, and Stereo-Energy are non-learnable methods that generate binaural audio by scaling mono audio using estimated energy values. Other methods are neural network-based approaches, such as AVGS [1] and SOAF [11]. Our approach outperforms both non-learnable and learnable approaches, setting new state-of-the-art performance on the RWAVS dataset. The results demonstrate that our $\pi$ -AVAS model achieves plausible novel-view audio synthesis quality. + +RAF Results. We present our $\pi$ -AVAS's performance on the RAF dataset in Tab. 3. AAC [16] and Opus [50] are traditional audio encoding methods. "Nearest" and "linear" refer to different interpolation modes. AVR [20] is the state-of-the-art method on this dataset. Our method surpasses + +AVR on most metrics and performs on par with it on the Envelope metric. Considering that AVR takes 24 hours to converge while our model trains in 5 hours, our $\pi$ -AVAS exhibits a better trade-off between training time and impulse response generation quality. + +# 4.3. Applicability Of Our Simulation Method + +The generalization capability of our method is empowered by our first stage, the vision-guided audio simulation module (Sec. 3.1). We find that this module not only improves the performance of our flow matching model but can also be applied to existing neural synthesis approaches to enhance their generalization ability. By replacing their input source audio with our simulated audio, we integrate our vision-guided audio simulation module into their framework. Experiment results shown in Tab. 4 demonstrate the applicability of our approach, with overall performance improvements across most methods. For example, we improve the MAG metric of the AV-NeRF model by 0.772 and reduce the ENV loss by 0.036. + +# 4.4. Ablation Studies + +We provide a thorough ablation study using the RWAVS-Gen dataset, with results shown in Tab. 5. + +Table 4. Applicability of our method. We apply our simulation method to other approaches to improve their generalization ability (denoted as "w/ sim"). We conduct experiments on the RWAVS-Gen dataset. The performance improvement is marked with a green triangle $\triangledown$ . + +
MethodsOfficeHouseApartmentOutdoorsOverall
MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓
INRAS [43]2.1260.1823.6050.2204.5350.2322.0580.1573.0810.198
w/ sim2.1250.1792.3290.1574.9070.2461.6910.1362.763 (▼ 0.318)0.180 (▼ 0.018)
NAF [26]2.2750.1812.8730.1864.8780.2311.5750.1352.9000.183
w/ sim2.2030.1842.2140.1544.9250.2411.5560.1312.724 (▼ 0.176)0.178 (▼ 0.050)
ViGAS [6]2.1370.1833.8780.2133.9460.2211.9670.1542.9820.193
w/ sim2.0740.1732.3170.1373.6830.2061.6790.1352.438 (▼ 0.544)0.163 (▼ 0.030)
AV-NeRF [21]2.0860.1803.7590.2214.5200.2302.3080.1653.1680.199
w/ sim2.0140.1741.9460.1364.3740.2211.2500.1222.396 (▼ 0.772)0.163 (▼ 0.036)
+ +Table 5. Ablation Studies. We conduct a comprehensive ablation study to verify the effectiveness of our proposed method. The term "pra+HRTF" refers to substituting our vision-guided acoustic simulation approach with pyroomacoustics and HRTF. "Regression" denotes training our audio refinement convolutional network without the flow matching loss. + +
MethodsOfficeHouseApartmentOutdoorsOverall
SimulationRefinementAugmentationMAG↓ENV↓MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓MAG↓ENV↓
pra[37]+HRTF[31]4.2590.2265.2640.2278.7620.2883.4540.2035.4350.236
✓2.6090.1862.7530.1707.3600.2683.0310.1773.9380.200
Regression1.9040.1673.8260.2044.0330.2191.6120.1432.8440.183
✓1.8800.1663.1440.1984.2090.2071.5770.1382.7020.177
✓✓1.8560.1632.1910.1483.8980.2091.4960.1342.3600.164
✓✓✓2.0020.1691.9460.1404.0710.2201.3260.1252.3360.164
+ +Vision-Guided Audio Simulation. First, we test the importance of our vision-guided audio simulation module (Sec. 3.1). We use pyroomacoustics [37] plus HRTF [31] as a baseline, which does not incorporate vision information. In this setup, we create a shoebox environment and use pyroomacoustics to estimate the room impulse response. We convolve the impulse response and the input audio to render mono audio at the target location. We then apply HRTF to generate binaural audio. Compared with this baseline, our module consistently outperforms it (see the first and the second rows), showing the importance of vision information in audio simulation. + +Audio Refinement Network. We proceed to evaluate our second stage — the audio refinement flow matching model (see Sec. 3.2). To establish a baseline, we remove the flow matching training objective from the second stage and train the audio refinement network with a regression loss function, labeled as "Regression" in the table. By incorporating the flow matching training objective, we achieve more precise audio refinement performance (compare the third and fourth rows). We hypothesize that the flow matching formula decomposes the challenging one-step estimation into several simpler steps, thereby progressively refining the simulated sound. By combining our first and second stages (see the fifth row), we achieve improved performance beyond either stage alone, demonstrating (1) the realism limitations of simulation-only approaches, (2) the generaliza + +tion challenges of neural rendering-only methods, and (3) the advantages of our two-stage approach. + +Augmentation Strategy. We also assess the impact of our data augmentation strategy (refer to the last row). By enhancing the audio refinement training with additional data, we achieve lower metric losses for both house and outdoor scenes; however, we observe no improvement for office and apartment scenes. Consequently, we apply data augmentation only to house and outdoor scenes. + +# 5. Conclusion + +In this paper, we study the limitations of existing approaches to the audio-visual acoustic synthesis problem. We design a two-stage, physics-integrated audio-visual acoustic synthesis framework to enhance both realism and generalization capabilities. The first stage of our framework is a vision-guided audio simulation module, followed by a flow-matching-based audio refinement module. To mitigate data scarcity in this task, we also propose a data augmentation strategy. Experimental results show the effectiveness of our proposed approach, achieving new state-of-the-art results on the RWAVS-Gen, RWAVS, and RAF datasets. We further show how our physics-integrated method improves existing approaches in terms of generalization. + +# References + +[1] Swapnil Bhosale, Haosen Yang, Diptesh Kanojia, Jiankang Deng, and Xiatian Zhu. Av-gs: Learning material and geometry aware priors for novel view acoustic synthesis. arXiv preprint arXiv:2406.08920, 2024. 2, 6, 7 +[2] Amandine Brunetto, Sascha Hornauer, and Fabien Moutarde. Neraf: 3d scene infused neural radiance and acoustic fields. arXiv preprint arXiv:2405.18213, 2024. 2 +[3] Chunxiao Cao, Zhong Ren, Carl Schissler, Dinesh Manocha, and Kun Zhou. Interactive sound propagation with bidirectional path tracing. ACM TOG, 35(6):1-11, 2016. 3 +[4] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In ECCV, 2020. 2, 4 +[5] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In CVPR, 2022. 2 +[6] Changan Chen, Alexander Richard, Roman Shapovalov, Vamsi Krishna Ithapu, Natalia Neverova, Kristen Grauman, and Andrea Vedaldi. Novel-view acoustic synthesis. arXiv preprint arXiv:2301.08730, 2023. 1, 2, 5, 6, 7, 8 +[7] Mingfei Chen and Eli Shlizerman. Av-cloud: Spatial audio rendering through audio-visual cloud splatting. Advances in Neural Information Processing Systems, 37:141021-141044, 2025. 2 +[8] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 2 +[9] Ziyang Chen, Israel D Gebru, Christian Richardt, Anurag Kumar, William Laney, Andrew Owens, and Alexander Richard. Real acoustic fields: An audio-visual room acoustics dataset and benchmark. In CVPR, pages 21886–21896, 2024. 1, 2 +[10] Roland W Fleming. Visual perception of materials and their properties. Vision research, 94:62-75, 2014. 4 +[11] Huiyu Gao, Jiahao Ma, David Ahmedt-Aristizabal, Chuong Nguyen, and Miaomiao Liu. Soaf: Scene occlusion-aware neural acoustic field. arXiv preprint arXiv:2407.02264, 2024. 2, 6, 7 +[12] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In CVPR, 2019. 2 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[14] Chao Huang, Dejan Marković, Chenliang Xu, and Alexander Richard. Modeling and driving human body soundfields through acoustic primitives. In European Conference on Computer Vision, pages 1-17. Springer, 2024. 2 +[15] Chao Huang, Ruohan Gao, JMF Tsang, Jan Kurcius, Cagdas Bilen, Chenliang Xu, Anurag Kumar, and Sanjeel Parekh. Learning to highlight audio by watching movies. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 23925-23935, 2025. 2 + +[16] International Organization for Standardization. Advanced audio coding (aac). ISO/IEC 13818-7:2006, 2006. 7 +[17] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 3 +[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):139-1, 2023. 2, 3 +[19] Erwin Kreyszig. Introductory functional analysis with applications. John Wiley & Sons, 1991. 3 +[20] Zitong Lan, Chenhao Zheng, Zhiwei Zheng, and Mingmin Zhao. Acoustic volume rendering for neural impulse response fields. NeurIPS, 37:44600-44623, 2025. 2, 6, 7 +[21] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. In Thirty-seventh Conference on Neural Information Processing Systems. 1, 2, 5, 6, 7, 8 +[22] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Neural acoustic context field: Rendering realistic room impulse response with neural fields. arXiv preprint arXiv:2309.15977, 2023. 1 +[23] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Language-guided joint audio-visual editing via one-shot adaptation. In Proceedings of the Asian Conference on Computer Vision, pages 1011–1027, 2024. 2 +[24] Susan Liang, Dejan Markovic, Israel D Gebru, Steven Krenn, Todd Keebler, Jacob Sandakly, Frank Yu, Samuel Hassel, Chenliang Xu, and Alexander Richard. Binauralflow: A causal and streamable approach for high-quality binaural speech synthesis with flow matching models. arXiv preprint arXiv:2505.22865, 2025. 2 +[25] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations. 2 +[26] Andrew Luo, Yilun Du, Michael J Tarr, Joshua B Tenenbaum, Antonio Torralba, and Chuang Gan. Learning neural acoustic fields. NeurIPS, 2022. 1, 2, 6, 7, 8 +[27] Simian Luo, Chuanhao Yan, Chenxu Hu, and Hang Zhao. Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[28] Robert J McCann. A convexity principle for interacting gases. Advances in mathematics, 128(1):153-179, 1997. 2 +[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3 +[30] Pedro Morgado, Nuno Nvasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. Advances in neural information processing systems, 31, 2018. 6 +[31] Markus Noisternig, Alois Sontacchi, Thomas Musil, and Robert Holdrich. A 3d ambisonic based binaural sound reproduction system. In Audio Engineering Society Conference: 24th International Conference: Multichannel Audio, The New Reality. Audio Engineering Society, 2003. 8 + +[32] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2 +[33] Anton Ratnarajah, Sreyan Ghosh, Sonal Kumar, Purva Chiniya, and Dinesh Manocha. Av-rir: Audio-visual room impulse response estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27164–27175, 2024. 2 +[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2 +[35] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 6 +[36] Lauri Savioja and U Peter Svensson. Overview of geometrical room acoustic modeling techniques. The Journal of the Acoustical Society of America, 138(2):708-730, 2015. 2 +[37] Robin Scheibler, Eric Bezzam, and Ivan Dokmanić. Pyroomacoustics: A python package for audio room simulation and array processing algorithms. In 2018 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 351-355. IEEE, 2018. 8 +[38] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 3 +[39] Gabriel Schwartz and Ko Nishino. Recognizing material properties from images. IEEE transactions on pattern analysis and machine intelligence, 42(8):1981-1995, 2019. 4 +[40] Samuel Siltanen, Tapio Lokki, Sami Kiminki, and Lauri Savioja. The room acoustic rendering equation. The Journal of the Acoustical Society of America, 122(3):1624-1635, 2007. 3 +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations. 2 +[42] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 6 +[43] Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. In NeurIPS, 2022. 1, 2, 6, 7, 8 +[44] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 33:7537-7547, 2020. 4 +[45] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modu + +lar framework for neural radiance field development. arXiv preprint arXiv:2302.04264, 2023. 3 +[46] Zhenyu Tang, Rohith Aralikatti, Anton Jeran Ratnarajah, and Dinesh Manocha. Gwa: A large high-quality acoustic dataset for audio processing. In SIGGRAPH, pages 1-9, 2022. 2, 4 +[47] Alexander Tong, Kilian FATRAS, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research. 2 +[48] Aaron Van Den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, Koray Kavukcuoglu, et al. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 12, 2016. 5 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 2 +[50] Xiph.Org Foundation. Xiph opus. https://opusc%cadc.org/, 2012.7 +[51] Xudong Xu, Hang Zhou, Ziwei Liu, Bo Dai, Xiaogang Wang, and Dahua Lin. Visually informed binaural audio generation without binaural audios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15485-15494, 2021. 6 +[52] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In ECCV, pages 570-586, 2018. 2 \ No newline at end of file diff --git a/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/images.zip b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e15e0ab9f4d3be89f4c9a415e209c8fdc1394efd --- /dev/null +++ b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e478109ea5a3c772da79399016e877da36f2fe43d721932d55d90bbb244b41c +size 630639 diff --git a/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/layout.json b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..45e791d64e79ae0c01d337d1c53ec651f902b0e8 --- /dev/null +++ b/pavascanphysicsintegratedaudiovisualmodelingboostneuralacousticsynthesis/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e823086c0a4b08f2db2be2b2887aeb2d5137d8b2415ba019916e0ec839d91fe +size 391172 diff --git a/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/4fc1e63c-b45e-4fff-aea4-5e3d73bc60ee_content_list.json b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/4fc1e63c-b45e-4fff-aea4-5e3d73bc60ee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a6351b8331dd6cf5494a8162241f223fb1d3157b --- /dev/null +++ b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/4fc1e63c-b45e-4fff-aea4-5e3d73bc60ee_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d12556b9dd3996479717383be9c5f8cda9de7bdc15e2e6d6167f1de7b4d0be5b +size 81506 diff --git a/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/4fc1e63c-b45e-4fff-aea4-5e3d73bc60ee_model.json b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/4fc1e63c-b45e-4fff-aea4-5e3d73bc60ee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a4e8a009b587d9d1b0fb9f5f35bee8b42fda93b0 --- /dev/null +++ b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/4fc1e63c-b45e-4fff-aea4-5e3d73bc60ee_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df6d1f0bba22b45cd3cdac9fdc1eeb95963359a30c7c54ccf9cea75a29462e0c +size 103096 diff --git a/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/4fc1e63c-b45e-4fff-aea4-5e3d73bc60ee_origin.pdf b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/4fc1e63c-b45e-4fff-aea4-5e3d73bc60ee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..75b5fa850441ebc806031fef42995549fc70245e --- /dev/null +++ b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/4fc1e63c-b45e-4fff-aea4-5e3d73bc60ee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dff5a4531ebfad9531692b82efe2bf0bf22446ecaefdedeab5bff882eb432e01 +size 1109101 diff --git a/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/full.md b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a2dd7065a664ed81c4f088a83ece6b0d31b812c --- /dev/null +++ b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/full.md @@ -0,0 +1,336 @@ +# $p$ -MoD: Building Mixture-of-Depths MLLMs via Progressive Ratio Decay + +Jun Zhang $^{1,\ast}$ , Desen Meng $^{1,\ast}$ , Zhengming Zhang $^{2}$ , Zhenpeng Huang $^{1}$ , Tao Wu $^{1}$ , Limin Wang $^{1,3,\text{图}}$ , $^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University $^{2}$ China Mobile Research Institute $^{3}$ Shanghai AI Lab https://github.com/MCG-NJU/p-MoD + +# Abstract + +Despite the remarkable performance of multimodal large language models (MLLMs) across diverse tasks, the substantial training and inference costs impede their advancement. In this paper, we propose $p$ -MoD, an efficient MLLM architecture that significantly reduces training and inference costs while maintaining model performance. The majority of computation in MLLMs stems from the overwhelming volume of vision tokens processed by the transformer-based LLM. Accordingly, we leverage the Mixture-of-Depths (MoD) mechanism, where each LLM layer selects essential vision tokens to process while skipping redundant ones. However, integrating MoD into MLLMs is non-trivial. To address the challenges of training and inference stability as well as limited training data, we adapt the MoD module with two novel designs: tanh-gated weight normalization (TanhNorm) and symmetric token reweighting (STRing). Moreover, we observe that vision tokens exhibit higher redundancy in deeper layers and thus design a progressive ratio decay (PRD) strategy, which gradually reduces the token retention ratio layer by layer, employing a shifted cosine schedule. This crucial design fully unleashes the potential of MoD, significantly boosting the efficiency and performance of our models. Extensive experiments on two baseline models across 15 benchmarks show that our model matches or even surpasses the performance of corresponding baselines, while requiring only $55.6\%$ TFLOPs and $53.7\%$ KV cache storage during inference, and $77.7\%$ GPU hours during training. + +# 1. Introduction + +Recently, both academia and industry have witnessed the rapid development of multimodal large language models (MLLMs) [1, 14, 24, 46, 47, 51, 55], which have demonstrated exceptional performance across various vision-language understanding tasks. Pioneering works in this field [8, 27, 29, 59] focused on processing single low-resolution image inputs. Subsequently, to meet the di + +![](images/d1b1aa3e3fd4303dfca5301dd74caf0d4cc0a96ee2835f148deef0bd9588aa08.jpg) +Figure 1. Comparison with the LLaVA-NeXT [28] baseline. Our model demonstrates comparable performance with the baseline model on 15 benchmarks across various domains, with $46.3\%$ fewer KV cache storage and $44.4\%$ fewer TFLOPs during inference. + +verse demands of real-world applications, increasing efforts have broadened the forms of visual inputs that MLLMs can support, including multiple high-resolution images and videos [6, 24, 45, 47, 55]. + +Current state-of-the-art MLLMs handle high-resolution images either by dividing the original image into multiple slices [6, 28, 55] which are independently processed by the vision encoder, or by using a stronger vision encoder with improved positional encoding which can process any image at its native resolution [1, 31, 47]. Consequently, when processing multiple high-resolution images or videos, the number of vision tokens increases dramatically, proportional to the number of pixels and the number of images or video frames. The overwhelming volume of vision tokens processed by the transformer-based LLM results in explosion of computational costs, which severely hampers further de + +velopment and broader application of MLLMs. Designing more efficient MLLM architectures with minimal performance degradation has thus become an urgent challenge for the community. + +Previous efforts primarily focus on compressing vision tokens before the LLM, either in the Vision Encoder or the multimodal projector [2, 6, 12, 38, 39, 42, 49, 50, 55, 56]. These approaches force the LLM to handle heavily compressed vision information, overlooking the fact that the LLM, with its enormous model capacity, has the potential to compress the vision tokens by itself within the transformer layer. + +In this paper, we focus on optimizing computation efficiency of MLLMs within the transformer layers of the LLM. We propose to build efficient MLLMs with Mixture-Depths (MoD) [41] mechanism, which selects the most important and informative vision tokens to be processed by each transformer layer, while skipping redundant ones to improve efficiency. However, integrating MoD mechanism into MLLMs is non-trivial and entails several significant challenges. + +First, different from training an MoD-based LLM from scratch, integrating MoD mechanism into a pre-trained vanilla LLM during multimodal training poses a substantial risk of disrupting the language abilities of the original LLM. To address this, we design tanh-gated weight normalization (TanhNorm), which not only ensures proper initialization of the newly added MoD module, but also enhances training stability and performance. It also mitigates numerical stability issues during inference. + +Second, MLLMs are trained on multimodal data [28] that are several orders of magnitude smaller in scale compared to the text data used for training MoD-based LLMs [41]. We enhance the MoD mechanism with symmetric token reweighting (STRing) module which fully leverage the language supervision signals during training, enabling MoD modules to learn to accurately assess token importance even with limited training data. + +Thanks to these two enhancements, our upgraded MoD layers can be seamlessly applied to MLLMs. However, we argue that setting a fixed ratio of retained tokens across different MoD layers [41] is a suboptimal design choice under multimodal scenario, as the degree of redundancy in vision tokens should vary across layers. We conducted a series of exploratory experiments by adjusting the ratio of different layers in our MoD-based MLLM, which demonstrate that vision tokens exhibit higher redundancy in deeper layers. Accordingly, we propose a progressive ratio decay (PRD) strategy, which follows a shifted cosine schedule to gradually reduce the token retention ratio layer by layer. Trained with this strategy, our model significantly outperforms models that use a constant retention ratio across all layers under the same computation budget. + +Building upon the above innovations, we present our model, $p$ -MoD, Mixture-of-Depth MLLMs equipped with our progressive ratio decay strategy and upgraded MoD layers (i.e. $p$ -MoD layers). Extensive experiments validate the effectiveness of our proposed model. As shown in Figure 1, across 15 benchmarks spanning various domains, our model matches or even outperforms the strong LLaVA-NeXT [28] baseline, with only $55.6\%$ TFLOPs and $53.7\%$ KV cache storage during inference, and $77.7\%$ GPU hours during training. + +# 2. Related Work + +Building Efficient LLMs. Considerable efforts have been made to build efficient LLMs. One representative approach is the Mixture of Experts (MoE) mechanism [7, 9, 19, 22, 43, 60], where a router directs each token to different MLP experts of the transformer block. MoE models achieve faster training and inference speeds compared to dense models of the same size, but at the cost of lower performance. The more recently proposed Mixture of Depths (MoD) [41] module assigns weights to tokens and processes only a portion of highest-weighted tokens, skipping lower-weighted tokens to save computation. MoD-based LLMs demonstrate strong performance under limited compute budgets. In this paper, we propose to build more efficient MLLMs by upgrading MoD with several innovative improvements. + +Building Efficient MLLMs. The main computational burden in MLLMs arises from the large number of vision tokens that the LLM must process. Previous works mainly focused on compressing vision tokens before they enter the LLM via convolution [6, 55], pooling [38, 50], query-based modules [2, 16] or token merging [3, 15, 25, 42, 49]. However, reducing vision-related computation within the LLM layers is less explored. LLaVolta [4] speeds up MLLM training by performing naive average pooling operation in intermediate LLM layers. FastV [5] and some subsequent works [13, 15, 17, 48, 57] improve MLLMs' efficiency by compressing vision tokens in intermediate transformer layers based on attention scores. However, these methods neglect the fact that tokens compressed in early layers might be crucial in deeper layers, potentially leading to performance degradation. In contrast, we utilize MoD for learnable layerwise vision token selection, improving model efficiency while avoiding potential performance loss caused by token dropping. + +# 3. Method + +In this section, we introduce our $p$ -MoD model. As illustrated in Figure 2, our model consists of $p$ -MoD layers which upgrades MoD architecture with tanh-gated weight + +![](images/80337e718331e5f93657a206491ff59e82eb53f78545271a4ea9cedba4120f7e.jpg) +Figure 2. Overview of $p$ -MoD. Middle: Our efficient MLLM model which consists of $p$ -MoD layers. Each layer independently selects important vision tokens to process while skipping redundant ones. The proportion of selected tokens for layer $i$ is specified by the token retention ratio $R_{i}$ . For simplicity, the vision encoder and connector are omitted. Left: Detailed architecture of $p$ -MoD layers. Given the input tokens, the weight predictor first assigns weights to each token. The top $R_{i} \%$ of tokens with highest weights are selected and processed by the transformer layer, while the rest of the tokens are skipped. The weights are normalized by TanhNorm module, and then both the selected and skipped tokens are symmetrically scaled by their corresponding weights in our STRing module. Right: Our crucial design is the progressive ratio decay (PRD) strategy, which gradually reduces the token retention ratio $R_{i}$ layer by layer, following a shifted cos schedule. + +normalization(TanhNorm) and symmetric token reweighting(STRing). Our crucial design is the progressive ratio decay(PRD) strategy which controls the token retention ratio across different layers. In the following sections, we first revisit MoD briefly and then explain each component of our $p$ -MoD model in detail. + +# 3.1. Revisiting Mixture-of-Depths + +The MoD layer consists of a weight predictor and a vanilla transformer layer. It assigns weights for each input token using the weight predictor and then selects the top $R\%$ tokens with the highest weights to be processed by the transformer layer. Formally, given an MoD layer $M$ with a token retention ratio $R$ , a transformer layer $T$ , and a linear weight predictor, the input tokens $X \in \mathbb{R}^{n \times d}$ is first passed through the linear predictor to generate a set of weights: + +$$ +w = \operatorname {L i n e a r} (X) \in \mathbb {R} ^ {n}, \tag {1} +$$ + +with $n$ denoting the sequence length and $d$ denoting the embedding dimension. + +Then, we compute the $R$ -th percentile of the router weights, denoted as $P_{R}(w)$ . Tokens with weights larger than $P_{R}(w)$ will be selected to be processed by the transformer layer $T$ while other tokens are skipped in this layer, which guarantees only $n \times R\%$ tokens are processed. The + +calculation process of an MoD layer can be formulated as: + +$$ +X _ {i} ^ {\prime} = \left\{ \begin{array}{l l} w _ {i} T \left(X _ {i}\right) + T \left(X _ {i}\right), & \text {i f} \quad w _ {i} > P _ {R} (w) \\ X _ {i}, & \text {i f} \quad w _ {i} \leq P _ {R} (w) \end{array} . \right. \tag {2} +$$ + +Notably, after being processed by the transformer layer, the selected tokens are then scaled by their corresponding weights. In this way, the weight predictor is engaged into the gradient path, enabling its parameters to be updated by backpropagation. We term this process token reweighting. + +In the original MoD module designed for LLMs [41], $X$ represents text tokens. In our multimodal scenario, $X$ represents vision tokens. We only apply MoD on vision tokens as they occupy the main computation load and exhibit high redundancy in LLM layers. + +# 3.2. Adapting Mixture-of-Depths Module + +Different from training MoD-based LLMs from scratch, integrating MoD into MLLMs presents several challenges. The insertion of new MoD modules into pre-trained LLMs can lead to instability during training and inference. The relatively small amount of multimodal data may not be sufficient to train the MoD modules. In this section, we introduce two enhancements to the MoD module to address these challenges. + +![](images/3d66317e841ddbfb43f3d5a1cbced070092b38fdf35c097a32a66cad744e0471.jpg) +Figure 3. Exploratory experiments on vision token redundancy. + +![](images/5f3ab454e65eb73904ca4d5c42b254d5f427df310a75e8f871b2dc3625b93ccd.jpg) +(a) Illustration of our exploratory experiment setup. Our baseline MoD model is trained with a token retention ratio of $70\%$ for all layers. We divide the layers into three groups: shallow, middle, and deep. In each set of experiments, we decrease the token retention ratio $(R\%)$ in a specific group of layers. +(b) Results of our exploratory experiments. Performance drops more slowly when the token retention ratio of deeper layers is decreased, indicating that vision tokens are more redundant in deeper layers. This inspires us to progressively reduce the token retention ratio layer by layer. + +# 3.2.1. Tanh-gated Weight Normalization + +The original MoD mechanism [41] is designed for training MoD-based LLMs from scratch. In our multimodal training scenario, we need to insert new MoD modules into a pre-trained vanilla LLM. One common approach is alternate training [58], where only the newly inserted modules are trained in the first stage, and all modules are simultaneously trained in the second stage. However, this approach is not suitable for MoD, as the MoD modules scale the tokens by their weights during the token reweighting process, and the subsequent frozen LLM layers are unable to handle scaled tokens. Our experiments show consistent results that training the MoD modules while freezing the LLM layers fails to converge. Consequently, the only feasible method is to directly address the challenge: inserting the MoD modules into the LLM and training both of them simultaneously. + +To address this challenge, we design tanh-gated weight normalization (TanhNorm), which employs a normalization function $f(w) = \alpha \tanh(w)$ to normalize the predicted weights before token reweighting. After applying TanhNorm to MoD, Equation 2 can be reformulated as: + +$$ +X _ {i} ^ {\prime} = \left\{ \begin{array}{l l} \alpha \tanh (w _ {i}) T (X _ {i}) \\ + T (X _ {i}), & \text {i f} w _ {i} > P _ {R} (w) \\ X _ {i}, & \text {i f} w _ {i} \leq P _ {R} (w) \end{array} \right.. \tag {3} +$$ + +Here $\alpha$ is a hyper-parameter which controls the range of the normalized weights. Although simple, our TanhNorm ensures: (1) the normalized weight distribution is zero-centered, (2) the range(variance) of the weight distribution can be easily controlled by adjusting the gating factor $\alpha$ . These properties offer the following benefits: + +- The normalized weights are closely around zero at the start of training, which ensures proper training initialization by keeping the LLM intact after inserting MoD modules. +- The zero-centered weight distribution reduced the risk that some tokens are repetitively scaled by positive(or negative) weights in all MoD layers. This ensures training stability and mitigates numerical stability issues during inference. + +Both of the above benefits contribute to improved model performance. We validate the effectiveness of $\mathrm{TanhNorm}$ in Section 4.3. + +# 3.2.2. Symmetric Token Reweighting + +Compared to the text data used for training MoD-based LLMs, MLLMs are trained on multimodal data [28] that are several orders of magnitude smaller in scale. It is challenging to train a weight predictor (the only learnable part in the MoD module) to accurately and robustly assess the importance of vision tokens and assign corresponding weights. + +As stated in Section 3.1, the gradient of the weight predictor module stems from the token reweighting process. The original MoD performs token reweighting only on selected tokens as in Equation 2. In this way, the language supervision signals only supervise the weight predictor to assign high weights to the selected tokens, while the process of predicting weights for the skipped tokens is not supervised (i.e. no gradient). + +To fully leverage the limited training data, we enhance the MoD mechanism by symmetrically applying the token reweighting process to both selected and skipped tokens. With our symmetric token reweighting (STRing) module, + +
ModelInference TFLOPs ↓Inference KV cache ↓Doc VQAChart QAText VQAInfo VQARW QAG QAOK VQAPO PEAI 2DSE EDAVG
LLaVA-v1.58.38100%28.118.246.025.855.661.953.485.955.266.249.6
+p-MoD4.92(-41.3%)53.7%27.616.844.826.855.762.256.085.556.266.549.8
LLaVA-NeXT39.46100%70.161.662.734.757.663.554.087.264.068.962.4
+p-MoD21.94(-44.4%)53.7%70.061.860.534.157.663.355.186.865.169.062.3
+ +Table 1. Comparison with baseline models on 10 benchmarks. Our models (marked in gray) achieves comparable or even better performance compared to baseline models, with only $55.6\%$ TFLOPs and $53.7\%$ KV cache storage during inference. All models are of 7B parameter scale. + +
ModelS QAMM BMM MUVQA v2MM EAVG
LLaVA-v1.569.764.136.676.61506.864.5
+ p-MoD69.365.436.376.91482.864.4
LLaVA-NeXT69.767.535.379.31519.365.6
+ p-MoD71.067.336.078.81495.565.6
+ +Table 2. Comparison with baseline models on more benchmarks. Our $p$ -MoD models (marked in gray) matches the performance of corresponding baseline models. SQA stands for ScienceQA-IMG. All models are 7B in size. + +
ModelDoc VQAChart QAText VQAG QASE EDMM BAVG
vanilla MoD61.154.055.861.866.963.160.4
+ TanhNorm61.854.556.462.567.165.961.4
+ STRing65.758.459.363.067.166.863.4
+ PRD70.061.860.563.369.067.365.3
+ +Table 3. Ablation study on our proposed innovations. Under fair comparison, all of our proposed modules (TanhNorm, STRing, and PRD) significantly improve model performance without any computational overhead, making indispensable contributions to the strong performance of our final model (marked in gray). + +Equation 3 can be further modified as: + +$$ +X _ {i} ^ {\prime} = \left\{ \begin{array}{c c c c} \alpha \tanh (w _ {i}) T (X _ {i}) & & \\ + T (X _ {i}), & \text {i f} & w _ {i} > P _ {R} (w) & . \\ \alpha \tanh (w _ {i}) X _ {i} + X _ {i}, & \text {i f} & w _ {i} \leq P _ {R} (w) \end{array} \right. +$$ + +In this way, the MoD module can fully leverage the language supervision signals during training and learn to accurately assess token importance with limited training data. + +# 3.3. Progressive Ratio Decay + +After upgrading MoD with TannahNorm and STRing modules, we are able to successfully train MoD-based MLLMs. However, the performance-efficiency trade-off exhibited by the model is far from satisfactory. When the token retention ratio is set below $70\%$ , the model's performance deteriorates sharply. + +Original MoD-based LLMs [41] adopt a fixed token retention ratio across all MoD layers. This strategy assumes that the tokens have the same degree of redundancy in different layers. We argue that this assumption does not hold in the multimodal scenario. Under the effect of self-attention in every transformer layer, vision tokens gradually aggregate information from each other and text tokens gather relevant information from vision tokens. Therefore, vision tokens are expected to become increasingly redundant in deeper layers. + +To validate our hypothesis, we conduct a series of exploratory experiments on our MoD-based MLLM trained with a token retention ratio of $70\%$ across all layers. As + +shown in Figure 3a, we first divide the layers into several groups: shallow, middle, and deep layers. In each set of experiment, we gradually decrease the token retention ratio $(\mathbb{R}\%)$ in a specific group of layers while keeping the other layers unchanged, and evaluate the model under this customized inference setting. The results in Figure 3b strongly support our hypothesis: performance drops more slowly when the token retention ratio of deeper layers is decreased. This suggests that vision tokens exhibit higher redundancy in deeper layers, which is consistent with observations in previous works [5, 57]. + +Accordingly, we design a progressive ratio decay (PRD) strategy. As shown on the right side of Figure 2, PRD gradually reduces the token retention ratio layer by layer, following a shifted cosine schedule. Suppose the model has $L$ layers, the token retention ratio for the $l$ -th layer is formulated as: + +$$ +R _ {l} = \frac {1}{2} \cos \frac {\pi l}{L} + \beta , \quad l = 1, 2, \dots , L. \tag {5} +$$ + +Here $\beta$ is a shift factor, which can be used to flexibly control the overall computational cost of the model by vertically shifting the cosine decay curve. + +Experiments in Section 4.3 demonstrates that our $p$ -MoD model with PRD strategy significantly outperforms models that use a constant retention ratio under the same computation budget. It also outperforms other kinds of ratio schedulers. + +
Ratio SchedulerDecayProgressiveDoc VQAChart QAText VQAG QASE EDMM BAVG
ConstantXX58.051.656.162.164.562.859.2
InterleavedXX60.354.857.062.365.362.160.3
Stepped✓X67.261.259.463.568.466.164.3
PRDLinear✓✓69.360.460.563.868.066.764.8
Cosine70.061.860.563.369.067.365.3
+ +![](images/1d05030f79188807c403f900b9fb3ffaedcde34542d0ca190aea54791e2aa4f4.jpg) +Figure 4. Illustration of different ratio schedule functions in Table 4. + +Table 4. Ablation on different token retention ratio schedules. Our default model is marked in gray. + +
ModelToken Compression Ratio ↓Infer. TFLOPsDoc VQAChart QAText VQAInfo VQARW QAGQASQAMM MUPO PEAI 2DVQA v2SE EDAVG
LLaVA-NeXT100%39.4670.161.662.734.757.663.569.735.387.264.079.368.962.9
+MQT [16]50.2%19.8649.944.053.529.853.562.369.136.185.864.077.265.857.6
+LLaVolta [4]53.1%21.3066.658.860.133.256.963.770.236.786.365.378.668.562.0
+FastV [5]53.1%22.7365.958.262.233.055.662.469.935.684.563.678.768.261.5
+p-MoD53.7%21.9470.061.860.534.157.663.371.036.086.865.178.869.062.8
+ +Table 5. Comparison with other vision token compression methods. $p$ -MoD significantly outperforms other compression methods, achieving the best performance on most benchmarks and the best average performance. All models are of 7B parameter scale. Results on 13B models are demonstrated in Table 8. + +# 4. Experiment + +# 4.1. Setups + +Models. To evaluate the effectiveness of $p$ -MoD, we select two representative open-source MLLMs: LLaVA-1.5 [27] and LLaVA-NeXT [28], as the baseline models in our experiments. LLaVA-1.5 resizes input images to the fixed resolution which aligns with the training setup of its CLIP [40] vision encoder, encoding an image into 576 tokens. LLaVA-NeXT divides high-resolution images into multiple slices which are independently processed by the vision encoder. This strategy enhances the visual perception capabilities, but also leads to a significantly larger number of vision tokens (up to 2880) and much higher computation costs. + +Training Recipe. Our $p$ -MoD models follow the two-stage training recipe of the baselines. In the pre-training stage, the MLP connector is trained on image caption data. In the fine-tuning stage, $p$ -MoD modules are insert into the LLM and updated together with the connector and the LLM. + +Benchmarks. We conduct comprehensive experiments across 15 benchmarks: VQAv2 [11], GQA [18], OK-VQA [34], AI2D [21] and ScienceQA-IMG [32] are traditional visual question answering benchmarks; DocVQA [36], TextVQA [44], ChartQA [35] and InfographicVQA [37] focus on fine-grained visual question answering; POPE [26] evaluates hallucination in MLLMs; SEED-Bench (Image) [23], RealWorldQA, MME [10], MMBench [30] and MMMU [53] are comprehensive benchmarks tailored for MLLMs. To ensure our results + +can be conveniently reproduced, we evaluate our model on all these benchmarks with the lmms-eval [54] evaluation framework. + +Evaluation Metrics. In addition to performance on benchmarks, we present various metrics that reflect training and inference efficiency. We report TFLOPs which measures the inference computation complexity, and the KV cache storage which constitutes the main memory bottleneck during inference. Furthermore, we measure GPU hours consumed during training, along with inference latency. + +# 4.2. Main Results + +We integrate $p$ -MoD with LLaVA-1.5 and LLaVA-NeXT baseline models and conduct comprehensive evaluation across 15 benchmarks, which measures comprehensive abilities of MLLMs across various aspects. The results are shown in Table 1 and Table 2. Both $p$ -MoD-LLaVv1.5 and $p$ -MoD-LLaVA-NeXT achieve comparable or even better performance across all benchmarks compared to their baseline models, with significant savings in inference TFLOPs and inference KV cache savings. + +In Table 1, it is noteworthy that text-rich and graph-based benchmarks like DocVQA, ChartQA, TextVQA, and InfoVQA require fine-grained visual perception and reasoning abilities. Models can only given a correct answer when they successfully identify answer-related regions that occupy only small portions of the entire image. Remarkably, $p$ -MoD-LLaVA-NeXT achieves negligible performance drop on DocVQA and InfoVQA, and even gains $0.2\%$ accuracy on ChartQA with substantial improvements + +in time and memory efficiency. + +The above results indicate that our approach substantially improves efficiency while maintaining performance. + +# 4.3. Ablation Study + +Effectiveness of proposed innovations. We conduct ablation in Table 3 to demonstrate the effectiveness of our proposed innovations. All experiments in the table are conducted under the same computational cost. TanhNorm, STRing and PRD all significantly enhance model performance, making indispensable contributions to the strong performance of our final $p$ -MoD model. + +Progressive Ratio Decay. Based on the exploratory experiments in Section 3.3, we conclude that vision tokens exhibit higher redundancy in deeper layers, and design the progressive ratio decay strategy with shifted cosine schedule. + +To further validate this conclusion and the effectiveness of PRD, we experiment with four other schedule functions which control the token retention ratio for each LLM layer. As illustrated in Figure 4, the functions include: (1) a constant function that uses the same ratio for all layers, (2) an interleaved function which interleaves a vanilla transformer layer and an MoD layer with low ratio, which is adopted in the original MoD paper [41], (3) a stepped decay function, (4) a linear decay function. To ensure a fair comparison, we fix the average retention ratios across all layers at approximately $54\%$ . Based on the results demonstrated in Table 4, we can draw the following conclusions: + +- The non-decaying functions (i.e. constant and interleaved) significantly underperform the decaying functions. This result strongly supports our conclusion that vision tokens exhibit higher redundancy in deeper layers. +- Our progressively decayed PRD functions (i.e. cosine and linear) notably outperforms the discontinuous stepped decay function, and the cosine function achieves the best results. This proves the effectiveness of our shifted cosine schedule. + +Tanh-gated Weight Normalization. To validate the effectiveness of $\mathrm{TanhNorm}$ , we compare different weight normalization methods in Table 6. First, we verify that using no weight normalization module will result in overflow during training, as shown in the first row (details are explained in appendix). Then, we validate the effectiveness of the two important properties of $\mathrm{TanhNorm}$ stated in Section 3.2.1: (1) the normalized weight distribution is zero-centered, (2) the range(variance) of the weight distribution can be easily controlled by adjusting the gating factor $\alpha$ . + +For the first property, we design two experiments with the Softmax function, a standard normalization function that is not zero-centered. Specifically, we experiment with the vanilla softmax function $f(w) = \alpha \cdot \text{Softmax}(w), \alpha = 0.2$ , and a shifted softmax function $f(w) = \alpha \cdot \text{Softmax}(w) + b, \alpha = 0.4, b = -0.2$ which + +
Norm +TypeDoc +VQAChart +QAText +VQAG +QASE +EDMM +BAVG
-OVERFLOW-
Softmax63.957.657.962.768.267.062.9
Shifted softmax62.557.057.362.968.366.462.4
TanhNorm(α=1)OVERFLOW-
TanhNorm(α=0.2)70.061.860.563.369.067.365.3
+ +Table 6. Ablation on tanh-gated weight normalization. Our default model is marked in gray. The rows with no performance reported indicates that the corresponding experiment faces overflow issue during training or inference. + +has the same range as our $\mathrm{TanhNorm}$ function $f(w) = \alpha \tanh (w),\alpha = 0.2$ + +Comparing the last row of Table 6 with the second and third rows, we verify that $\mathrm{TanhNorm}$ significantly outperforms the vanilla softmax function and the shifted softmax function. It proves that being zero-centered is key to $\mathrm{TanhNorm}$ 's strong performance. The shifted softmax function, despite having the same range as $\mathrm{TanhNorm}$ , results in worse performance due to the lack of zero-centered property. + +For the second property, we experiment with a naive choice of the gating factor $\alpha$ by setting it to 1, which can be interpreted as no gating is applied. As shown in the fourth row of Table 6, this experiment results in the overflow problem during inference. This validates the importance of controlling the gating factor $\alpha$ to guarantee training and inference stability. + +# 4.4. Comparison with Related Works + +In Table 5, we compare $p$ -MoD with several strong token compression methods. MQT [16] employs a query transformer to compress vision tokens. LLaVolta [4] performs average pooling in intermediate LLM layers, achieving remarkable results despite its simplicity. FastV [5] drops vision tokens in intermediate LLM layers based on the attention score from text tokens. $p$ -MoD significantly outperforms other compression methods, achieving the best performance on most benchmarks and the best average performance. + +# 4.5. Efficiency-Performance Trade-off + +Thanks to the shifted cosine schedule used by PRD in Section 3.3 to control the token retention ratio across all layers, we can change the shift factor $\beta$ to control the overall computational cost of the model. In this way, we can train and deploy $p$ -MoD models under different performance-efficiency trade-offs based on our actual needs. + +To showcase the versatility of our approach, we conducted experiments with $\beta = 0.3, 0.4, 0.5$ , respectively. The results are presented in Table 7. Our default $p$ -MoD + +
ModelEfficiencyBenchmark
Training GPU hours ↓Inference TFLOPs ↓Inference Latency (ms) ↓Inference KV cache storage ↓Doc VQAChart QAText VQAG QASE EDMM BAVG
LLaVA-NeXT-7B56039.46519.1100%70.161.662.763.568.967.565.7
p-MoD-0.3403 (-28.0%)17.78 (-54.9%)326.7 (-37.1%)42.3%65.557.858.163.168.667.063.4
p-MoD-0.4415 (-25.9%)19.56 (-50.4%)347.1 (-33.1%)47.5%66.559.658.463.868.566.663.9
p-MoD-0.5435 (-22.3%)21.94 (-44.4%)368.0 (-29.1%)53.7%70.061.860.563.369.067.365.3
+ +Table 7. The Efficiency-Performance trade-off experiments. To showcase the versatility of our approach, we conducted experiments with $\beta = 0.3, 0.4, 0.5$ , respectively. Inference latency is measured on TextVQA dataset. + +![](images/c5ddfab8b80d5dafd9c3b8aefb1a2c7c611ec59196be0f36552f40090a5b53f3.jpg) +layer 11 + +![](images/776af6911f614c325a5e3edb7f4c992dd43d696d661b7ce56307e9c5caf4be1f.jpg) +layer 16 + +![](images/6af051c2f639d9efe19d01c87def112116974df56d7eebaee277c4c8bbb74d37.jpg) +layer 21 + +![](images/28805ef3a563275838b3a1fe5a6fbb5f708b2e97adbe53de5888ffca936a3486.jpg) +layer 26 + +![](images/b7ae22df1f018e5a3b1c0ff91927bb5b8949ff8b77b7963146162d776de70693.jpg) +layer 11 + +![](images/0a939700d20d873e0e4816c7257fbe974d375e8f06c25b72f03a319eb5c8f5c8.jpg) +layer 16 + +![](images/6954fff18735dc6c051dd7b67116ba7b79285f859a01f399245063a0c7c57b3e.jpg) +layer 21 +Figure 5. Visualization of tokens selected by different $p$ -MoD layers. The selected tokens are colored in red. Please zoom in for a clearer view. As the number of selected tokens gradually decreases in deeper layers (due to PRD strategy), the model gradually concentrates on vision tokens corresponding to regions with rich semantic information. + +![](images/f78a7d03cdb3804349c8ffd4e009bfbdda21b768640a9066c24c185dc75bc695.jpg) +layer 26 + +model with $\beta = 0.5$ reduced training GPU hours by $22.3\%$ , inference TFLOPs by $44.4\%$ , inference latency by $29.1\%$ , and KV cache storage by $46.3\%$ . Using a lower $\beta$ can further improve efficiency, with an acceptable performance drop. + +# 4.6. Visualization: Which tokens are selected? + +In Figure 5, we visualize the tokens selected by $p$ -MoD-LLaVA-NeXT at different layers. Guided by the PRD strategy, the number of tokens processed by the MoD layers gradually decreases layer by layer. It can be observed that the selected vision tokens progressively concentrate on key regions in the image with rich semantic information. For the upper row of images in Figure 5, the selected tokens gradually converge to those corresponding to text, drawings, and the person on the right. Tokens corresponding to the white background are skipped in deep layers. In the lower row of images in Figure 5, the selected tokens + +gradually converge towards the artifact and the measurement markings on the ruler. These results indicate that our model effectively compresses visual information by selecting the informative tokens, enabling it to maintain performance while significantly reducing training and inference costs. + +# 5. Conclusion and Future Work + +In this paper, we explore building efficient MLLMs by adapting the Mixture-of-Depths mechanism. Our proposed model, termed $p$ -MoD, features three key designs: tanh-gated weight normalization, symmetric token reweighting module, and the progressive ratio decay strategy. Our model achieves comparable or even superior results to the baseline models on a diverse set of 15 benchmarks, with substantial improvements on training and inference efficiency. We hope $p$ -MoD can serve as a strong and efficient baseline for future research on developing efficient MLLMs. + +Acknowledgements. This work is supported by the National Key R&D Program of China (No. 2022ZD0160900), Jiangsu Frontier Technology Research and Development Program (No. BF2024076), the Collaborative Innovation Center of Novel Software Technology and Industrialization, and Nanjing University-China Mobile Communications Group Co., Ltd. Joint Institute. + +# References + +[1] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 1 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2 +[3] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jenq-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 2 +[4] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Llavolta: Efficient multi-modal models via stage-wise visual context compression. arXiv preprint arXiv:2406.20092, 2024. 2, 6, 7 +[5] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2025. 2, 5, 6, 7 +[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1, 2 +[7] Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y Wu, et al. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models. arXiv preprint arXiv:2401.06066, 2024. 2 +[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 1 +[9] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 2 +[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2024. 6 + +[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 6 +[12] Kai Han, Jianyuan Guo, Yehui Tang, Wei He, Enhua Wu, and Yunhe Wang. Free video-llm: Prompt-guided visual perception for efficient training-free video llms. arXiv preprint arXiv:2410.10441, 2024. 2, 12 +[13] Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. Zipvl: Efficient large vision-language models with dynamic token sparsification and kv cache compression. arXiv preprint arXiv:2410.08584, 2024. 2 +[14] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. CogvIm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 1 +[15] Lianyu Hu, Fanhua Shang, Liang Wan, and Wei Feng. illava: An image is worth fewer than 1/3 input tokens in large multimodal models. arXiv preprint arXiv:2412.06263, 2024. 2, 12 +[16] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024. 2, 6, 7 +[17] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviating the semantic saw-tooth effect for lightweight mllms via complementary image pyramid. arXiv preprint arXiv:2408.02034, 2024. 2 +[18] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6 +[19] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 2 +[20] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 12 +[21] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 6 +[22] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 2 + +[23] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 6 +[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1 +[25] Xinhao Li, Yi Wang, Jiashuo Yu, Xiangyu Zeng, Yuhan Zhu, Haian Huang, Jianfei Gao, Kunchang Li, Yinan He, Chenting Wang, et al. Videochat-flash: Hierarchical compression for long-context video modeling. arXiv preprint arXiv:2501.00574, 2024. 2 +[26] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 6, 14 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.1,2, 4,6,14 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 14 +[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6 +[31] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 1 +[32] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 6 +[33] Yaxin Luo, Gen Luo, Jiayi Ji, Yiyi Zhou, Xiaoshuai Sun, Zhiqiang Shen, and Rongrong Ji. $\gamma$ -mod: Exploring mixture-of-depth adaptation for multimodal large language models. arXiv preprint arXiv:2410.13859, 2024. 12 +[34] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 6 +[35] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 6 +[36] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceed- + +ings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 6 +[37] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 6 +[38] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611, 2024. 2 +[39] Yasmine Omri, Parth Shroff, and Thierry Tambe. Token sequence compression for efficient multimodal computing. arXiv preprint arXiv:2504.17892, 2025. 2 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[41] David Raposo, Sam Ritter, Blake Richards, Timothy Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models. arXiv preprint arXiv:2404.02258, 2024. 2, 3, 4, 5, 7 +[42] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 12 +[43] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2 +[44] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6 +[45] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1 +[46] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 1 +[47] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1 +[48] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large + +vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 12 +[49] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 19792-19802, 2025. 2, 12 +[50] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2 +[51] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024. 1 +[52] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European conference on computer vision, pages 69-85. Springer, 2016. 12 +[53] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 6 +[54] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, et al. Lmms-eval: Reality check on the evaluation of large multimodal models. arXiv preprint arXiv:2407.12772, 2024. 6 +[55] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 1, 2 +[56] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv eprints, pages arXiv-2412, 2024. 2, 12 +[57] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2, 5, 12 +[58] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 4 +[59] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1 +[60] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St + +moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 2 \ No newline at end of file diff --git a/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/images.zip b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..690b59adf856369c3196ad37e1fcacca7a87a75c --- /dev/null +++ b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1113471755e2c47d5e4de14b7d8f3cbe5fd3306836152d3ecb7c3389365bb478 +size 579264 diff --git a/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/layout.json b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f08371997a9c07c39eb8fbdcfe6bfdd825dde700 --- /dev/null +++ b/pmodbuildingmixtureofdepthsmllmsviaprogressiveratiodecay/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1832b3e4847a4edee396a783301627c3aa7072f8b77470b2c05bfc0a9527178 +size 424023 diff --git a/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/a7078303-ba33-4c42-a787-cdfd0a7d7b84_content_list.json b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/a7078303-ba33-4c42-a787-cdfd0a7d7b84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..42fd9b24592d887d189e4a69c399ea0c853cdb28 --- /dev/null +++ b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/a7078303-ba33-4c42-a787-cdfd0a7d7b84_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5b5c4b9dcac62534d573104144e5f26c998e383f3ba592336135ba9b024394d +size 101152 diff --git a/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/a7078303-ba33-4c42-a787-cdfd0a7d7b84_model.json b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/a7078303-ba33-4c42-a787-cdfd0a7d7b84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..36b93fae008f83d84a796bed9ebf09f09118d3ef --- /dev/null +++ b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/a7078303-ba33-4c42-a787-cdfd0a7d7b84_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70345196344f7b772ab770d2acf92a010e49a08da9dfe4913d6361ab241e8312 +size 133546 diff --git a/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/a7078303-ba33-4c42-a787-cdfd0a7d7b84_origin.pdf b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/a7078303-ba33-4c42-a787-cdfd0a7d7b84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..047e79642d7c9458237d64ca802b22c046844528 --- /dev/null +++ b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/a7078303-ba33-4c42-a787-cdfd0a7d7b84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81f4bb8cbae5f9e90873d315a68d2bf201ccd98efcf6f23bbb7e2d42c226a552 +size 3739604 diff --git a/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/full.md b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cbb85beeb99d714630ad604630a1b815c8a3f14f --- /dev/null +++ b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/full.md @@ -0,0 +1,420 @@ +# Why LVLMs Are More Prone to Hallucinations in Longer Responses: The Role of Context + +Ge Zheng $^{1,2*}$ Jiaye Qian $^{2*}$ Jiajin Tang $^{2}$ Sibei Yang $^{1\dagger}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University $^{2}$ ShanghaiTech University + +Project Page: https://github.com/SooLab/HalTrapper + +# Abstract + +Large Vision-Language Models (LVLMs) have made significant progress in recent years but are also prone to hallucination issues. They exhibit more hallucinations in longer, free-form responses, often attributed to accumulated uncertainties. In this paper, we ask: Does increased hallucination result solely from length-induced errors, or is there a deeper underlying mechanism? After a series of preliminary experiments and findings, we suggest that the risk of hallucinations is not caused by length itself but by the increased reliance on context for coherence and completeness in longer responses. Building on these insights, we propose a novel "induce-detect-suppress" framework that actively induces hallucinations through deliberately designed contexts, leverages induced instances for early detection of high-risk cases, and ultimately suppresses potential object-level hallucinations during actual decoding. Our approach achieves consistent, significant improvements across all benchmarks, demonstrating its efficacy. The strong detection and improved hallucination mitigation not only validate our framework but, more importantly, re-validate our hypothesis on context. Rather than solely pursuing performance gains, this study aims to provide new insights and serves as a first step toward a deeper exploration of hallucinations in LVLMs' longer responses. + +# 1. Introduction + +Recently, Large Vision-Language Models (LVLMs) [3, 7, 8, 12, 18, 48, 98] have made significant strides in developing general-purpose foundation models, achieving new, unprecedented capabilities. These models facilitate dynamic, context-driven interactions centered on the image content through open-ended conversations with users, given the input image and user instructions. Their impressive generative capabilities allow them to address various traditional + +![](images/19e28f9336ad5e136fafd6188f913f1d7f12479e2a453cb896aff2db720fda91.jpg) +Figure 1. Left: Our three main findings and the three steps of our HalTrapper method. Right: The distribution of hallucination locations detected by our HalTrapper is close to the true distribution of hallucinations, indicating that our method, to some extent, captures the essence of LVLM hallucinations. + +vision tasks [5, 20, 34, 35, 38, 49, 56, 58, 63-65, 88, 95, 96, 99, 100] within a unified framework and seamlessly handle more comprehensive tasks [15, 16, 23, 51, 60, 79, 86, 93] that require world knowledge and complex reasoning, such as visual question answering [2, 26, 59, 62], video-based reasoning [6, 9, 37, 41] and mathematical reasoning [54, 74]. However, LVLMs also grapple with the hallucination issue [27, 57, 92, 97], a serious and well-recognized challenge in deploying them in real-world scenarios [21, 36, 45, 52], due to their propensity for erroneous generation. + +Hallucination in LVLMs specifically refers to the discrepancy between the generated textual responses and the actual visual content and user instruction received, resulting in the production of irrelevant or non-existent objects, attributes, and other details. Various approaches have been proposed to reduce hallucinations, including filtering more reliable training data [46, 90, 97] or using specialized contrastive training materials [28] to re-fine-tune the model, thereby minimizing factually incorrect outputs. Rather than relying on costly, data-intensive solutions, recent approaches propose training-free strategies, such as + +contrastive decoding to contrastive model responses with their error-prone versions [31, 33, 76], rolling back uncertain outputs [25], or enhancing attention to visual content [50]. This has significantly mitigated the hallucination phenomenon, particularly in answering visual questions and identifying specific object hallucinations. However, most of these efforts primarily focus on short responses, while hallucinations in long-form generation remains underexplored. + +In this paper, we explore a seemingly straightforward—even widely taken for granted—phenomenon: LVLMs are more prone to hallucinations in longer, freeform textual responses compared to shorter answers. As shown in Fig. 1, the frequency of hallucinated objects correlates with their position in the output token sequence, with a higher likelihood of appearing at later positions. Previous work [97] has also observed similar phenomena, simply attributing the issue to autoregressive text generation, where increasing length leads to accumulated hallucinations and greater uncertainties. However, beneath the intuitive manifestation of length (like an iceberg), deeper factors (beneath the surface) have yet to receive adequate attention: Is the increased hallucination merely a result of the cumulative errors due to length itself, or does it arise from a deeper underlying mechanism? + +Motivated by this, this paper presents the first and preliminary attempt to explore the underlying factors through a three-step analysis approach: + +- Phenomenon discovery to propose hypotheses (Sec. 3). +- Preliminary statistics to analyze hypotheses (Sec. 4). +- Hypothesis application to detect and mitigate hallucinations, thereby re-validating it (Sec. 5). + +Phenomenon Discovery: Context may be a potential factor. Since free-form textual responses lack a predefined answer set or clear response forms, LVLMs rely heavily on context, including user instructions, visual input, and especially prior textual outputs. Consequently, we investigate the effect of context (see Sec. 3), specifically by modifying either the image or text context and observing marked shifts in the distribution of the hallucination-length curve, which indicates that hallucinations appear at earlier positions. + +Hypothesis Analysis: Contextual coherence and completeness induce hallucinations. Based on this observation, we hypothesize that contextual cues influence hallucinations along two key dimensions: + +- Contextual coherence drives LVLMs to maintain consistency with prior outputs while avoiding redundancy through distinct generation. The former focuses attention on contextual image content, while the latter shifts it to new information, potentially leading to dispersed attention, confusion, and hallucinations (see Sec. 4.1). Non-hallucinated tokens exhibit clear, focused attention, whereas hallucinated tokens show dispersed patterns. Notably, hallucinated tokens share highly similar attention + +distributions (see Fig. 3), suggesting LVLMs may be forced to attend to the same ungrounded, fragmented regions when balancing contextual and distinct content fails. + +- Contextual completeness requires responses to incorporate comprehensive content while maintaining a logically coherent linguistic structure. However, when available recognized content is insufficient, LVLMs may employ contextual extrapolation as a compensatory strategy, potentially leading to hallucinated outputs (see Sec. 4.2). As contextual completeness increases, hallucinations tend to appear earlier in the response (see Fig. 4). Furthermore, contextual extrapolation seems to follow inherently fixed patterns, with different sets of prompts repeatedly generating overlapping hallucinated tokens. + +Application and Re-validation. To further validate the hypotheses, we propose HalTrapper—a novel "induce-detect-suppress" framework that directly induces hallucinations by applying the two hypotheses, leverages the induced instances to detect high-risk cases early to nip them in the bud, and ultimately suppress potential hallucinations during the actual decoding stage. + +- Induction: (1) Imposing new, coherent outputs on an already complete response induces intra-response hallucinations. (2) Explicitly guiding imagination both based on and beyond recognized objects induces external expansion hallucinations. +- Detection: (1) Building on our coherence findings in Fig. 3, we identify hallucinations by analyzing attention similarity with induced intra-response hallucinations. (2) Building on our completeness findings in Fig. 4, we collect potential hallucinations by identifying objects that frequently appear under different imagination prompts. (3) Interestingly, our detection results align with the original hallucination distribution in Fig. 1, suggesting that context-induced and detected hallucinations mirror those seemingly driven by length, re-validating context is one of the potential factors beneath the iceberg of length. +- Suppression: Given the detected potential hallucinations, we can directly suppress their likelihood to mitigate hallucinations. Inspired by contrastive decoding [31, 32, 76], we innovatively treat detected hallucinated objects as contrastive context tokens to their probability in the contrastive branch, thereby reducing their likelihood in the original decoding branches. + +To sum up, our contributions are as follows: + +- We are the first to explore the underlying factors beneath the intuitive length-hallucination correlations, and identify context as the potential factor. +- We introduce a novel hypothesis based on coherence and completeness, and validate it through statistical analysis, hallucination detection, and suppression. +- Our exploration reveals novel insights, including the sim + +ilarity in image attention patterns of hallucinated objects and the repetition of hallucinations across prompts. + +- Building on the hypothesis, we propose a novel "inducedetect-suppress" framework, which re-validates our hypothesis while achieving competitive performance on public benchmarks. + +# 2. Related Work + +# 2.1. Large Vision-Language Models + +The success of large language models (LLMs) [1, 4, 13, 71] establishes the foundation for the development of large visual-language models (LVLMs) [3, 17, 47, 98]. Recent approaches typically adopt a unified framework, where a pre-trained visual encoder extracts visual features, which are then mapped to the LLM embedding space via either linear layers [12, 47] or Q-Former [3, 17, 98], and subsequently processed with text inputs. While LVLMs demonstrate remarkable capabilities in visual understanding [2, 10, 14, 26, 43, 55, 59, 61, 67–69, 80–85, 94] and reasoning tasks [29, 53, 89] through supervised fine-tuning [22, 24, 42, 47, 91], hallucinations remains a prominent challenge [33, 40, 57, 97]. Existing studies [19, 30, 70, 77] on the internal mechanisms of LVLMs have yet to provide a thorough explanation of the nature of hallucinations, particularly in long-form responses. This work sheds light on hallucinations in long-form generation in LVLMs. + +# 2.2. Hallucinations in LVLMs + +Unlike hallucination in LLMs, which refers to the generation of factually incorrect or meaningless content, hallucinations in LVLMs are more concerned with discrepancies between the generated content and the provided visual inputs. Early studies [40, 57] adapt the definition of hallucinations from the captioning task to the context of LVLMs. Subsequent research [25, 32, 46, 97] conduct preliminary analyses of hallucinations, investigating factors such as language priors [32, 46], co-occurrence patterns [32, 97], uncertainty [97], and positional dependencies [97]. + +Several approaches [25, 28, 31, 32, 46, 50, 76, 87, 90, 97] are proposed to mitigate hallucinations in LVLMs through training. These methods include curating high-quality training datasets [97], integrating specialized contrastive training signals [28], and employing revisor models designed to correct hallucinated outputs [46, 87]. In contrast, other studies [25, 31, 32, 50, 76] explore training-free strategies as alternatives to resource-intensive training approaches. VCD [32] introduces the contrastive decoding (CD) [39] method to suppress hallucinations, gaining significant attention in the field. Subsequent methods [31, 50, 76] further design various contrastive conditions to induce hallucinations from new perspectives. Additionally, OPERA [25] identifies the overreliance on knowledge aggregation posi + +tions within the text attention mechanism as a key cause of hallucinations and suggests a rollback strategy to address this issue. Furthermore, PAI [50] strengthens the impact of image attention on model outputs, effectively reducing hallucinations. + +# 3. Is Context a Deeper Underlying Factor? + +In this section, we conduct exploratory experiments to investigate the underlying factor influencing hallucination beyond generation length. We first introduce PoScore to represent hallucination positions and reproduce the widely recognized phenomenon that hallucinations tend to occur in longer responses (Sec. 3.1). Subsequently, we modify either image or text context and analyze their effects on hallucination distribution, thereby identifying context as a potential underlying factor (Sec. 3.2). + +Default Experimental Settings. Our default experimental setup (in Sec. 3 and Sec. 4) evaluates the LLaVA v1.5 7B [47], Qwen VL Chat [3], and MiniGPT-4 [98] on a randomly sampled set of 500 COCO [44] images for statistical analysis. Additional experimental details are presented in Appendix A. + +# 3.1. Hallucinations Linked to Length. + +When leveraging LVLMs for dialogue or question-answering tasks, a notable phenomenon is that hallucinations tend to occur more frequently in the later positions of the response. To quantitatively analyze this phenomenon, we define the relative position score for each generated object as follows, consistent with previous work [97]: + +$$ +\operatorname {P o S c o r e} _ {s, i} = \frac {\operatorname {I n d e x} \left(o _ {s , i}\right)}{N _ {s}} \tag {1} +$$ + +where $o_{s,i}$ denotes the $i^{th}$ object in the response of the $s^{th}$ sample, and $N_{s}$ represents the length of the $s^{th}$ sample. We visualize the PoScore distributions for hallucinated and non-hallucinated objects for the LLaVA model in Fig. 1, with additional results from other models provided in Fig. 7 in Appendix. The results reveal a marked increase in the frequency of hallucinations as the response lengthens, aligning with findings from previous studies [78, 97]. + +# 3.2. Hallucinations Beyond Length. + +Moving beyond these prior observations, we delve deeper by posing a critical question: Is the increased hallucination merely a result of the cumulative errors due to length itself, or does it arise from a deeper underlying mechanism? In light of the critical role that context plays in free-form responses, we design the following two context modification strategies and analyze the changes in hallucination positions (PoScore) to investigate the effect of context: + +![](images/aef323b2cc227a52c852fbde021f5e108f2ed0788997ac9b4bbc058b971ec02a.jpg) +Figure 2. Statistical analysis of hallucination positions under context modifications. Both cropping the image and enriching the prompt lead to earlier hallucination occurrences. + +- Crop the image input into centered squares, retaining approximately one-third of the original area, and re-annote accordingly. +- Enrich the text input by adding two sentences that describe the image, and then prompt to describe other details. + +The results in Fig. 2 show that hallucinations tend to occur earlier in the generation process across both settings, challenging the widely held belief that they are more likely to appear in the later stages. These findings underscore the complexity of hallucinations, revealing that context plays a significant role in their occurrence, rather than attributing them solely to generation length. + +# 4. Coherence and Completeness + +This section delve into the mechanisms through which context influences hallucinations by employing a hypothesis-verification framework. Our analysis focuses on two key aspects: contextual coherence (Sec. 4.1) and contextual completeness (Sec. 4.2). Finally, we link back to text and image manipulation experiments in Sec. 3.2, providing explanations with these factors (Sec. 4.3). + +# 4.1. Coherence: Avoidance of Internal Repetition + +Contextual coherence drives the model to maintain consistency with previous outputs while avoiding redundant repetition of both the input and prior content. Based on this, we propose and validate a hypothesis on hallucination occurrence. + +Hypothesis. The two aspects of contextual coherence in image attention are conflicting: attention is required to focus on relevant regions for consistency with previous outputs, while also shifting to new areas to avoid repetition. This tension leads to dispersed attention and hallucinations. + +Experimental settings. To validate our hypothesis, we analyze both individual attention and pairwise attention comparisons. Specifically, we analyze the image attention maps of hallucinated objects $\mathcal{H}$ and non-hallucinated objects $\mathcal{N}$ , with representative results shown in Fig. 3 (right). Addi + +![](images/315c81e7cf8431ed411a65a08f7687bad7e3b51b3f095e4ed4eb44ecec51896a.jpg) +Figure 3. Statistical analysis related to contextual coherence. Within the same caption, hallucinated object pairs exhibit higher attention similarity scores than non-hallucinated pairs. + +tionally, we quantify the intra-set attention similarity of objects within the same response, denoted by $S_{\mathcal{H}}$ and $S_{\mathcal{N}}$ , as follows: + +$$ +S _ {\mathcal {H}} = \left\{\operatorname {s i m} \left(A _ {s, i}, A _ {s, j}\right) \mid o _ {s, i}, o _ {s, j} \in \mathcal {H} \right\}, \tag {2} +$$ + +$$ +S _ {\mathcal {N}} = \left\{\operatorname {s i m} \left(A _ {s, i}, A _ {s, j}\right) \mid o _ {s, i}, o _ {s, j} \in \mathcal {N} \right\} +$$ + +where $A_{s,i}$ and $A_{s,j}$ represent the image attention maps of the $i^{th}$ and $j^{th}$ objects in the response for the $s^{th}$ image, and $\mathrm{sim}(\cdot ,\cdot)$ denotes the cosine similarity function. Fig. 3 (left) illustrates the distributions of $S_{\mathcal{H}}$ and $S_{\mathcal{N}}$ . + +Results. Qualitative analysis (right panel of Fig. 3) indicates that when the model successfully identifies real objects, it concentrates on the relevant regions. Conversely, if the model fails to recognize a novel object, its attention disperses and distracting information, leading to hallucinations. Quantitative results (left panel of Fig. 3) show a clear difference between the distributions of $S_{\mathcal{H}}$ and $S_{\mathcal{N}}$ . Specifically, hallucinated objects exhibit higher attention similarity, while real objects show lower values. This further indicates that hallucinated objects typically manifest diffuse, noisy attention patterns, making attention similarity a robust metric for their detection. + +# 4.2. Completeness: External Extrapolation + +Contextual completeness comprises two key dimensions: the informational dimension, which demands a thorough and comprehensive response, and the structural dimension, which ensures the response is logically coherent and grammatically sound. Building on this, we propose the following hypotheses regarding the occurrence mechanism and inherent tendency of hallucination. + +Hypothesis. (a) Occurrence: When a response includes correctly identified objects but remains incomplete in informative or structural aspect, the model compensates by ex + +![](images/e079918b105aff38e8a0ec9b10fc78e774377596d91eae82cc89062df5e4e77c.jpg) + +![](images/303cbd00fd7678906836349eb6c563c0dc708f82e31b68e783ad1e0c516d540a.jpg) +Figure 4. Statistical analysis related to contextual completeness: (a) Hallucination positions shift progressively earlier as more image information is included in the prompts. (b) Similar hallucinations consistently recur across varied prompts for the same image. + +panding imagined details, i.e., hallucinations. (b) Tendency: These hallucinations from external extrapolation rely on multimodal context, particularly visual inputs. + +Experimental settings. We conduct two separate experiments for validation as follows: + +(a) We validate the role of completeness by analyzing its correlation with hallucination positions. Specifically, we extend text manipulation experiment in Sec. 3.2 by incrementally adding image descriptions to the prompt and visualizing the average PoScore in Fig. 4(a). +(b) We further investigate the consistency and image-related properties of hallucinated objects across different prompts. Specifically, we apply five prompts to each image and compute the proportion of repeated hallucinated objects. Formally, let $\mathcal{H}_{s_k}$ represent the set of hallucinated objects generated by the $k^{th}$ prompt for the $s^{th}$ sample, with the complete hallucination set given by $\mathcal{H}_s = \bigcup_{k=1}^5 \mathcal{H}_{s_k}$ . We count the occurrence of each hallucinated object $h \in \mathcal{H}_s$ as $c_s(h) = \sum_{k=1}^5 \mathbb{1}(h \in \mathcal{H}_{s_k})$ , where $\mathbb{1}$ is the indicator function. Then we calculate $N(k)$ , the number of hallucinated objects that appear $k \in [1,2,3,4,5]$ times over all samples, along with its proportion $R(k)$ shown in Fig. 4(b): + +$$ +\begin{array}{l} N (k) = \sum_ {s} \sum_ {\substack {h \in \mathcal {H} _ {s} \\ N (1)}} k \cdot \mathbb {1} \left(c _ {s} (h) = k\right), \tag{3} \\ R (k) = \frac {N (k)}{\sum_ {k = 1} ^ {5} N (k)} \\ \end{array} +$$ + +Results. (a) The results in Fig. 4(a) indicate that as more enriched sentences are incorporated, leading to a more comprehensive context, hallucinations occur at earlier positions. This is because the diminishing content available for generation makes it increasingly challenging for LVLMs to accurately identify details for a complete and coherent response. + +(b) The proportion presented in Fig. 4(b) demonstrate that all models exhibit a high degree of repetitiveness in hallucinated objects, with objects appearing in only one response accounting for merely $30\%$ on average. Given the variations in both questions and preceding responses, the repeated hallucinated objects are often closely tied to the image context, aligning with our qualitative analysis in Appendix E. + +# 4.3. Link Back to Phenomenon in Sec. 2.2 + +Explaining Text Manipulation Experiments. Revisiting the text manipulation experiments, we find that contextual coherence and completeness provides an intuitive explanation for this behavior. When additional descriptions of real objects are incorporated, the model tends to avoid redundancy and maintain coherence, thereby reducing the number of objects to describe. Consequently, the model turns to uncertain or unverified objects more quickly to ensure completeness, leading to earlier hallucinations. + +Explaining Image Manipulation Experiments. Contextual completeness offers a compelling explanation for the image manipulation experiments. Similarly, cropping images systematically reduces the number of recognizable objects, forcing the model to hallucinate earlier in order to maintain contextual completeness. + +# 5. Re-Validation via Detection and Suppression + +To rigorously validate our hypothesis, we extend the findings from Section 4 to practical application of hallucination detection and suppression. Specifically, we propose HalTrapper, which introduces a novel "induce-detect-suppress" strategy (see Fig. 5). The induce-detect stages leverage Internal Grounding (IG) and External Expansion (EE) techniques for hallucination detection (Sec. 5.1), and can be easily adapted with Contrastive Contextual Decoding (CCD) for suppression (Sec. 5.2). + +# 5.1. Hallucination Induction-Detection + +# 5.1.1. Internal Grounding + +In Sec. 4.1, we demonstrate that the attention similarity between paired objects serves as an effective indicator for distinguishing hallucinated pairs from non-hallucinated ones. Building on this insight, we propose the Internal Grounding (IG) method, which adopts an induce-then-detect paradigm to detect hallucinated objects in model responses. + +Induction. A key component of IG is the selection of reference objects, which serve as anchors for similarity computation. Instead of using naturally generated objects, we induce the model to generate additional objects following the initial response, which are more prone to hallucination. Specifically, given an input image and the model's initial response, we replace the EOS token in the generated output with an additional cue, "There is also". Since the initial responses inherently covers a considerable number of + +![](images/09f50547f09bdaa3aaf2d94edb40b0e9ec43d3c0724036ec44aadc4745eedb1c.jpg) +Figure 5. Overview of HalTrapper: It consists of two branches leveraging coherence and completeness insights. One generates captions with an appended "There is also" prompt to induce potential hallucinated objects, detected via high attention similarity between caption and induced tokens. The other prompts the LVLM to imagine surrounded content beyond the image to identify consistent hallucinations. With detected hallucinated objects, HalTrapper further suppresses hallucinations through Contrastive Contextual Decoding. + +the identified objects, when completeness is compromised, the model tends to externally extrapolate to compensate, thereby restoring completeness (see Sec. 4.2). The resulting object serves as the reference object and is denoted as $\sigma_{\mathrm{s}}^{\mathrm{ref}}$ for the $s^{th}$ sample. + +Detection. We then compute the attention similarity scores IGScore between the induced hallucinated objects $o_s^{ref}$ and the preceding objects, filtering out potential hallucination candidates $S_{IG}$ with high similarity: + +$$ +\operatorname {I G S c o r e} _ {s, i} = \operatorname {s i m} \left(A _ {s} ^ {r e f}, A _ {s, i}\right) \tag {4} +$$ + +$$ +S _ {I G} = \left\{o _ {s, i} \mid \operatorname {I G S c o r e} _ {s, i} > \theta_ {I G} \right\} +$$ + +where $\theta_{IG}$ denotes the threshold. Notably, the proposed method remains robust even when the reference object is real, as the similarity scores between non-hallucinated objects are typically low, effectively preventing real objects from being misclassified as hallucinations. + +# 5.1.2. External Expansion + +Another observation is that hallucinated objects exhibit consistency across identical visual inputs (Sec. 4.2). Based on this property, we propose the External Expansion (EE) method, explicitly inducing the imagination related to the image, treating them as detected potential hallucinations. + +Induction. Considering that hallucinations from external extrapolation rely on image context, we first prompt with "Please imagine what object might be outside the frame" to induce image-related associations and capture potential hallucinations. However, directly extracting hallucinated objects from the response leads to false positives, as the model might imagine objects present in the image. To address this, we design a reason-then-imagine prompt to filter out such existing objects (see Appendix C.2). It explicitly guides the model in distinguishing between recognized objects and imagined ones. Furthermore, it utilizes reliable intermediate steps to enable context-driven reasoning, thereby improving response fidelity. + +Detection. We introduce EEScore, based on the principle that an object's presence in the imagination set improves the + +likelihood of it being perceived as a hallucination, while its presence in the reason set reduces this likelihood. Specifically, we define the imagination set and the reason set at direction $d \in \mathcal{D}$ as $S_{I,d}$ and $S_{R,d}$ , respectively. The final set of potential hallucinations is formulated as follows: + +$$ +\operatorname {E E S c o r e} _ {s, i} = \sum_ {d \in \mathcal {D}} \left[ \mathbb {1} \left(o _ {s, i} \in S _ {I, d}\right) - \mathbb {1} \left(o _ {s, i} \in S _ {R, d}\right) \right] \tag {5} +$$ + +$$ +S _ {E E} = \left\{o _ {s, i} \mid \operatorname {E E S c o r e} _ {s, i} > \theta_ {E E} \right\} +$$ + +Finally, we combine the potential hallucinations detected by the IG and EE methods as follows: + +$$ +S _ {\text {i n d u c t i o n}} = S _ {I G} \cup S _ {E E} \tag {6} +$$ + +# 5.2. Hallucination Suppression + +Preliminaries. Let $\theta$ denote the parameters of an LVLM. Given an input image $v$ and a text prompt $x$ , the model autogressively generates a response $y$ of length $L$ . Formally, the decoding process can be formulated as follows: + +$$ +p _ {\theta} (y | v, x) = \prod_ {i = 1} ^ {L} p _ {\theta} \left(y _ {i} | v, x, y _ {< i}\right) \tag {7} +$$ + +where $y_{i}$ and $y_{ModelMetricAUROCTPR5%FPRF1maxAcc.LLaVA v1.5PoScore70.74.338.370.7Top Logit64.013.032.261.9Logits'Entropy67.716.636.671.4Image Attn. Ratio44.96.027.332.0IG Score82.343.354.886.3EE Score77.5-46.172.9MiniGPT 4PoScore70.512.235.466.2Top Logit65.622.937.076.5Logits'Entropy65.522.135.375.9Image Attn. Ratio64.37.731.957.9IG Score76.634.048.680.7EE Score60.5-30.046.5Qwen VLPoScore71.14.834.465.8Top Logit71.519.636.177.7Logits'Entropy70.723.336.673.9Image Attn. Ratio57.36.826.941.4IG Score76.233.343.884.6EE Score81.3-46.373.0 + +Table 1. Quantitative results for hallucination detection. The best performances within each setting are bolded. + +puts unchanged, is expressed as follows: + +$$ +\begin{array}{l} p _ {c d} \left(y _ {i} \mid v, x, y _ {< i}\right) = \operatorname {s o f t m a x} [ (1 + \alpha) \operatorname {l o g i t} _ {\theta} \left(y _ {i} \mid v, x, y _ {< i}\right) \\ - \alpha \operatorname {l o g i t} _ {\theta^ {\prime}} \left(y _ {i} | v, x, y _ {< i}\right) ] \tag {8} \\ \end{array} +$$ + +where $p_{\theta^{\prime}}(x_i|v,x,y_{< i})\propto \exp \log \mathrm{it}_{\theta^{\prime}}(x_i|v,x,y_{< i})$ . It also employs a truncation of the probability distribution following [32]. + +Contrastive Contextual Decoding (CCD). Building on the previously introduced induce-detect stages, a simple CD-based extension CCD enables hallucination suppression. Unlike previous CD methods, CCD explicitly integrates a prior for potential hallucination objects, aiming to reduce their likelihood in response. Specifically, we encode potential hallucinated objects as text tokens, referred to as Contrastive Contextual Tokens (CCT) $x_{cct}$ . We then concatenate CCT with the image input to construct a contrastive branch, with model parameters and other inputs unchanged. The CCD process can be formally expressed as follows: + +$$ +p _ {c c d} \left(y _ {i} \mid v, x, y _ {< i}\right) = \prod_ {i = 1} ^ {L} p _ {c c d} \left(y _ {i} \mid v, x _ {c c t}, x, y _ {< i}\right) \tag {9} +$$ + +We then detail the modifications applied to the CD process as follows: + +$$ +\begin{array}{l} p _ {c c d} \left(y _ {i} \mid v, x _ {c c t}, x, y _ {< i}\right) = \\ \operatorname {s o f t m a x} \left[ (1 + \alpha) \log \mathrm {i t} _ {\theta} \left(y _ {i} \mid v, x, y _ {< i}\right) \right. \tag {10} \\ - \alpha \operatorname {l o g i t} _ {\theta} \left(y _ {i} | v, x, x _ {c c t}, y _ {< i}\right) ] \\ \end{array} +$$ + +By treating CCT tokens as complementary to image content, the model naturally increases the likelihood of potential hallucinated objects and their associated terms in the contrastive branch, thereby effectively reducing their occurrence in the final generation. + +# 6. Experiments + +Datasets and Benchmarks. To demonstrate the effectiveness of our HalTrapper, we use images from COCO [44] + +![](images/a644c3cd01cf746cf86bf98b749df0ac51045305877b4da68e25e8ebb686f6fb.jpg) +Figure 6. Comparison between the positional distribution of hallucinations detected by our method and the overall hallucination distribution, demonstrating a high degree of alignment. + +and AMBER [72] datasets. Detailed descriptions can be found in the Appendix C.1. + +Base Models. We select LLaVA v1.5 7B [47], MiniGPT-4 [98], and Qwen VL Chat [3] as our main baselines for our study. We also evaluate more recent models Qwen2 VL 7B [75] and Janus Pro 7B [11] on AMBER, which has higher annotation qualities. + +Implementation Details. For all experiments, the maximum number of newly generated tokens is set to 512. Following prior mainstream studies on CD [32, 76], we adapt $\alpha = 1.0$ and $\beta = 0.1$ . See Appendix C.3 and C.4 for details on CCT construction and more hyperparameters. + +# 6.1. Detection + +Metrics. Inspired by [66], we adapt AUROC (Area Under the ROC Curve) and TPR@5%FPR (the True Positive Rate at 5% False Positive Rate) as our primary metrics for hallucination detection. AUROC quantifies the model's overall discriminative ability across all classification thresholds, while TPR@5%FPR is suitable for scenarios with strict requirements on the false positive rate. We also report the F1 Score and Accuracy at the threshold that maximizes the F1. Baseline Methods. For each generated object $o_{s,i}$ , we first employ PoScore [97] as a basic metric. We also propose two uncertainty-based metrics: Top Logit and Logits' Entropy. The Top Logit is the maximum value of the logits when generating $o_{s,i}$ , while Logits' Entropy refers to the entropy of the logits at that moment. Additionally, we employ an Attention-based metric called the Image Attention Ratio, defined as the ratio of the model's attention score on the image to its total attention score when generating $o_{s,i}$ . + +Results. The quantitative results of hallucination detection are presented in Table 1. As shown, our approach demonstrates significant improvements across all evaluation settings. For IG, in terms of the AUROC metric, our method outperforms the best baseline PoScore by $5\% - 12\%$ . This indicates that our method enhances performance across the entire classification curve. Considering that our IG method originates from within the model, this indicates that the model indeed exhibits significant similar attention pat + +
DecodingMethodLLaVA v1.5 7B [47]
CS↓CI↓Prec.RecallF1Len
GreedyICD [76]51.414.773.481.077.0102.1
CODE [31]50.013.775.876.976.488.3
Vanilla52.214.673.780.376.9100.8
Ours41.611.978.780.179.4100.0
10.6↓2.7↓5.0↑0.2↓2.5↑
NucleusVCD [32]58.216.970.878.874.6103.2
ICD [76]55.016.570.977.974.2102.1
CODE54.216.472.376.274.291.6
Vanilla58.618.868.176.472.0105.2
Ours48.614.574.677.776.1100.9
10.0↓4.3↓6.5↑1.3↓4.1↑
Beam SearchOPERA [25]53.615.772.477.674.998.8
Vanilla55.615.872.881.076.7104.2
Ours45.212.178.981.280.0101.8
10.4↓3.7↓6.1↑0.2↓3.3↑
+ +tern in certain hallucination scenarios. Additionally, for TPR@5%FPR, our method improves by at least $10\%$ compared to the baselines. This highlights the substantial potential of the EE metric in inducing hallucinations. Given that MiniGPT 4 is trained only on the image interface, its ability to follow instructions is relatively limited, which may account for the lack of improvement in the EE metric. + +We also visualized the qualitative results of hallucination positions distribution detected by our method, with the overall distribution of hallucination positions, as shown in Fig. 6. It demonstrates that our method accurately captures the hallucination distribution, closely aligning with the overall pattern observed in captions. This further indicates that although we claim that our method is designed for long-text scenarios, its effectiveness is not merely dependent on the length of the generated text. Instead, our approach effectively captures an intrinsic mechanism underlying LVLM hallucinations, which is beyond text length. Therefore, our study not only validates the applicability of our method but also provides a new perspective for understanding the formation mechanism of LVLM hallucinations. + +# 6.2. Suppression + +Metrics. CHAIR [57] is commonly used to quantify hallucinations in model-generated captions based on COCO. Besides CHAIR, we also report several classic metrics, including Precision, Recall, F1, and the average length of the captions. For AMBER [72], following the approach outlined in their paper, we report CHAIR, Cover, Hal, and Cog. As we primarily focus on long context scenarios, we conduct full evaluations only on its generative subset and reported the results accordingly. We also conduct experiments on POPE and GPT-4o, please refer to the Appendix D.2 and D.4. + +Baseline Methods. We compare our HalTrapper with VCD [32], ICD [76], CODE [31], and OPERA [25]. + +CHAIR Evaluation. As shown in Tables 2 and 3. HalTrapper significantly reduces CHAIR while maintaining Recall + +Table 2. Results on CHAIR. Lower $\mathrm{CHAIR}_S$ , $\mathrm{CHAIR}_I$ , and higher precision, recall and F1 indicate fewer hallucinations. The best performances within each setting are bolded. + +
DecodingMethodMiniGPT 4 [98]Qwen VL Chat [3]
CS↓CI↓Prec.CS↓CI↓Prec.
GreedyVanilla39.614.776.643.413.575.8
ICD [76]42.614.776.350.414.473.7
CODE [31]32.813.681.240.412.578.9
Ours28.610.783.138.610.280.9
NucleusVanilla37.214.677.144.813.676.3
VCD [32]39.614.976.647.414.174.3
ICD41.414.976.152.615.073.0
CODE [31]36.614.079.543.614.575.4
Ours29.011.582.142.411.379.3
Beam SearchVanilla38.813.878.041.411.679.0
OPERA [25]43.014.975.842.812.576.9
Ours37.613.778.334.29.782.7
+ +Table 3. More results on CHAIR with MiniGPT-4 and Qwen VL. + +
Model / MethodCHAIR↓Cover↑Hal↓Cog↓
LLaVA v1.5 7B [47]11.250.247.94.6
+ VCD [32]8.951.238.14.4
+ ICD [76]8.651.137.33.9
+ CODE [31]9.051.139.54.3
+ Ours8.0 (3.2↓)51.5 (1.3↑)36.3 (11.6↓)3.8 (0.8↓)
Qwen2 VL [75]6.671.850.34.6
+ VCD7.370.653.24.6
+ ICD8.274.974.99.1
+ CODE7.671.656.35.1
+ Ours5.6 (1.0↓)70.946.1 (4.2↓)3.8 (0.8↓)
Janus Pro 7B [11]6.365.637.52.0
+ VCD5.566.232.52.1
+ ICD6.167.136.32.5
+ CODE6.065.333.61.6
+ Ours5.4 (0.9↓)66.5 (0.9↑)32.7 (4.8↓)1.8 (0.2↓)
+ +Table 4. Results on AMBER [73] generative task. $\downarrow$ indicates lower is better. + +with minimal negative impact. Across all experiments on $\mathrm{CHAIR}_S$ and $\mathrm{CHAIR}_I$ , HalTrapper achieves significant improvements. Notably, in Table 2, our approach consistently improves $\mathrm{CHAIR}_S$ by over $10\%$ and $\mathrm{CHAIR}_I$ by $2.5\%$ . This demonstrates that the hallucination candidates identified by our IG and EE metrics are of high quality, enabling the inclusion of a large number of hallucinated objects while minimizing the presence of non-hallucinated ones. This, in turn, provides validation of the effectiveness of our IG and EE metrics in detecting hallucinations, further highlighting the universality and practical significance of our findings. + +AMBER Evaluation. As shown in the Table 4, HalTrapper continues to demonstrate performance improvements on latest models. Ablation Study. See Appendix D.1 for more details on the ablation study. + +# 7. Conclusion + +In this paper, we propose a novel method for eliminating hallucinations in Large Vision-Language Models through two mechanisms: external spatial expansion and internal visual grounding. Our HalTrapper introduces a simple, zero-shot hallucination detection and suppression technique that achieves significant improvements across all benchmarks, with no additional training required. Our approach consistently delivers substantial improvements across all benchmarks, validating its effectiveness. + +Acknowledgment. This work is supported by the National Natural Science Foundation of China (No.62206174). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In International Conference on Computer Vision (ICCV), 2015. 1, 3 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1, 3, 7, 8 +[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020.3 +[5] Chaoqi Chen, Yushuang Wu, Qiyuan Dai, Hong-Yu Zhou, Mutian Xu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. A survey on graph neural networks and graph transformers in computer vision: A task-oriented perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46 (12):10297-10318, 2024. 1 +[6] Guo Chen, Yin-Dong Zheng, Jiahao Wang, Jilan Xu, Yifei Huang, Junting Pan, Yi Wang, Yali Wang, Yu Qiao, Tong Lu, et al. Videollm: Modeling video sequence with large language models. arXiv preprint arXiv:2305.13292, 2023. 1 +[7] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1 +[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1 +[9] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 1 +[10] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 3 +[11] Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Januspro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025. 7, 8 +[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer + +Vision and Pattern Recognition, pages 24185-24198, 2024. 1, 3 +[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 3 +[14] Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13711-13722, 2024. 3 +[15] Qiyuan Dai and Sibei Yang. Free on the fly: Enhancing flexibility in test-time adaptation with online em, 2025. 1 +[16] Qiyuan Dai, Hanzhuo Huang, Yu Wu, and Sibei Yang. Adaptive part learning for fine-grained generalized category discovery: A plug-and-play enhancement, 2025. 1 +[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. 3 +[18] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 1 +[19] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In European Conference on Computer Vision, pages 148-166. Springer, 2025. 3 +[20] Weifeng Ge, Sibei Yang, and Yizhou Yu. Multi-evidence filtering and fusion for multi-label classification, object detection and semantic segmentation based on weakly supervised learning, 2018. 1 +[21] Xiang He, Sibei Yang, Guanbin Li, Haofeng Li, Huiyou Chang, and Yizhou Yu. Non-local context encoder: Robust biomedical image segmentation against adversarial attacks. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):8417-8424, 2019. 1 +[22] Zijian He, Yuwei Ning, Yipeng Qin, Guangrun Wang, Sibei Yang, Liang Lin, and Guanbin Li. Vton 360: High-fidelity virtual try-on from any viewing direction, 2025. 3 +[23] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. arXiv preprint arXiv:2309.14494, 2023. 1 +[24] Hanzhuo Huang, Yuan Liu, Ge Zheng, Jiepeng Wang, Zhiyang Dou, and Sibei Yang. MVTokenflow: High-quality 4d content generation using multiview token flow. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[25] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and + +retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024. 2, 3, 8 +[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 3 +[27] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1-38, 2023. 1 +[28] Chaoya Jiang, Haiyang Xu, Mengfan Dong, Jiaxing Chen, Wei Ye, Ming Yan, Qinghao Ye, Ji Zhang, Fei Huang, and Shikun Zhang. Hallucination augmented contrastive learning for multimodal large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27036-27046, 2024. 1, 3 +[29] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 3 +[30] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's" up" with vision-language models? investigating their struggle with spatial reasoning. arXiv preprint arXiv:2310.19785, 2023. 3 +[31] Junho Kim, Hyunjun Kim, Yeonju Kim, and Yong Man Ro. Code: Contrasting self-generated description to combat hallucination in large multi-modal models. arXiv preprint arXiv:2406.01920, 2024. 2, 3, 6, 8 +[32] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv preprint arXiv:2311.16922, 2023. 2, 3, 6, 7, 8 +[33] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13872-13882, 2024. 2, 3 +[34] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 1 +[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1 +[36] Jinpeng Li, Haiping Wang, Jiabin chen, Yuan Liu, Zhiyang Dou, Yuexin Ma, Sibei Yang, Yuan Li, Wenping Wang, Zhen Dong, and Bisheng Yang. Cityanchor: City-scale 3d visual grounding with multi-modality LLMs. In The Thirteenth International Conference on Learning Representations, 2025. 1 + +[37] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1 +[38] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 1 +[39] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint arXiv:2210.15097, 2022. 3, 6 +[40] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 3 +[41] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1 +[42] Han Liang, Jiacheng Bao, Ruichi Zhang, Sihan Ren, Yuecheng Xu, Sibei Yang, Xin Chen, Jingyi Yu, and Lan Xu. Omg: Towards open-vocabulary motion generation via mixture of controllers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 482-493, 2024. 3 +[43] Liang Lin, Pengxiang Yan, Xiaogian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for referring image segmentation. IEEE Transactions on Multimedia, 24:1922-1932, 2022. 3 +[44] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll'a r, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014. 3, 7 +[45] Zhenxiang Lin, Xidong Peng, Peishan Cong, Ge Zheng, Yujing Sun, Yuenan Hou, Xinge Zhu, Sibei Yang, and Yuexin Ma. Wildrefer: 3d object localization in large-scale dynamic scenes with multi-modal visual data and natural language. In ECCV (46), pages 456-473, 2024. 1 +[46] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations, 2023. 1, 3 +[47] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 7, 8 +[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1 + +[49] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1 +[50] Shi Liu, Kecheng Zheng, and Wei Chen. Paying more attention to image: A training-free method for alleviating hallucination in lvlms. arXiv preprint arXiv:2407.21771, 2024. 2, 3 +[51] Xuyang Liu, Bingbing Wen, and Sibei Yang. Ccq: Cross-class query network for partially labeled organ segmentation. Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):1755-1763, 2023. 1 +[52] Yumeng Liu, Yaxun Yang, Youzhuo Wang, Xiaofei Wu, Ji-amin Wang, Yichen Yao, Soren Schwertfeger, Sibei Yang, Wenping Wang, Jingyi Yu, Xuming He, and Yuexin Ma. Realdex: towards human-like grasping for robotic dexterous hand. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024. 1 +[53] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35: 2507-2521, 2022. 3 +[54] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 1 +[55] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 3 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1 +[57] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. arXiv preprint arXiv:1809.02156, 2018. 1, 3, 8 +[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[59] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European Conference on Computer Vision, pages 146–162. Springer, 2022. 1, 3 +[60] Cheng Shi and Sibei Yang. Spatial and Visual Perspective-Taking via View Rotation and Relation Reasoning for Em + +bodied Reference Understanding, page 201-218. Springer-Verlag, Berlin, Heidelberg, 2022. 1 +[61] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3 +[62] Cheng Shi and Sibei Yang. Logoprompt:synthetic text images can be good visual prompts for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1 +[63] Cheng Shi and Sibei Yang. The devil is in the object boundary: Towards annotation-free instance segmentation using foundation models. In The Twelfth International Conference on Learning Representations, 2024. 1 +[64] Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. arXiv preprint arXiv:2407.10084, 2024. +[65] Cheng Shi, Yuchen Zhu, and Sibei Yang. Plain-det: A plain multi-dataset object detector. In Computer Vision – ECCV 2024, pages 210–226, Cham, 2025. Springer Nature Switzerland. 1 +[66] Gaurang Sriramanan, Siddhant Bharti, Vinu Sankar Sadasivan, Shoumik Saha, Priyatham Kattakinda, and Soheil Feizi. LLM-check: Investigating detection of hallucinations in large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 7 +[67] Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Contrastive grouping with transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23570-23580, 2023. 3 +[68] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. +[69] Jiajin Tang, Ge Zheng, Jingyi Yu, and Sibei Yang. Cotdet: Affordance knowledge prompting for task driven object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3068-3078, 2023. 3 +[70] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 3 +[71] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[72] Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Ming Yan, Ji Zhang, and Jitao Sang. Amber: An llm-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv preprint arXiv:2311.07397, 2023. 7, 8 + +[73] Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Ming Yan, Ji Zhang, and Jitao Sang. An llm-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv preprint arXiv:2311.07397, 2023.8 +[74] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024. 1 +[75] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 7, 8 +[76] Xintong Wang, Jingheng Pan, Liang Ding, and Chris Biemann. Mitigating hallucinations in large vision-language models with instruction contrastive decoding. arXiv preprint arXiv:2403.18715, 2024. 2, 3, 6, 7, 8 +[77] Yifan Wang, Yifei Liu, Yingdong Shi, Changming Li, Anqi Pang, Sibei Yang, Jingyi Yu, and Kan Ren. Discovering influential neuron path in vision transformers. In International Conference on Representation Learning, pages 25244-25272, 2025. 3 +[78] Hongliang Wei, Xingtao Wang, Xianqi Zhang, Xiaopeng Fan, and Debin Zhao. Toward a stable, fair, and comprehensive evaluation of object hallucination in large vision-language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3 +[79] Yu Wu, Yana Wei, Haozhe Wang, Yongfei Liu, Sibei Yang, and Xuming He. Grounded image text matching with mismatched relation reasoning, 2023. 1 +[80] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4643-4652, 2019. 3 +[81] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4140-4149, 2019. +[82] Sibei Yang, Guanbin Li, and Yizhou Yu. Propagating over phrase relations for one-stage visual grounding. In Computer Vision - ECCV 2020, pages 589-605, Cham, 2020. Springer International Publishing. +[83] Sibei Yang, Guanbin Li, and Yizhou Yu. Graph-structured referring expression reasoning in the wild. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9949–9958, 2020. +[84] Sibei Yang, Guanbin Li, and Yizhou Yu. Relationship-embedded representation learning for grounding referring expressions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2765-2779, 2021. +[85] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11261-11270, 2021. 3 + +[86] Chunlin Yu, Hanqing Wang, Ye Shi, Haoyang Luo, Sibei Yang, Jingyi Yu, and Jingya Wang. Seqafford: Sequential 3d affordance reasoning via multimodal large language model. arXiv preprint arXiv:2412.01550, 2024. 1 +[87] Qifan Yu, Juncheng Li, Longhui Wei, Liang Pang, Wentao Ye, Bosheng Qin, Siliang Tang, Qi Tian, and Yueting Zhuang. Hallucidobot: Mitigating hallucinatory toxicity in visual instruction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12944–12953, 2024. 3 +[88] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1 +[89] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6720-6731, 2019. 3 +[90] Jinrui Zhang, Teng Wang, Haigang Zhang, Ping Lu, and Feng Zheng. Reflective instruction tuning: Mitigating hallucinations in large vision-language models. arXiv preprint arXiv:2407.11422, 2024. 1, 3 +[91] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. ACM Transactions on Graphics, 42:1-16, 2023. 3 +[92] Muru Zhang, Ofir Press, William Merrill, Alisa Liu, and Noah A Smith. How language model hallucinations can snowball. arXiv preprint arXiv:2305.13534, 2023. 1 +[93] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. arXiv preprint arXiv:2310.16436, 2023. 1 +[94] Hong-Yu Zhou, Chixiang Lu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. Preservational learning improves self-supervised medical image models by reconstructing diverse contexts. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3479-3489, 2021. 3 +[95] Hong-Yu Zhou, Chixiang Lu, Sibei Yang, and Yizhou Yu. Convnets vs. transformers: Whose visual representations are more transferable? In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 2230-2238, 2021. 1 +[96] Hong-Yu Zhou, Chixiang Lu, Chaoqi Chen, Sibei Yang, and Yizhou Yu. A unified visual information preservation framework for self-supervised pre-training in medical image analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8020-8035, 2023. 1 +[97] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754, 2023. 1, 2, 3, 7 +[98] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language + +understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3, 7, 8 +[99] Yuchen Zhu, Cheng Shi, Dingyou Wang, Jiajin Tang, Zhengxuan Wei, Yu Wu, Guanbin Li, and Sibei Yang. Rethinking query-based transformer for continual image segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4595-4606, 2025. 1 +[100] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 1 \ No newline at end of file diff --git a/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/images.zip b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3dcd3f483ffcc77f3f8388f8858fb6e47987d10a --- /dev/null +++ b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19e293716740b7cecae700dec57c5b57e335b72f3e5f7006c49d484db9c2ec58 +size 512591 diff --git a/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/layout.json b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8e87ab6551ebdcc6c9e6ed37ed1a75298d336c0f --- /dev/null +++ b/whylvlmsaremorepronetohallucinationsinlongerresponsestheroleofcontext/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08fe75fb5bb04e33ff5e91ef1553e64d0d102776ce3a8a43bad3f177ce39167e +size 490819 diff --git a/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/baa830f2-e8bf-491c-a264-552b980530f2_content_list.json b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/baa830f2-e8bf-491c-a264-552b980530f2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8ae6468f896d426b06621eba68e24e59b1a184dd --- /dev/null +++ b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/baa830f2-e8bf-491c-a264-552b980530f2_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaa2af55e141b8d9fb8e6947d32334b50b282390af3ca9ef2893822c3612c251 +size 68339 diff --git a/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/baa830f2-e8bf-491c-a264-552b980530f2_model.json b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/baa830f2-e8bf-491c-a264-552b980530f2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b94565eec7c011a2c9d462b89c3dff85d588fd12 --- /dev/null +++ b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/baa830f2-e8bf-491c-a264-552b980530f2_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6063254d70182024acd5a37001085ed2e17cbe9bd61e7e29cbf27fe39a12808a +size 81169 diff --git a/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/baa830f2-e8bf-491c-a264-552b980530f2_origin.pdf b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/baa830f2-e8bf-491c-a264-552b980530f2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..93a2b26ccf9dc72d1a694d3363af4bc667df820e --- /dev/null +++ b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/baa830f2-e8bf-491c-a264-552b980530f2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9324ebca6c3315a49c3634b48e0599b2fe617df79d4f890fa84ae6449b988156 +size 2554954 diff --git a/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/full.md b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6b59e1312e6b25b5b94e73afc875cd627d9f95d4 --- /dev/null +++ b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/full.md @@ -0,0 +1,312 @@ +# Wide2Long: Learning Lens Compression and Perspective Adjustment for Wide-Angle to Telephoto Translation + +Soumyadipta Banerjee + +Jiaul H. Paik + +Debashis Sen + +soumyadiptobnrj071@gmail.com jiaul@ai.iitkgp.ac.in dsen@ece.iitkgp.ac.in + +Indian Institute of Technology Kharagpur, India + +# Abstract + +A translation framework that produces images as if they were captured with a telephoto lens, from images captured with a wide-angle lens, will help in reducing the necessity of complex, expensive and bulky lenses on smartphones. To this end, we propose an image-to-image translation pipeline to simulate the lens compression and perspective adjustment associated with this reconstruction, where the size of the main subject in the images remains the same. We judiciously design depth-based image layering, layer-wise in-painting, redundancy reduction and layer scaling modules to construct the desired tele-photo image, where the pipeline parameters are estimated by a convolutional network. Our approach is compatible with the related optical transformation, and hence, contents behind the main subject are enlarged and that before are diminished, achieving lens compression with appropriate perspective adjustment. + +Our pipeline performs well qualitatively and quantitatively on several source-target image pairs we have captured solely for this task, and also on images in-the-wild. We show that it can simulate the different amounts of lens compression associated with targeted $2 \times$ , $4 \times$ , $8 \times$ changes in the focal length. Further, the pipeline is demonstrated to be effective for a sub-class of the lens-compression problem - portrait perspective distortion correction. We also provide an ablation study to show the significance of the various components in the pipeline. + +# 1. Introduction + +There is a common understanding among professional photographers that different camera lenses capture the same subject in different ways [5]. Camera lenses differ primarily in their focal lengths, which define their field of view [11]. Wide-angle lenses, typically with focal lengths under $35\mathrm{mm}$ , capture broader scenes, making them ideal for landscape photography. On the other hand, telephoto (or long) lenses, with focal lengths usually over $35\mathrm{mm}$ , pro + +![](images/735ab14cc01650e727788b76ac9fc2da8807437aee01b8e03e96bbabf71fecef.jpg) +Figure 1. A brief outline of our Wide2Long (W2L) pipeline. + +vide a narrower field of view and are better suited for portraits or distant subjects. Smartphones, which make photography widely accessible, usually feature only wide-angle lenses with fewer lens elements due to space and cost constraints [12]. Higher-end phones may offer both wide-angle and telephoto lenses for enhanced versatility but are often bulkier and more expensive [1], and physical movement may be required while using them for proper capture on switching lenses. + +As smartphones have become the norm for capturing photographs by the general public, a scope has emerged for investigating systems that allow different kinds of photography on a variety of phones, with or without expert knowledge. In this regard, we focus on producing telephoto shots from images captured using a camera with a wide-angle lens. Such an image-to-image translation can be achieved by simulating lens compression and the associated perspective adjustment. Lens (or depth) compression [2] in photography lingo, refers to the effect due to simultaneous increase in the lens' focal length and it's distance from a subject. The effect is such that the size of the subject at a particular depth remains the same, while the background content at greater depths appears enlarged and the foreground content + +at shallower depths appears diminished, with all of them seem to have drawn closer to the subject. This cannot be simulated through digital zoom, which can merely act as a substitute for optical zoom [17] (where only focal length increases without camera displacement). + +We formulate the implementation of targeted lens compression as a learning-based problem and propose an image translation pipeline in this paper that learns lens compression and perspective adjustment to generate an image that would have been captured using a telephoto lens, from an image captured with a wide-angle lens. Our pipeline can be used to simulate different amounts of lens compression associated with targeted $2^{p} \times$ change in the focal length $(p \in \mathbb{Z}^{+})$ . Our approach contains meticulously designed image layering, layer-wise in-painting, inter-layer redundancy reduction, and layer scaling and flattening modules, whose optimal parameters are estimated through different convolutional networks. While the image layering module decomposes the input image into layers of different depths, layer-wise in-painting merges every layer with those at greater depths and fills the holes remaining at the regions of shallower depths. The inter-layer redundancy reduction module diminishes the redundancy among the in-painting at different layers, and finally, the layer scaling and flattening module scales the contents in the various layers differently and fuses them seamlessly to produce the desired image. We also form a new dataset comprising of image pairs captured using two different focal lengths at different corresponding distances from a subject for our use and for future reference. + +We demonstrate our approach's effectiveness qualitatively and quantitatively using the new dataset. An ablation study of our work and a comparison with naive image-to-image translation are also performed establishing the importance of the proposed components of our pipeline. A related area of work is perspective distortion correction in portraits, where the primary focus is on correcting barrel distortion in a portrait image captured using wide-angle lenses [6]. Even though barrel distortion correction does not usually involve foreground and background scaling, we show that our pipeline can also reduce the distortion in wide-angle portraits as good as the state-of-the-art. Therefore, our main contributions are: + +1. A novel image translation pipeline "Wide2Long" (W2L) to convert wide-angle to telephoto shots by simulating lens compression, using simple mathematical formulations and a convolutional network. +2. A new dataset1 with image pairs of corresponding wide-angle and telephoto shots. + +The paper is organized as follows. Related work and necessary background is provided in Sections 2 and 3. Our proposed image translation pipeline is described in detail in + +Section 4. Section 5 discusses the new dataset, the evaluation of our approach's performance and the importance of its components. Section 6 concludes the paper. + +# 2. Related Work + +To the best of our knowledge, the problem of controlled lens compression and perspective adjustment for wide-angle to telephoto translation has not been explicitly addressed. A related problem is the generation of Dolly-Zoom from a single image, which involves simultaneously zooming out (or in) while moving ("dollying") the camera forward (or backward). This causes the subject to retain its size while everything behind and in front of it gradually expand or compress, dramatically altering perspective [19]. Contrary to lens compression, existing works such as [10] and [15] have tackled the problem of simultaneous zooming out and dollying in, where an image with a larger field of view is generated from an image with a smaller field of view, keeping the size of a subject same. Importantly, these approaches are not designed to have any control on the amount of lens compression on the output image. + +There have, however, been several proposals [6, 21, 24] on correcting perspective (barrel) distortion in wide-angle portrait images, which is simply the lens compression problem applied to face images. Portraits shot with wide-angle lenses suffer commonly from barrel distortion, where facial features close to the lens (forehead, chin, etc.) exhibit exaggerated proportions, whereas features farther away (such as ears) are diminished or occluded, making the face bulge outward in a barrel-like shape [23]. This occurs because wide-angle lenses, owing to their large field of view, need to be brought very close to the subject (in this case, the face) to take a close shot resembling a portrait [8]. + +Traditional solutions to correct wide-angle distortions in portraits include Fried et al.'s method [6], which used a fitted 3D head model to adjust perspective, allowing for angle and distance corrections without multiple images. Zhao et al. [24] later developed a cascaded neural network predicting distortion corrections and in-painting missing features. Their approach, trained on synthetic and real datasets, significantly improved upon traditional 3D model-based methods, especially in handling extreme distortions and facial expressions. More recently, Wang et al. [21] introduced a 3D-GAN inversion approach to correct portrait distortion by adjusting facial geometry and camera parameters, such as distance and focal length, to achieve more flattering images. Unlike traditional 2D warping, this method maintains accurate face geometry and reveals occluded areas. Optimization enhancements, like initializing camera distance and applying geometric constraints, enable this model to outperform prior techniques both quantitatively and qualitatively on portrait datasets. + +# 3. Background and Relevance + +Consider a hypothetical scenario where an image of a bottled beverage is captured using a smartphone with a wide angle lens of $f$ mm, placed at a distance $D$ from the beverage bottle. Settings such as aperture size, ISO, shutter speed, etc. have been adjusted as required. Figure 2(a) shows the image captured for $f = 25$ mm. We see that along with the bottled beverage, there are a lot of background details visible. The ray diagram for this capture is shown in Figure 3(a). + +![](images/ab0d4ca7af22c5dbee14882bcec0fb0d31f4a173489c08b7c81d3ee2a3b21a02.jpg) +(a) + +![](images/44330c1ff7e387a7992c4a578ebc9f53b6433adbdde125f94d6af7807b29128a.jpg) +(b) + +![](images/5830d683d2a3c1adabec319c4a834747669e1e74a606e42423412135e3ea097d.jpg) +(c) + +Now, consider that the goal is to capture the image with a lens of focal length $2f \mathrm{~mm}$ , such that the size of the bottle remains the same. If all the camera settings and the distance $D$ from the bottle are kept unchanged, the entire image will be an optically zoomed-in version of the first image. Figure 2(b) shows the image captured for the $2f = 50 \mathrm{~mm}$ telephoto lens. The corresponding ray diagram is shown in Figure 3(b). To achieve our goal, the camera has to be moved away (backward) to a distance $D' (> D)$ from the bottle retaining all the other camera settings. Figure 2(c) shows such a capture using the $50 \mathrm{~mm}$ long lens. As we see, the bottle retains the original size as in Figure 2(a). But the size and apparent displacement of objects in the background with respect to the bottle have changed significantly illustrating the effect of lens or depth compression. Figure 3(c) shows the related ray diagram. + +We design our image translation pipeline to reconstruct an image captured using a wide-angle lens (at $D$ ) into an estimate of that captured using a telephoto lens (at $D'$ ), emulating the above phenomena, so that we can bypass the need for a longer lens as well as positional adjustments for shooting. As mentioned in Section 1, the image layering module at the beginning of our pipeline (See Figure 1) ensures that the contents at different depths are processed differently, which, from the above discussion, is evident to be a characteristics of the desired translation. Further, the layer scaling module at the end ensures that the object sizes in the output is appropriately scaled, which, as discussed above, is also an attribute of the translation. + +![](images/554b37042248bc5a60dbb02ba9cb6369ddf66b38e4adbea410018c936ba30529.jpg) + +![](images/4917f2cee1173a776da896f803b1caef8fa0cfdc53f60bda3cf63f1ab5b3c3ef.jpg) +Figure 2. The image on the left is captured with a $25\mathrm{mm}$ lens at a distance D from the bottle. The center image is taken with a 50 mm lens from the same distance D. The image on the right is taken with a $50\mathrm{mm}$ lens by moving away to a distance $\mathrm{D}^{\prime}(>\mathrm{D})$ +Figure 3. Ray diagrams for all the three scenarios of Figure 2. In (b), the magnified image $S^{*}S^{\prime *}$ is formed far from the lens. In (c), notice that the rays from the top of $B^{*}B^{\prime *}$ strike the sensor at a higher location than in (a), resulting in an increased size. + +# 4. Wide2Long - The Proposed Approach + +Our image translation pipeline "Wide2Long" (W2L) takes a single RGB image captured with a wide-angle lens as input, and outputs the target image of the same scene using a telephoto lens. The input is sequentially processed through the following four modules: depth-based layering, layer-wise in-painting, inter-layer redundancy reduction, and layer scaling and flattening. + +# 4.1. Depth-based Layering (DBL) + +To emulate lens compression, we first compute the depth at every image pixel in our image translation pipeline. For this, we utilize the well known model MiDaS by Ranftl et al. [13], which has been pre-trained on diverse datasets with scale and shift-invariant losses, and has been demonstrated to produce robust performance on unseen data. + +Given an RGB image $I \in \mathbb{R}^{H \times W \times 3}$ , we obtain its depth map $I_{depth} \in \mathbb{R}^{H \times W}$ , where each pixel's $((i,j))$ depth $d_{i,j} \in [0, d_{\max}]$ represents its relative distance from the camera. We define the set of $N + 1$ $(N \in \mathbb{Z}^+)$ discrete intervals over the range $[0, d_{\max}]$ , denoted as + +$$ +\Delta_ {n} = \left[ \frac {n}{N + 1} d _ {\max }, \frac {n + 1}{N + 1} d _ {\max } \right] \tag {1} +$$ + +where $n = 0,1,\ldots ,N$ . We then create depth-layers $L_{n}$ (as set of pixels) based on the interval to which each pixel depth + +$d_{i,j}$ belongs as follows: + +$$ +L _ {n} = \{(i, j) \mid d _ {i, j} \in \Delta_ {n} \} \tag {2} +$$ + +whose corresponding binary depth-masks are: + +$$ +M _ {n} (i, j) = \left\{ \begin{array}{l l} 1, & \text {i f} d _ {i, j} \in \Delta_ {n} \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +Thus the image region corresponding to a depth layer $L_{n}$ is $I\odot M_n$ , where $\odot$ is color channel-wise Hadamard product. + +# 4.2. Layer-wise Inpainting (LI) + +Breaking the image into $N + 1$ layers leaves out areas of occlusion, shown in blue in Figure 4. In order to fill these occluded regions, an in-painting approach is required. We leverage the pre-trained in-painting model $(\mathcal{I})$ of Suvorov et al. [18], which uses a fast Fourier convolution (FFC) to invoke large receptive fields early in their network, allowing it to consider the global image context adequately in all the layers. + +![](images/d15cd75a60f36198cf48d307cfc234e61fff86c1ea0ebbe768d1ca2b7fcc3683.jpg) +Figure 4. Image channel decomposed into depth-layers, with $N = 9$ taken for visual clarity. As we can see, there are occluded regions (blue) left out by parts of objects in shallower depth-layers. + +However, we simply cannot supply a layer's occlusion region for in-painting. The issue with this approach is shown in Figure 5, where availability of very little visual context for estimation has resulted in an incongruous inpainted pattern. So, we define a composite context layer $C_n$ for each $n$ as follows: + +$$ +C _ {n} = \bigcup_ {k = 0} ^ {n} L _ {k} \tag {4} +$$ + +The corresponding context $\zeta_{n}$ and in-painting $\mu_{n}$ masks are obtained from the depth masks in Equation (3) as follows: + +$$ +\zeta_ {n} (i, j) = \sum_ {k = 0} ^ {n} M _ {k} (i, j), \mu_ {n} (i, j) = \sum_ {k = n + 1} ^ {N + 1} M _ {k} (i, j) \tag {5} +$$ + +![](images/215c756fa2e05fc1def4cd6bca3cf6503b8d52b9bfcf0b1bc5295d67ed02b96b.jpg) +Figure 5. As shown in a depth layer region (left) from an image channel, there is very little information to draw context for inpainting in the area circled out near the occluded area. Therefore, the in-painting (right) creates undesirable structures and patterns. + +We have $\mu_n(i,j) \in 0,1$ as $M_k(i,j) = 1$ for exactly one value of $k$ . This mask $\mu_n$ is then dilated by a factor $\lambda \in \mathbb{Z}^+$ to get: + +$$ +\mu_ {n} ^ {\prime} = \operatorname {D i l} _ {\lambda} (\mu_ {n}) \tag {6} +$$ + +where $\mathrm{Dil}_{\lambda}$ being the morphological binary dilation operator [7]. Dilation helps in deriving context from the surroundings for the in-painting. We find empirically that $\lambda$ depends on the input image $I_{S}$ . The suitable value of $\lambda$ is thus provided by a trained convolutional neural network that takes as $I$ and $I_{depth}$ inputs. This convolutional network, which we call Image2Params (I2P), outputs the values of 4 quantities $\lambda$ , $\gamma$ , $\alpha$ and $T$ . The remaining three quantities are explained in the following sub-sections. The details of the I2P network and its training procedure is discussed in Section 4.5. + +We apply the in-painting model $\mathcal{I}$ for every $n$ to produce the corresponding in-painted image as follows: + +$$ +\iota_ {n} = \mathcal {I} \left(I, \zeta_ {n}, \mu_ {n} ^ {\prime}\right) \tag {7} +$$ + +where $I\odot \mu_n^\prime$ is filled using $I\odot \zeta_{n}$ . The corresponding overall in-painted layer is obtained as: + +$$ +\tilde {C} _ {n} = \bigcup_ {k = n} ^ {N + 1} L _ {k} \tag {8} +$$ + +whose corresponding binary overlay mask is $O_{n} = \mu_{n}(i,j) + M_{n}(i,j)$ and the relevant in-painted image region is $\iota_{n} \odot O_{n}$ . + +# 4.3. Inter-layer Redundancy Reduction (ILRR) + +An unwanted effect in the above stage is that in-painted pixels in one layer might repeat themselves around the same location in regions of adjacent layers, as shown in Figure 6. This is a consequence of combining $0^{\mathrm{th}}$ to $(n - 1)^{\mathrm{th}}$ depth-layers with the $n^{\mathrm{th}}$ depth-layer to form context layer prior to in-painting (Equation 4). This may sometimes lead to redundant context among multiple layers, which may register as a significant distortion, after scaling and flattening. + +![](images/b52e8a42a36ed26049a6cf354e73300b8c45318cf9d80df55480bc7e0a9e6217.jpg) +Figure 6. Redundant visual information around the same location across several successive depth-regions due to in-painting (right) shows up as a distortion on the final output image (left). + +To handle this issue we compute a difference map $D_{n},\forall n$ as follows: + +$$ +D _ {n} (i, j) = \left\{ \begin{array}{l l} 1, & \text {i f} s i m _ {i, j} < T \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {9} +$$ + +where $sim_{i,j}$ is the grayscale similarity score computed between $\iota_{n-1}^{O}(i,j)$ and $\iota_{n}^{O}(i,j)$ using the Structural Similarity Index Measure (SSIM) [20], and $\iota_{n}^{O} = \iota_{n} \odot O_{n}$ . We use a threshold $T \in [0,1]$ and a steepness parameter $\alpha \in \mathbb{R}^{+}$ , both provided by our $I2P$ network. A sigmoidal transformation, using $\alpha$ , is applied to make a smooth transition between masked and un-masked pixels as shown below: + +$$ +D _ {n} ^ {\prime} (i, j) = \left\{ \begin{array}{l} 0. 5, \text {s i m} _ {i, j} < T \\ \frac {1}{1 + e ^ {- \alpha (T - \operatorname {s i m} _ {i , j})}}, \text {o t h e r w i s e} \end{array} \right. \tag {10} +$$ + +$D_{n}^{\prime}$ is then used to diminish the pixels values in $\iota_{n}^{O}$ which are similar to that in $\iota_{n - 1}^{O}$ , so that they do not create distortion of the kind shown in Figure 6. Therefore, we get: + +$$ +I _ {n} ^ {\prime} = \iota_ {n} ^ {O} \odot D _ {n} ^ {\prime} \tag {11} +$$ + +where $I_{n}^{\prime}$ is subsequently used for layer scaling. + +# 4.4. Layer Scaling and Flattening (LSF) + +Now that we have all the in-painted regions $I_{n}^{\prime}$ for the $(N + 1)$ layers without significant redundancies, we need to scale them to appropriate size and combine them to flatten out a single output image. Note that in lens compression, the objects or object parts at a particular distance from the camera retain the same size in the images captured using both the wide-angle and telephoto lenses. + +Suppose we have a pixel location $(x, y)$ provided by the user inside the region of such an object (or part of it) in the + +input image $I$ (captured using a wide-angle lens). Then $L_{n}$ for a particular $n = n_f$ is the depth-layer in $I$ , where $(x,y)$ belongs (see Equation 2). We then calculate a scaling factor for every image pixel $(i,j)$ as follows: + +$$ +\rho (i, j) = e ^ {\gamma \cdot \left(n _ {f} - n _ {(i, j)}\right)} \tag {12} +$$ + +where $n_{(i,j)}$ is such that the pixel $(i,j)$ lies in the depth layer $L_{n_{(i,j)}}$ , and therefore $L_{n_f} = L_{n_{(x,y)}}$ . This equation primarily models the lens compression phenomenon in our W2L pipeline. Contents in layers at greater depths than that of $L_{n_f}$ will be magnified $\forall n < n_f$ and those in a layer at shallower depth than that of will be diminished in size $\forall n > n_f$ . This is in accordance with the optical ray diagrams we discussed in Section 3. The $\gamma (\in \mathbb{R}^{+})$ here is a learnable data driven parameter, supplied by our $I2P$ network. As there are $N + 1$ number of depth layers, we will have $N + 1$ number of different $\rho (i,j)$ values, which we represent as $\rho_n,n = 0,1,\ldots ,N$ . + +The $I_{n}^{\prime}, n = 0,1,\dots ,N$ regions from Equation (11) are then scaled (resized through interpolation/decimation), using respective $\rho_{n}$ and superimposed in that order with the center as pivot. The resultant is cropped centrally to produce the desired output image $I_{out}$ with same dimensions as $I$ . + +# 4.5. Convolutional Network 'Image2Param' (I2P) for Parameter Estimation + +To estimate the parameters $\gamma$ , $\lambda$ , $T$ , and $\alpha$ of our Wide2Long approach described in Sections 4.1 - 4.4, we use a convolutional neural network trained on our dataset explained later in Section 5.1. Each image-ground truth pair in its training set is associated with target values for the above parameters. The target values are determined by searching within a set range of values for the parameters to maximize the SSIM between the output and target images. + +The Image2Param (I2P) network processes a 4-channel input image (comprising RGB and depth-map) of size $4 \times 512 \times 512$ . The input is fed through four convolutional layers: the first has 32 filters, followed by layers with 64, 128, and 256 filters respectively, all using $3 \times 3$ kernels and ReLU activations, each followed by a max-pooling operation. The feature maps are then flattened into a 1D vector, which is passed through two fully connected layers with 512 and 128 units respectively, both using ReLU activations. The final output layer consists of a single unit with ReLU activation, used to predict a parameter's value. + +For the four parameters, four separate $I2P$ instances are trained using the Adam optimizer [9], with the hyperparameters listed in Table 1. Note that we assign a value to $N$ in Section 4.1 and do not make it data-driven. Our approach can learn the data-driven $\gamma$ differently in Equation (12) for different values of $N$ . Since the capability of learning $\gamma$ may depend on the particular $N$ value assigned, + +we find $(N + 1) = 20$ suitable empirically in terms of performance. As shown in the supplementary, this also means we can conveniently vary $(N + 1)$ during inference as per our choice and the nature of the data, while scaling $\gamma$ by $(1 / \tau)$ . + +
ParameterEpochsBatch SizeLearning Rate
γ5050.001
λ50100.001
T4050.0001
α2550.001
+ +# 5. Experiments and Results + +# 5.1. Dataset Acquisition + +We create a source-target image pair dataset for experiments using both real-world (camera-based) and synthetic (ray-tracing) methods. Each source image is captured with a focal length $f_{s} (= 25 \, \mathrm{mm})$ lens at distance $D$ , and each target image with a focal length $f_{t} = 2^{p} f_{s} (= 50 \, \mathrm{mm}, p = 1)$ at distance $D' > D$ to maintain the subject's size. For any source-target pair, we keep the camera and light settings constant. Figure 7 shows a few samples from our dataset. We include 500 pairs for training and 100 for testing, with each $512 \times 512$ -pixel image accompanied by $(x, y)$ coordinates for estimating $n_{f}$ in Equation (12). Note that the source or target images in the train and test sets cannot be modeled to be from the same data distribution, as no correspondence has been enforced on their contents. As our pipeline is focal length-agnostic (does not know about the absolute focal length of the input), use of our dataset enables our W2L pipeline to simulate lens compression for a generic $2 \times$ change in focal length, which can be used iteratively for the $2^{p} \times$ (integer $p > 1$ ) cases. Therefore, we also capture additional image pairs with $p \geq 1$ for testing. + +![](images/ffd472219c6968f15c822e8d07cc25aad614f34377a2a286558b70ea1f1261a9.jpg) +Figure 7. A couple of training pairs from our acquired dataset of $f_{t} = 2f_{s}$ lens compression. The left shows a real-world image pair, while the one on the right is made synthetically. + +![](images/18e088a24afa87e37e753dfba1a5ea94e19988432df743a68cd124da6c6a60fe.jpg) + +# 5.2. Lens Compression on Our Dataset + +For evaluating the performance of our W2L pipeline, and in turn that of our I2P network, we use the test set consisting + +of 100 image pairs. For each image pair in the set, we use the I2P network to predict the corresponding values of $\gamma$ , $\lambda$ , $T$ , and $\alpha$ , and generate the output image. This output image is compared to the target image in terms of image reproducibility metrics, such as PSNR, SSIM and LPIPS [22]. Table 2 gives the average value of each of the three metrics calculated on the test set. + +As a naive off-the-shelf baseline, we train a U-Net [14] encoder-decoder network for a total of 800 epochs on our dataset with a batch size of 10 using the Adam optimizer and an SSIM-based loss. We use an initial learning rate of 0.001, which is decreased by a factor of 0.1 at the $200^{th}$ and $400^{th}$ epochs. Table 2 shows that the U-net is highly ineffective for this task. + +In Section 2, we saw that Dolly-Zoom is the area of work closest to lens compression. So, we turn to the recent efforts of Shih et al. (which, to the best of our knowledge, is the state-of-the-art for Dolly-Zoom generation from 2D photos) for an indirect comparison, as there is no prior work to explicitly address lens compression. Frames generated from Dolly-Zoom sequences are matched with the corresponding target images and the ones having the highest SSIM value are considered, being essentially a brute-force approach to solving targeted lens compression, unlike ours. + +Table 1. Hyper-parameter values for training ${I2P}$ instances + +
MetricU-Net [14]Shih [15]W2L
SSIM (↑)0.1830.7350.768
PSNR (↑)6.2917.68219.873
LPIPS (↓)0.7150.1470.139
+ +Table 2. Quantitative results on our W2L test dataset. + +Figure 8 illustrates performances on a few examples from our dataset. We see that the outputs of [15] possess artifacts which we explicitly address in Section 4.3. Further, there are no direct means to control the extent of lens compression, unlike in our approach. These comparisons along with the superiority observed in Table 2 attest the potential of Wide2Long in effectively simulating lens compression, and in avoiding artifacts that state-of-the-art Dolly-Zoom techniques may suffer from. + +For lack of space, we show results on higher orders of lens compression (associated with $4 \times$ and $8 \times$ change in focal length) using our proposed pipeline in the supplementary (Figure 4). + +# 5.3. Perspective Distortion Correction in Portraits + +Additionally, the W2L pipeline is capable of correcting perspective distortion in wide-angle portraits by simulating lens compression, where increasing focal length and moving the camera back preserves the size of key facial features (e.g., the nose) while enlarging background features like ears and the jawline, thereby reducing distortion. To evaluate, we use the Caltech Multi-Distance Portraits (CMDP) + +![](images/75d815e55e7d9177c18f4db68d531a4f7ce4df95914dc351a282ae8284ddca56.jpg) +Figure 8. Output samples from our Wide2Long pipeline, compared against the corresponding target samples. We also show the corresponding outputs of Shih et al. [15] (third row), which generates Dolly-Zoom sequences from 2D photos. This is an indirect comparison with our results (last row), as no prior work has explicitly addressed targeted lens compression. We see the results of [15] suffer from significant artifacts around object edges, which our pipeline explicitly addresses to reasonably diminish them (see Section 4.3). + +dataset [4] with 53 subjects photographed at seven distances using a $28 - 300\mathrm{mm}$ lens. We select 41 samples without artifacts, pairing two portraits for each sample such that $f_{t} = 2f_{s}$ . We ensure proper translation by manually aligning nose ridges and note tips (features not undergoing lens compression) between source and target images. To refine target approximations, we retrain our I2P model instances on 31 CMDP samples following this focal-length relationship. Figure 9 shows W2L outputs, and Table 3 provides the average SSIM, PSNR, and LPIPS values, alongside the metrics for other approaches provided by Wang et al. [21] for comparison3. We see that W2L produces results at par with the state-of-the-art. + +
MetricFried [6]Shih [15]Wang [21]W2L
SSIM (↑)0.7240.6960.7470.716
PSNR (↑)15.4113.0817.5220.38
LPIPS (↓)0.1880.2680.1670.117
+ +Table 3. Quantitative results from the Wide2Long pipeline for the CMDP dataset. + +# 5.4. Results on Images In-the-Wild + +We apply our "Wide2Long" pipeline to several images collected from the Internet. We give the W2L outputs for some of these images in Figure 10. To maintain uniformity, we simply choose the object/object-feature that is closest to the camera to be the subject (whose scale is unchanged). + +# 5.5. Ablation Study + +We conduct ablation studies on W2L, both on the CMDP dataset and ours. As seen from Table 4, the in-painting stage's absence affects the pipeline's performance the most. We also ablate the dilation parameter $\lambda$ , and the redundancy removal parameters $\alpha$ and $T$ . We find that the network generalizes poorly as shown by the results of Table 4. The dilation parameter and inter-layer redundancy reduction have roughly about the same performance effect even in the case when they are removed together. Overall, they affect the performance albeit by large margin, but not as large as the removal of in-painting. Figure 11 shows examples from our dataset under the above-mentioned ablation conditions. + +Table 5 shows that a variant of I2P, modified to output all 4 parameters at once, trained with a weighted loss function (weights being the inverse of the ranges of values taken by each parameter in train-set) till convergence performs + +![](images/3db4ccc9e1e9abfb0e376125a028bb45e441f77b669e5683eb9d36e816f44c90.jpg) +Figure 9. Output samples from the Wide2Long pipeline for the CMDP samples. To ensure proper translation from wide to long, all image pairs have nose ridges and tips aligned between source and target. + +![](images/6831bca07b3e239be3a18e494e876a92f540a8f610cfad7fd208a5226e708a23.jpg) +Figure 10. Results from the Wide2Long pipeline for In-the-Wild images performing lens compression for $2 \times$ focal length increase. + +
Ablation TypeSSIM (↑), PSNR (↑)
CMDPOurs
w/o LI (4.2)0.315, 7.290.463, 9.28
w/o λ (4.2)0.624, 16.210.671, 16.47
w/o ILRR (4.3) (α, T)0.492, 15.020.680, 17.58
w/o λ, ILRR (4.2, 4.3)0.527, 15.730.664, 15.82
Entire W2L pipeline0.716, 20.380.768, 19.87
+ +poorly on the W2L dataset. Also, using a larger network (e.g. VGG-16 [16]) modified to output 1 parameter (per + +![](images/75b6834140eb2630ce232d63de09d8088eabc45d5477a9fcd9578e53843d2aec.jpg) +Figure 11. Qualitative examples of W2L's ablation study on our dataset showing the significance of the various components. + +model instance) or 4 parameters given the 4-channel input, does not fare as well as the I2P model given in the paper. + +Table 4. Ablation study of our W2L on CMDP and our dataset. + +
NetworksSSIM (↑)PSNR (↑)LPIPS (↓)
I2P Net, 4-output0.48313.5290.371
VGG-16, 4-output0.71417.3510.164
VGG-16, 1-output0.64916.2740.296
+ +Table 5. Comparison with other I2P and VGG-16 variants. + +# 6. Conclusion + +The proposed Wide2Long (W2L) translation pipeline simulates long-lens effects on wide-angle images, creating lens compression and associated perspective adjustment to enhance smartphone photography without extra lenses. It can thus increase the versatility of wide-angle cameras to be able to implement various types of shots and visual effects such as forced-perspective [3]. + +The pipeline involves four stages: depth-based layering, layered in-painting, redundancy removal, and layer scaling attempting to mimic long-lens qualities. A wide range of experiments for qualitative and quantitative performance evaluation demonstrate the effectiveness of the proposed approach in performing different amounts of lens compression. Our pipeline also successfully works for the allied problem of perspective distortion correction in portraits. Further, unlike existing works such as [10] and [15] on novel-view rendering for different purposes (say, DollyZooming from 2D photos), our pipeline does not rely on estimating virtual camera projections and movements, making it easier to implement. Ablation study on our multicomponent pipeline reveals that layer in-painting plays the most critical role in the model. + +However, Wide2Long may face some challenges. As our pipeline intends to mimic both the increase in focal length and camera-subject distance, it may fail when the movement of the camera backwards needs to expose new objects or object parts in the line of the lens with increased focal length, which are originally not in the source. This is seen in the last column of Figure 8, where W2L's estimate of the target does not show the portion of the oven in the frame accurately, as it draws information only from the input. Our analysis shows that even within such a limitation, W2L performs reasonably well both visually and in terms of reproducibility metrics. + +# References + +[1] Ziv Attar. Smartphone camera design basics: Optics and form factor. Glass Imaging Journal, https://www.glass-imaging.com/journal/smartphone-camera-design-basics-optics-and-form-factor, 2023.1 +[2] Martin Banks, Emily Cooper, and Elise Piazza. Camera focal length and the perception of pictures. Ecological psychology: a publication of the International Society for Ecological Psychology, 26:30-46, 2014. 1 +[3] Keith M. Booker. Historical Dictionary of American Cinema. Scarecrow Press, 2011. 8 +[4] Xavier P. Burgos-Artizzu, Matteo Ruggero Ronchi, and Pietro Perona. Distance estimation of an unknown person from a portrait. In Computer Vision - ECCV 2014, pages 313-327, Cham, 2014. Springer International Publishing. 7 +[5] Michael Freeman. The Photographer's Eye: Composition and Design for Better Digital Photos. Focal Press, Burlington, MA, 2007. 1 +[6] Ohad Fried, Eli Shechtman, Dan B Goldman, and Adam Finkelstein. Perspective-aware manipulation of portrait photos. ACM Transactions on Graphics (Proc. SIGGRAPH), 2016. 2, 7 +[7] Rafael C. Gonzalez and Richard E. Woods. Digital Image Processing. Prentice Hall, 3rd edition, 2008. 4 +[8] R. Johnson. Lens Design Fundamentals, Second Edition. 2009. 2 +[9] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learning Representations, 2014. 5 +[10] Yangwen Liang, Rohit Ranade, Shuangquan Wang, Dongwoon Bai, and Jungwon Lee. The "vertigo effect" on your smartphone: Dolly zoom via single shot view synthesis. pages 1407-1415, 2020. 2, 8 +[11] Sidney Lumet. Making Movies. Vintage, New York, 1995. 1 +[12] The Smartphone Photographer. Understanding focal length on smartphone cameras. https : / / thesmartphonephotographer . com / smartphone - camera - focal - length/, 2020. 1 +[13] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer, 2020. 3 +[14] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241. Springer, 2015. 6 +[15] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6, 7, 8 +[16] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 8 + +[17] Warren J. Smith. Modern Lens Design: A Resource Manual. McGraw-Hill Professional, 2nd edition, 2004. 2 +[18] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.4 +[19] Timotheus Vermeulen. Flat film: Strategies of depthlessness in pleasantville and la haine. Film-Philosophy, 22(2):168–183, 2018. 2 +[20] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5 +[21] Zhixiang Wang, Yu-Lun Liu, Jia-Bin Huang, Shin'ichi Satoh, Sizhuo Ma, Guru Krishnan, and Jian Wang. disco: Portrait distortion correction with perspective-aware 3d gans. International Journal of Computer Vision, 2024. 2, 7 +[22] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[23] Z. Zhang. A flexible new technique for camera calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(11):1330-1334, 2000. 2 +[24] Yajie Zhao, Zeng Huang, Tianye Li, Weikai Chen, Chloe Legendre, Xinglei Ren, Ari Shapiro, and Hao Li. Learning perspective undistortion of portraits. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7848-7858, 2019. 2, 7 \ No newline at end of file diff --git a/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/images.zip b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..46688447688d46ecc3864aa9b973ca170adaa040 --- /dev/null +++ b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f5c58b6cd0c4b9516edf710aacb6742d224fbaaf1745c3992185735af862b29 +size 551817 diff --git a/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/layout.json b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a858a3b76cdc5aa700e5c4493ece73b5841b3732 --- /dev/null +++ b/wide2longlearninglenscompressionandperspectiveadjustmentforwideangletotelephototranslation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:accd2e1188236bf5da8bacb6b584ebd2929b8a9b2bcb538cc76479e434846a07 +size 398282 diff --git a/wikiautogentowardsmultimodalwikipediastylearticlegeneration/707bfef2-e682-4352-8e81-2af4733b9fe7_content_list.json b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/707bfef2-e682-4352-8e81-2af4733b9fe7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e5e4f716e088a83359553c01d78c30cda27c9dfb --- /dev/null +++ b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/707bfef2-e682-4352-8e81-2af4733b9fe7_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1491d82deacc1d3547c5125883935568fd38747f4dcde3332d9c1a32d3a9508f +size 78693 diff --git a/wikiautogentowardsmultimodalwikipediastylearticlegeneration/707bfef2-e682-4352-8e81-2af4733b9fe7_model.json b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/707bfef2-e682-4352-8e81-2af4733b9fe7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ed648b08149b5530e2f23ab4984db560d8dfd541 --- /dev/null +++ b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/707bfef2-e682-4352-8e81-2af4733b9fe7_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1e6fcfbad213108ec5b7b08a10850ad6d22a28327de794643047092880b44e4 +size 96579 diff --git a/wikiautogentowardsmultimodalwikipediastylearticlegeneration/707bfef2-e682-4352-8e81-2af4733b9fe7_origin.pdf b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/707bfef2-e682-4352-8e81-2af4733b9fe7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..761b648c7cf2b31b6c93a6aeea63da6a950bda39 --- /dev/null +++ b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/707bfef2-e682-4352-8e81-2af4733b9fe7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74ec17a1e695169c05a68068da285694ea22c65223459a6f67940d798ff91912 +size 1865254 diff --git a/wikiautogentowardsmultimodalwikipediastylearticlegeneration/full.md b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..02693a53bdba15379f2cc3693baf3cbeb06da49f --- /dev/null +++ b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/full.md @@ -0,0 +1,271 @@ +# WikiAutoGen: Towards Multi-Modal Wikipedia-Style Article Generation + +Zhongyu Yang\(^{1,2*}\), Jun Chen\(^{3,1*}\), Dannong Xu\(^{1,4}\), Junjie Fei\(^{1}\) +Xiaogian Shen\(^{1}\), Liangbing Zhao\(^{1}\), Chun-Mei Feng\(^{5}\), Mohamed Elhoseiny\(^{1}\) +\(^{1}\)King Abdullah University of Science and Technology, Lanzhou University \(^{3}\)Meta AI \(^{4}\)The University of Sydney \(^{5}\)IHPC, A\*STAR \(\left\{\text { zhongyu.yang, jun.chen, junjie fei, xiaogian.shen liangbing.zhao, mohamed.elhoseiny } \right\} @kaust.edu.sa daxu8019@uni.sydney.edu.au, fengcm.ai@gmail.com + +![](images/02c3cd499a1ec15b4f9dd57d653bdb8b923c7b7475b181f6534b62538f88fb3e.jpg) +Figure 1. Comparison of existing text-only article generation methods and our proposed WikiAutoGen. Existing approaches [14, 28] rely exclusively on textual sources, often producing inconsistent or inaccurate results. For example, in (a), the target topic is 'Benzoxonium Chloride', yet the baseline incorrectly generates information about 'Benzalkonium Chloride'. In contrast, our WikiAutoGen framework integrates both visual and textual modalities to generate coherent multimodal content. Additionally, WikiAutoGen employs a multiperspective self-reflection mechanism, significantly improving content accuracy and reliability, as illustrated in (b). + +# Abstract + +Knowledge discovery and collection are intelligence-intensive tasks that traditionally require significant human effort to ensure high-quality outputs. Recent research has explored multi-agent frameworks for automating Wikipedia-style article generation by retrieving and synthesizing information from the internet. However, these methods primarily focus on text-only generation, overlooking the importance of multimodal content in enhancing informativeness and engagement. In this work, we introduce WikiAutoGen, a novel system for automated multimodal Wikipedia-style article generation. Unlike prior approaches, WikiAutoGen retrieves and integrates relevant images alongside text, enriching both the depth and visual appeal of generated content. To further improve fac + +tual accuracy and comprehensiveness, we propose a multiperspective self-reflection mechanism, which critically assesses retrieved content from diverse viewpoints to enhance reliability, breadth, and coherence, etc. Additionally, we introduce WikiSeek, a benchmark comprising Wikipedia articles with topics paired with both textual and image-based representations, designed to evaluate multimodal knowledge generation on more challenging topics. Experimental results show that WikiAutoGen outperforms previous methods by $8\% -29\%$ on our WikiSeek benchmark, producing more accurate, coherent, and visually enriched Wikipedia-style articles. Our code and examples are available at https://wikiautogen.github.io/ + +# 1. Introduction + +Knowledge discovery and content generation are essential for organizing and disseminating information, but they re + +main time-consuming and intelligence-intensive, requiring substantial human effort to collect, structure, and verify information. With the advent of large-scale AI models like large language models (LLMs) [2, 5, 8, 22, 35], there is growing potential to automate knowledge collection, synthesis, and organization in a more efficient and scalable manner [14, 28]. Such automation not only accelerates knowledge discovery but also enhances accessibility, making information more readily available and up to date. + +Recently, several methods, such as Storm [28] and CoStorm [14], have been proposed to automate Wikipedia-style article generation. While they can produce Wikipedia-like content, they still face key limitations: (1) they are limited to text-only generation and lack the ability to incorporate multimodal content such as relevant images; (2) the generated articles often lack breadth, depth, and reliability, reducing their overall informativeness and credibility. + +To address these challenges, we introduce WikiAutoGen, a multi-agent framework designed to generate high-quality, multimodal Wikipedia-style articles automatically. Unlike prior works, WikiAutoGen can directly search both textual and visual information and generate multimodal content (see Figure 1), enriching article content with relevant and diverse modalities. Additionally, we propose a novel multi-perspective self-reflection module, which enables the system to self-regulate, refine, and critically evaluate its generated content. This mechanism enhances the reliability, depth, and breadth of the information by encouraging iterative improvement and multi-source validation. + +To advance the development and evaluation of multimodal knowledge generation, we introduce WikiSeek, a new benchmark designed to tackle challenging topics comprising both textual and visual components. Existing benchmarks primarily focus on text generation or cover only straightforward topics (see Table 1). In contrast, WikiSeek is multimodal and specifically targets more complex subjects with limited coverage on Wikipedia, making it significantly harder for current methods to retrieve and synthesize comprehensive information. This increases the challenge of content generation, pushing models to explore deeper, enhance their retrieval capabilities, and improve their ability to handle underexplored subjects. + +Extensive experiments demonstrate that WikiAutoGen significantly outperforms existing methods in generating high-quality textual and visual content. We evaluated text quality across nine key dimensions and image quality across four essential criteria. Experimental results show that WikiAutoGen outperforms prior methods by $8\% - 29\%$ in textual quality and $11\% - 14\%$ in image quality, demonstrating consistent gains across the input topics of text-only, image-only, and image-text tasks. + +Our contributions can be summarized as follows: + +- We introduce WikiSeek, a new multimodal benchmark + +
DATASETSDataset Statistics
TypeRetrieval ModalityDifficulty Levels
Surfer100 [18]国国easy
FreshWiki [28]国国easy
IRP [3]国国n/a
WildSeek [14]国国n/a
WikiSeek (Ours)国国high
+ +Table 1. Comparison of WikiSeek with existing benchmarks. Modalities are indicated by (text) and (images). Difficulty levels are categorized based on the average number of characters in corresponding Wikipedia pages: difficult (<500 characters), medium (500-2000 characters), and easy (>2000 characters). + +designed for evaluating Wikipedia-style article generation, featuring challenging topics with limited existing coverage, represented through both text and images. + +- We propose WikiAutoGen, a multimodal article generation framework that synthesizes comprehensive content by effectively integrating textual and visual inputs. On WikiSeek, WikiAutoGen achieves $8\% - 29\%$ improvements over the previous best model in textual generation. +- We develop a novel multi-perspective self-reflection module, which iteratively enhances readability, informativeness, and coherence by incorporating feedback from diverse roles, including reader, writer, and editor. + +# 2. Related Work + +Automatic Expository Writing. LLMs have shown strong performance in automatic expository writing, particularly in generating Wikipedia-style articles [4, 17, 21, 33, 37, 41]. Despite their strength in traditional Natural Language Generation (NLG), LLMs still struggle to produce long-form text that is coherent and logically structured [10, 26, 29, 34, 36]. To address this, [32] proposed using domain-specific keywords that are progressively refined into full passages through multi-stage generation. Notably, Shao et al. [14, 28] highlighted the crucial role of pre-writing strategies, identifying them as a key factor in improving article quality. Recently, some automatic writing systems expanded their knowledge boundaries through mindmaps and tree-based methods [3, 19, 29, 40]. However, while these iterative content planning methods effectively leverage widely accessible information for common topics, they remain largely ineffective in less-explored or niche domains due to the scarcity of structured prior knowledge [11, 42]. Meanwhile, existing methods are limited to generating text-based articles and fail to incorporate other modalities, such as visual elements. This inherent constraint in data input inevitably leads to incomplete information and reduced readability, making knowledge acquisition more challenging. Conversely, our WikiAutoGen is the first multimodal writing system that integrates visual content and retrieves knowledge across multiple modalities, allowing visual in + +formation to complement textual content by capturing details that may otherwise be overlooked. + +Self-reflection. Recent progress in optimizing LLMs through self-reflection mechanisms is significant. The core idea is to enable models to analyze and refine their outputs through self-generated feedback. Existing approaches construct feedback sources from three strategies: (1) The LLM conducts iterative self-evaluation [30]; (2) A separately trained critic module provides specialized feedback [12]; (3) External knowledge from sources like Wikipedia and browsers is integrated [1, 20]. Specifically, REFINER [24] demonstrated that a trained critic module can enhance reasoning without fine-tuning the reasoning module, supporting feedback mechanism optimization. Further, some methods [7, 45] introduced feedback mechanisms based on error-type templates and context-hypothesis mappings. Recent studies [25, 31, 38, 39, 44] focus on LLMs' in-context learning. They design prompt templates to help models generate feedback from historical outputs or patterns. However, the multi-perspective self-reflection method proposed by [43], which relies on a Navigator-Reasoner heuristic interaction, is limited to closed-loop LLM discussions without external knowledge acquisition, resulting in restricted information richness and verifiability. In contrast, our approach enhances multi-perspective self-reflection with multi-web search, addressing complex topic exploration and retrieval challenges while enabling multi-dimensional control over topic and article quality. + +# 3. WikiSeek benchmark + +# 3.1. Task Definition + +Given a topic as text $(T)$ , image $(I)$ , or a combination of both $(T, I)$ , the objective is to generate a Wikipedia-style article $(A)$ that integrates relevant knowledge and is supported by verifiable references $(R)$ . This task is particularly important in domains such as investigative journalism, scientific research, and market analysis, where the generation of accurate, well-sourced content is essential. + +# 3.2. WikiSeek Construction + +A key challenge in this task is the lack of a suitable benchmark that effectively encompasses multimodal topics. Existing benchmarks, however, remain largely text-centric, thus failing to adequately reflect the complexity of multimodal content generation. To address this limitation, we introduce WikiSeek, a new benchmark specifically designed to evaluate more challenging topics that incorporate both text and images. WikiSeek establishes a robust evaluation framework, enabling a more comprehensive and reliable assessment of multimodal knowledge generation in practice. In the following sections, we detail the construction process of this benchmark. + +Benchmark construction pipeline. Our WikiSeek benchmark is designed with two key objectives: (1) to evaluate multimodal article generation where both the input and output include text and images; and (2) to target underexplored topics on Wikipedia that present greater challenges for retrieval and synthesis. + +We select topics from the WikiWeb2M dataset [6], which comprises approximately 2 million English Wikipedia articles containing both text and images. To identify challenging topics, we focus on articles where the main content has fewer than 500, 300, and 100 characters, categorizing them as hard, very hard, and extremely hard, respectively. Typically, these more challenging or rare topics have minimal coverage on Wikipedia, representing less-documented and lesser-known subjects. + +Topic filtering and quality control. To curate a high-quality benchmark that includes both challenging and meaningful topics, we implement a rigorous topic filtering process. First, we retain only Wikipedia articles with fewer than 500 characters, ensuring a focus on underexplored and more difficult topics. We then sample hundreds of topics associated with images from the WikiWeb2M dataset [6]. + +However, some topics, such as “1997 in Japan” and “.bh”, are either overly general or semantically underspecified, making them unsuitable for evaluation. To address this, we manually verify all selected topics and remove those that lack meaningful content or pose evaluation challenges. After this filtering process, we obtain a final set of 300 topics, evenly distributed across three difficulty levels (hard, very hard, and extremely hard), with 100 topics per level. These topics are represented in one of three formats: text-only, image-only, or a combination of both. + +# 4. Method + +Writing high-quality multimodal Wikipedia-style articles usually requires a coordinated multi-agent system that effectively breaks down the process into distinct stages, including outline generation, web-based material retrieval, and article synthesis. Beyond these stages, maintaining quality necessitates collaboration across roles, ensuring the article is well-structured, accurate, and engaging. To address these challenges, we propose WikiAutoGen framework. This framework facilitates structured collaboration among specialized agents, enabling comprehensive topic exploration, multi-modal content generation, and coherent generation. In the following sections, we provide a detailed breakdown of WikiAutoGen and its core components. + +# 4.1. WikiAutoGen Framework + +Our WikiAutoGen framework consists of several key components that work collaboratively to generate high-quality multimodal Wikipedia-style articles, as illustrated in Figure 2. Each module plays a distinct role in the article genera- + +![](images/a9c73be7027b0bc8e10665cc3a6dfdd8779dff75b5111708c5ba987e6b4a590d.jpg) +Figure 2. Overview of WikiAutoGen, our multimodal framework for Wikipedia-style article generation. The pipeline includes three main stages: (1) an Outline Proposal module that structures the article outline based on the multimodal topic input (image and text); (2) a Textual Article Writing module involving persona generation, multi-agent collaborative exploration, and article drafting; and (3) a Multimodal Article Writing module that incorporates relevant images through positioning proposals, retrieval, selection, and final polishing. The entire generation process is enhanced by a Multi-Perspective Self-Reflection module, leveraging supervisory and agent-specific feedback (writer, reader, editor) to iteratively improve article quality in terms of coherence, readability, and engagement, etc. + +tion process: 1). Outline Proposal Module. This module converts the given text and image topics into structured outline proposals, laying the foundation for content organization. 2). Textual Article Writing Module. This stage involves multiple sub-components, including a persona generator, a multi-agent discussion system, and the article generation process, ensuring the content is well-structured and contextually rich. 3). Multi-Perspective Self-reflection Module. This component evaluates the generated text from multiple viewpoints, including those of a writer, reader, and editor, providing constructive feedback to refine and enhance the article. 4). Multimodal Article Writing Module. This final stage integrates visual content, consisting of image positioning proposals, image retrieval, image selection, and multimodal refinement, ensuring a cohesive and well-balanced article presentation. These components collectively enable WikiAutoGen to produce high-quality and multi-modal Wikipedia-style articles. + +# 4.2. Outline Proposal + +Given a multimodal topic, the first step is to interpret the input and generate topic-related outlines to guide knowledge exploration and information retrieval. We achieve it by leveraging LLMs [13] and external search tools. + +For text-based topics, the LLM analyzes the input, identifies relevant subtopics, and generates a structured outline to facilitate knowledge exploration. For image-based topics, we utilize Google Vision Search to retrieve metadata, including descriptions and contextual information. Named entity recognition (NER) [9] is then applied to extract the top 10 most frequent entities as query keywords. These + +keywords, combined with the original topic, are then fed into the LLM to generate a well-structured outline. In the case of image-text topics, we combine insights from both modalities by extracting subtopics from the text and retrieving image metadata. The LLM then refines the topic and synthesizes a cohesive outline, integrating textual and visual information to guide comprehensive knowledge retrieval. + +# 4.3. Textual Article Writing + +The textual article writing process involves multiple components working together to generate a well-structured and informative article. It consists of a persona generator, multiagent knowledge exploration, and article generation. + +Persona generator. Given a draft outline, the LLM generates $n$ distinct personas relevant to the topic (where $n$ is a customizable parameter, $n \geq 1$ ), each acting as an independent agent. The LLM assigns specific objectives to each agent based on their role. These agents are equipped with access to external web search tools and contribute to a more comprehensive and well-supported article. + +Multi-agent knowledge exploration. The knowledge exploration stage involves a fixed agent, the asker, and n LLM-generated agents who are assigned specific roles. The asker iterates through the outline, posing targeted questions, while the other agents search the internet for relevant information. They then share findings, discuss them, and refine their understanding. During the discussion, they also interact with the multi-perspective self-reflection module, which provides feedback and improvement advice from a writer's perspective on reliability, engagement, consistency, and informativeness. This iterative process ensures a well- + +rounded knowledge base before article generation. + +Article generation. After gathering web knowledge, the next step is to summarize the collected content and generate the textual article using an LLM-based writing agent. Once the initial draft is produced, the agent iteratively sends each generated section to the multi-perspective self-reflection module for feedback and refines each paragraph. This module evaluates the article from a writer's perspective, providing suggestions to enhance coherence and readability. The writing agent incorporates these refinements, progressively improving the text to produce the final article. + +# 4.4. Multi-Perspective Self-reflection. + +Writing a high-quality Wikipedia-style article requires addressing multiple aspects, including topic consistency, readability, engagement, and informativeness, to provide an optimal reading experience. Therefore, we introduce a multiperspective self-reflection module that systematically evaluates and refines content across these dimensions. This module takes four distinct viewpoints and assesses the article from seven perspectives. + +Perspectives. Our multi-perspective self-reflection focuses on the seven key criteria to improve the paper writing quality. They include reliability, engagement, informativeness, coherence, readability, consistency and helpfulness. We provide a detailed explanation for them in the Appendix. + +Supervisor viewpoint. The supervisor assesses whether the generated content fully addresses the questions posed by the asker, evaluates the article's depth, breadth, and coherence, and reviews the effectiveness of the multi-agent discussion. Additionally, it evaluates whether the generated content aligns with the topic and proposed outlines. Based on these criteria, the supervisor provides an evaluation and passes the feedback to the next role. Depending on the specific needs, the next role can be the writer, reader, or editor. + +Writer viewpoint. From the writer's viewpoint, the primary focus is on the multi-agent knowledge exploration and article generation stages. The writer evaluates whether the generated content maintains coherence, ensures engagement, verifies factual accuracy, and upholds logical consistency. Based on these assessments, the writer provides targeted improvement suggestions. + +For instance, to enhance coherence, it may be recommended to rearrange sentences or add transitional words and phrases. To improve readability, it might suggest simplifying complex concepts. Finally, the writer responds with a set of targeted and refined suggestions. + +Reader viewpoint. To create a high-quality multi-modal article, it is essential to effectively integrate textual content with relevant images to enhance reader engagement and readability. To achieve this, our framework first employs an LLM to propose suitable image placements within the article and generate descriptive content specifying the types of + +images to include. This initial proposal is then reviewed by the multi-perspective self-reflection module, which evaluates the image positioning and the generated image descriptions from the perspective of readability, engagement, and helpfulness. Based on this assessment, the module provides constructive feedback, ensuring that visual content is effectively integrated to enrich the overall reader experience. + +Editor viewpoint. After inserting images into the generated article, there may still be discrepancies or inconsistencies between the visual content and the corresponding textual descriptions. To address this, our framework sends the images along with their related text segments to the multiperspective self-reflection module. This module evaluates the alignment and coherence between the images and their accompanying texts from an editorial viewpoint. It provides targeted suggestions, such as refining the image captions, adjusting image placement, or enhancing textual explanations to better reflect visual content. This final step ensures enhanced relevance, coherence, and readability between visual and textual elements. + +# 4.5. Multi-modal Article Writing + +Following textual generation, we incorporate relevant visual content to enhance the article's readability and expressiveness. This multimodal integration process involves several stages: image positioning proposal, image retrieval, image selection, and finally, an article refinement step that seamlessly integrates the images and text. + +Image positioning proposal. After generating the complete textual article, an LLM-based agent is employed to propose appropriate placements and corresponding descriptions for relevant images. These initial proposals are then refined through interaction with the multi-perspective self-reflection module, which evaluates their relevance, coherence, and engagement from the reader's perspective. + +Image retrieval. We then retrieve relevant images by performing searches based on multiple sources, including general image search engines, Wikipedia, and the websites mentioned in the references. After that, we can obtain a list of relevant image candidates. + +Image selection. We first use the CLIP model [27] to rank retrieved images based on semantic similarity to the generated captions, selecting the top-3 candidates. Subsequently, we leverage a multi-modal model [13] to further evaluate these candidates and select the most contextually appropriate image for inclusion in the article. + +Article Polishing. After finalizing image selection, we integrate the chosen images into the article and proceed to a polishing stage. Due to potential discrepancies between textual content and visual figures, we employ a multi-modal model to revise the entire article, enhancing coherence and consistency across modalities. Additionally, this multimodal model interacts with the multi-perspective self-reflection + +
MethodsContent QualityInformativenessReliabilityEngagementAverage
AlignmentConsistencyRelevanceRepetitionBreadthDepthVerifiabilityEngagementNovelty
Text as Topic
oRAG [1]61.3573.9676.0471.1163.3052.4245.4757.5145.2460.71
Storm [28]72.4979.1371.2269.4765.6262.6152.4158.8055.5865.26
Co-Storm [14]78.0584.1075.1175.4068.4267.7058.2061.0261.6169.96
OmniThink [42]70.5379.6772.4169.2663.5561.2148.5757.3953.2163.98
WikiAutoGen (Ours)81.6890.8788.0283.6279.6473.7370.6971.1469.2178.73
Image as Topic
oRAG50.1072.1650.9265.4742.0143.2633.9040.6636.9148.38
Storm45.8059.6045.9946.3842.6939.9234.2342.3835.1743.57
Co-Storm47.0061.4044.8547.7641.9841.8335.0341.9937.6944.39
OmniThink43.6158.2943.0345.6738.6342.2629.1838.3133.5741.39
WikiAutoGen (Ours)82.5788.7587.2080.2277.2474.9968.4169.3668.6977.49
Image-Text as Topic
oRAG60.0875.1670.9472.2458.3850.5742.4755.0143.9558.76
Storm67.2075.2966.6464.3361.6158.2649.2156.2551.2761.12
Co-Storm70.1579.2967.3168.8961.9061.2252.4457.0555.6863.77
OmniThink64.5675.3364.6363.6457.4456.3843.0454.1448.8658.67
WikiAutoGen (Ours)85.2690.6388.4482.1179.3175.2068.5968.7971.0678.82
+ +Table 2. Comparison of article generation performance for textual content. We evaluate content quality, informativeness, reliability, and engagement under three input modalities (Text-only, Image-only, and Image-Text) on our WikiSeek benchmark. + +module, obtaining editorial feedback to further refine the integrated content. + +# 5. Experiment + +# 5.1. Experiment setup + +Implementation Details. For the language model (LM) components of WikiAutoGen, we employ zero-shot prompting implemented using the DSPy framework [15] in conjunction with GPT-4o [13], GPT-4o-mini, and GPT-o3-mini [23]. Specifically, we use GPT-o3-mini for the multiperspective self-reflection module due to its strong reasoning capabilities, GPT-4o for the multimodal knowledge exploration tasks, and GPT-4o-mini for all other remaining tasks. WikiAutoGen retrieves real-time web information via Serper's $\mathsf{API}^1$ , with each query returning up to 5 web pages. Throughout the experiments, we maintain consistent settings by fixing the LM temperature at 1.0 and the top- $p$ value at 0.9. For evaluation, we utilize GPT-4o as the evaluator. To further validate the results, we conduct more evaluations using two distinct evaluators, Gemini2.5-Flash [8] and Prometheus2 [16], as detailed in Appendix B. + +Evaluation metrics. We evaluate the generated multimodal articles through separate assessments of their textual and visual content. + +- Text quality evaluation. Following prior evaluation frameworks [14, 28], we utilize GPT-4o as the evaluator to assess generated articles across nine criteria grouped into four main aspects: + +- Content quality: alignment, relevance, repetition, and consistency; + +- Informativeness: breadth and depth; +- Reliability: verifiability; +- Engagement: engagement and novelty. + +- Image quality evaluation. As there are no existing benchmarks specifically designed for evaluating multimodal Wikipedia-style article generation, we propose an evaluation method inspired by the textual evaluation frameworks [14, 28] to evaluate image quality. Specifically, we assess image quality based on four criteria: image-text coherence, engagement, helpfulness, and information supplement (the image's ability to provide additional useful information beyond the textual context). + +Baselines We compare WikiAutoGen with four representative LLM-based baseline frameworks for automated expository writing on our WikiSeek benchmark: + +- Outline-driven RAG (oRAG) generates articles guided by outlines produced by Self-RAG [1]. +- Storm [28] leverages LLM-driven conversations and outlines from diverse perspectives to generate Wikipedia-style content. +- Co-STORM [14] utilizes collaborative discourse among multiple LLM agents to explore and discover unknown information. +- OmniThink [42] enhances article quality through iterative expansion and reflection, emulating human slow-thinking to increase knowledge density. + +Since these baselines originally lack multimodal capabilities, we equip them with image retrieval functionalities to enable a fair multimodal comparison. Specifically, each baseline can also retrieve images via Google image search, extract relevant metadata, and search for images based on generated textual descriptions. + +
ModulesContent QualityInformativenessReliabilityEngagementAverage
Multi-agentOutline proposalSelf-reflectionAlignmentConsistencyRelevanceRepetitionBreadthDepthVerifiabilityEngagementNovelty
xxx50.6362.8753.1051.1948.5844.3543.6445.5439.0848.78
✓xx71.8176.3872.6768.2564.3669.1664.2057.0455.1766.56
x✓x79.1186.2080.9377.2172.6562.1369.1964.4564.1172.88
xx✓75.9184.0882.2075.5673.2466.6665.4163.0358.3671.60
x✓✓77.6885.5784.7378.4975.0868.8565.5965.7366.4573.80
✓✓✓82.5788.7587.2080.2277.2474.9968.4169.3668.6977.49
+ +Table 3. Ablation study. We study the impact of individual modules (Multi-agent knowledge exploration, outline proposal, and self-reflection) on article generation performance for text content. The input modality is image-only. + +
MethodCoherenceEngagementHelpfulnessInfo. Sup.Average
Text as Topic
oRAG57.3656.2663.6151.9057.28
Storm55.2045.9751.8943.9749.26
Co-Storm57.6248.6454.1945.0751.38
OmniThink58.8249.3654.9347.5552.67
WikiAutoGen (Ours)70.1266.3174.7664.7868.99
Image as Topic
oRAG61.2152.0758.03945.9654.32
Storm52.5943.9849.5341.8446.99
Co-Storm54.3245.6151.5541.5648.26
OmniThink59.8850.6256.1347.2353.47
WikiAutoGen (Ours)71.9061.6977.6363.8868.78
Image-Text as Topic
oRAG66.3856.3162.9449.7858.85
Storm59.6750.2655.7846.7853.12
Co-Storm54.2845.6551.2942.8848.53
OmniThink61.7651.4057.8849.9755.25
WikiAutoGen (Ours)72.2470.2972.1169.2970.98
+ +Table 4. Comparison of article generation performance for image content. We evaluate textual generation for image-text coherence, image engagement, image helpfulness, and information supplement on our WikiSeek benchmark. + +# 5.2. Experimental results + +Textual content comparison on WikiSeek. We demonstrate our results in the Table 2. The results indicate that WikiAutoGen consistently achieves the highest average scores across all evaluation dimensions—content quality, informativeness, reliability, and engagement—highlighting its comprehensive effectiveness in article generation. + +For text-only inputs, our WikiAutoGen achieves an average score of 78.73, significantly outperforming the best baseline (Co-Storm, 69.96) by approximately 8.8 points, demonstrating superior coherence, alignment, and informativeness. In image-only scenarios, the improvement is even more pronounced, with WikiAutoGen obtaining 77.49, surpassing the strongest baseline (oRAG, 48.38) by approximately 29.1 points, reflecting its exceptional capability to extract meaningful textual insights from visual content. For combined image-text topics, our model maintains its advantage with an average score of 78.82, showing a clear improvement (+ 15.05 points) over the next-best baseline (Co-Storm, 63.77). Overall, the substantial performance gains demonstrate that WikiAutoGen excels at synthesizing cohesive and engaging content across diverse inputs. + +Image content comparison on WikiSeek. Table 4 compares the performance of image content across different article generation methods. Our WikiAutoGen consistently achieves the highest scores across all image evaluation metrics for all three input modalities. + +Specifically, our method significantly improves upon baseline methods, outperforming the next-best method by + +![](images/efcdbf0ca94ac7f9aecf53350432089a7689749dbd7508acd7db77ea23fe27fd.jpg) +Figure 3. Ablation study across different data difficulty levels. We compare our WikiAutoGen with Storm and OmniThink on three difficulty categories: hard (300-500 characters), very hard (100-300 characters), and extremely hard (<100 characters). + +approximately 11.34 points on image-text coherence (72.24 vs. Storm's 59.67) under the image-text topics. Similarly, for image-only topics, our approach excels particularly in helpfulness (77.63) and coherence (71.90), demonstrating WikiAutoGen's superior capability in selecting images that meaningfully complement textual content and provide additional useful information. Overall, these results highlight WikiAutoGen's effectiveness in integrating images to enhance coherence, engagement, and informativeness, substantially advancing the overall quality and readability of multimodal articles. + +# 5.3. Ablation studies + +Ablation on different components of WikiAutoGen. We conduct an ablation study to analyze the individual contributions of three core components to textual article generation from image-only topics. Specifically, we examine the impact of components, including multi-agent knowledge exploration, outline proposal, and multi-perspective self-reflection. Results are shown in Table 3. Specifically, without the outline proposal, the system simply retrieves a single image and extracts its description from metadata. Without multi-agent knowledge exploration, only a single static agent responds to the asker. + +Incorporating the multi-agent module improves performance from 48.78 (baseline without modules) to 66.56 (+17.78 points), highlighting its effectiveness in collaborative knowledge exploration. Using the outline proposal alone further increases performance to 72.88 (+24.10 points), underscoring its importance for content structur + +ing. The self-reflection module individually achieves 71.60 (+22.82 points), indicating its strength in refining coherence and consistency. Combining all three modules results in the highest performance (77.49), validating their complementary roles in generating coherent, informative, and engaging articles from image-only inputs. + +Ablation on different difficulty levels. In our WikiSeek benchmark, topics are grouped into three difficulty levels based on article length (character count): hard (300–500 characters), very hard (100–300 characters), and extremely hard (fewer than 100 characters), with 100 examples per level. We evaluate text-only inputs and present the average textual evaluation results in Figure 3. The results indicate that our WikiAutoGen consistently outperforms all baseline methods across all three difficulty levels. Notably, as the topic difficulty increases (from hard to extremely hard), the performance gap between WikiAutoGen and Storm widens from 12.35 points to 15.35 points, with similar trends observed against other baselines. These results demonstrate the robustness and stability of WikiAutoGen in effectively handling highly challenging and underexplored topics. + +5.4. Compared with Commercial Deep Research + +
MethodText as TopicImage as TopicImage-Text as TopicAverageCompute Time
OpenAI92.7591.9594.5093.06~ 30 min
Google81.91———~ 12 min
Grok88.3581.7086.1385.06~ 10 min
WikiAutoGen88.5889.5588.8989.01~ 8 min
+ +Table 5. Comparison of commercial models' article generation performance with Prometheus2 [16]. + +We evaluate the performance of WikiAutoGen against leading commercial models on the WikiSeek benchmark, spanning three input modalities. As shown in Table 5, WikiAutoGen achieves an impressive average score of 89.01, closely approaching OpenAI (93.06), while significantly outperforming Grok (85.06) and Google (81.91). WikiAutoGen delivers consistently strong results across text (88.58), image (89.55), and image-text (88.89) inputs, demonstrating robust cross-modal capabilities. In terms of efficiency, WikiAutoGen generates articles in just 8 minutes, making it over $3.75 \times$ faster than OpenAI (30 minutes), the fastest among the commercial baselines. This substantial speed advantage underscores its practical scalability and suitability for real-world, time-sensitive applications. + +# 5.5. Human evaluation + +To evaluate the quality of the generated Wikipedia-style articles, we conduct a human evaluation study via Amazon Mechanical Turk (AMT) $^2$ . We randomly sample 100 text-only topics from the WikiSeek benchmark dataset and perform pairwise comparisons between articles generated by our method (WikiAutoGen), Storm, and OmniThink. Each + +topic is evaluated by three independent participants in randomized order. Participants first answer the question: "Do you think adding images would improve comprehension of the topic?" As shown in Figure 4 (left), $97.7\%$ of participants agree that images improve topic comprehension. + +Additionally, participants answer multiple-choice questions, including: (1) Which article is the easiest to understand? (2) Which article is the most engaging in terms of narrative, examples, or overall presentation? (3) Which article provides the most comprehensive background information and in-depth analysis? (4) Which article is your overall favorite? The question order is randomized to mitigate potential evaluation bias. As illustrated in Figure 4 (right), participants consistently prefer articles generated by WikiAutoGen over those by Storm and OmniThink across all evaluation criteria. + +![](images/146ec0fe704c523c265e9a4a4ea4caa12c3c4eaa6664a43349ba0a569a4d87ec.jpg) +Figure 4. Human evaluation study. We randomly sample text-only topics and conduct comparative evaluations between WikiAutoGen, Storm, and OmniThink. Left: Participants respond to the question, "Do you think adding images would improve comprehension of the topic?". Right: Participants answer multiple-choice questions evaluating readability, engagement, informativeness, and overall preference. + +# 6. Conclusion + +In this paper, we introduced WikiAutoGen, a comprehensive multi-agent framework designed for automated multimodal Wikipedia-style article generation. WikiAutoGen integrates both visual and textual content, significantly enhancing the depth, informativeness, and engagement of generated articles. To address key limitations in prior work, we proposed a novel multi-perspective self-reflection mechanism, which systematically improves article coherence, reliability, and overall quality. Furthermore, we presented WikiSeek, a challenging multimodal benchmark specifically crafted to evaluate the performance of models in generating content for less-explored topics. Experimental results demonstrated that WikiAutoGen substantially outperforms existing state-of-the-art baselines across both textual and visual evaluation metrics. Particularly notable were the improvements in content quality, informativeness, and reader engagement, validating the effectiveness of integrating multimodal inputs and iterative self-reflection. + +# References + +[1] Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection. ArXiv, abs/2310.11511, 2023. 3, 6 +[2] Jinze Bai et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. Accessed: 2025-03-04. 2 +[3] Nishant Balepur, Jie Huang, and Kevin Chen-Chuan Chang. Expository text generation: Imitate, retrieve, paraphrase. ArXiv, abs/2305.03276, 2023. 2 +[4] Maciej Besta, Florim Memedi, Zhenyu Zhang, Robert Gerstenberger, Guangyuan Piao, Nils Blach, Piotr Nczyk, Marcin Copik, Grzegorz Kwa'sniewski, Jurgen Muller, Lukas Gianinazzi, Ales Kubcek, Hubert Niewiadomski, Aidan O'Mahony, Onur Mutlu, and Torsten Hoefler. Demystifying chains, trees, and graphs of thoughts. In ArXiv, 2024. 2 +[5] DeepSeek-AI: Xiao Bi et al. Deepseek llm: Scaling opensource language models with longtermism. arXiv preprint arXiv:2401.02954, 2024. Accessed: 2025-03-04. 2 +[6] Andrea Burns, Krishna Srinivasan, Joshua Ainslie, Geoff Brown, Bryan A. Plummer, Kate Saenko, Jianmo Ni, and Mandy Guo. Wikiweb2m: A page-level multimodal wikipedia dataset. ArXiv, abs/2305.05432, 2023. 3 +[7] Steven Coyne. Template-guided grammatical error feedback comment generation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 94-104, Dubrovnik, Croatia, 2023. Association for Computational Linguistics. 3 +[8] Google DeepMind. Gemini: Multimodal language model. https://deepmind.google/technologies/gemini/, 2023. Accessed: 2025-03-04. 2, 6 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018. 4 +[10] Zican Dong, Tianyi Tang, Junyi Li, Wayne Xin Zhao, and Ji-Rong Wen. BAMBOO: A comprehensive benchmark for evaluating long text modeling capacities of large language models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 2086-2099, Torino, Italia, 2024. ELRA and ICCL. 2 +[11] Fan Gao, Hang Jiang, Rui Yang, Qingcheng Zeng, Jinghui Lu, Moritz Blum, Dairui Liu, Tianwei She, Yuang Jiang, and Irene Li. Evaluating large language models on wikipia-style survey generation. In Annual Meeting of the Association for Computational Linguistics, 2023. 2 +[12] Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. Critic: Large language models can self-correct with tool-interactive critiquing. ArXiv, abs/2305.11738, 2023. 3 +[13] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4, 5, 6 + +[14] Yucheng Jiang, Yijia Shao, Dekun Ma, Sina J. Semnani, and Monica S. Lam. Into the unknown unknowns: Engaged human learning through participation in language model agent conversations. In Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2, 6 +[15] Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. DSPy: Compiling declarative language model calls into self-improving pipelines. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR), 2024. 6 +[16] Seungone Kim, Juyoung Suk, Shayne Longpre, Bill Yuchen Lin, Jamin Shin, Sean Welleck, Graham Neubig, Moontae Lee, Kyungjae Lee, and Minjoon Seo. Prometheus 2: An open source language model specialized in evaluating other language models. arXiv preprint arXiv:2405.01535, 2024. 6, 8 +[17] Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan A Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, et al. Longlamp: A benchmark for personalized long-form text generation. arXiv preprint arXiv:2407.11016, 2024. 2 +[18] Irene Li, Alexander R. Fabbri, Rina Kawamura, Yixin Liu, Xiangru Tang, Jaesung Tae, Chang Shen, Sally Ma, Tomoe Mizutani, and Dragomir R. Radev. Surfer100: Generating surveys from web resources, wikipedia-style. In International Conference on Language Resources and Evaluation, 2021. 2 +[19] Yi Liang, You Wu, Honglei Zhuang, Li Chen, Jiaming Shen, Yiling Jia, Zhen Qin, Sumit K. Sanghai, Xuanhui Wang, Carl Yang, and Michael Bendersky. Integrating planning into single-turn long-form text generation. ArXiv, abs/2410.06203, 2024. 2 +[20] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Sean Welleck, Bodhisattwa Prasad Majumder, Shashank Gupta, Amir Yazdanbakhsh, and Peter Clark. Self-refine: Iterative refinement with self-feedback. ArXiv, abs/2303.17651, 2023. 3 +[21] Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wentau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12076-12100, Singapore, 2023. Association for Computational Linguistics. 2 +[22] OpenAI. Chatgpt: Optimizing language models for dialogue. https://openai.com/blog/chatgpt, 2022. Accessed: 2025-03-04. 2 +[23] OpenAI. Openai o3-mini, 2025. 6 +[24] Debjit Paul, Mete Ismayilzada, Maxime Peyrard, Beatrix Borges, Antoine Bosselut, Robert West, and Boi Faltings. Refiner: Reasoning feedback on intermediate representations. In Conference of the European Chapter of the Association for Computational Linguistics, 2023. 3 + +[25] Alexandre Piché, Aristides Milios, Dzmitry Bahdanau, and Chris Pal. Llms can learn self-restraint through iterative self-reflection. ArXiv, abs/2405.13022, 2024. 3 +[26] Haoran Que, Feiyu Duan, Liquin He, Yutao Mou, Wangchunshu Zhou, Jiaheng Liu, Wenge Rong, Zekun Moore Wang, Jian Yang, Ge Zhang, et al. Hellobench: Evaluating long text generation capabilities of large language models. arXiv preprint arXiv:2409.16191, 2024. 2 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 5 +[28] Yijia Shao, Yucheng Jiang, Theodore A. Kanell, Peter Xu, Omar Khattab, and Monica S. Lam. Assisting in writing wikipedia-like articles from scratch with large language models. In North American Chapter of the Association for Computational Linguistics, 2024. 1, 2, 6 +[29] Zejiang Shen, Tal August, Pao Siangliulue, Kyle Lo, Jonathan Bragg, Jeff Hammerbacher, Doug Downey, and Joseph Chee Chang. Beyond summarization: Designing ai support for real-world expository writing tasks. ArXiv, abs/2304.02623, 2023. 2 +[30] Noah Shinn, Federico Cassano, Beck Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflection: language agents with verbal reinforcement learning. In Neural Information Processing Systems, 2023. 3 +[31] Zhiqing Sun, Yikang Shen, Qinhong Zhou, Hongxin Zhang, Zhenfang Chen, David D. Cox, Yiming Yang, and Chuang Gan. Principle-driven self-alignment of language models from scratch with minimal human supervision. ArXiv, abs/2305.03047, 2023. 3 +[32] Bowen Tan, Zichao Yang, Maruan Al-Shedivat, Eric Xing, and Zhiting Hu. Progressive generation of long text with pretrained language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4313-4324, Online, 2021. Association for Computational Linguistics. 2 +[33] Haochen Tan, Zhijiang Guo, Zhan Shi, Lu Xu, Zhili Liu, Yunlong Feng, Xiaoguang Li, Yasheng Wang, Lifeng Shang, Qun Liu, et al. Proxyqa: An alternative framework for evaluating long-form text generation with large language models. arXiv preprint arXiv:2401.15042, 2024. 2 +[34] Chen Tang, Hongbo Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. Enhancing dialogue generation via dynamic graph knowledge aggregation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4604-4616, Toronto, Canada, 2023. Association for Computational Linguistics. 2 +[35] Hugo Touvron et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2 +[36] Qiyao Wang, Shiwen Ni, Huaren Liu, Shule Lu, Guhong Chen, Xi Feng, Chi Wei, Qiang Qu, Hamid Alinejad + +Rokny, Yuan Lin, and Min Yang. Autopatent: A multiagent framework for automatic patent generation. ArXiv, abs/2412.09796, 2024. 2 +[37] Yidong Wang, Qi Guo, Wenjin Yao, Hongbo Zhang, Xin Zhang, Zhen Wu, Meishan Zhang, Xinyu Dai, Min Zhang, Qingsong Wen, Wei Ye, Shikun Zhang, and Yue Zhang. Autosurvey: Large language models can automatically write surveys. ArXiv, abs/2406.10252, 2024. 2 +[38] Yifei Wang, Yuyang Wu, Zeming Wei, Stefanie Jegelka, and Yisen Wang. A theoretical understanding of self-correction through in-context alignment. ArXiv, abs/2405.18634, 2024. 3 +[39] Yutong Wang, Jiali Zeng, Xuebo Liu, Fandong Meng, Jie Zhou, and Min Zhang. Taste: Teaching large language models to translate through self-reflection. In Annual Meeting of the Association for Computational Linguistics, 2024. 3 +[40] Yilin Wen, Zifeng Wang, and Jimeng Sun. Mindmap: Knowledge graph prompting sparks graph of thoughts in large language models. In Annual Meeting of the Association for Computational Linguistics, 2023. 2 +[41] Yuhao Wu, Ming Shan Hee, Zhiqing Hu, and Roy KWei Lee. Spinning the golden thread: Benchmarking long-form generation in language models. arXiv preprint arXiv:2409.02076, 2024. 2 +[42] Zekun Xi, Wenbiao Yin, Jizhan Fang, Jialong Wu, Runnan Fang, Ningyu Zhang, Jiang Yong, Pengjun Xie, Fei Huang, and Huajun Chen. Omnithink: Expanding knowledge boundaries in machine writing through thinking. ArXiv, abs/2501.09751, 2025. 2, 6 +[43] Hanqi Yan, Qinglin Zhu, Xinyu Wang, Lin Gui, and Yulan He. Mirror: A multiple-perspective self-reflection method for knowledge-rich reasoning. ArXiv, abs/2402.14963, 2024. 3 +[44] Tianjun Zhang, Aman Madaan, Luyu Gao, Steven Zheng, Swaroop Mishra, Yiming Yang, Niket Tandon, and Uri Alon. In-context principle learning from mistakes. ArXiv, abs/2402.05403, 2024. 3 +[45] Yangqiaoyu Zhou, Haokun Liu, Tejes Srivastava, Hongyuan Mei, and Chenhao Tan. Hypothesis generation with large language models. ArXiv, abs/2404.04326, 2024. 3 \ No newline at end of file diff --git a/wikiautogentowardsmultimodalwikipediastylearticlegeneration/images.zip b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3a29a05f7c2d0e3d040355d67588949b06c10590 --- /dev/null +++ b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2be0991abcd0c811b2abc90b95f734d147e60c79b94dc5c4775d9ef329c1d1ef +size 481684 diff --git a/wikiautogentowardsmultimodalwikipediastylearticlegeneration/layout.json b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dcca00b133bebf1fca4b4c5c67fca11ef628fc89 --- /dev/null +++ b/wikiautogentowardsmultimodalwikipediastylearticlegeneration/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:661cfc17bce51ba7779fcb664919b8be2741f9d77ec5cf4ea4d3636433a1b945 +size 297687 diff --git a/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/ffaa63f6-2b3f-41d0-a1eb-88d48e346985_content_list.json b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/ffaa63f6-2b3f-41d0-a1eb-88d48e346985_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..59dca0b539b87493182983199da2461207e55f08 --- /dev/null +++ b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/ffaa63f6-2b3f-41d0-a1eb-88d48e346985_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57af9fb5e517a6f2ebbd491068820d5ae9bfab990200c69432f5a59495c28b72 +size 93830 diff --git a/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/ffaa63f6-2b3f-41d0-a1eb-88d48e346985_model.json b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/ffaa63f6-2b3f-41d0-a1eb-88d48e346985_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f489a1e2c0575a239ff673baf3df69884636d0c7 --- /dev/null +++ b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/ffaa63f6-2b3f-41d0-a1eb-88d48e346985_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e8b44f3ab3244bdbcba75d45b37966f503d8c60e24520f64f760fd5d7f596d8 +size 117722 diff --git a/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/ffaa63f6-2b3f-41d0-a1eb-88d48e346985_origin.pdf b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/ffaa63f6-2b3f-41d0-a1eb-88d48e346985_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e77b069072a4448f9a1d34093f0d765a6ee6034 --- /dev/null +++ b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/ffaa63f6-2b3f-41d0-a1eb-88d48e346985_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b588ecefba7863a0c311e63dcefa317151e4557ac66792f1ab6916a9fc88ac9 +size 10601423 diff --git a/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/full.md b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e9456a6a56d4c3045cb4bda689bcd2f32a91e675 --- /dev/null +++ b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/full.md @@ -0,0 +1,338 @@ +# WildSAT: Learning Satellite Image Representations from Wildlife Observations + +Rangel Daroya1 Elijah Cole2 Oisin Mac Aodha3 Grant Van Horn1 Subhransu Maji1 +1University of Massachusetts, Amherst 2GenBio AI 3University of Edinburgh + +![](images/fb74549684e71abda8371868f130b55783ba57304165446bdaf9385368a5dade.jpg) +Figure 1. Wildlife observations can provide valuable supervision for learning satellite image representations. Known wildlife locations derived from human observations, coupled with descriptive information on species range, habitat, and other ecological attributes on Wikipedia, serve as a rich source of contextual information for satellite imagery. Our WildSAT approach leverages these additional data sources to (i) learn robust satellite image representations for downstream tasks, and (ii) complement and further improve existing models using continual pre-training. + +# Abstract + +Species distributions encode valuable ecological and environmental information, yet their potential for guiding representation learning in remote sensing remains underexplored. We introduce WildSAT, which pairs satellite images with millions of geo-tagged wildlife observations readily-available on citizen science platforms. WildSAT employs a contrastive learning approach that jointly leverages satellite images, species occurrence maps, and textual habitat descriptions to train or fine-tune models. This approach significantly improves performance on diverse satellite image recognition tasks, outperforming both ImageNet-pretrained models and satellite-specific baselines. Additionally, by aligning visual and textual information, WildSAT enables zero-shot retrieval, allowing users to search geographic locations based on textual descriptions. WildSAT surpasses recent cross-modal learning methods, including approaches that align satellite images with ground imagery or wildlife photos, demonstrating the advantages of our approach. Finally, we analyze the impact of key design choices and highlight the broad applicability of WildSAT to remote sensing and biodiversity monitoring. + +# 1. Introduction + +The growth in the number of satellites with imaging capabilities deployed over the past 50 years has provided an unprecedented ability to monitor the surface of the earth [37, 80, 83]. The image data derived from these remote sensors has been shown to be highly effective for diverse tasks such as estimating global tree canopy height [39, 68], detecting illegal fishing activity [22, 36, 56], crop monitoring [19, 33, 75], disaster management [57, 66, 73], among others. Central to building computer vision models for these tasks is the need for mechanisms for learning effective representations from image data. As a result of the distribution shift between remote sensing imagery and web-sourced images, a large body of work has emerged exploring the merits and trade-offs between different sources of supervision. + +Direct supervision in the form of paired images and labels (e.g. image tiles with labels denoting land cover type) can be prohibitively expensive to obtain at a global scale [28]. To address this, there is growing interest to develop methods that learn remote sensing representations from self-supervision [32, 46, 48], multiple paired modalities [15, 47, 63], or other auxiliary sources [16, 72]. A + +useful supervision source needs to be globally distributed, correlated with the local landscape as viewed from an image, and able to discriminate regions at a fine spatial scale. + +A promising auxiliary supervision source is provided by locations where different species of plants and animals can be found around the world. For example, Mountain Goats (Oreamnos americanus) are found in rugged mountainous areas, while habitat specialists like the Cactus Wren (Campylorhynchus brunneicapillus) are typically found in deserts nesting in spiny cacti (Fig. 1). Species location data offer a rich source of supervision, reflecting the local natural environment around each observation. It is also readily available from citizen science platforms such as iNaturalist [2] and eBird [64] which host hundreds of millions of wildlife observations. While species location data has improved fine-grained species classification [6, 10, 44], its potential for learning remote sensing representations remains unclear. Prior works have largely relied on anthropogenic labels (e.g. human-made features like roads, buildings, industrial areas) to learn satellite image representations [41, 47, 69], whereas we explore using wildlife observations as a complementary and potentially valuable signal. + +We introduce a new approach that uses signals derived from species location observations. We take inspiration from recent work that attempt to fuse multi-modal ecological data and remote sensing imagery into a shared common embedding space [29, 62, 63]. WildSAT uses a contrastive learning objective to align satellite image, text, and location based on species observation data, bringing embeddings from the same area closer together and pushing those from different areas further apart. Through this method, we utilize information about the preferred habitats of species to improve satellite image representations. + +We make the following contributions: (i) We introduce WildSAT, a new approach to learning remote sensing representations using species observation locations as a supervisory signal. (ii) We show WildSAT-derived representations are competitive with state-of-the-art satellite representations, while enabling zero-shot satellite image retrieval. (iii) We present a thorough evaluation that highlights WildSAT-derived representations not only outperform but also complement existing methods focused on anthropogenic labels by incorporating wildlife information. (iv) We perform ablation studies to show the impact of each component of our approach, and show WildSAT outperforms recent cross-modal methods like GRAFT [47] and TaxaBind [63]. The code and dataset are available at https://github.com/cvl-umass/wildsat. + +# 2. Related Work + +Previous works learn satellite image representations by training on large-scale remote sensing datasets from programs like Landsat [50], Sentinel [18, 30], or NAIP [51]. + +These methods range from using self-supervised [13, 32, 48], supervised [5, 54, 65], and cross-modal [15, 23, 47, 53, 58, 63] learning to learn rich image representations for downstream satellite-based tasks. + +Self-supervised learning. These methods learn representations by taking advantage of spatio-temporal invariance or by predicting missing image patches from satellite images. SeCo [48] collects data from the same location across different seasons and uses a contrastive objective to force the image embeddings of samples to be closer if they are from the same location, and farther otherwise. Other works [46, 82] extend this by selecting points in time based on the level of similarity between satellite images, or by synthetically generating images that are variants of the same location. Vision transformers trained using masked autoencoders [13, 27, 32] have also been adopted owing to their success in learning natural image representations. A notable example is Prithvi [32], a transformer network with 100 million parameters pretrained on 1TB of satellite imagery from around the world, achieving strong performance across a variety of Earth observation tasks. + +Supervised learning. These methods leverage labels from tasks like object detection [77], instance segmentation [74], and classification [28, 65] in the satellite domain. Some works focus solely on image classification [54, 65, 74, 76, 77], while others learn from multiple label types. For example, SatlasPretrain [5] curated a large dataset with over 300 million labels across 137 categories, using domain experts, crowdsourced workers, and publicly available datasets (e.g. OpenStreetMap [24], WorldCover [70]). It is a unified model jointly trained for multiple tasks (e.g. segmentation, object detection) and showed improved performance on various downstream remote-sensing tasks, significantly outperforming ImageNet-pretrained models and other baselines. + +Cross-modal learning. Recently, other works have explored adding other modalities while training on satellite images [15, 23, 29, 34, 35, 47, 62, 63, 72]. A common approach uses geo-tagged images and pre-trained image-text encoders like CLIP [59], aligning new modalities to their embedding space using contrastive learning [15, 31, 35, 41, 47, 72]. This strategy has been used for various tasks: satellite image localization in GeoCLIP [72], bird species classification and mapping in BirdSAT [62], and improving plant species image representations in CRISP [29]. Models like GRAFT [47], TaxaBind [63], RemoteCLIP [41], and GeoBind [15] align multiple modalities at the same time for cross-modal retrieval and zero-shot tasks. Zermatten et al. [81] have also demonstrated the benefits of aligning satellite imagery with species observation data for zero-shot classification. TaxaBind was the first to use species geographic locations and satellite imagery, but it focuses on ecological tasks rather than satellite image tasks. We also expand on their approach by leveraging open-source Wikipedia text + +![](images/7e53afd9ab3606994b91755132862257d85ca8c5d7f531a2611d1f1e8b5c9a33.jpg) +(a) Training Pipeline + +![](images/39ab0134cbf50ab02cc5a5dbb86e03ac050f1e31f80eedada5eac10a20ea7325.jpg) +(b) Inference Pipeline +Figure 2. Architecture for training and evaluating the satellite image encoder. (a) The training pipeline uses the location of a species, the satellite images at those locations, the environmental covariates, and the Wikipedia text associated with the species. In addition to the alignment of image, text, and location modalities, the encoder is encouraged to learn additional image features by using temporal and geometric image transformations on the input satellite image. (b) Downstream tasks use the frozen satellite image encoder with an additional trainable layer (or layers). Alternatively, the predicted image embeddings can be used for zero-shot retrieval via text queries. + +instead of taxonomic hierarchy data, offering a more diverse supervision for satellite image representations. Beyond contrastive learning, other methods use supervised learning to fuse embeddings of different modalities for predicting species range maps and encounter rates [16, 25, 67]. Most similar to our work is WikiSatNet [69] which uses location-aligned Wikipedia articles and satellite images to improve satellite image representations. However, while WikiSatNet and previous works [35, 41, 47, 69] primarily focus on anthropogenic data, we explore the impact of wildlife observations. Specifically, we investigate how species distributions—capturing habitat preferences, climate, and environmental factors—can serve as powerful signals. + +While previous work has focused on improving species distribution modeling [16, 43, 62, 67] or fine-grained image classification [15, 45, 63] using satellite images, our work improves satellite image representations using wildlife observations. Our experiments show that both randomly initialized models and strong baselines, such as Prithvi [32], SatlasNet [5], and SeCo [48], benefit from this supervision on a wide range of satellite image tasks (Fig. 3, Tab. 1). + +# 3. Method + +We define the problem as follows: given an image encoder $f_{\theta}:\mathbf{I}\rightarrow \mathbf{z}$ with parameters $\theta$ , we want to find an optimal set of parameters $\theta^{*}$ that improves the performance of $f$ on various remote sensing tasks through a robust satellite image feature representation $\mathbf{z}$ . It takes an image $\mathbf{I}\in \mathbb{R}^{W\times H\times 3}$ as input and outputs an embedding $\mathbf{z}\in \mathbb{R}^d$ . We propose to optimize $\theta$ using our WildSAT framework using data consisting of satellite images, locations, environmental covariates, and text. We hypothesize that leveraging known envi + +ronmental context around each species observation (Fig. 1) allows for more effective optimization of model parameters. + +# 3.1. WildSAT + +To supplement satellite images, we take advantage of additional modalities that naturally align based on the distribution of species throughout the globe. Information on species habitat can provide a rich source of supervision for improving satellite image representations, and we describe how to leverage this through our proposed WildSAT framework. Fig. 2 shows the overall architecture used to train a satellite image encoder $f_{\theta}$ . The encoder $f$ can be any architecture (e.g. a ResNet50 [26], ViT-B/16 [17], etc.). The initial parameters $\theta$ can be randomly initialized, pre-trained on a different domain (e.g. ImageNet [14]), or pre-trained on a related dataset (e.g. SatlasPretrain [5]). The output embedding $\mathbf{z}$ can be used for downstream remote sensing tasks such as classification and zero-shot image retrieval. + +WildSAT aims to improve the model by training on additional modalities related to species observation data. To incorporate other modalities, we use pre-trained models, e.g. SINR [12] for location and GritLM [52] for text. + +Location encoder. SINR [12] is mainly used to predict the presence and absence of species at a location by training on large collections of species observation data. It takes location $\mathbf{a}_i = (lat,lon) \in \mathbb{R}^2$ and, optionally, the corresponding environmental covariates (e.g. weather and climate data) to produce a location embedding in $\mathbb{R}^{256}$ . + +Large language model. GritLM [52] is a large language model (LLM) that outputs a fixed-length embedding in $\mathbb{R}^{4096}$ given a text input. It is trained to handle text of arbitrary length, making it suitable for longer and varying lengths of text obtained from sources such as Wikipedia [4]. + +Satellite image encoder. Given an initial image encoder $f$ , we add three sets of linear layers to predict embeddings for images $(\mathbf{z}_{I^t})$ , text $(\mathbf{z}_{\mathrm{txt}})$ , and locations $(\mathbf{z}_{\mathrm{loc}})$ . Similarly, both the pre-trained LLM (GritLM) and location encoder (SINR) have an added trainable linear layer to project their respective feature embeddings to $\mathbb{R}^d$ as $\mathbf{e}_{\mathrm{txt}}$ and $\mathbf{e}_{\mathrm{loc}}$ , respectively. In addition to text and location, we also fine-tune the model on image-specific features by forcing embeddings of similar satellite images to be close to each another. For two satellite images $\mathbf{I}^{t_1}$ and $\mathbf{I}^{t_2}$ taken at the same location but at different times, their corresponding feature representations should be similar. We also apply geometric augmentations $T$ such as flipping and random cropping on the latter image such that $f(\mathbf{I}^{t_1}) \approx f(T(\mathbf{I}^{t_2}))$ . Doing this encourages the model to learn meaningful image features by distinguishing between similar and dissimilar images while refining the same features through text, location, and environment. + +# 3.2. Training + +Our framework uses a contrastive learning objective to improve satellite image encoder embeddings. We jointly optimize the parameters of the model $f_{\theta}$ and the additional linear layers through the training objective in Eqn. 1. These loss terms correspond to a contrastive objective over image embeddings $(\mathcal{L}_{\mathrm{img}})$ , text embeddings $(\mathcal{L}_{\mathrm{txt}})$ , and location embeddings $(\mathcal{L}_{\mathrm{loc}})$ of $f$ . The embeddings $\mathbf{z}_{\mathbf{I}^t}, \mathbf{z}_{\mathrm{txt}}, \mathbf{z}_{\mathrm{loc}}$ are linear projections of the image embedding for each of the modalities (see Fig. 2). Eqn 1 shows the objective. + +$$ +\min _ {\theta} \left[ \underbrace {\mathcal {L} \left(\mathbf {Z} _ {\mathbf {I} ^ {t _ {1}} , \mathbf {Z} _ {T \left(\mathbf {I} ^ {t _ {2}}\right)}}\right)} _ {\mathcal {L} _ {\mathrm {i m g}}} + \underbrace {\mathcal {L} \left(\mathbf {Z} _ {\mathrm {t x t}} , \mathbf {E} _ {\mathrm {t x t}}\right)} _ {\mathcal {L} _ {\mathrm {t x t}}} + \underbrace {\mathcal {L} \left(\mathbf {Z} _ {\mathrm {l o c}} , \mathbf {E} _ {\mathrm {l o c}}\right)} _ {\mathcal {L} _ {\mathrm {l o c}}} \right] \tag {1} +$$ + +We compute the distance between two sets of embeddings $\mathbf{Z}$ and $\mathbf{E}$ using a minibatch of $n$ samples with the $i$ -th embedding in $\mathbf{Z}$ aligned with the $i$ -th embedding of $\mathbf{E}$ . That is, given two sets of embeddings $\mathbf{Z} = \{\mathbf{z}_1,\dots ,\mathbf{z}_n\}$ and $\mathbf{E} = \{\mathbf{e}_1,\dots ,\mathbf{e}_n\}$ , the embedding $\mathbf{z}_i$ is matched to $\mathbf{e}_i$ against other embeddings $\mathbf{e}_{1,\dots,n}$ using the loss in Eqn. 2. + +$$ +\mathcal {L} (\mathbf {Z}, \mathbf {E}) = \frac {1}{2 n} \sum_ {i = 1} ^ {n} \left(\mathcal {L} _ {\text {c o n}} \left(\mathbf {z} _ {i}, \mathbf {e} _ {1, \dots , n}\right) + \mathcal {L} _ {\text {c o n}} \left(\mathbf {e} _ {i}, \mathbf {z} _ {1, \dots , n}\right)\right) \tag {2} +$$ + +Based on InfoNCE loss [55, 59], Eqn. 3 matches the embedding $\mathbf{z}_i$ with the corresponding embedding $\mathbf{e}_i$ by minimizing the distance between them with respect to the other embeddings in $\mathbf{E}$ , with temperature hyperparameter $\tau$ . + +$$ +\mathcal {L} _ {\text {c o n}} \left(\mathbf {z} _ {i}, \mathbf {e} _ {1, \dots , n}\right) = - \log \frac {\exp \left(\mathbf {z} _ {i} \cdot \mathbf {e} _ {i} / \tau\right)}{\sum_ {j = 1} ^ {n} \exp \left(\mathbf {z} _ {i} \cdot \mathbf {e} _ {j} / \tau\right)} \tag {3} +$$ + +# 3.3. Implementation Details + +During training, we fine-tune all satellite image encoders and added linear layers on the species observation dataset + +using Eqn. 1. For models pre-trained on out-of-domain datasets (e.g. ImageNet1K [14]), we apply parameter-efficient fine-tuning (PEFT) tailored to each architecture: ResNet50 uses scale and shift fine-tuning [21, 40], tuning only BatchNorm parameters, while ViT and Swin [42] use DoRa [49] on the attention layers. These techniques enable gradual parameter updates, allowing models to learn new satellite image features without forgetting those from their original domain. For a randomly initialized model or a model pre-trained in the same domain (i.e. satellite images), we fine-tune all parameters. + +# 4. Dataset + +To train the model, we combine images, text, location, and environmental covariates from publicly available datasets [4, 12, 18, 20, 25]. For a given species, we obtain its corresponding observation data through iNaturalist [71], and a text description of its preferred habitat from its corresponding Wikipedia [4] page (Fig. 1). Satellite images are then retrieved based on the species observation locations. We describe each component of the dataset below. + +Location observation data. iNaturalist [71] observations consist of latitude and longitude values denoting the locations where a species has been observed. We use the dataset from SINR [12] that contains 35.5 million observations of 47,375 different species. It is composed of $\{(\mathbf{a}_i,b_i)\}_{i = 1}^N$ pairs where $\mathbf{a}_i = (lat,lon)$ is the location and $b_{i}\in \{s_{1},s_{2},\ldots ,s_{M}\}$ is an integer encoding the species ID of the observed species. The locations of each species are used as the basis for matching other sources of data such as environmental covariates, satellite images, and text. + +Environmental covariates. Environmental covariates are obtained from WorldClim2 [20] for a location $\mathbf{a}_i$ . The data aggregate temperature and precipitation values and are often used for ecological applications. It returns a vector for each location $\mathrm{env}(\mathbf{a}_i) \in \mathbb{R}^{20}$ . We use five-minute resolution data and bilinearly interpolate between data points to get values for specific locations. + +Satellite images. Sentinel-2 satellite images are used with a resolution of ten meters per pixel and a size of $512 \times 512$ . The satellite provides 13 spectral bands at different resolutions and has a revisit frequency of five days near the equator. We remove satellite images with significant cloud cover. Satellite images taken at the same locations but at a different time are collected as a form of augmentation for training. A total of 305,689 satellite images are collected [5, 18]. + +Text. For each species, we use readily available text data. The text is taken from Wikipedia [4] as in LE-SINR [25]. The corresponding page typically has several sections describing different aspects of the species such as its habitat, range, behavior, taxonomy, etc. Each section is processed separately so that a text embedding corresponds to a single section of the Wikipedia text. Similar to [25], we remove + +sections that do not provide information about the species (e.g. references, bibliography, links). The final text dataset contains 127,484 sections from 37,889 species. + +For a single satellite image-species location match, there could be multiple text embeddings that correspond to the multiple sections of the species Wikipedia [4] page. Taking these into account, there are a total of 980,376 training samples with location, satellite image, and text. During training, we randomly sample one section of text for every satellite image-species location match, resulting in effectively 134,890 iterations per epoch. We show the spatial distribution of the data in Fig. A1 (Appendix), and provide more training details in the Appendix. + +# 5. Experiments + +We evaluate the representations learned by WildSAT via linear probing experiments. Starting with different models and different parameter initializations (either random or pre-trained), we evaluate the performance before and after fine-tuning. When probing for each downstream dataset, the trained satellite image encoder is frozen and a randomly initialized decoder is added (Fig. 2b.1). For all tasks except segmentation, a linear layer is used for the decoder. Segmentation tasks use a convolutional-based decoder with a U-Net architecture [61]. Only the decoder is trained for each downstream task to assess the impact of the image embedding $\mathbf{z}$ without diluting its representation. + +# 5.1. Satellite Image Classification and Segmentation + +Nine remote sensing classification datasets were used as downstream tasks to evaluate the performance of the image embeddings on classification and segmentation. For classification, we evaluate on UCM [78], AID [76], RE-SISC45 [8], FMoW [9], EuroSAT [28], So2Sat20k [38, 84], and BigEarthNet20k (BEN20k) [38, 65]. For segmentation, we evaluate on Cashew1k [79] and SAcrop3k [3]. Classes vary from man-made structures (e.g. airplanes, buildings) to land type (e.g. forest, vegetation). Of the seven datasets, one (BEN20k) is multi-label, and the rest are single-label. We report accuracy for single-label classification datasets, micro-F1 for BEN20k, and IoU for both multi-class segmentation datasets. Datasets are described in the Appendix. + +# 5.2. Bird Species Encounter Rate Prediction + +We further evaluate our satellite image representations by predicting species encounter rates from satellite images. SatBird [67] introduces a benchmark for predicting bird species encounter rates for an area given the satellite image. Encounter rates for a location are given as a vector $\mathbf{r} \in [0,1]^S$ for $S$ species, where the $i$ -th element is the probability of observing species $i$ at that location. The benchmark includes three subsets: (1) USA summer, (2) USA winter, and (3) Kenya. Models are evaluated using top- $k$ + +accuracy, which compares the species with the top $k$ predicted encounter rates with the actual species observed in an area. $k$ is the actual number of species present. + +# 5.3. Base Models + +Base models refer to the different pre-trained encoders before we fine-tune with WildSAT. We experiment on 12 pre-training methods spanning random initialization, indomain pre-training, and out-of-domain pre-training. These cover different architectures ResNet50, Swin-T, Swin-B, ViT-B/16, and ViT-L/16 for a total of 20 base models. + +ImageNet [14] models are pre-trained with supervision on the updated ImageNet1K V2 dataset [60]. Previous works use V1 of the dataset [14], but we opted for the updated version which improves performance on the ImageNet benchmark by $3 - 4\%$ . We include results on the ImageNet1K V1 pre-trained models in Tab. A2 (Appendix). + +MoCov3 [7] models are pre-trained using self-supervision with InfoNCE [55] on ImageNet1K. + +CLIP [59] uses a contrastive objective to train an image and text encoder on image-text pairs from the Internet. + +Prithvi-100M [32] models are pre-trained using self-supervision (MAE [27]) on the Harmonized Landsat Sentinel-2 (HLS) [11] data. + +SatCLIP [35] is a self-supervised approach that uses paired location and satellite images from Sentinel-2. While the model has both a location and image encoder, we only use their image encoder for further fine-tuning and evaluation. + +SatlasNet [5] models are pre-trained using supervised learning on the SatlasPretrain dataset. The supervision spans a variety of label types ranging from segmentation to object detection and image classification. + +SeCo [48] models are self-supervised on Setinel-2 images augmented in time. + +TaxaBind [63] uses contrastive learning to learn a common embedding space for multiple modalities including satellite image, ground image, audio, and taxonomic text. + +GRAFT [47] uses contrastive learning with the CLIP encoders to align ground images to satellite images and text. + +RemoteCLIP [41] uses contrastive learning with CLIP to align various annotations of aerial and satellite images. + +SatMAE [13] is a self-supervised model that uses MAE [27] on temporal multi-spectral satellite imagery. + +Random denotes randomly initialized models. + +# 6. Results and Discussion + +# 6.1. Classification and Segmentation Performance + +Fig. 3 and Tab. 1 display the results on the 7 downstream classification datasets and 2 segmentation datasets, respectively, across 15 different architectures and pre-training methods. The addition of WildSAT improves 108 of the 115 settings with an overall average improvement ranging + +![](images/e4023e411a25d161f88e2e75701c303ca9b93b3105e389211e0e3d8296fbf64e.jpg) + +![](images/13956d764e4bb6e6734dcf94abeff39270c2ad2072798cf29d771447828b736e.jpg) + +![](images/de616b871be97ec169624e694df23fb65880a1f48fa7d9e8c009d3a58838f6e9.jpg) + +![](images/bb5ab801e295fa622078ebd1561d05788a0002626da6659ce9c6ac4159b64d40.jpg) + +![](images/d0367d616b18a8b9262d49ce9aff18d32c5c2e919f24db5b9c709c44ec160a25.jpg) +Figure 3. Linear probing performance improvement on seven downstream datasets without (Base) and with WildSAT (+WS) fine-tuning. Accuracy is visualized for all dataset plots except BEN20k that visualizes micro F1 score. For each architecture and pre-training combination, the horizontal line marker represents the performance of the original model, while the triangle marker indicates performance after additional training on species observation data (WildSAT). The tables on the right summarize average performance across all seven datasets: the top table includes models with random weights, and the bottom table excludes them. Across the board, fine-tuning with species observation data leads to notable performance gains over most base models. We include the raw numbers in Tab. A1 (Appendix). *Both Prithvi-100M and SatCLIP are pre-trained with multispectral images, but for consistency across downstream datasets and models, only RGB bands are used here. We show that WildSAT also improves on multispectral images in Tab. A4 (Appendix). + +![](images/fbf1dfdc73d7777b2832b7c350422d3603d61586c8b6af2e3e2e742b2f2688e2.jpg) + +![](images/a17ab47de8c4645140d6d5ce7a13dd96825061a37629d96d4c0ffc426ce7c663.jpg) + +![](images/6c0ade3e794677977729ae5290a68ad43a24c772aa2ece80b854d34c77f5d2e1.jpg) + +
Average (w/ random)
DatasetBase+WS
AID [76]61.277.0
BEN20k [38, 65]38.553.1
EuroSAT [28]80.293.8
FMOW [9]33.441.1
RESISC45 [8]65.381.0
So2Sat20k [38, 84]32.647.6
UCM [78]68.886.1
+ +
Average (no random)
DatasetBase+WS
AID [76]72.779.4
BEN20k [38, 65]45.753.4
EuroSAT [28]88.994.3
FMoW [9]39.043.3
RESISC45 [8]77.883.5
So2Sat20k [38, 84]37.948.2
UCM [78]81.887.9
+ +
Cashew1k [79]SAcrop3k [3]
Base+WSBase+WS
ImageNet [14]70.3%70.6%24.3%25.0%
MoCov3 [7]71.4%73.3%22.9%24.9%
SeCo [48]62.6%73.3%22.3%22.8%
SatlasNet [5]55.2%71.0%19.4%20.5%
Random40.1%72.6%18.0%20.3%
Average59.9%72.2%21.4%22.7%
+ +Table 1. Downstream satellite image segmentation results, reported using IoU, show WildSAT (+WS) improving on existing models. Qualitative results are available in Fig. A3 (Appendix). + +
TaxaBind [63]GRAFT [47]RemoteCLIP [41]CLIP [59]WildSAT (Ours)
Average Performance59.8%65.0%71.0%71.6%76.6%
+ +from $7.7\%$ to $17.4\%$ in the different datasets (4.3% to 10.4% without the randomly initialized models). + +WildSAT improves satellite image representations. The results in Fig. 3 and Tab. 1 highlight the performance improvements WildSAT contributes. These improvements may be attributed to our use of diverse supervision—integrating images, text embeddings, and species data at scale. This strategy ultimately helps in downstream + +Table 2. Average linear probing performance across all seven satellite image classification datasets of models based on CLIP. TaxaBind, GRAFT, and RemoteCLIP fine-tune CLIP and use additional modalities (e.g. text, satellite images, ground images) for cross-modal tasks. We show that fine-tuning CLIP with WildSAT improves performance on CLIP compared to other CLIP-based models. We include accuracy per dataset in Tab. A3 (Appendix). + +
KenyaUSA SummerUSA Winter
Base+WSBase+WSBase+WS
SatlasNet [5]23.9024.4048.5950.0354.0255.01
SatMAE [13]23.6623.7546.0948.6952.4053.86
TaxaBind [63]23.8324.2048.3149.8753.7655.00
Average23.8024.1247.6649.5353.3954.62
+ +Table 3. Top- $k$ accuracy for bird species encounter rate prediction on the SatBird [67] dataset. Top- $k$ is defined as the accuracy of predicting the $k$ species present in an area. + +tasks, particularly for both increasing true positive rates on classes related to habitats (e.g. forests, deserts), while reducing false positives on the same types of classes. We show how classes related to wildlife habitats improve performance in the Appendix. Meanwhile, Tab. 2 provides a comparison of WildSAT (CLIP ViT-B/16 as base) with TaxaBind [63], GRAFT [47], and RemoteCLIP [41] all of which also fine-tune CLIP using cross-modal supervision. While the latter three methods show improvements on cross-modal tasks, their performance when linear probed on satellite imagery tasks suffers compared to CLIP, indicating a degree of "forgetting" of the representations. In contrast, our method not only outperforms standard CLIP but also achieves the best overall results. + +Satellite pre-trained models see a larger boost in performance. Fig. 3 demonstrates that training with WildSAT can improve performance by as much as $10\%$ on satellite pre-trained models such as SeCo and SatlasNet. While models pre-trained on ImageNet and MoCov3 also see performance improvements, we see less improvements on the + +![](images/5571438aca839965587195425ae9960aaa0588448ad385b785b2cee99cb733cd.jpg) +Figure 4. Zero-shot results for text-based satellite image retrieval. The columns show the top 4 images returned given the text query on top. A model can be queried using general landscape descriptions (e.g. 'desert', 'ocean', 'forest'). In addition, specific wildlife text such as 'ibex' and 'house finch' can be used as queries to view the types of environment they inhabit. The 'ibex' inhabits mountains and high elevation areas, 'american alligators' are found in swamps and the coast, the 'house finch' is commonly found in urban areas—consistent with the retrieved satellite images shown. More examples are available in Fig. A4 (Appendix). + +AID, RESISC45, and FMoW datasets. This could be attributed to the three datasets having less categories related to species habitats (e.g. storage tank, airport, church). Additionally, since ImageNet and MoCov3 pre-trained models are already performing effectively, there is little room for improvement. WildSAT is trained on a dataset that is specifically geared towards habitats and land characteristics. Thus, we see more improvements in So2Sat20k and BEN20k, which cover climate zones and land cover types. + +WildSAT reduces errors on habitat-related classes. We show through a confusion matrix of So2Sat20k classes in Fig. A2 (Appendix) that WildSAT achieves higher true positive counts across all types of classes (including anthropogenic classes) by reducing false positives on habitat-related categories (e.g. 'Scattered trees', 'Dense trees'). This reduction in misclassifications contributes to overall performance gains, explaining the improvements observed across satellite image tasks. + +Larger performance gains on ViTs than ResNets. While the ResNet50 ImageNet and MoCov3 models show smaller performance gains, their ViT counterparts show significantly larger improvements (Fig. 3). Consistent with observations from previous work [59], we see better performance with the addition of other modalities when using transformers. This could be attributed to the more flexible representations of ViTs, unlike CNNs that inherently incorporate a strong inductive bias from the use of convolution. Using attention layers in ViTs likely makes their embeddings more adaptable to other modalities such as text and location. + +# 6.2. Bird Species Encounter Rate Prediction + +Tab. 3 shows WildSAT complementing existing methods on bird species encounter rate prediction. We take models SatlasNet (Swin-B) [5] and SatMAE (ViT-L/16) [13] presented in SatBird [67] that use satellite images to predict encounter rates of all bird species in an area. TaxaBind [63] is also added as a baseline. Using WildSAT improves these existing methods. In addition, since most of the data used in WildSAT are from the USA and Europe, we see more performance improvement in USA bird encounter rates. + +# 6.3. Zero-shot Image Retrieval + +When trained using our WildSAT framework, we observe that models learn wildlife-specific attributes. By using the frozen satellite image encoder and a large language model, a user can input text to query satellite images. The top $k$ images with the most similar embeddings to the text embeddings (computed using cosine similarity) can be retrieved (Fig. 2b.2). Fig. 4 displays examples of satellite images retrieved given different text queries. General descriptions of landscapes or locations can be used for querying such as 'desert', 'ocean', 'forest', or 'grassland'. At the same time, specific wildlife text can also be used as queries such as 'ibex' or 'gull'. We see that zero-shot retrieval returns images of the habitat of the corresponding wildlife. This feature could potentially be used to find habitats for species with limited observation data, and can serve as a reference for species distribution studies and biodiversity monitoring. + +# 6.4. Ablations + +Our WildSAT framework is composed of multiple components: a satellite image encoder, location encoder, and text encoder. We investigate the contribution of each of these components in an ablation study. In Tab. 4 we ablate two models: a randomly initialized ResNet50 and an ImageNet pre-trained ViT-B/16. Columns with check marks on 'loc' and/or 'env' indicate the use of the location encoder ( $\mathcal{L}_{\mathrm{loc}}$ term in Eqn. 1). The location encoder can use either only location as input or use both location and environmental covariates. The 'text' and 'img-a' columns indicate the use of the large language model ( $\mathcal{L}_{\mathrm{txt}}$ ) and contrastive loss on the augmented satellite images ( $\mathcal{L}_{\mathrm{img}}$ ), respectively. A complete ablation of all combinations of WildSAT components for other models is also provided in Tab. A7-A9 (Appendix). + +Each modality enhances performance. Larger improvements are observed for the randomly initialized model, given its lower starting accuracy and greater potential for improvement (Tab. 4). Nonetheless, similar improvement trends are seen with both types of models. Improvements from location and environmental covariates can be from associating satellite images with particular locations and corresponding environmental characteristics (i.e. satellite images are mapped to specific latitudes and longitudes that + +
locenvtextimg-aUCM [78]AID [76]RESISC45 [8]So2Sat20k [84]Average
Random ResNet50ImageNet ViT-B/16Random ResNet50ImageNet ViT-B/16Random ResNet50ImageNet ViT-B/16Random ResNet50ImageNet ViT-B/16
24.3%93.2%25.2%84.4%25.5%88.2%5.9%41.8%48.6%
✓44.2%95.0%41.6%85.1%43.0%89.3%18.3%43.4%57.5%
✓✓60.0%95.0%48.7%86.2%48.2%88.8%25.2%44.2%62.0%
✓✓✓70.0%95.4%55.6%86.2%58.9%89.7%27.9%45.0%66.1%
✓✓✓✓79.9%97.2%68.2%88.9%74.7%93.0%42.3%55.2%74.9%
+ +Table 4. Ablation of various components of WildSAT. The best performance is a combination of all components: location (loc), environmental covariates (env), text (text), and satellite image augmentations (img-a). Results are presented for a randomly initialized ResNet50 model and an ImageNet pre-trained ViT-B/16 model with the numbers referring to linear probing accuracy. + +are along the coast, or in the mountains). The addition of text further improves the performance while enabling zero-shot image retrieval capabilities. Text provides a rich source of information with more detailed descriptions of areas. Adding the satellite image loss term likely improves general image understanding, such as the model learning the rotation invariance of satellite images. Overall, each component of WildSAT contributes to performance improvements. + +Different modalities can strengthen models by covering model-specific gaps. We hypothesize that models trained on satellite imagery datasets benefit primarily from location and text supervision associated with wildlife observation. This hypothesis is supported by observations that self-supervised models such as SeCo, SatMAE, and Prithvi—trained with objectives similar to WildSAT's image self-supervision term on satellite datasets—still achieve significant gains (Fig. 3). Similarly, models like SatlasNet, which are trained with large-scale supervised learning on satellite images, also benefit. On the other hand, ImageNet pre-trained models benefit from the additional satellite image supervision. These results highlight the complementary nature of WildSAT's supervision compared to existing datasets, which primarily focus on anthropogenic labels. Tab. A7 (Appendix) further supports this finding. + +PEFT preserves out-of-domain pre-training representations. Applying PEFT to out-of-domain pre-trained models (e.g. ImageNet, CLIP) helps preserve their original representations, preventing performance degradation—a crucial advantage for large models like ViT, where fine-tuning all weights risks shifting parameters suboptimally (Tab. 2). In contrast, in-domain pre-trained models (e.g. SeCo) benefit more from full fine-tuning, as WildSAT's similar data leads to a non-disruptive shift as shown in Tab. A9 (Appendix). + +SINR location encoder achieves the best performance. We explore different location encoders in Tab. A8 (Appendix). Comparing no location encoder (i.e. using position encoded latitude and longitude), SINR, and SatCLIP, SINR delivers best overall performance, likely due to its more informative location representations that better incorporate species habitat information. + +# 6.5. Limitations + +While we show that WildSAT improves satellite image representations and has promising zero-shot performance, the datasets we train on contain inherent limitations that could affect model use. The US and Europe are overrepresented in the data as most citizen contributed data currently come from these locations. Underrepresented areas in Asia, Africa, Australia, and South America are likely to see less accurate results especially on satellite image retrieval. LLMs are further prone to generating hallucinations, which could impact model output reliability. + +# 7. Conclusion + +While satellite images are often used to interpolate sparse wildlife observations to create species range maps, our work demonstrates that these observations also provide a rich source of supervision for learning satellite image representations. WildSAT can not only learn high-quality representations from scratch but also improve performance of strong pre-trained models, such as those trained on ImageNet and satellite imagery datasets, across a range of satellite imagery tasks. We attribute this success to the global scale of community efforts, which document diverse wildlife observations through platforms like iNaturalist and eBird and record detailed species attributes on sources like Wikipedia. This supervisory signal complements existing satellite datasets, which often focus on anthropogenic labels, by introducing a broader ecological perspective. As these resources continue to expand in both observational scale (e.g. geographical and taxonomic scope) and modality (e.g. incorporating sound, aerial imagery), they offer even greater potential for improving WildSAT's representations. + +Acknowledgements. We thank the iNaturalist community for providing the data used for training. Experiments were performed on the University of Massachusetts GPU cluster funded by the Mass. Technology Collaborative. RD and SM were supported in part by NASA grant 80NSSC22K1487 and NSF grant 2329927. + +# References + +[1] eBird. https://www.ebird.org. Accessed on 2025-03-03. 13 +[2] iNaturalist. https://www.inaturalist.org. Accessed on 2025-03-03. 2, 13 +[3] Planet, Radiant Earth Foundation, Western Cape Department of Agriculture, German Aerospace Center (DLR): A fusion dataset for crop type classification in Western Cape, South Africa. https://source.coop/esa/fusion-competition. Accessed on 2025-03-03. 5, 6, 13, 17 +[4] Wikipedia. https://www.wikipedia.org. Accessed on 2025-03-03. 3, 4, 5, 14 +[5] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinando, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 2, 3, 4, 5, 6, 7, 14, 17, 19 +[6] Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N Belhumeur. Bird-snap: Large-scale fine-grained visual categorization of birds. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2014. 2 +[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 5, 6, 14, 17 +[8] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 5, 6, 8, 13, 14, 15, 18, 19 +[9] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6172-6180, 2018. 5, 6, 13, 14, 15, 19 +[10] Grace Chu, Brian Potetz, Weijun Wang, Andrew Howard, Yang Song, Fernando Brucher, Thomas Leung, and Hartwig Adam. Geo-aware networks for fine-grained recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019. 2 +[11] Martin Claverie, Junchang Ju, Jeffrey G Masek, Jennifer L Dungan, Eric F Vermote, Jean-Claude Roger, Sergii V Skakun, and Christopher Justice. The harmonized landsat and sentinel-2 surface reflectance data set. Remote sensing of environment, 219:145–161, 2018. 5 +[12] Elijah Cole, Grant Van Horn, Christian Lange, Alexander Shepard, Patrick Leary, Pietro Perona, Scott Loarie, and Oisin Mac Aodha. Spatial implicit neural representations for global-scale species mapping. In International Conference on Machine Learning, pages 6320-6342. PMLR, 2023. 3, 4, 18, 19 +[13] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 2, 5, 6, 7, 15 + +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 3, 4, 5, 6, 14, 15, 17, 18 +[15] Aayush Dhakal, Subash Khanal, Srikumar Sastry, Adeel Ahmad, and Nathan Jacobs. Geobind: Binding text, image, and audio through satellite images. In International Geoscience and Remote Sensing Symposium, 2024. 1, 2, 3 +[16] Johannes Dollinger, Philipp Brun, Vivien Sainte Fare Garnot, and Jan Dirk Wegner. Sat-sinr: High-resolution species distribution models through satellite imagery. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2024. 1, 3 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3 +[18] ESA. Sentinel-1-missions-sentinel online-sentinel online. Eur. Sp. Agency, 2022. 2, 4, 14 +[19] Joshua Fan, Junwen Bai, Zhiyun Li, Ariel Ortiz-Bobea, and Carla P Gomes. A gnn-rnn approach for harnessing geospatial and temporal information: application to crop yield prediction. In Proceedings of the AAAI conference on artificial intelligence, pages 11873-11881, 2022. 1 +[20] Stephen E Fick and Robert J Hijmans. Worldclim 2: new 1-km spatial resolution climate surfaces for global land areas. International journal of climatology, 37(12):4302-4315, 2017. 4 +[21] Jonathan Frankle, David J Schwab, and Ari S Morcos. Training batchnorm and only batchnorm: On the expressive power of random features in cnns. In International Conference on Learning Representations, 2021. 4, 19 +[22] Rollan C Geronimo, Erik C Franklin, Russell E Brainard, Christopher D Elvidge, Mudjekeewis D Santos, Roberto Venegas, and Camilo Mora. Mapping fishing activities and suitable fishing grounds using nighttime satellite images and maximum entropy modelling. Remote Sensing, 10(10):1604, 2018. 1 +[23] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, Huimei He, Jian Wang, Jingdong Chen, Ming Yang, Yongjun Zhang, and Yansheng Li. Skysense: A multimodal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27672-27683, 2024. 2 +[24] Mordechai Haklay and Patrick Weber. Openstreetmap: User-generated street maps. IEEE Pervasive computing, 7(4):12-18, 2008. 2 +[25] Max Hamilton, Christian Lange, Elijah Cole, Alexander Shepard, Samuel Heinrich, Oisin Mac Aodha, Grant Van Horn, and Subhransu Maji. Combining observational data and language for species range estimation. In Advances in Neural Information Processing Systems, 2024. 3, 4 + +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[27] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 5 +[28] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 1, 2, 5, 6, 13, 14, 15, 16, 19 +[29] Andy V Huynh, Lauren E Gillespie, Jael Lopez-Saucedo, Claire Tang, Rohan Sikand, and Moises Exposito-Alonso. Contrastive ground-level image and remote sensing pretraining improves representation learning for natural world imagery. In European Conference on Computer Vision, 2024. 2 +[30] Markus Immitzer, Francesco Vuolo, and Clement Atzberger. First experience with sentinel-2 data for crop and tree species classifications in central europe. Remote sensing, 8(3):166, 2016. 2 +[31] Pallavi Jain, Diego Marcos, Dino Ienclo, Roberto Interdonato, Aayush Dhakal, Nathan Jacobs, and Tristan Berchoux. Aligning geo-tagged clip representations and satellite imagery for few-shot land use classification. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, pages 319-323. IEEE, 2024. 2 +[32] Johannes Jakubik, S Roy, CE Phillips, P Fraccaro, D Godwin, B Zadrozny, D Szwarcman, C Gomes, G Nyirjesy, B Edwards, et al. Foundation models for generalist geospatial artificial intelligence, 2023. URL https://arxiv.org/abs/2310.18660, 2023. 1, 2, 3, 5, 14, 15 +[33] Priyabrata Karmakar, Shyh Wei Teng, Manzur Murshed, Shaoning Pang, Yanyu Li, and Hao Lin. Crop monitoring by multimodal remote sensing: A review. Remote Sensing Applications: Society and Environment, 33:101093, 2024. 1 +[34] Subash Khanal, Srikumar Sastry, Aayush Dhakal, and Nathan Jacobs. Learning tri-modal embeddings for zero-shot soundscape mapping. In BMVC, 2023. 2 +[35] Konstantin Klemmer, Esther Rolf, Caleb Robinson, Lester Mackey, and Marc Rußwurm. Satclip: Global, general-purpose location embeddings with satellite imagery. In AAAI Conference on Artificial Intelligence, 2024. 2, 3, 5, 14, 15, 18, 19 +[36] Andrey A Kurekin, Benjamin R Loveday, Oliver Clements, Graham D Quartly, Peter I Miller, George Wiafe, and Kwame Adu Agyekum. Operational monitoring of illegal fishing in ghana through exploitation of satellite earth observation and ais data. Remote Sensing, 11(3):293, 2019. 1 +[37] Nataliia Kussul, Mykola Lavreniuk, Sergii Skakun, and Andrii Shelestov. Deep learning classification of land cover and crop types using remote sensing data. IEEE Geoscience and Remote Sensing Letters, 14(5):778-782, 2017. 1 + +[38] Alexandre Lacoste, Nils Lehmann, Pau Rodriguez, Evan Sherwin, Hannah Kerner, Björn Lütjens, Jeremy Irvin, David Dao, Hamed Alemohammad, Alexandre Drouin, et al. Geobench: Toward foundation models for earth monitoring. Advances in Neural Information Processing Systems, 36, 2023. 5, 6, 13, 16 +[39] Nico Lang, Walter Jetz, Konrad Schindler, and Jan Dirk Wegner. A high-resolution canopy height model of the earth. Nature Ecology & Evolution, 2023. 1 +[40] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 4, 19 +[41] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 3, 5, 6, 15, 16 +[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 4 +[43] Titouan Lorieul, Elijah Cole, Benjamin Deneu, Maximilien Servajean, Pierre Bonnet, and Alexis Joly. Overview of geolifeclef 2022: Predicting species presence from multi-modal remote sensing, bioclimatic and pedologic data. In CLEF (Working Notes), pages 1940-1956, 2022. 3 +[44] Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presence-only geographical priors for fine-grained image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. 2 +[45] Gengchen Mai, Ni Lao, Yutong He, Jiaming Song, and Stefano Ermon. Csp: Self-supervised contrastive spatial pre-training for geospatial-visual representations. In International Conference on Machine Learning, pages 23498–23515. PMLR, 2023. 3 +[46] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 1, 2, 15 +[47] Utkarsh Mall, Cheng Perng Phoo, Meilin Kelsey Liu, Carl Vondrick, Bharath Hariharan, and Kavita Bala. Remote sensing vision-language foundation models without annotations via ground remote alignment. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 15, 16 +[48] Oscar Manas, Alexandre Lacoste, Xavier Giroi-nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 1, 2, 3, 5, 6, 14, 15, 17, 18, 19 +[49] Yulong Mao, Kaiyu Huang, Changhao Guan, Ganglin Bao, Fengran Mo, and Jinan Xu. Dora: Enhancing parameter-efficient fine-tuning with dynamic rank distribution. arXiv preprint arXiv:2405.17357, 2024. 4, 19 + +[50] Jeffrey G Masek, Michael A Wulder, Brian Markham, Joel McCorkel, Christopher J Crawford, James Storey, and Del T Jenstrom. Landsat 9: Empowering open science and applications through continuity. Remote Sensing of Environment, 248:111968, 2020. 2 +[51] Aaron E Maxwell, Timothy A Warner, Brian C Vanderbilt, and Christopher A Ramezan. Land cover classification and feature extraction from national agriculture imagery program (naip) orthoimagery: A review. Photogrammetric Engineering & Remote Sensing, 83(11):737-747, 2017. 2 +[52] Niklas Muennighoff, SU Hongjin, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. Generative representational instruction tuning. In ICLR 2024 Workshop: How Far Are We From AGI, 2024. 3 +[53] Vishal Nedungadi, Ankit Kariryaa, Stefan Oehmcke, Serge Belongie, Christian Igel, and Nico Lang. Mmearth: Exploring multi-modal pretext tasks for geospatial representation learning. In European Conference on Computer Vision, 2024. 2 +[54] Maxim Neumann, Andre Susano Pinto, Xiaohua Zhai, and Neil Houlsby. In-domain representation learning for remote sensing. arXiv preprint arXiv:1911.06721, 2019. 2 +[55] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4, 5 +[56] Fernando Paolo, Tsu-ting Tim Lin, Ritwik Gupta, Bryce Goodman, Nirav Patel, Daniel Kuster, David Kroodsma, and Jared Dunnmon. xview3-sar: Detecting dark fishing activity using synthetic aperture radar imagery. Advances in Neural Information Processing Systems, 2022. 1 +[57] Gustavo Perez, Subhransu Maji, and Daniel Sheldon. Discount: counting in large image collections with detector-based importance sampling. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 22294-22302, 2024. 1 +[58] Lukas Picek, Christophe Botella, Maximilien Servajean, César Leblanc, Rémi Palard, Théo Larcher, Benjamin Deneu, Diego Marcos, Pierre Bonnet, and Alexis Joly. Geoplant: Spatial plant species prediction dataset. Advances in Neural Information Processing Systems - Datasets and Benchmark, 2024. 2 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4, 5, 6, 7, 14, 15, 16, 19 +[60] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International conference on machine learning, pages 5389-5400, 2019. 5, 15, 19 +[61] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI, pages 234-241. Springer, 2015. 5, 17 +[62] Srikumar Sastry, Subash Khanal, Aayush Dhakal, Di Huang, and Nathan Jacobs. Birdsat: Cross-view contrastive masked + +autoencoders for bird species classification and mapping. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7136-7145, 2024. 2, 3 +[63] Srikumar Sastry, Subash Khanal, Aayush Dhakal, Adeel Ahmad, and Nathan Jacobs. Taxabind: A unified embedding space for ecological applications. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2025. 1, 2, 3, 5, 6, 7, 15, 16 +[64] Brian L Sullivan, Christopher L Wood, Marshall J Iliff, Rick E Bonney, Daniel Fink, and Steve Kelling. ebird: A citizen-based bird observation network in the biological sciences. Biological conservation, 142(10):2282-2292, 2009. 2 +[65] Gencer Sumbul, Marcela Charfuelan, Begüm Demir, and Volker Markl. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium, pages 5901-5904. IEEE, 2019. 2, 5, 6, 13, 14, 15, 19 +[66] Wenjuan Sun, Paolo Bocchini, and Brian D Davison. Applications of artificial intelligence for disaster management. Natural Hazards, 103(3):2631-2689, 2020. 1 +[67] Mélisande Teng, Amna Elmustafa, Benjamin Akera, Yoshua Bengio, Hager Radi, Hugo Larochelle, and David Rolnick. Satbird: a dataset for bird species distribution modeling using remote sensing and citizen science data. Advances in Neural Information Processing Systems - Datasets and Benchmarks, 36, 2023. 3, 5, 6, 7 +[68] Catherine Torres de Almeida, Jéssica Gerente, Jamerson Rodrigo dos Prazeres Campos, Francisco Caruso Gomes Junior, Lucas Antonio Providelo, Guilherme Marchiori, and Xinjian Chen. Canopy height mapping by sentinel 1 and 2 satellite images, airborne lidar data, and machine learning. Remote Sensing, 14(16):4112, 2022. 1 +[69] Burak Uzkent, Evan Sheehan, Chenlin Meng, Zhongyi Tang, Marshall Burke, David Lobell, and Stefano Ermon. Learning to interpret satellite images using wikipedia. In Proceedings of the International Joint Conference on Artificial Intelligence, 2019. 2, 3 +[70] Ruben Van De Kerchove, Daniele Zanaga, Wanda Keersmaecker, Niels Souverijns, Jan Wevers, Carsten Brockmann, Alex Grosu, Audrey Paccini, Oliver Cartus, Maurizio Santoro, et al. Esa worldwide: Global land cover mapping at $10\mathrm{m}$ resolution for 2020 based on sentinel-1 and 2 data. In AGU Fall Meeting Abstracts, pages GC45I-0915, 2021. 2 +[71] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 4 +[72] Vicente Vivanco Cepeda, Gaurav Kumar Nayak, and Mubarak Shah. Geoclip: Clip-inspired alignment between locations and images for effective worldwide geolocation. Advances in Neural Information Processing Systems, 2024. 1, 2 +[73] Stefan Voigt, Thomas Kemper, Torsten Riedlinger, Ralph Kiefl, Klaas Scholte, and Harald Mehl. Satellite image anal + +ysis for disaster and crisis-management support. IEEE transactions on geoscience and remote sensing, 45(6):1520-1528, 2007. 1 +[74] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 2 +[75] Bingfang Wu, René Gommes, Miao Zhang, Hongwei Zeng, Nana Yan, Wentao Zou, Yang Zheng, Ning Zhang, Sheng Chang, Qiang Xing, et al. Global crop monitoring: a satellite-based hierarchical approach. Remote Sensing, 7(4): 3907-3933, 2015. 1 +[76] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 2, 5, 6, 8, 13, 14, 15, 18, 19 +[77] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3974-3983, 2018. 2 +[78] Yi Yang and Shawn Newsam. Bag-of-visual-words and spatial extensions for land-use classification. In Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems, pages 270–279, 2010. 5, 6, 8, 13, 14, 15, 18, 19 +[79] Leikun Yin, Rahul Ghosh, Chenxi Lin, David Hale, Christoph Weigl, James Obarowski, Junxiong Zhou, Jessica Till, Xiaowei Jia, Nanshan You, et al. Mapping smallholder cashew plantations to inform sustainable tree crop expansion in benin. Remote Sensing of Environment, 295:113695, 2023. 5, 6, 13, 17 +[80] Qiangqiang Yuan, Huanfeng Shen, Tongwen Li, Zhiwei Li, Shuwen Li, Yun Jiang, Hongzhang Xu, Weiwei Tan, Qianqian Yang, Jiwen Wang, et al. Deep learning in environmental remote sensing: Achievements and challenges. Remote sensing of Environment, 241:111716, 2020. 1 +[81] Valerie Zermatten, Javiera Castillo-Navarro, Pallavi Jain, Devis Tuia, and Diego Marcos. Ecowikirs: Learning ecological representation of satellite images from weak supervision with species observations and wikipedia. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2275–2285, 2025. 2 +[82] Zhuo Zheng, Stefano Ermon, Dongjun Kim, Liangpei Zhang, and Yanfei Zhong. Changen2: Multi-temporal remote sensing generative change foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[83] Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep learning in remote sensing: A comprehensive review and list of resources. Geoscience and remote sensing magazine, 5 (4):8-36, 2017. 1 + +[84] Xiao Xiang Zhu, Jingliang Hu, Chunping Qiu, Yilei Shi, Jian Kang, Lichao Mou, Hossein Bagheri, Matthias Haberle, Yuansheng Hua, Rong Huang, et al. So2sat lcz42: A benchmark data set for the classification of global local climate zones [software and data sets]. IEEE Geoscience and Remote Sensing Magazine, 8(3):76-89, 2020. 5, 6, 8, 13, 14, 15, 16, 18, 19 \ No newline at end of file diff --git a/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/images.zip b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..53c2e8bf384addf78d34367989a788663e4032eb --- /dev/null +++ b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94dad771d6657bb01e661194e61b4355e512d22cd63524d193b25158106a1714 +size 490245 diff --git a/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/layout.json b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7adc67f8d485a5f9936a53c0083f8bf43fbd6415 --- /dev/null +++ b/wildsatlearningsatelliteimagerepresentationsfromwildlifeobservations/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d65065554c6dd1182e00a71c6c351e491dd3e77e0591fbaa97edf95876e5870 +size 423073 diff --git a/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/c727212a-10ef-4c0d-a1c9-d78f5484a70b_content_list.json b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/c727212a-10ef-4c0d-a1c9-d78f5484a70b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b77d13762d7358c97ac5fd182408d773fd40ba27 --- /dev/null +++ b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/c727212a-10ef-4c0d-a1c9-d78f5484a70b_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffefe0f43ca013ed6391175012c410f82e87a24211afeffec60ea56fc933911b +size 94806 diff --git a/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/c727212a-10ef-4c0d-a1c9-d78f5484a70b_model.json b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/c727212a-10ef-4c0d-a1c9-d78f5484a70b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6799fd1fdb8f35c72ef5515ce11f738c8af4ddb4 --- /dev/null +++ b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/c727212a-10ef-4c0d-a1c9-d78f5484a70b_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:734bd764298b6def73e08aeb7e07ad89baf885afb67f1c42fd2078b83e96f119 +size 122882 diff --git a/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/c727212a-10ef-4c0d-a1c9-d78f5484a70b_origin.pdf b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/c727212a-10ef-4c0d-a1c9-d78f5484a70b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e6818ca06e3b0597fce7fce76cedfe19153dee22 --- /dev/null +++ b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/c727212a-10ef-4c0d-a1c9-d78f5484a70b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91fadd936d41b35336a842ea8a7236553de67e0ee8399c5e6ff8ca82deb8971c +size 4722948 diff --git a/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/full.md b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/full.md new file mode 100644 index 0000000000000000000000000000000000000000..57a2291c8b8a2b7c4b57a18225482c6835c32b99 --- /dev/null +++ b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/full.md @@ -0,0 +1,439 @@ +# WildSeg3D: Segment Any 3D Objects in the Wild from 2D Images + +Yansong Guo $^{1}$ , Jie Hu $^{2}$ , Yansong Qu $^{1}$ , Liujuan Cao $^{1\dagger}$ + +1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University + +$^{2}$ National University of Singapore, Singapore + +# Abstract + +Recent advances in intuitive 3D segmentation from 2D images have demonstrated impressive performance. However, current models typically require extensive scene-specific training to accurately reconstruct and segment objects, which limits their applicability in real-time scenarios. In this paper, we introduce WildSeg3D, an efficient approach that enables the segmentation of arbitrary 3D objects across diverse environments using a feed-forward mechanism. A key challenge of this feed-forward approach lies in the accumulation of 3D alignment errors across multiple 2D views, which can lead to inaccurate 3D segmentation results. To address this issue, we propose Dynamic Global Aligning (DGA), a technique that improves the accuracy of global multi-view alignment by focusing on difficult-to-match 3D points across images, using a dynamic adjustment function. Additionally, for real-time intuitive segmentation, we introduce Multi-view Group Mapping (MGM), a method that utilizes an object mask cache to integrate multi-view segmentations and respond rapidly to user prompts. WildSeg3D demonstrates robust generalization across arbitrary scenes, thereby eliminating the need for scene-specific training. Specifically, WildSeg3D not only attains the accuracy of state-of-the-art (SOTA) methods but also achieves a $40\times$ speedup compared to existing SOTA models. Code will be released at https://github.com/Ethan16162/WildSeg3D. + +# 1. Introduction + +Intuitive 3D segmentation from 2D images plays a critical role in 3D scene understanding and remains a fundamental challenge in computer vision, attracting significant attention from the research community [11, 18, 21, 38, 47-49]. This technology has broad applications across various fields, including virtual and augmented reality, real-time intuitive systems, and automatic labeling. Recent advancements in intuitive 3D segmentation have demonstrated remarkable + +performance, particularly based on Neural Radiance Fields (NeRF) [33] and 3D Gaussian Splatting (3DGS) [17]. For instance, models such as SA3D [4] and SANeRF-HQ [27] integrate NeRF with foundational segmentation models like Segment Anything Model (SAM) [19], aligning semantic information with 3D representations to enable effective 3D object segmentation. Similarly, 3DGS-based approaches [5, 15, 40, 59, 67] address the high training demands of NeRF by constructing Gaussian feature fields in combination with SAM, facilitating faster model training. However, both NeRF-based and 3DGS-based methods typically rely on extensive scene-specific training to obtain accurate 3D priors, which significantly hinders their applicability in real-time scenarios. + +To overcome this limitation, we propose WildSeg3D, a novel approach with a feed-forward manner, eliminating the need for scene-specific training. Inspired by Dust3r [51] and Mast3r [24], our feed-forward approach represents 3D scenes as pointmaps and performs scene reconstruction via global alignment. A key challenge in this process is the influence of redundant background points and the difficulty of matching 3D points across different views, which accumulates 3D alignment errors across multiple 2D views. This can lead to inaccurate 3D segmentation with confusion between background and target objects. To address this, we introduce Dynamic Global Alignment (DGA), a method that dynamically adjusts attention on view-specific points during matching, minimizing alignment errors during the global registration of pointmaps across multiple views. For real-time intuitive segmentation, we also propose a mask cache constructed during preprocessing, leveraging the multi-frame segmentation capabilities of SAM2 [43]. This cache stores consistent object masks across all viewpoints, providing offline data that can be used for subsequent real-time segmentation. We further introduce Multiview Group Mapping (MGM), which combines the mask cache with the DGA strategy to integrate multi-view segmentation results into an aligned global coordinate system. The MGM module retrieves target feature points from the mask cache based on user inputs and generates a unified 3D + +![](images/3c27aa345c3db0116396bc280aaa226053532fb436223eeddf2c8b21a270bf07.jpg) +Figure 1. Framework of WildSeg3D. WildSeg3D operates in three stages. First, during the pre-processing stage, 2D feature masks are constructed from multi-view images using SAM2, providing offline support for intuitive segmentation. Second, in the Dynamic Global Alignment (DGA) stage, a dynamic weight adjustment strategy is applied to achieve global alignment of the pointmaps generated by MASt3r, thereby improving the accuracy of object reconstruction. Finally, in the Multi-view Group Mapping (MGM) stage, multi-view masks for the target are retrieved from the mask cache based on user input, and these masks are transformed into an aligned 3D space in real time, where "P2W" refers to the transformation from pixel coordinates to the aligned world coordinates. + +mask for the target across all viewpoints. + +WildSeg3D demonstrates robust generalization across diverse scenes, transforming 2D target masks into an aligned 3D space, thus enabling accurate 3D segmentation without the need for scene-specific training. We conduct extensive experiments on WildSeg3D using complex real-world scenes to evaluate both segmentation accuracy and efficiency. Our results show that WildSeg3D not only matches the accuracy of state-of-the-art (SOTA) methods [4, 5, 11, 35, 44, 45, 48, 61] but also achieves a $40 \times$ speedup compared to the SOTA models. Specifically, WildSeg3D completes scene reconstruction in under 30 seconds, a significant reduction compared to the fastest model, SA3D, which requires 780 seconds. Additionally, real-time interaction results are delivered in just 5-20 milliseconds. + +Our main contributions can be summarized as follows: + +- We introduce WildSeg3D, the first feed-forward 3D segmentation model that operates directly from 2D views, eliminating the need for scene-specific training and enabling immediate segmentation of arbitrary 3D objects in diverse environments. +- we propose Dynamic Global Alignment (DGA), a novel method for addressing 3D alignment errors and enhanc + +ing segmentation accuracy. We also propose Multi-view Group Mapping (MGM) that enables real-time intuitive 3D segmentation with robust generalization across diverse scenes. + +- Extensive experiments on complex real-world scenes demonstrate that WildSeg3D achieves both high segmentation accuracy and significant efficiency, processing 2D views to 3D segmentation over $40 \times$ faster than existing SOTA models. + +# 2. Related Work + +# 2.1. 2D Image-based 3D Reconstruction + +Recent advancements in image-based 3D reconstruction using neural networks have led to significant progress. Innovations like Neural Radiance Fields (NeRF) [33] have shown strong performance in generating realistic novel viewpoints for view synthesis. However, their reliance on neural networks leads to long training and rendering times. To improve surface reconstruction, [10, 13, 28, 50, 55] leverage the signed distance function (SDF) for surface representation and introduce an innovative volume rendering technique to learn an SDF model. Kerbl et al. intro + +duced 3D Gaussian Splatting (3DGS) [17], which provides an explicit representation of 3D scene information, bypassing the time-consuming implicit reconstruction process of NeRF via MLP. Some studies [16, 18, 40, 41, 59, 67], incorporate semantic information into 3DGS, equipping it with semantic awareness through training. Other researches [30, 52, 56-58] extend 3DGS by incorporating deformation fields to track the positions of 3D Gaussians at each timestamp, capturing dynamic 3D environments. DUSt3R [51] presents a novel approach for dense, unconstrained 3D reconstruction from arbitrary image collections without requiring camera calibration or viewpoint poses, unlike 3DGS and NeRF, which depend on dense viewpoints for scene construction. Building on DUSt3R, MASt3R [24] reframes image matching as a 3D task, achieving superior 3D reconstruction performance. We extended MASt3R's feed-forward mechanism to enable 3D scene perception and address alignment errors. + +# 2.2. 2D Foundation Models + +Foundation models (FMs) have emerged as a transformative paradigm in AI. These models are typically trained on extensive datasets, possess a large number of parameters, and demonstrate adaptability across a broad spectrum of downstream tasks. Specifically, 2D visual foundation models (VFMs) [12, 19, 25, 37, 42, 43, 54] have gained significant attention due to their ability to process and understand visual data. Kirillov et al. proposed the Segment Anything Model (SAM) [19], a 2D segmentation foundation model for prompt-based segmentation. SAM generates segmentation masks based on prompts that identify target objects in an image, allowing it to generalize across unseen categories. As the successor to SAM, Segment Anything 2 (SAM2) [43] unifies video and image segmentation by utilizing a larger training dataset and incorporating architectural enhancements to improve performance across a wide range of tasks. Our method leverages SAM2 to segment 2D images, maintaining consistent masks across multiple viewpoints, thereby addressing the challenges of real-time 3D intuitive segmentation. + +# 2.3. Intuitive 3D Segmentation + +Few approaches [11, 16, 22, 23, 39, 63] support intuitive segmentation directly in 3D space. For example, [22] directly segment 3D point clouds based on user clicks. Meanwhile, representations in 2D image-based 3D reconstruction have driven progress in intuitive 3D segmentation from 2D images. Inspired by the advancements in 3D Neural Scene Representation [17, 33], several studies [2, 4, 7, 26, 36, 46, 62, 64] have explored 3D segmentation within these frameworks. With SAM [19], SA3D [4] proposed an automatic strategy for cross-view prompt collection, leveraging SAM to obtain 2D masks and guide 3D + +feature training. Within the context of 3DGS, Feature3DGS [67] converts features from SAM's encoder into 3D space and uses SAM's decoder to generate masks. Mask-lifting-based methods directly map 2D segmentation masks from SAM into 3D space. Notable examples include SAGA [5], Gaussian Grouping [59], SAGS [15], Click-Gaussian [8], and FlashSplat [45]. These approaches computational costs of feature-to-mask conversion but still suffer from slow reconstructing. Our method distinguishes itself by achieving feed-forward segmentation, without scene-specific training. + +# 3. Methods + +# 3.1. Preliminary: Feed-Forward Mechanism + +The feed-forward mechanism for 3D scene reconstruction was initially proposed in DUSt3R [51] and further refined in MASt3R [24]. Unlike methods such as NeRF [33] and 3DGS [17], which require scene-specific pre-training, the feed-forward approach enables general 3D scene reconstruction through two key steps: pointmap prediction and global alignment. + +Pointmap Prediction. The process of pointmap prediction can be described as a network function $\mathcal{F}$ : $(I^n,I^m)\to (X^{n,e},C^{n,e},F^{n,e},X^{m,e},C^{m,e},F^{m,e})$ , the inputs are two RGB images $I^n,I^m\in \mathbb{R}^{W\times H\times 3}$ from different views of the scene, the outputs include two corresponding pointmaps, $X^{n,e},X^{m,e}\in \mathbb{R}^{W\times H\times 3}$ confidence maps $C^{n,e},C^{m,e}\in \mathbb{R}^{W\times H}$ , and dense local features $F^{n,e},F^{m,e}\in \mathbb{R}^{W\times H\times d}$ . Note that $e = (n,m)$ refers to the image pair formed by $I^n$ and $I^m$ , and both pointmaps are positioned in the camera coordinate system of $I^n$ . The predicted pointmaps locate the 3D positions for every pixel of the input 2D images. + +Global Alignment. Global alignment is used as a postprocess that optimizes the pointmaps from multiple views into an aligned 3D coordinate system. Given a set of images $\{I^1, I^2, \dots, I^N\}$ from a scene, a connectivity graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ is constructed, where the vertices $\mathcal{V}$ represent the $N$ images, and each edge $e = (n, m) \in \mathcal{E}$ connects an image pair $I^n$ and $I^m$ . By traversing the connected graph $\mathcal{G}$ , globally aligned pointmaps $\{\chi^n \in \mathbb{R}^{W \times H \times 3}\}$ are recovered for all pixel coordinates $(i, j) \in \{1 \dots W\} \times \{1 \dots H\}$ and all cameras for different views $n = 1, \dots, N$ . The global optimization process is formulated as follows: + +$$ +\chi^ {*} = \arg \min _ {\chi , P, \sigma} \sum_ {e \in \mathcal {E}} \sum_ {v \in e} \sum_ {i = 1} ^ {H W} C _ {i} ^ {v, e} \| \chi_ {i} ^ {v} - \sigma_ {e} P _ {e} X _ {i} ^ {v, e} \|, \tag {1} +$$ + +where $P_{e}\in \mathbb{R}^{3\times 4}$ represents the pairwise pose, which is a rigid transformation used to align the pointmaps $X^{n,e},X^{m,e}$ with the world-coordinate pointmaps $\chi^n,\chi^m$ . Additionally, $\sigma_{e}$ is a scale factor, subject to the constraint that $\prod_{e}\sigma_{e} = 1$ for all $e\in \mathcal{E}$ + +# 3.2. Task Formulation: Segment with Pointmaps + +Although the feed-forward approach offers the advantage of generalizing across various scenes without the need for scene-specific pre-training, its application to 3D segmentation often results in lower accuracy. This is primarily due to the accumulation of 3D alignment errors during the global alignment stage. Given multiple 2D images from different viewpoints and a user-provided prompt specifying a target object in 2D, our task is to optimize the alignment of 3D object segmentation across these views. The goal is to transform the 2D object masks, related to the provided prompt, into an aligned 3D space. Two key challenges need to be addressed in this task: (1) obtaining real-time segmentation masks of the target across all viewpoints based on user prompts, and (2) refining the alignment loss function to enhance segmentation accuracy, which requires a strategy to minimize discrepancies between predicted pointmaps from multiple views and globally aligned pointmaps. + +# 3.3. WildSeg3D + +As illustrated in Figure 1, WildSeg3D is a feed-forward framework designed for real-time intuitive 3D segmentation from 2D views. The framework operates in three stages: 2D mask pre-processing, 3D point dynamic global alignment, and multi-view group mapping. In the 2D mask pre-processing stage, segmentation masks are generated by SAM2 from the input multi-view images and stored in a mask cache for efficient access during real-time segmentation. In the 3D point dynamic global aligning stage, the proposed DGA refines the alignment of 3D pointmaps by focusing on challenging pixel correspondences across different views. This approach ensures more accurate 3D scene reconstruction by addressing misalignments caused by complex or occluded regions. In the multi-view group mapping stage, the stored masks are retrieved from the mask cache and the transform matrix learned through DGA is applied to map the multi-view pointmaps into an aligned 3D coordinate system in real time. This process allows for the accurate 3D segmentation results based on user prompts. + +# 3.3.1. 2D Mask Pre-processing + +In the 2D mask pre-processing stage, we eliminate the need for online computation during 2D segmentation by introducing a mask cache, enabling rapid, accurate segmentation across multiple viewpoints. First, by leveraging the video tracking capabilities of SAM2, we perform panoptic segmentation on a single viewpoint, generating precise masks for each target object within the scene. These object masks are then stored offline in the mask cache, creating a repository of segmentation data that can be efficiently accessed during real-time processing. In subsequent stages, these pre-generated masks serve as accurate prompts for SAM2's tracking functionality, allowing consistent tracking + +and segmentation of each object across multiple viewpoints in later frames. This approach ensures continuity and coherence in the segmentation process, even as the viewpoint changes. By pre-generating and storing the masks offline, the mask cache reduces the computational burden at runtime. Furthermore, this offline caching mechanism not only accelerates the segmentation process but also enhances its robustness. + +# 3.3.2. 3D Point Dynamic Global Aligning + +While the segmentation masks from the first step can be globally aligned to unify the 3D pointmaps into a single world coordinate system, challenges remain in achieving precise 3D segmentation across different viewpoints. Specifically, misalignment and loss of detail can occur due to cluttered and inconsistent backgrounds in images from various viewpoints, weakening the alignment of target objects during the global alignment process. To address these issues, we propose the Dynamic Global Aligning (DGA) method, which introduces soft masks to minimize background interference and dynamically adjusts aligning weights. This adjustment enhances the focus on challenging sample points, thus improving the overall alignment effectiveness for the target objects. + +Soft-mask and Confidence Aggregation. As described in Eq. 1, the original global alignment approach considers all pixels from all viewpoints for alignment. However, due to significant background differences between images from different viewpoints, the alignment process suffers from two issues: (1) background features are difficult to be aligned, and (2) the large proportion of the background in the images weakens the alignment focus on the objects to be segmented. To mitigate this, we propose to soften the masks generated by SAM2 and store them in the mask cache. Concretely, begin by multiplying the confidence of the pointmaps, $C \in \mathbb{R}^{W \times H}$ , with the corresponding soft masks, $S \in \mathbb{R}^{W \times H}$ , to obtain a weighted confidence. A sigmoid function, $\sigma(\cdot)$ , is then applied to the weighted confidence to map it to a confidence score in the range of 0 to 1. For all views $v = 1, \dots, N$ , this process is expressed as follows: + +$$ +F _ {i} ^ {v, e} = \sigma \left(S _ {i} ^ {v} \times C _ {i} ^ {v, e}\right) = \frac {1}{1 + e ^ {- S _ {i} ^ {v} \times C _ {i} ^ {v , e}}}, \tag {2} +$$ + +where $F \in \mathbb{R}^{W \times H}$ represents the adjusted confidence after applying the soft mask. + +Transform Matrix. We initialize a transform matrix for each view, composed of camera intrinsics, extrinsics, and depth information, enabling projection from pixel coordinates to world coordinates. Given a 2D point $(x,y)$ from a single view, with depth map $\mathbf{D} \in \mathbb{R}^{W \times H}$ , camera intrinsic matrix $\mathbf{K} \in \mathbb{R}^{3 \times 3}$ , and extrinsic matrix $\mathbf{P}$ , its corresponding 3D point $\mathbf{p} = (x,y,z)^T$ in world coordinates can be + +![](images/5173892d36778f4f670d30d68eddca18109eeea2e4065e1d2fd76bc93a20e04f.jpg) +(a) Sparse View Reconstruction: horns + +![](images/3580f315ef6f99837e814f2748a89148d42cfe578d363b51d965eb856baf0498.jpg) +Time Cost: 20 s +(b) Sparse View Reconstruction: trex + +![](images/6877564df2a68d3c8e70fa5629e093e6f50390b6453cb725c5bbaca38170281c.jpg) +Time Cost: 7 s +(c) Sparse View Reconstruction: orchids + +![](images/da3b553782705a6a2da3bfba9e6ba999aa11377c6b8cdd27dd75ccbeeafcf85a.jpg) +Time Cost: 13 s +(d) Sparse View Reconstruction: fortress +Time Cost: 11 s + +![](images/a2b9287bc1706c91ca401bff07d91a35ddd8151a11dcc1f4d643250f34859c84.jpg) +WildSeg3D (ours) +3D Segmentation: 6 ms + +![](images/2d194a3459a32d8965df439d6955a216783a7d6a45523918fa39de7077d74c3b.jpg) +WildSeg3D (ours) +3D Segmentation: 12 ms + +![](images/7fe0c2b3a4434c248c8b5718de1121373e5698115195af2cd83a24f8294d50ea.jpg) +OmniSeg3D + +![](images/706b996cfac98895d08d45855879d0b5239424d9577df70eee9a8704d46092a9.jpg) +3D Segmentation: 1400 ms +OmniSeg3D +3D Segmentation: 1500 ms + +![](images/f019e4e85a81e43a5167664bce13f00e2bf313455a60fe116d6c51a25b61a640.jpg) +SAGA + +![](images/a2480b4f2cf3360c4624bf6cca12e9b07ba66d9de2746bb7b99b7830f855ee04.jpg) +3D Segmentation: 290 ms +SAGA +3D Segmentation: 250 ms + +![](images/f22d7199376a515723f5fbc326cce0648e85780ed211f31051d5ca0ba67b4927.jpg) +SA3D + +![](images/3fa9e3b16e0361e4900b0e7877c599aea00d3aba993752975913a50e05c3b017.jpg) +3D Segmentation: 70000 ms +SA3D +3D Segmentation: 68000 ms +3D Segmentation: 5500 ms + +![](images/1fe88942f7d3c8b620caa0888279111886c23cfda9cbdff07f80b560bcef1c40.jpg) +ISRF + +![](images/26a5e516dda87440288db6d54dc86b04102bcd588a1ceb2348a3a9cd4b6e3851.jpg) +ISRF +3D Segmentation: 3000 ms + +![](images/1c15223285384088ed870cb2100f1730ad4e45f3dfb1d40c240cc75321cd673a.jpg) +WildSeg3D (ours) +3D Segmentation: $2.9\mathrm{ms}$ +3D Segmentation: 1800 ms + +![](images/6d8a0755380dc0bbd33d2e36450c61430c9aecdf884de1dac78bb9a1581f15ba.jpg) +OmniSeg3D + +![](images/65c4261232e4cae1bc46fcdfeb990df88cdfedf96822b038b4a166a43db19168.jpg) +SAGA +3D Segmentation: 240 ms + +![](images/c9df30e82d41fba9b32fe3b673c2284f848b7f9d93582bf6d0c8097c574119e0.jpg) +SA3D +3D Segmentation: 31000 ms + +![](images/5d91a3b99ac0eb3687e2f9ea76971cfb673f088fe7b92acecace20563d848215.jpg) +ISRF + +![](images/888ce0e716eb8eb0f7de7406153a5d5a8ac98fafd01781430f3647d483fe2426.jpg) +WildSeg3D (ours) +3D Segmentation: 8 ms +Figure 2. Visualization on the NVOS dataset. (a)-(d) show the sparse view reconstruction and timing results on horns, trex, orchids, and fortress scenes, including preprocessing and DGA-based scene reconstruction. Target objects for segmentation are marked with red dashed lines in the first column. From left to right: segmentation results and elapsed time from prompt input to segmentation acquisition across models. + +![](images/c5d0ca9029327f0c741b8e36c7c88f304248d10acf810c2b6de6b49aceb7f83b.jpg) +OmniSeg3D +3D Segmentation: 1400 ms +3D Segmentation: 310 ms + +![](images/643f61d8d5caecf21239b9f52d7148fcfe9a2524be998a3695a3e08f05ab1da6.jpg) +SAGA + +![](images/23c24266113e8e5b57b71dfb1cecc38ca67e118cbb3357af0b117f4d87c5e91d.jpg) +SA3D +3D Segmentation: 42000 ms + +![](images/f99ce2dc9a2e5fdb0bfd955a4c6faa2acff12c9b6eeea718fa6f639daaaf7617.jpg) +3D Segmentation: 4300 ms +ISRF +3D Segmentation: 2200 ms + +estimated as: + +$$ +\mathbf {p} = \mathbf {P} ^ {- 1} \mathbf {K} ^ {- 1} \left( \begin{array}{c} x \cdot \mathbf {D} (x, y) \\ y \cdot \mathbf {D} (x, y) \\ \mathbf {D} (x, y) \end{array} \right). \tag {3} +$$ + +For all views $v = 1, \ldots, N$ , each 3D point $\chi_i^v$ in the world coordinate, where $i \in \mathbb{R}^{HW}$ , can be derived using the above transform matrix. + +Dynamic Aligning Loss. To handle challenging target sample points, we perform point pre-matching on all image pairs in the connectivity graph $\mathcal{G}$ . For each image pair, we generate a matching map $\Phi^{v,e} \in \mathbb{R}^{W \times H}$ , which indicates whether each sample point $i \in \Phi^{v,e}$ has a matched point. Based on the confidence scores derived from Eq. 2, we define a dynamic adjustment function as follows: + +$$ +A _ {i} ^ {v, e} = \frac {F _ {i} ^ {v , e} + \alpha_ {i} ^ {v , e} \cdot F _ {i} ^ {v , e} \cdot \left(1 - F _ {i} ^ {v , e}\right)}{1 + \left| \alpha_ {i} ^ {v , e} \right| \cdot F _ {i} ^ {v , e} \cdot \left(1 - F _ {i} ^ {v , e}\right) + \epsilon}, \tag {4} +$$ + +where $F_{i}^{v,e}\cdot (1 - F_{i}^{v,e})$ enhances attention on difficult-to-match points, with confidence scores approaching 0.5, the axis of symmetry of the quadratic function. The denominator ensures that the adjusted confidence score $A_{i}^{v,e}$ remains + +within an effective range. The adjustment factor $\alpha$ is defined as: + +$$ +\alpha_ {i} ^ {v, e} = \left\{ \begin{array}{l l} \alpha_ {p} & \text {i f} \Phi_ {i} ^ {v, e} = 1 \\ - \alpha_ {n} & \text {i f} \Phi_ {i} ^ {v, e} = 0 \end{array} , \right. \tag {5} +$$ + +where $\alpha_{p}$ is the positive adjustment factor for matching points, and $\alpha_{n}$ is the negative adjustment factor for non-matching points. Both are used to differentiate the weights between matching and non-matching points. + +To map the dynamically adjusted confidence score back to its original value range, we apply the inverse of the sigmoid function, $\sigma^{-1}(\cdot)$ , as follows: + +$$ +W _ {i} ^ {v, e} = \sigma^ {- 1} \left(A _ {i} ^ {v, e}\right) = - \ln \left(\frac {1}{A _ {i} ^ {v , e}} - 1\right). \tag {6} +$$ + +Finally, we compute the L2 distance between pointmaps of all image pairs and their corresponding pointmaps $\chi^v$ in world coordinates. The final Dynamic Global Aligning is defined as: + +$$ +\chi^ {*} = \arg \min _ {\chi , P, \sigma} \sum_ {e \in \mathcal {E}} \sum_ {v \in e} \sum_ {i = 1} ^ {H W} W _ {i} ^ {v, e} \| \chi_ {i} ^ {v} - \sigma_ {e} P _ {e} X _ {i} ^ {v, e} \|, \tag {7} +$$ + +
MethodScene-specific trainingmIoU (%)mAcc (%)Total
NVOS [44]†need70.192.0-
ISRF [11]need83.896.4840 s
SGISRF [48]†need84.597.2-
SA3D [4]need90.398.2780 s
SAGA [5]need90.998.32280 s
OmniSeg3D [61]need91.798.48220 s
FlashSplat [45]†need91.898.61500 s
WildSeg3D (ours)no need94.199.030 s
+ +Table 1. Quantitative results on NVOS dataset. Boldface highlights the best results and underline the second-best. "Total" denotes the overall time from scene reconstruction to completing an intuitive 3D segmentation. The symbol $\dagger$ denotes data taken from the respective references, as the code was not fully released. + +
MethodScene-specific trainingmIoU (%)mAcc (%)Time
Single view[4]need74.695.5-
MVSeg [35]need90.998.9180-360 s
ISRF [11]need77.493.462-3 s
SA3D [4]need92.498.9120-600 s
SAGA [5]need88.098.50.08-0.9 s
OmniSeg3D [61]need94.399.31-2 s
WildSeg3D (ours)no need94.098.60.005-0.02 s
+ +Table 2. Quantitative results on SPIn-NeRF dataset. + +where $\chi_i^v$ is estimated from the transform matrices as described in Eq. 3. After the above training, we obtain an optimized transform matrix for each viewpoint, which can effectively minimize the aligning errors introduced by the feed-forward mechanism. + +# 3.3.3. Multi-view Group Mapping + +WildSeg3D enables real-time 3D segmentation of target objects by using single-view prompts as input. To efficiently map multi-view object masks into an aligned 3D space based on user prompts, we propose a multi-view group mapping method designed to optimize mask retrieval and integration. The process begins by sorting all object masks within the mask cache for the current viewpoint in ascending order based on their area, with priority given to smaller, fine-grained objects. Based on the user's prompts, we then sequentially filter the relevant masks from the cache, appending the corresponding object IDs to the result set $O$ . Once the relevant masks are identified, we compute the union of these masks across all viewpoints in the dataset. Specifically, the unified mask for each viewpoint is defined as: + +$$ +M = \left\{M ^ {v} \mid M ^ {v} = \bigcup_ {o \in O} m _ {o} ^ {v}, v \in V \right\}, \tag {8} +$$ + +where $V$ represents the set of viewpoints, $m_{o}^{v}$ denotes the mask of object $o$ in viewpoint $v$ , and $M$ is the collection of masks for each viewpoint after prompt-based retrieval. + +We apply the transform matrices learned by DGA to convert the segmentation masks from all viewpoints in $M$ from pixel coordinates to aligned world coordinates. The result + +ing 3D segmentation is denoted as $\mathcal{P}$ + +$$ +\mathcal {P} = \bigcup_ {v \in V} \left\{\mathbf {P} _ {v} ^ {- 1} \mathbf {K} _ {v} ^ {- 1} \left( \begin{array}{c} x \mathbf {D} _ {v} (x, y) \\ y \mathbf {D} _ {v} (x, y) \\ \mathbf {D} _ {v} (x, y) \end{array} \right) \Bigg | (x, y) \in M _ {v}, M _ {v} (x, y) = 1 \right\}, \tag {9} +$$ + +where $\mathcal{P}$ represents a set of 3D point cloud coordinates for the object. Through this framework, WildSeg3D efficiently aggregates 2D masks from multiple viewpoints, enabling real-time intuitive segmentation via the P2W (pixel to world coordinates) strategy. + +# 4. Experiments + +# 4.1. Datasets + +To evaluate the effectiveness of our method, we conducted experiments on multiple benchmark datasets [1, 20, 35, 44]. The NVOS dataset provides a reference view with scribble annotations for the segmented targets, as well as a target view with the corresponding segmentation mask, both captured from frontal perspectives. The SPIn-NeRF dataset, a 3D scene dataset, is annotated using the widely-adopted NeRF datasets [9, 20, 31, 32, 60]. It is used to assess the performance of intuitive 3D segmentation methods, including the evaluation of segmentation quality in more complex 3D environments. For qualitative experiments, we used the NVOS[44], SPIn-NeRF [35], T&T[20], and Mip-NeRF360 [1] datasets. These datasets were chosen to compare our method with existing approaches and to showcase the segmentation results produced by our method across different 3D scenes and viewpoints. To evaluate segmentation accuracy and facilitate comparisons, we use mean Intersection over Union (mIoU) and mean Accuracy (mAcc) as primary metrics. Additionally, we assess both the training duration and the time required for intuitive 3D segmentation on a single NVIDIA RTX 3090 GPU to evaluate the models' efficiency and real-time performance. + +# 4.2. Quantitative Results + +NVOS Dataset. To ensure experimental fairness, we adopted the evaluation approach used by models such as SAGA [5], utilizing the scribble annotations provided by the NVOS dataset [44] as input to generate 2D masks for the SAM model [19]. Additionally, we performed random point sampling on the scribble annotations of the reference view to acquire point prompts for segmentation. The results of our experiments on the NVOS dataset are presented in Table 1, where we compare the performance of our WildSeg3D framework against other state-of-the-art methods. As shown in the table, WildSeg3D outperforms existing approaches in both mIoU and mAcc. In terms of runtime efficiency, NeRF-based methods, including NVOS [44], ISRF [11], SGISRF [48], and SA3D [4], as well as 3DGS-based methods such as SAGA [5], OmniSeg3D [61], and FlashSplat [45], typically require scene-specific training. + +
DatasetsSceneD-GAD-DGAM-GAM-SGAM-DGA
NVOSfern82.5%85.1%94.1%71.7%94.2%
flower90.3%90.9%73.8%91.3%94.3%
fortress95.5%96.3%95.5%96.0%96.8%
horns (center)88.4%93.6%93.0%92.9%95.9%
horns (left)89.9%95.2%94.8%95.2%95.0%
leaves65.8%61.9%88.0%93.5%96.7%
orchids78.2%82.6%84.1%84.1%93.5%
trex79.8%80.2%85.3%61.0%86.4%
average83.8%85.7%88.6%85.7%94.1%
SPIn-NeRFfern82.5%85.1%94.1%71.7%94.2%
fork85.4%88.7%89.8%88.3%88.9%
fortress95.5%96.3%95.5%96.0%96.8%
horns88.4%93.6%93.0%92.9%95.9%
leaves65.8%61.9%88.0%93.5%96.7%
lego77.5%79.4%84.3%81.5%92.8%
orchids78.2%82.6%84.1%84.1%93.5%
pinecone84.2%88.6%85.2%89.2%95.7%
room90.4%90.3%88.7%74.1%91.5%
truck82.5%82.6%90.9%90.5%93.7%
average83.1%84.9%89.4%86.2%94.0%
+ +Table 3. Effect of different alignment strategies for DUSt3R-based and MAST3R-based 3D segmentation models on NVOS and SPIn-NeRF datasets. "D" and "M" represent the pointmaps from DUSt3R and MAST3R, respectively. "GA" refers to global alignment, "SGA" denotes sparse global alignment in MAST3R, and "DGA" stands for our proposed dynamic global alignment. + +In contrast, our approach demonstrates strong generalization across diverse scenes, eliminating the need for scene-specific training. This advantage leads to significant efficiency improvements, with our model requiring only $3.8\%$ of the computation time of SA3D (NeRF-based) and $2\%$ of FlashSplat (3DGS-based), while maintaining high segmentation accuracy. + +SPIn-NeRF Dataset. The quantitative results of our experiments on the SPIn-NeRF dataset are presented in Table 2. We evaluate the accuracy of our method by projecting the 3D segmentation masks onto the reference views and comparing them against the ground truth. In the MVSeg method [14], an intuitive segmentation model is first employed to acquire object masks from a single view. These masks are then used to segment training views, which are treated as a video sequence and processed through a video instance segmentation model [3, 53] to generate 3D masks. These masks are further refined using a semantic NeRF model [34, 65, 66]. For SA3D [4], intuitive segmentation requires traversing all training views. Cross-view self-prompting is used to guide SAM in generating 2D masks, which are subsequently projected into 3D space using Mask Inverse Rendering. This process incurs significant time costs due to the need to process each view individually. SAGA [5] incorporates low-dimensional 3D features for each Gaussian, which are rendered into 2D feature maps through differentiable rasterization during training. This approach allows for contrastive training with the segmentation results generated by SAM. In comparison to MVSeg [35] and SA3D [4], our method achieves comparable accuracy + +![](images/b5b79459b3461cd9566a1a0e4b755a9c67a9432c6be74f5152f88872bba3fc77.jpg) + +![](images/c7b3c111ca1f93eb5570f7edd715b08c6ec71b8fb34475c91ab92954fab7998a.jpg) + +![](images/572eebb42b26853036c652d3580a10b95fd7eaa46adeb563f4140082dfae0706.jpg) + +![](images/2ff3e56ffcd593ea621da5ede049edd87be307c24484103bacb924c387d91b3f.jpg) + +![](images/66f8bf10567a361251b37cf01c456b2b6af2d3453504d4afa9a9ffedec50a467.jpg) +1&1-museum + +![](images/1d2ed9117d0237efac12d5740a67367b0f7a76f0bdf64b040c9bebd40cb9d6bd.jpg) + +![](images/c6792fe00c8132f916c236ae0b210c9387455067e63d8b291ef4163267e57759.jpg) + +![](images/47a5c28aab97d193ecee62480626ce216bd471ff59da8c0dbc4e5d529c3b48ec.jpg) + +![](images/0b821fd622b0487b391715cf2e53a1c8b43e56ebc1c35704fd63554e91749482.jpg) + +![](images/174dc25c6ad48732ae4a8762ff7db196a77979eb5eff83f6ffa684694d652995.jpg) + +![](images/9d28cafc6233bca7c4d55936d51e3a564fb2b60f9fda68bb782371696576d2bf.jpg) + +![](images/158ec852e6e7dda8e48eafff07b93b01d57027bb3ae8ae67ab37cfe8a5b30227.jpg) + +![](images/f2cc29db28d241fd2d8e5f09fbd62f564b41d6932dcfc3c5710aa296d312c1d7.jpg) +Eiffel Tower + +![](images/078a941bf03146f773f5230ce57f39f8658f6578b27c780070a674c571ed98d3.jpg) + +![](images/1c6c80062ab41bcf82e094b46480b34f998445fc3118e3b81b1508ca1dd076cb.jpg) +Figure 3. Performance of WildSeg3D on indoor and outdoor scenes. For each scene, the left side shows the sparse views for reconstruction, with the segmentation target indicated by red dashed lines in the first view as prompts. + +while requiring only a fraction—approximately one tenthousandth—of the processing time. Moreover, when compared to SAGA, our approach not only demonstrates superior accuracy but also outperforms SAGA in terms of processing time. Overall, our method delivers exceptional performance, making real-time intuitive 3D segmentation feasible in practical applications. + +# 4.3. Qualitative Results + +We conduct a comparative analysis of our method against existing approaches and present qualitative experimental results. As shown in Figure 2, the first column illustrates the reconstruction time of our model for various scenes based on sparse views. The subsequent columns compare the segmentation results produced by our method with those from other approaches, along with the time taken to obtain 3D segmentation results after user input. The final column displays the segmentation results of ISRF [11] across different scenes. ISRF utilizes the TensoRF representation [6] for scene training and rendering, incorporating DINO features [3] within each voxel to enable 2D-to-3D semantic matching. Segmentation is performed via nearest neighbor feature matching (NNFM). However, ISRF struggles with distinguishing semantically similar objects, as demonstrated in the Trex and Orchids scenes, where the method faces challenges in accurate segmentation. SA3D [4], upon receiving user prompts, employs a cross-view self + +
Number of Views34579
NVOS89.8%92.9%94.1%94.0%93.8%
SPIn-NeRF87.8%92.2%94.0%94.9%94.7%
Total Time15 s18 s23 s35 s50 s
+ +Table 4. Ablation Experiments on the Impact of Viewpoint Quantity on mIoU. + +prompting mechanism and SAM [19] to generate 2D masks for each view. These masks are then mapped to 3D space via Mask Inverse Rendering, with the process repeated for each user interaction. This approach incurs high computational costs due to the need for repeated processing. SAGA [5] and OmniSeg3D [61] transform 2D masks into 3D features and require additional training for each 3D Gaussian. SAGA relies on a collection of loss functions during training, whereas OmniSeg3D uses hierarchical contrastive learning to map discontinuous multi-view segmentations to consistent 3D features. In contrast, our method achieves superior performance with high-quality transform matrices learned through DGA, which map 2D views into an aligned 3D coordinate system. The MGM module then enables fast retrieval of 2D masks from multiple views, which are subsequently transformed into the unified 3D space. This design results in significantly improved computational efficiency, allowing our method to achieve faster intuitive segmentation compared to existing approaches. + +As demonstrated in Figure 3, our method does not require scene-specific training, enabling near real-time 3D intuitive segmentation for arbitrary scenes. We showcase this capability using indoor and outdoor scenes from the MipNeRF360 [1], T&T [20], SPIn-NeRF [35] datasets, and two additional scenes in the wild. By leveraging only sparse views, our method is able to perform both scene reconstruction and near real-time 3D segmentation, underscoring its versatility and ability to segment objects in diverse, uncontrolled environments. + +# 4.4. Ablation Studies + +Comparison of Different Viewpoint Counts. Table 4 presents an ablation study on the number of viewpoints input into WildSeg3D. Considering overall performance, we randomly select five viewpoints in prior experiments. On the SPIn-NeRF dataset, our method outperforms OmniSeg3D when seven viewpoints are selected, despite the results in Table 2. + +Effect of Dynamic Global Alignment. To further validate the effectiveness of DGA, we also integrate it into DUSt3R, replacing the matching maps with SIFT [29]. As shown in Table 3, incorporating DGA significantly improves performance in both DUSt3R and MAST3R models compared to traditional GA training methods. + +Furthermore, as shown in Figure 4, comparative visualizations on the NVOS and LERF [18] datasets demonstrate + +![](images/ea97018c5a1d9ea8df2fc83a0bfac9fe3c095a896f69407cffbeb9132b33071f.jpg) + +![](images/9f38e8f49f0fbe8bc57d969c1f72f48d1d64a81392830c8257d076eae1556fb4.jpg) +Figure 4. Visualization of ablation experiments on DGA. + +![](images/9036cafcddecb526450cdcb7d03a703a415f6dd9e51961fedaf3b44826f39038.jpg) + +![](images/233aa0ae7e761063186e3db887e56c95ec23c4f4b7a72f9b60c4d1ebeb8f87ff.jpg) + +![](images/226a2b74b1e2ca53c4f0740796f9d1268690068afa045922a52229ac43f0aa08.jpg) + +![](images/bb3d13fa42ef6460ebd6129f49c47778bac99e09d310b57e3463aa2f5e585499.jpg) + +![](images/393496fe6eef9533316a968e13520e6287bc13fe9549629ced0b89e732e21a80.jpg) +Figure 5. Ablation Experiments on NVOS Dataset. Left: visualization of different confidence adjustment functions for DGA. Right: Results of these functions. + +![](images/ec14a67c1604f09004ea3f92f06abf64d8bc527cb0583b58d5975626fb8c9ea7.jpg) + +that DGA effectively reduces 3D alignment errors, minimizing blurred details and background confusion. + +Different Confidence adjustment functions. Figure 5 compares the confidence adjustment function in DGA with other adjustment functions, defined as: $A_{\theta}(x) = \frac{\arctan(\theta \cdot (x - 0.5))}{\pi} + 0.5$ , $S_{\delta}(x) = \frac{1}{1 + e^{-\delta \cdot (x - 0.5)}}, L_{\beta}(x) = \frac{\log(1 + \beta \cdot x)}{\log(1 + \beta)}$ , where $\theta, \delta$ , and $\beta$ control the scaling of the Arc-tan, Sigmoid, and Logarithmic functions, respectively. + +Our function performs adaptive adjustments centered at 0.5 confidence, improving the weight of hard-to-match points while ensuring smooth transitions across both low and high confidence levels, achieving the best performance. + +# 5. Conclusion + +We introduce WildSeg3D, a feed-forward method that enables real-time intuitive 3D segmentation from 2D images without scene-specific pre-training. With Dynamic Global Aligning, WildSeg3D learns high-quality transform matrix for each view, aligning 2D images to an aligned coordinate system. Additionally, mask cache stores masks from multiple views consistently. For real-time intuitive segmentation, MGM uses a search strategy to retrieve target masks and map them into an aligned 3D space, responding promptly to user inputs. Through extensive experiments, WildSeg3D demonstrates a significant speedup over existing methods while maintaining high accuracy. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 6, 8 +[2] Yash Bhalgat, Iro Laina, Joao F Henriques, Andrew Zisserman, and Andrea Vedaldi. Contrastive lift: 3d object instance segmentation by slow-fast contrastive fusion. arXiv preprint arXiv:2306.04633, 2023. 3 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 7 +[4] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, et al. Segment anything in 3d with nerfs. Advances in Neural Information Processing Systems, 36:25971-25990, 2023. 1, 2, 3, 6, 7 +[5] Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians, 2024. 1, 2, 3, 6, 7, 8 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 7 +[7] Hanlin Chen, Chen Li, Mengqi Guo, Zhiwen Yan, and Gim Hee Lee. Gnesf: Generalizable neural semantic fields. Advances in Neural Information Processing Systems, 36: 36553-36565, 2023. 3 +[8] Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do. Click-gaussian: Interactive segmentation to any 3d gaussians. arXiv preprint arXiv:2407.11793, 2024. 3 +[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 6 +[10] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2 +[11] Rahul Goel, Dhawal Sirikonda, Saurabh Saini, and PJ Narayanan. Interactive segmentation of radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4201-4211, 2023. 1, 2, 3, 6, 7 +[12] Yubin Gu, Honghui Xu, Yueqian Quan, Wanjun Chen, and Jianwei Zheng. Orsi salient object detection via bidimensional attention and full-stage semantic guidance. IEEE Transactions on Geoscience and Remote Sensing, 61:1-13, 2023. 3 +[13] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, + +and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2 +[14] Yuying Hao, Yi Liu, Zewu Wu, Lin Han, Yizhou Chen, Guowei Chen, Lutao Chu, Shiyu Tang, Zhiliang Yu, Zeyu Chen, et al. Edgeflow: Achieving practical interactive segmentation with edge-guided flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1551-1560, 2021. 7 +[15] Xu Hu, Yuxi Wang, Lue Fan, Junsong Fan, Junran Peng, Zhen Lei, Qing Li, and Zhaoxiang Zhang. Semantic anything in 3d gaussians. arXiv preprint arXiv:2401.17857, 2024. 1, 3 +[16] Chi Huang, Xinyang Li, Shengchuan Zhang, Liujuan Cao, and Rongrong Ji. Nerf-dets: Enhancing multi-view 3d object detection with sampling-adaptive network of continuous nerf-based representation. arXiv e-prints, pages arXiv-2404, 2024. 3 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 3 +[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 1, 3, 8 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 3, 6, 8 +[20] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6, 8 +[21] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 1 +[22] Theodora Kontogianni, Ekin Celikkan, Siyu Tang, and Konrad Schindler. Interactive object segmentation in 3d point clouds. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2891-2897. IEEE, 2023. 3 +[23] Itai Lang, Fei Xu, Dale Decatur, Sudarshan Babu, and Rana Hanocka. iseg: Interactive 3d segmentation via interactive attention. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3 +[24] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 1, 3 +[25] Jianghang Lin, Yunhang Shen, Bingquan Wang, Shaohui Lin, Ke Li, and Liujuan Cao. Weakly supervised open-vocabulary object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3404-3412, 2024. 3 + +[26] Yichen Liu, Benran Hu, Junkai Huang, Yu-Wing Tai, and Chi-Keung Tang. Instance neural radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 787–796, 2023. 3 +[27] Yichen Liu, Benran Hu, Chi-Keung Tang, and Yu-Wing Tai. Sanerf-hq: Segment anything for nerf in high quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3216–3226, 2024. 1 +[28] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 2 +[29] G Lowe. Sift-the scale invariant feature transform. Int. J, 2 (91-110):2, 2004. 8 +[30] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3 +[31] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 6 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 6 +[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3 +[34] Ashkan Mirzaei, Yash Kant, Jonathan Kelly, and Igor Gilitschenski. Laterf: Label and text driven object radiance fields. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 7 +[35] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669-20679, 2023. 2, 6, 7, 8 +[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3 +[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3 +[38] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. + +Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 1 +[39] Zhipeng Qian, Yiwei Ma, Jiayi Ji, and Xiaoshuai Sun. X-refseg3d: Enhancing referring 3d instance segmentation via structured cross-modal graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4551-4559, 2024. 3 +[40] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 3 +[41] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. 3 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[43] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 3 +[44] Zhongzheng Ren, Aseem Agarwala, Bryan Russell, Alexander G Schwing, and Oliver Wang. Neural volumetric object selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6133-6142, 2022. 2, 6 +[45] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Flashsplat: 2d to 3d gaussian splattering segmentation solved optimally. arXiv preprint arXiv:2409.08270, 2024. 2, 3, 6 +[46] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Buló, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043-9052, 2023. 3 +[47] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 1 +[48] Songlin Tang, Wenjie Pei, Xin Tao, Tanghui Jia, Guangming Lu, and Yu-Wing Tai. Scene-generalizable interactive segmentation of radiance fields. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6744-6755, 2023. 2, 6 +[49] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443-453. IEEE, 2022. 1 + +[50] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2 +[51] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 3 +[52] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 3 +[53] Wenguan Wang, Tianfei Zhou, Fatih Porikli, David Crandall, and Luc Van Gool. A survey on deep learning technique for video segmentation. arXiv e-prints, pages arXiv-2107, 2021. 7 +[54] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1130-1140, 2023. 3 +[55] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems, 35:1966-1978, 2022. 2 +[56] Yuze Wang, Junyi Wang, Ruicheng Gao, Yansong Qu, Wantong Duan, Shuo Yang, and Yue Qi. Look at the sky: Sky-aware efficient 3d gaussian splattering in the wild. IEEE Transactions on Visualization and Computer Graphics, 2025. 3 +[57] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. +[58] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3 +[59] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732, 2023. 1, 3 +[60] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. Nerf-supervision: Learning dense object descriptors from neural radiance fields. In 2022 international conference on robotics and automation (ICRA), pages 6496-6503. IEEE, 2022. 6 +[61] Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20612-20622, 2024. 2, 6, 8 +[62] Hong-Xing Yu, Leonidas J Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. arXiv preprint arXiv:2107.07905, 2021. 3 +[63] Yuanwen Yue, Sabarinath Mahadevan, Jonas Schult, Francis Engelmann, Bastian Leibe, Konrad Schindler, and Theodora Kontogianni. Agile3d: Attention guided interactive multi- + +object 3d segmentation. arXiv preprint arXiv:2306.00977, 2023.3 +[64] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 3 +[65] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 7 +[66] Shuaifeng Zhi, Edgar Sucar, Andre Mouton, Iain Haughton, Tristan Laidlow, and Andrew J Davison. ilabel: Interactive neural scene labelling. arXiv preprint arXiv:2111.14637, 2021. 7 +[67] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 1, 3 \ No newline at end of file diff --git a/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/images.zip b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6b0b2b6abb63ab13e858bda8dad4bd3fca95b56f --- /dev/null +++ b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c139fd1e1efc3798a92c6610a7259488575d718f315e76325b49611e6b95145 +size 660935 diff --git a/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/layout.json b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ed9c1381a94c6e9ddc310fe3f24afb6fb748203 --- /dev/null +++ b/wildseg3dsegmentany3dobjectsinthewildfrom2dimages/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8ac61121b58424e5d1afd0441413340fd13c2dc1515a9e0bfd05e6c31dc7b67 +size 517753 diff --git a/wonderplaydynamic3dscenegenerationfromasingleimageandactions/59b8ccd1-d5e7-4d4c-9298-ce3e755fe5ed_content_list.json b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/59b8ccd1-d5e7-4d4c-9298-ce3e755fe5ed_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..023de76f4f39b21f7e048d692f214eb507066978 --- /dev/null +++ b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/59b8ccd1-d5e7-4d4c-9298-ce3e755fe5ed_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea232110048f15ce6757162fe77312dec1a3b3fb22f159cf3208c4b5e0e6b760 +size 86252 diff --git a/wonderplaydynamic3dscenegenerationfromasingleimageandactions/59b8ccd1-d5e7-4d4c-9298-ce3e755fe5ed_model.json b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/59b8ccd1-d5e7-4d4c-9298-ce3e755fe5ed_model.json new file mode 100644 index 0000000000000000000000000000000000000000..60d2028e535f78276b61dbeec4232c6ce0746df7 --- /dev/null +++ b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/59b8ccd1-d5e7-4d4c-9298-ce3e755fe5ed_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f9fee8abd3635eb3146f2b71e01e6beee7c9ed88c923618e652984ec9e82724 +size 108142 diff --git a/wonderplaydynamic3dscenegenerationfromasingleimageandactions/59b8ccd1-d5e7-4d4c-9298-ce3e755fe5ed_origin.pdf b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/59b8ccd1-d5e7-4d4c-9298-ce3e755fe5ed_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..afa1db9d514c10b65c02c6db25748b3853ef9fc2 --- /dev/null +++ b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/59b8ccd1-d5e7-4d4c-9298-ce3e755fe5ed_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a83c455f3398753dd4e25afc66a94c48777fbb66148ca5649956e5a3a8b304bd +size 12266684 diff --git a/wonderplaydynamic3dscenegenerationfromasingleimageandactions/full.md b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0f713be2a41fc3a5062d66afbe3dfb5c3e25b8fd --- /dev/null +++ b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/full.md @@ -0,0 +1,379 @@ +# WonderPlay: Dynamic 3D Scene Generation from a Single Image and Actions + +Zizhang Li $^{1*}$ Hong-Xing Yu $^{1*}$ Wei Liu $^{1}$ +Yin Yang $^{2}$ Charles Herrmann $^{1}$ Gordon Wetzstein $^{1}$ Jiajun Wu $^{1}$ $^{1}$ Stanford University $^{2}$ University of Utah + +https://kyleleey.github.io/WonderPlay/ + +![](images/c8536f783d989a65c24d6da4df047148fb75029258f826ea781340e8ba43a867.jpg) +Figure 1. We propose WonderPlay, a framework that takes a single image and actions as inputs, and then generates dynamic 3D scenes that depict the consequence of the actions. WonderPlay allows users to interact with various scenes of diverse physical materials, e.g., the hat and wine glass (rigid body), the hair (thin strands), the steam (gas), the mushroom (elastic), honey (liquid), and more. See https://kyleleey.github.io/WonderPlay/ for interactive video results. + +# Abstract + +WonderPlay is a novel framework integrating physics simulation with video generation for generating action-conditioned dynamic 3D scenes from a single image. While prior works are restricted to rigid body or simple elastic + +dynamics, WonderPlay features a hybrid generative simulator to synthesize a wide range of 3D dynamics. The hybrid generative simulator first uses a physics solver to simulate coarse 3D dynamics, which subsequently conditions a video generator to produce a video with finer, more realistic motion. The generated video is then used to update the simulated dynamic 3D scene, closing the loop between the physics solver and the video generator. This approach + +enables intuitive user control to be combined with the accurate dynamics of physics-based simulators and the expressivity of diffusion-based video generators. Experimental results demonstrate that WonderPlay enables users to interact with various scenes of diverse content, including cloth, sand, snow, liquid, smoke, elastic, and rigid bodies - all using a single image input. Code will be made public. + +# 1. Introduction + +Recent years have seen rapid progress in image, video, and 3D and 4D scene generation, culminating in models that achieve great visual quality [6, 46] and dynamic realism [11]. This has directly motivated recent interest in generative world models, which, beyond their relevance to AR/VR and embodied AI [51], can also be created and explored as standalone experiences. However, while significant efforts have been devoted to enhancing the generation quality [12, 14], relatively little attention has been paid to enabling action-based interaction. In this work, we study action-conditioned dynamic 3D scene generation from a single image: given an input image and a 3D action, such as wind or a point force, we aim to generate the resulting dynamic 3D scene in the near future. In particular, we focus on three types of actions: gravity, force fields like wind, and point forces like pushes or pulls. + +Existing methods often rely exclusively on physics simulation for computing dynamic 3D scenes given user action input [59, 67, 73]. These methods face two critical limitations. First, they require accurate physics solvers for all types of dynamics involved in the scene. Nevertheless, accurate physics solvers such as solid-fluid two-way coupling [16] still remain an open problem. Second, they require full reconstruction of physical states from limited observations. However, reconstructing complete physical states for materials like snow, sand, cloth, and fluids from a single image is often infeasible. Consequently, existing methods are constrained to a narrow range of dynamics types, primarily rigid body dynamics [43] and simple elasticity [59, 73]. + +This motivates us to incorporate data priors from video generation models [8, 11, 70], which are trained on extensive real-world videos of diverse physical phenomena. However, video generation models cannot accept precise 3D actions as inputs and simulate the resulting dynamics. In this work, we reconsider the relationship between physics simulators and video generation models for action-conditioned dynamic 3D scene generation. Our novel framework, WonderPlay, enables users to interact with 3D scenes encompassing diverse materials—including rigid bodies, cloth, liquids, gases, and granular substances—from a single input image, as shown in Figure 1. + +Our core technical idea is a hybrid generative simulator. First, we let the physics simulator provide a coarse sim + +ulation of action-induced dynamic consequences to a video generator. Conditioned on the coarse simulation, the video generator synthesizes a video with realistic motion. Finally, the synthesized video is used to update the coarse simulation. + +In the conditional video generation stage, we explore a novel strategy to optimally use the simulator conditioning signal: a motion-appearance bimodal control scheme, designed to improve the quality and realism of the dynamics in the generated video. Additionally, to reduce the video generator hallucination in simulator-trustable spatial regions such as static backgrounds, we introduce a spatially varying masking scheme for the bimodal control. + +In summary, our contributions are three-fold: + +- We tackle the challenging problem of single-image, action-conditioned dynamic 3D scene generation with diverse physical materials. +- We propose WonderPlay, featuring a hybrid generative simulator that integrates a physics solver and video diffusion to acquire both high simulation fidelity in response to actions and high visual quality. +- We demonstrate that WonderPlay significantly outperforms both pure physics-based methods and adapted video generation models in terms of visual quality and physical plausibility under various interactions. + +# 2. Related Work + +Action-conditioned dynamic scene generation. Early work on generating action-conditioned dynamic scenes approached the problem by extracting modal bases of vibrating objects in 2D image space [19, 20], essentially representing motion as a series of vibrations with different frequencies and intensities. Following the advent of generative diffusion modeling [25, 56, 57], this approach was later extended by retaining the same motion representation but generating the modal basis with a diffusion model [38]. While this representation can be effective for motions similar to vibrations, modal basis functions struggle to represent more general motions, prompting the emergence of an alternative line of research that explicitly uses physics solvers. For example, PhysGen [43] focused on the 2D domain using a rigid-body physics solver to handle colliding objects. + +Recently, several physics-based approaches have been developed to synthesize dynamic 3D scenes [2, 15, 17, 28, 34, 39, 41, 67, 73, 78]. However, due to the requirements of physics solvers, all of these techniques require complete 3D geometric reconstructions of the scene, requiring complex, multi-view captures. For example, Virtual Elastic Objects [17] reconstructs the geometry, appearances, and physical parameters of elastic objects from a multi-view capture setup. Later work, such as PAC-NeRF [36], PhysGaussian [67], and PhysDreamer [73], integrates physics-based + +![](images/b560fe8a742ea6a7d0a17c9403d016fe220b84b3be93e2507fcc913e457ca665.jpg) +Figure 2. Overview of WonderPlay. Given a single image, we first reconstruct the 3D scene and estimate material properties. Then our hybrid generative simulator uses physics solver and input actions to infer coarse 3D dynamics. The simulated appearance and motion signals are used to condition the video generator through spatially varying bimodal control to synthesize the realistic motion. The dynamic 3D scene is refined using the synthesized video, finishing the hybrid generative simulation. + +![](images/db56ce6e64e0548678b1bf663433ccc77484eadc9d21934b4a404e2c2e1c75e4.jpg) + +simulations with NeRF or 3D Gaussians from multi-view reconstruction. A concurrent approach, PhysMotion [59], is closest to our work. Both approaches take a single image as input and use a combination of a 3D physics solver and a video generation model. Unlike ours, however, PhysMotion [59] relies on a physics solver to compute the dynamics for the entire scene, only using the video generator to refine the appearance. Due to the restrictive assumption that the physics solver will specify all the dynamics, PhysMotion is limited to rigid and elastic dynamics. In contrast, WonderPlay uses both a physics solver and the video generator to compute dynamics, enabling realistic action-conditioned dynamics for various types of physical phenomena, and generates a dynamic 3D scene as opposed to a video. + +Controllable video generation. In recent years, video generation has rapidly improved, making significant strides in both visual quality and realistic dynamics [8, 10, 11, 22, 26, 54, 70]. Recent video generation methods, such as Sora [11], have demonstrated great promise in generating diverse real-world physical phenomena. However, despite this promise, these models are conditioned with text and/or images and lack controllability regarding general actions and other physical inputs. While there has been considerable work on adding controls to video models, most of this work has focused on camera control [24, 27, 58, 66, 68, 77] and various types of motion control [18, 21, 37, 49, 50, 53, 62, 65, 71, 74], including drag-based, trajectory-based, and optical flow-based approaches. However, most of these motion control models require the resulting dynamics of an action as input to generate a trajectory-following video. Concurrently, Motion Prompting [21] uses temporally sparse trajectories as conditioning to generate videos that adhere to an initial trajectory and then continue using a generative video prior. Nevertheless, many actions, such as those involving fluids or wind, are difficult or impossible to represent as trajectory signals. In contrast, we aim to take physics-based 3D actions as input and model a dynamic 3D + +scene, rather than just a video. + +World models. Along with the rise in video models, there has been a growing interest in interactive world models [23], which recurrently generate world states from prior state and actions. While this area has seen considerable research, the focus has been on video game domains [12, 14, 60] due to the data availability of action-video pairs. Consequently, both the generated worlds and the actions considered are centered around those found in video games. Only a few concurrent works have explored this for the real world (e.g., [1, 7]) and none provides interaction beyond camera control or text. In contrast, we focus on physics-based actions in realistic worlds. + +Dynamic 3D scene generation. Text-conditioned generation of 3D scenes with motion has primarily been tackled by distilling video generation into dynamic 3D representations [55]. Most of the existing works focus on objects or simple scenes composed of a few objects [4, 5, 40, 52, 75], while recent methods also attempt to deal with more complex scenes including nature [35, 42]. Recent work [63, 64, 76] has focused on training generators which model 4D itself, conditioning on time and camera pose. Yet, they do not have the ability to simulate dynamics in response to actions. + +# 3. WonderPlay + +**Formulation.** Our goal is action-conditioned 3D scene dynamics synthesis. The input is a single image $\mathbf{I}$ and actions. We model actions as three types of forces: gravity $\mathbf{f}_{\mathrm{g}}$ , 3D force fields $\mathbf{f}_{\mathrm{w}}(x,y,z,t)$ such as wind, and 3D point forces $\mathbf{f}_{\mathrm{p}}(t)$ which is defined on a point of an object. The output is a dynamic 3D scene $\{S_{t}\}_{t=0}^{T}$ that is the consequence of applying the actions to the input scene, where $S_{0}$ denotes our initial 3D scene representation recovered from the input image $\mathbf{I}$ , and $T$ denotes the total simulation time steps. + +Overview. We aim to simulate the dynamics of diverse ma- + +terials including rigid, elastic, cloth, smoke, liquid, granular, and their interactions. To this end, we propose WonderPlay. As illustrated in Figure 2, we first reconstruct the 3D scene $S_0$ from the input image $\mathbf{I}$ (top left of Figure 2). Our main technical innovation is the hybrid generative simulator (middle left of Figure 2). It takes the 3D scene $S_0$ and the actions as input and predicts the 3D dynamics $\{S_t\}_{t=1}^T$ (middle right of Figure 2). + +# 3.1. 3D Scene Reconstruction + +Our 3D scene representation $S_{t} = \mathcal{B}_{t}\cup \mathcal{O}_{t}$ consists of a background $\mathcal{B}_t$ and objects $\mathcal{O}_t$ at a timestep $t$ . Our first step is to reconstruct/generate $S_0$ from the input image $\mathbf{I}$ . We reconstruct the 3D background $\mathcal{B}_0$ and the 3D objects $\mathcal{O}_0$ separately to jointly form $S_0 = \mathcal{B}_0\cup \mathcal{O}_0$ . + +Background. We represent the background with Fast Layered Gaussian Surfels (FLAGS) [72]. Formally, the background $\mathcal{B}_t = \{\mathbf{p}^{\mathrm{B}},\mathbf{q}^{\mathrm{B}},\mathbf{s}^{\mathrm{B}},\mathbf{o}^{\mathrm{B}},\mathbf{c}_t^{\mathrm{B}}\}$ consists of $N_{\mathrm{B}}$ Gaussian surfels, parameterized by 3D spatial positions $\mathbf{p}^{\mathrm{B}}\in \mathbb{R}^{3N_{\mathrm{B}}}$ , orientation quaternions $\mathbf{q}^{\mathrm{B}}$ , scales $\mathbf{s}^{\mathrm{B}}$ , opacities $\mathbf{o}^{\mathrm{B}}$ , and view-independent RGB colors $\mathbf{c}_t^{\mathrm{B}}$ . We detail the reconstruction of background in Appendix B.2 and treat it as a static boundary in simulation. + +Objects. An "object" in WonderPlay refers to a dynamic entity we simulate in the physics solver, including rigid object, cloth, granular material, and fluids. To represent a simulatable object that is compatible with our physics solvers, we build a simulation-ready representation on top of the Gaussian surfels by adding connectivity to them, turning them into "topological Gaussian surfels". Formally, the topological Gaussian surfels consist of $N_{\mathrm{O}}$ Gaussian surfels with edges and velocities, $\mathcal{O}_t = \{\mathbf{E}, \mathbf{v}_t, \mathbf{p}_t^{\mathrm{O}}, \mathbf{q}_t^{\mathrm{O}}, \mathbf{s}_t^{\mathrm{O}}, \mathbf{o}_t^{\mathrm{O}}, \mathbf{c}_t^{\mathrm{O}}\}$ , where the edge matrix $\mathbf{E} \in \{0,1\}^{N_0 \times N_0}$ indicates the topological connectivity of the surfels, and $\mathbf{v}_t \in \mathbb{R}^{3N_0}$ denotes the velocity. + +We create the initial topological Gaussian surfels $\mathcal{O}_0$ by first generating an object mesh from an image segment of the object using an image-to-mesh model InstantMesh [69]. Then, we bind a Gaussian surfel to each of the mesh vertices. We detail this process in Appendix B.3. + +Materials. Besides the geometry and appearance representation $\mathcal{O}$ , an object also has material properties $\mathbf{m}$ . The definition of object material depends on the object type, which follows a 6-way classification: rigid, elastic, cloth, smoke, liquid, and granular. We detail the material properties in Appendix B.4. + +# 3.2. Hybrid Generative Simulator + +Main idea. The reconstructed scene geometry $S_0$ and estimated material properties $\mathbf{m}$ are inherently inaccurate and incomplete, and accurate physics solvers for all materials and their complex interactions are still an open prob + +![](images/d5cdb40400442fa0b44a023790788c1b93942136ef9d9db13491af0414fd9de0.jpg) +Figure 3. Illustration on our spatially varying bimodal control, which drives the video generator with input image $\mathbf{I}$ , pixel-space flow $\mathbf{F}$ and simulation rendered $\tilde{\mathbf{V}}$ . + +lem. Therefore, existing methods are limited to simple rigid/elastic simulations [43, 59, 73]. Our main idea to address this challenge is extracting the dynamics knowledge from a video generator which has been trained on numerous videos of real-world physics. + +In particular, we use physics solvers to estimate a coarse and incomplete dynamic scene $\{\tilde{S}_t\}_{t=1}^T$ given initial scene $S_0$ and actions $\mathbf{f_g}$ , $\mathbf{f_w}$ , $\mathbf{f_p}$ . The coarse dynamic scene is used to drive the video generator to synthesize a video $V$ that has realistic dynamics. We obtain the output dynamic scene $\{\mathcal{S}_t\}_{t=1}^T$ by updating $\{\tilde{S}_t\}_{t=1}^T$ to match the video $V$ through differentiable rendering. + +Physics solvers. At each simulation time step, a physics solver takes the current scene $\tilde{S}_t$ and forces $\mathbf{f}_{\mathrm{g}}$ , $\mathbf{f}_{\mathrm{w}}(t)$ , $\mathbf{f}_{\mathrm{p}}(t)$ as input, and solves for the object dynamics attributes including the velocity $\mathbf{v}_{t+1}$ , position $\mathbf{p}_{t+1}^{\mathrm{O}}$ , and orientation $\mathbf{q}_{t+1}^{\mathrm{O}}$ at the next time step: + +$$ +\mathbf {v} _ {t + 1}, \mathbf {p} _ {t + 1} ^ {\mathrm {O}}, \mathbf {q} _ {t + 1} ^ {\mathrm {O}} = \operatorname {s o l v e r} (\tilde {\mathcal {S}} _ {t}, \mathbf {f} _ {\mathrm {g}}, \mathbf {f} _ {\mathrm {w}} (t), \mathbf {f} _ {\mathrm {p}} (t)), \quad (1) +$$ + +where $\tilde{S}_0 = S_0$ . Then, we construct the coarse scene at the next time step $\tilde{S}_{t + 1}$ by + +$$ +\tilde {\mathcal {S}} _ {t + 1} = \mathcal {B} _ {0} \cup \left\{\mathbf {E}, \mathbf {v} _ {t + 1}, \mathbf {p} _ {t + 1} ^ {\mathrm {O}}, \mathbf {q} _ {t + 1} ^ {\mathrm {O}}, \mathbf {s} _ {0} ^ {\mathrm {O}}, \mathbf {o} _ {0} ^ {\mathrm {O}}, \mathbf {c} _ {0} ^ {\mathrm {O}} \right\}, \tag {2} +$$ + +where we keep all non-dynamics attributes the same as $S_0$ . To compute the dynamics attributes of various materials, we employ multiple types of physics solvers. These solvers are coupled to tackle multi-physics scenes, e.g., fluid and rigid as shown in Figure 2. Please see details for different solvers in Appendix B.4. + +Conditioning the video generator. Given the coarse dynamic scene $\{\tilde{S}_t\}_{t=0}^T$ , we condition a video generator to synthesize a video $\mathbf{V} \in \mathbb{R}^{(T+1) \times H \times W \times 3}$ that has more detailed motion while adhering to the action consequence + +![](images/150a37dfc5c2dedb46de60861bacf0b4f4f89bd75f3894f2a61839d36ad29822.jpg) +"A rubber duck toy dropping into the sea water" + +![](images/292e4102a2d4384f0447f916d1551228f431825db0216bfcdf353b23183c763f.jpg) +"The red boat drifts right and pushes the blue boat" + +![](images/5e7486dec1aac75c6f09c40a080aa99211154a0869e7cf1f21c51e9c16b56906.jpg) + +![](images/0113de9cbbd72d3a0e6f61391dde5d954e99c32224e06895e8dac0e6cc235fb0.jpg) + +![](images/8434b2c1d26600982b1dd8434e190cd8c9773a555d39867d86322fe0fc69b215.jpg) + +![](images/a4d41163e7431885f6c28df0b4f9abbb9c598b48baad75f714c2f7c93c33f648.jpg) + +![](images/52b3f50ba0dd899df8b202d6e3b36cdc1917d10276e55949a3cae8a572bcc147.jpg) +Figure 4. Qualitative comparisons between WonderPlay (ours) and the baseline methods. The top row shows the input images, actions, and the texts describing the actions for CogVideoX [70]. + +![](images/25cede3c5b758c7fc9aba64f7b0efd38a4e87d165189019c283f9bfce976e175.jpg) + +![](images/249d2321d52fe0295a4478842070b4071ef669292e61010de18d78ab70d4061e.jpg) + +![](images/c14d67cf9895ac28bd31c03f375031b409aa5fab1661ef2442e38aac85fcebba.jpg) + +![](images/c54922efab4eee67a491ea6096ba3bcfab316bd282888a54133c1dc9164dc109.jpg) + +![](images/67323a2b99e8c439a10cd3dc3c26916a6a56472fed51d26b3fc6ab01b0197c98.jpg) + +depicted by the coarse dynamic scene. To this end, we introduce a bimodal control scheme that uses two modalities for control: motion (represented by flow) and appearance (RGB). In particular, + +$$ +\mathbf {V} = g (\mathbf {F}, \tilde {\mathbf {V}}, \mathbf {I}), \tag {3} +$$ + +where $g$ denotes the video generator, $\mathbf{F} \in \mathbb{R}^{T \times H \times W \times 2}$ denotes the pixel-space flow rendered using the velocity $\{\mathbf{v}_t\}_{t=1}^T$ , $\tilde{\mathbf{V}} \in \mathbb{R}^{(T+1) \times H \times W \times 3}$ denotes the video rendered from the coarse scene $\{\tilde{\mathcal{S}}_t\}_{t=0}^T$ , and $\mathbf{I}$ denotes the input image. We show an illustration in Figure 3. + +Motion control. We leverage a pre-trained motion-controlled image-to-video diffusion model, Go-with-the-Flow [13], as our $g$ . The motion control is based on noise warping. In short, instead of using an unstructured random Gaussian noise distribution, it uses a warping-based structured noise $\mathbf{N}(\mathbf{F}) \in \mathbb{R}^{(T + 1) \times H \times W \times 3}$ . $\mathbf{N}(\mathbf{F})$ is created by first sampling a random Gaussian $\mathbf{N}_0 \in \mathbb{R}^{H \times W \times 3}$ and then iteratively doing warping such that $\mathbf{N}_{t + 1} = \mathrm{warp}(\mathbf{N}_t, \mathbf{F}_{t + 1})$ where $\mathbf{F}_{t + 1} \in \mathbb{R}^{H \times W \times 2}$ denotes the flow at $t + 1$ . The structured noise $\mathbf{N}(\mathbf{F})$ is then fused with some random noise to improve visual quality, + +controlled by a degradation factor $\gamma$ [13].1 + +RGB control. To incorporate additional control with RGB frames $\tilde{\mathbf{V}}$ , we use SDEdit [45]. Specifically, the diffusion-based generation process is gradually denoising $\mathbf{N}(\mathbf{F})$ , such that $\mathbf{V}_{s-1} = \text{Denoise}(\mathbf{V}_s, s)$ where $s = S, S-1, \dots, 1$ denotes the diffusion timestep, $\mathbf{V}_S = \mathbf{N}(\mathbf{F})$ is the initial noise, and the generated video is given by $\mathbf{V} = \mathbf{V}_0$ . We control this process by skipping first several steps and directly starting the denoising from step $s_1 < S$ with + +$$ +\mathbf {V} _ {s _ {1}} = \alpha_ {s _ {1}} \tilde {\mathbf {V}} + \sqrt {1 - \alpha_ {s _ {1}} ^ {2}} \mathbf {N} (\mathbf {F}), \tag {4} +$$ + +where $\alpha_{i}$ denotes the diffusion coefficient at timestep $i$ . This has been shown to control the main content of the generation, while allowing details to be synthesized [45]. + +Discussion. With our bimodal control, we pass the coarse motion and appearance information from the physics simulator to the video generator. This allows not only gener + +ating more realistic motion, but also fixes appearance artifacts caused by imperfect 3D scene reconstruction and the lack of lighting information to resolve appearance changes. The question is how much we trust the generator to overwrite the coarse information. Intuitively, this is dictated by $s_1$ : If $s_1$ is close to 0, then the generator only modifies the coarse video $\tilde{\mathbf{V}}$ a bit; if $s_1$ is close to $S$ , then it can overwrite $\tilde{\mathbf{V}}$ more and hallucinate new contents. Thus, $s_1$ positively corresponds to the responsibility of the video generator. + +Spatially varying responsibility. The responsibility of the video generator is inherently uneven across spatial regions in every frame. For example, most of our background remains static in the dynamic process, which we want to trust the simulator output $\tilde{\mathbf{V}}$ more, rather than the video generator, because the video generator may hallucinate incorrect details such as ghost objects. To this end, we introduce the spatially varying bimodal control. + +In this work, we consider two responsibility levels in our spatially varying bimodal control for the background and the dynamic objects, respectively, such that we set a lower responsibility $s_2 < s_1$ of the video generator on modifying the background. Specifically, at the step $s_2$ , we compute + +$$ +\hat {\mathbf {V}} _ {s _ {2}} = \mathbf {M} \odot \mathbf {V} _ {s _ {2}} + (\mathbf {1} - \mathbf {M}) \odot \left(\alpha_ {s _ {2}} \tilde {\mathbf {V}} + \sqrt {1 - \alpha_ {s _ {2}} ^ {2}} \mathbf {N} (\mathbf {F})\right), \tag {5} +$$ + +where $\mathbf{V}_{s_2}$ is computed from gradually denoising $\mathbf{V}_{s_1}$ . $\mathbf{M} \in \{0,1\}^{(T + 1)\times H\times W\times 3}$ denotes the binary mask that uses 1 to mark a pixel of dynamic objects and 0 to mark a pixel of the background, which is rendered from the coarse scene. After this computation, we set $\mathbf{V}_{s_2} \gets \hat{\mathbf{V}}_{s_2}$ and continue the denoising to generate $\mathbf{V}$ . + +Updating scene dynamics. Finally, we use the generated video $\mathbf{V}$ as a supervision to update the coarse dynamic scene $\{\tilde{S}_t\}_{t=0}^T$ . This is done by minimizing a photometric L1 loss: $\min_{\{\mathbf{c}_t^{\mathrm{B}}, \mathcal{O}_t\}_{t=0}^T\| \mathbf{V} - \mathbf{V}\|_1$ over the foreground object's motion trajectory and appearance $\{\mathcal{O}_t\}_{t=0}^T$ . We also update the background color $\mathbf{c}_t^{\mathrm{B}}$ for shading effects. This optimization yields the final dynamic scene $\{\mathcal{S}_t\}_{t=0}^T$ . + +# 4. Experiments + +Implementation details. For physics simulation, we adopt the Genesis [3] framework, which unifies several different physics solvers. For all scenes, we run physical simulation for 960 steps and render one frame for each 20 steps. We include additional implementation details in Appendix B.1. + +Baselines. We compare against two types of baselines for action-conditioned 3d dynamic scene generation: physics-based and conditional video generation methods. For physics-based methods, we compare with PhysGen [43] and PhysGaussian [67]. PhysGen decomposes an image into 2D rigid bodies and run rigid simulation given certain action. PhysGaussian models the 3D scene as elastic objects + +
Physics PlausibilityMotion FidelityVisual Quality
Over PhysGen [43]78.0%78.0%80.1%
Over PhysGaussian [67]80.2%81.2%85.2%
Over Tora [74]77.0%72.0%71.0%
OverCogVideoX-I2V [70]80.2%73.0%74.6%
+ +Table 1. Human study 2AFC results of favor rate of WonderPlay (Ours) over baseline methods. + +
MethodsImaging (↑) Aesthetic (↑) Motion (↑) Consistency (↑) PhysReal (↑)
PhysGen0.6920.5930.9920.2120.545
PhysGaussian0.4920.5640.9940.2060.350
CogVideoX0.6860.5740.9930.2190.670
Tora0.6440.6200.9920.2100.530
Ours0.6950.6100.9950.2170.700
+ +Table 2. Quantitative comparison to baselines on 15 scenes. + +with the MPM [30] framework. Since PhysGen only requires a single image as the input, we directly follow their preprocessing for image decomposition. PhysGaussian requires multiview images to reconstruct the underlying scene first, so we provide as input our reconstructed 3D scene and then run simulation with given actions. For conditional video generation methods, we compare against two methods: CogVideoX-I2V [70] with text prompts and Tora [74] with drag-based conditioning. For Tora, we use the trajectories from our simulation as the drag input. + +Metrics. We render videos from the input viewpoint to compute quantitative metrics. We adopt the imaging, aesthetic, motion quality, and consistency metrics from VBench [29]. We also adopt the GPT-4o-based physical realism metric [15]. We curate 15 examples, including 7 real photos and 8 realistic synthetic images, covering diverse types of scenes contents including cloth, rigid body, elastic objects, liquid, gas, granular substance, etc. + +# 4.1. Results + +Comparison to baselines. We show side-by-side comparisons on two scenes in Figure 4. The top row shows input images, actions (gravity for duck dropping, force to pull the red boat towards the right) and the text prompt for CogVideoX-I2V [70], followed by the action-conditioned generated dynamics from our method and the baselines. + +Despite their ability to produce plausible visual quality, video generation methods struggle to adhere to the actions. In the duck-dropping-into-water scene, Tora [74] submerges the duck under the water and then changes its shape after it re-emerges. CogVideoX-I2V struggles to generate realistic dynamics for the duck's drop and adds undesirable dynamics by moving the duck to the left. Both models also struggle with the boats scene. Tora completely alters the scene mid-video, while CogVideoX-I2V fails to generate meaningful dynamics. + +As for the physics-based method, PhysGen [43] is limited to rigid body simulation in 2D space, making it hard to + +![](images/3ad22bb3ea5c395104c7f8cb600f4f91ec051e175f6beaf4fa0aba74189f6e1f.jpg) + +![](images/e04d5cf93e82e92eb7032949794c0c2b3c0b0abb6d5670ca243a963c3b80bb75.jpg) + +![](images/8f66dd3cd8763044e4417859e2f39458bbd6c41ebd126f6f44342c2c86cc4b9f.jpg) + +![](images/76fe9f2c421d7417b7a7328e7efb7e66b11f5af729e2de146e80d7068c2338a9.jpg) + +![](images/fa447f95dd14f1b4903d551bb8167798a289a5ad82bff51214e59e5ce4a0f9f8.jpg) + +![](images/7b512cb619d3022ddd12e06aafd26ef31e0a270042514699d1fac464f281e43d.jpg) + +![](images/82856af6b82bba2573cdd2b98dc4cf6eb9589f1c61755dc4ac88f53b53e4bb4f.jpg) + +![](images/4711e8f9267802b682da32060a809a8c832ddafc10c64f4e9ec7ae88be327b0f.jpg) + +![](images/84ded896d991bebc318f6c11b6ab03230d60eb5ea7a5d660d5930abf2c70147e.jpg) + +![](images/d05bad49b02ee3f60dc0c363e9349b74fe9b2273c9215a04e1a0d650cefe9d57.jpg) + +![](images/6847a04397b7a7f961e725d60ee3017938202841d63d17a0d44a4b393a5481fc.jpg) +Input Images & Actions + +![](images/9d31d120d8102f4fd6ce56c696c7ef796d825cb50f2f658686fcb3c23f4f850a.jpg) +Action-conditioned Dynamic 3D Scenes + +![](images/817b4740b8881656f3d9dc201db85843014713e8c0ad5f5267aea8da9f61b51b.jpg) +Figure 5. Qualitative results of the proposed WonderPlay. In the left column we show the input scene image and actions, where $\rightarrow$ indicate gravity action, wind field action and 3D point force action, respectively. + +![](images/9d85cd79976315769dd977d8d7b6e807c60f44bafa83f81ea1f5853295fdcccb.jpg) + +![](images/79371a0035f52383aa473063cab756abcefef8b107e68ce5fb1d44c032b48387.jpg) + +![](images/6e9c1c83c0ee74483ba3654db0f9961c2b49f34e703b5b560eb421a4c51f31d7.jpg) + +![](images/45c8f87a4af04ae3f13bd63d2e45bccd117b8278ef90dc1ce15da54d4908d7f1.jpg) + +![](images/7a7c9860b8e16f50a7a9d1cc30e5be5bee531e3c93e82450c21e90ba50c76533.jpg) + +![](images/21c46afdb8b83f819412b31c167f7a30bd8de7eec4810c4ff653fc9222405b71.jpg) +Figure 6. Different actions on the same scene. WonderPlay supports to use different 3D actions on the same scene. Here we show four different scenes and the corresponding dynamics from two different actions within each scene. + +![](images/ba9805bd6ed56262dda9c9e9ce3d25d680e77e96807bffacce25e765e8a5baf2.jpg) + +![](images/2b9efe429050db97df6621e68582068133ae82353ab00fb94142c71039712066.jpg) + +![](images/4f427f8d0bf21d19460d5c0913304d4fd85af40a72b076280cc707e8d970addb.jpg) + +![](images/42715b113df5edd6aad47f02083cdaa8ecd98d79bc7c68a01471450b86c23759.jpg) + +![](images/96c5e5abe4a211ebbfb5662699851a29770b8921136c55e1fd8ec7f99ee0cbcb.jpg) + +![](images/55d62599c2766867c09e0d75f328f9982cbb3947d3b07ea5688412ac6a894898.jpg) + +![](images/a3476e57a608d46ec18a7ec46a4a9f46c39647c23010223d727d0823b7bd5397.jpg) + +![](images/c56b27e0c421671110c77a30c3dad8929848064b38ae0334df857aaa1d09c82a.jpg) + +![](images/85bbed40aa889da6aa54040e48b8b7a83f5228dd5ccef30ca18336e40663c7b8.jpg) + +handle scene with complex materials such as water. Phys-Gaussian [43] typically requires multiview images and as a result, struggles to produce a reasonable 3D representation with only one input image. Also due to the lack of + +a complete 3D physical state, both physics-based methods fail to properly handle the shading effect in the boats scene and make the reflections move with the boats. Our method, in contrast, offers the advantages of both physical simul + +![](images/cdc70325b97af84a151704181aa8289b9a5f4dd120a1e3fa3de8a6e933680785.jpg) +Figure 7. Ablation on hybrid generative simulator. Top row: Coarse simulation (i.e., only physics solver is used without video generator for refinement). Bottom row: Refined dynamic scene. + +![](images/9e2dcc46b985ab19afc3954ae5b35bead5d20c1afe7349ea816a8a51090e5acc.jpg) +Figure 8. Ablation on the motion signal and the appearance signal to condition the video generator. + +tion and video generation: the physical simulation handles a wide range of materials and ensures the desired dynamics, and the video generation model provides visual realism by successfully synthesizing water waves and bubbles surrounding the duck, as well as the following reflections. Also shown in Table 2, WonderPlay (ours) achieve the best or second-best performance across all metrics, showing strong motion quality, visual quality, and physical plausibility. + +User study. To evaluate the generation results with human preference, we recruit 200 participants and conduct a user study. We employ a Two-alternative Forced Choice (2AFC) protocol. Each participant evaluates 10 scenes. The participants view an action description alongside a randomly ordered side-by-side comparison video: one from our method and one from a baseline. Participants then select which video demonstrates superior performance in one of three criteria: physics plausibility which measures the correctness of the predicted motion in response to the action, motion fidelity that reflects the quality and naturalness of the generated motion, and visual quality. + +We show the averaged results on all scenes in Table 1. In comparison to all baselines, about $70\%$ to $80\%$ of the participants prefer WonderPlay (ours) across all three aspects, proving the superior performance of combining the physical simulator and video generator for dynamics with fidelity in response to actions and realistic visual appearance. + +Diverse scenes and materials. In Figure 5, we present the generated dynamic 3D scenes on a variety of input images with diverse actions. It is important to note that achieving + +realistic visual quality in simulations of complex materials from a single image input with limited physical state information is extremely challenging. However, with the aid of the video generator, the sticky jam appears vivid as it pours onto the cake, and the river waves look natural in response to the boat's movement. Notably, the underlying physical simulator ensures that all dynamics follow the input actions. For example, the roses are initially blown to the right by the wind and then move back due to their elasticity. + +Condition on different actions. A significant advantage of our method is that it enables generating different interactions with different actions in the same scene. In Figure 6, we present four scenes, each with two different actions and their corresponding output dynamic 3D scenes. + +# 4.2. Ablation on Hybrid Generative Simulator + +In the following we discuss an ablation study on the hybrid generative simulator. We leave quantitative numbers and further ablation in Appendix A. + +Video generator refines both dynamics and appearance. In Figure 7, we compare a dynamic scene created solely with the physics simulator, i.e., the coarse simulation (top row), and the refined dynamic scene created by our full model (bottom row). In the coarse simulation, we observe unrealistic motion: the motion of the smoke looks too sticky due to numerical viscosity that exists for almost all fluid solvers. The video model refines it so that the fluid motion looks smooth with swirls. There are also appearance artifacts in the coarse simulation such as the grainy smoke, where the video model can also refine them. + +Both signals to condition the video generator are necessary. To demonstrate the benefit of conditioning the video generator on both motion and appearance signals, we show ablation results in Figure 8. The top row shows the synthesized video from our full model with both signals; "w/o RGB" uses motion but no appearance; and "w/o flow" uses appearance but no motion. Using only RGB conditioning ("w/o flow"), the video model fails to retain or improve detailed dynamics in the sand grains. Using only a motion signal ("w/o RGB") leads to unexpected hallucinations beyond user action input, e.g., it hallucinates a pile of sand standing in the back and the background texture unexpectedly changes. In contrast, using both signals produces the best results. + +# 5. Conclusion + +In this work, we propose WonderPlay, a novel framework for action-conditioned dynamic 3D scene generation from a single image. WonderPlay features a hybrid generative simulator for simulation fidelity and visual quality. We showcase superior performance of WonderPlay on diverse scenes with various interactions. + +Acknowledgments. We thank Guandao Yang, Yunzhi Zhang, and Zhehao Li for the comments and fruitful discussions, and Hadi Alzayer for help reviewing the draft. This work is in part supported by the Stanford Institute for Human-Centered AI (HAI), the Okawa Foundation Research Grant, NSF RI #2211258 and #2338203, ONR YIP N00014-24-1-2117, and ONR MURI N00014-22-1-2740. + +# References + +[1] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 3 +[2] Luca Savant Aira, Antonio Montanaro, Emanuele Aiello, Diego Valsesia, and Enrico Magli. Motioncraft: Physics-based zero-shot video generation. arXiv preprint arXiv:2405.13557, 2024. 2 +[3] Genesis Authors. Genesis: A universal and generative physics engine for robotics and beyond, 2024. 6 +[4] Sherwin Bahmani, Xian Liu, Wang Yifan, Ivan Skorokhodov, Victor Rong, Ziwei Liu, Xihui Liu, Jeong Joon Park, Sergey Tulyakov, Gordon Wetzstein, Andrea Tagliasacchi, and David B. Lindell. Tc4d: Trajectory-conditioned text-to-4d generation. arXiv, 2024. 3 +[5] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996-8006, 2024. 3 +[6] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 2 +[7] Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. arXiv preprint arXiv:2412.03572, 2024. 3 +[8] Bar-Tal et al. Lumiere: A space-time diffusion model for video generation. arXiv:2401.12945, 2024. 2, 3 +[9] Jan Bender, Matthias Müller, and Miles Macklin. Position-based simulation methods in computer graphics. In Eurographics (tutorials), 2015. S2 +[10] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[11] Tim Brooks, Bill Peebles, Connor Homes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators, 2024. 2, 3 +[12] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, + +Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3 +[13] Ryan Burgert, Yuancheng Xu, Wenqi Xian, Oliver Pilarski, Pascal Clausen, Mingming He, Li Ma, Yitong Deng, Lingxiao Li, Mohsen Mousavi, et al. Go-with-the-flow: Motion-controllable video diffusion models using real-time warped noise. arXiv preprint arXiv:2501.08331, 2025. 5 +[14] Haoxuan Che, Xuanhua He, Quande Liu, Cheng Jin, and Hao Chen. Gamegen-x: Interactive open-world game video generation. arXiv preprint arXiv:2411.00769, 2024. 2, 3 +[15] Boyuan Chen, Hanxiao Jiang, Shaowei Liu, Saurabh Gupta, Yunzhu Li, Hao Zhao, and Shenlong Wang. Physgen3d: Crafting a miniature interactive world from a single image. CVPR, 2025. 2, 6 +[16] Duowen Chen, Zhiqi Li, Junwei Zhou, Fan Feng, Tao Du, and Bo Zhu. Solid-fluid interaction on particle flow maps. ACM Transactions on Graphics (TOG), 2024. 2 +[17] Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne Vouga, and Christoph Lassner. Virtual elastic objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[18] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 3 +[19] Abe Davis, Katherine L Bouman, Justin G Chen, Michael Rubinstein, Fredo Durand, and William T Freeman. Visual vibrometry: Estimating material properties from small motion in video. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2015. 2 +[20] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG), 2015. 2 +[21] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 3 +[22] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 3 +[23] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 3 +[24] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation, 2024. 3 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2 +[26] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3 + +[27] Chen Hou, Guoqiang Wei, Yan Zeng, and Zhibo Chen. Training-free camera control for video generation. arXiv preprint arXiv:2406.10126, 2024. 3 +[28] Tianyu Huang, Haoze Zhang, Yihan Zeng, Zhilu Zhang, Hui Li, Wangmeng Zuo, and Rynson WH Lau. Dreamphysics: Learning physical properties of dynamic 3d gaussians with video diffusion priors. arXiv preprint arXiv:2406.01476, 2024. 2 +[29] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6 +[30] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. Acm siggraph 2016 courses, 2016. 6, S2 +[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. S1 +[32] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. S1 +[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. S1 +[34] Simon Le Cleac'h, Hong-Xing Yu, Michelle Guo, Taylor Howell, Ruohan Gao, Jiajun Wu, Zachary Manchester, and Mac Schwager. Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters, 2023. 2 +[35] Yao-Chih Lee, Yi-Ting Chen, Andrew Wang, Ting-Hsuan Liao, Brandon Y Feng, and Jia-Bin Huang. Vividdream: Generating 3d scene with ambient dynamics. arXiv preprint arXiv:2405.20334, 2024. 3 +[36] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. arXiv preprint arXiv:2303.05512, 2023. 2 +[37] Yaowei Li, Xintao Wang, Zhaoyang Zhang, Zhouxia Wang, Ziyang Yuan, Liangbin Xie, Yuexian Zou, and Ying Shan. Image conductor: Precision control for interactive video synthesis. arXiv preprint arXiv:2406.15339, 2024. 3 +[38] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[39] Jiajing Lin, Zhenzhong Wang, Shu Jiang, Yongjie Hou, and Min Jiang. Phys4dgen: A physics-driven framework for controllable and efficient 4d content generation from a single image. arXiv preprint arXiv:2411.16800, 2024. 2 +[40] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d + +with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8576-8588, 2024. 3 +[41] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338, 2024. 2 +[42] Jinxiu Liu, Shaoheng Lin, Yinxiao Li, and Ming-Hsuan Yang. Dynamicscaler: Seamless and scalable video generation for panoramic scenes. arXiv preprint arXiv:2412.11100, 2024.3 +[43] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imaged-to-video generation. In ECCV, 2024. 2, 4, 6, 7, S1 +[44] Miles Macklin and Matthias Müller. Position based fluids. ACM Transactions on Graphics (TOG), 2013. S2 +[45] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. ICLR, 2022. 5 +[46] Midjourney. https://www.midjourney.com/, 2023. 2 +[47] Matthias Müller, Bruno Heidelberger, Matthias Teschner, and Markus Gross. Meshless deformations based on shape matching. ACM transactions on graphics (TOG), 2005. S2 +[48] Matthias Müller, Bruno Heidelberger, Marcus Hennix, and John Ratcliff. Position based dynamics. Journal of Visual Communication and Image Representation, 2007. S2 +[49] Koichi Namekata, Sherwin Bahmani, Ziyi Wu, Yash Kant, Igor Gilitschenski, and David B Lindell. Sg-i2v: Self-guided trajectory control in image-to-video generation. arXiv preprint arXiv:2411.04989, 2024. 3 +[50] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adaptations in frozen image-to-video diffusion model. arXiv preprint arXiv:2405.20222, 2024. 3 +[51] Mohammad Nomaan Qureshi, Sparsh Garg, Francisco Yandun, David Held, George Kantor, and Abhisesh Silwal. Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting. arXiv preprint arXiv:2409.10161, 2024. 2 +[52] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3 +[53] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[54] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3 + +[55] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3 +[56] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2 +[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, S1 +[58] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 3 +[59] Xiyang Tan, Ying Jiang, Xuan Li, Zeshun Zong, Tianyi Xie, Yin Yang, and Chenfanfu Jiang. Physmotion: Physics-grounded dynamics from a single image. arXiv preprint arXiv:2411.17189, 2024. 2, 3, 4 +[60] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. arXiv preprint arXiv:2408.14837, 2024. 3 +[61] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. S1 +[62] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, 2024. 3 +[63] Daniel Watson, Saurabh Saxena, Lala Li, Andrea Tagliasacchi, and David J Fleet. Controlling space and time with diffusion models. arXiv preprint arXiv:2407.07860, 2024. 3 +[64] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 3 +[65] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, 2024. 3 +[66] Zeqi Xiao, Yifan Zhou, Shuai Yang, and Xingang Pan. Video diffusion models are training-free motion interpreter and controller. arXiv preprint arXiv:2405.14864, 2024. 3 +[67] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In CVPR, 2024. 2, 6, S1, S2 +[68] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 3 +[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large + +reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 4, S1 +[70] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5, 6 +[71] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 3 +[72] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 4, S1 +[73] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In ECCV, 2024. 2, 4 +[74] Zhenghao Zhang, Junchao Liao, Menghao Li, Zuozhuo Dai, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 3, 6 +[75] Yuyang Zhao, Zhiwen Yan, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Animate124: Animating one image to 4d dynamic scene. arXiv preprint arXiv:2311.14603, 2023. 3 +[76] Yuyang Zhao, Chung-Ching Lin, Kevin Lin, Zhiwen Yan, Linjie Li, Zhengyuan Yang, Jianfeng Wang, Gim Hee Lee, and Lijuan Wang. Genxd: Generating any 3d and 4d scenes. arXiv preprint arXiv:2411.02319, 2024. 3 +[77] Guangcong Zheng, Teng Li, Rui Jiang, Yehao Lu, Tao Wu, and Xi Li. Cami2v: Camera-controlled image-to-video diffusion model. arXiv preprint arXiv:2410.15957, 2024. 3 +[78] Licheng Zhong, Hong-Xing Yu, Jiajun Wu, and Yunzhu Li. Reconstruction and simulation of elastic objects with spring-mass 3d gaussians. In European Conference on Computer Vision, 2024. 2 \ No newline at end of file diff --git a/wonderplaydynamic3dscenegenerationfromasingleimageandactions/images.zip b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7705446c7c04281c587be41a98b5ac6cb3ddccd1 --- /dev/null +++ b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb889c127119b5f9247654ba7ca9dae3e3870e040567e9c389151095b8137e40 +size 1005010 diff --git a/wonderplaydynamic3dscenegenerationfromasingleimageandactions/layout.json b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2eb119812efd280d9b1843e97a98dec38e962485 --- /dev/null +++ b/wonderplaydynamic3dscenegenerationfromasingleimageandactions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a7c651462400b22dd5895b74fd60add4f83c10c544b360ee1cc0deeee93763d +size 494976 diff --git a/wonderturbogeneratinginteractive3dworldin072seconds/004277e4-cc32-4a24-bc57-88909bd31122_content_list.json b/wonderturbogeneratinginteractive3dworldin072seconds/004277e4-cc32-4a24-bc57-88909bd31122_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..67ddffe3671dc3f040ae00f7ca90d9cdaebd897c --- /dev/null +++ b/wonderturbogeneratinginteractive3dworldin072seconds/004277e4-cc32-4a24-bc57-88909bd31122_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc9d06f8c29988d328b49caadf772f5d7bb431d4db7a352ef10f65672eb16770 +size 83362 diff --git a/wonderturbogeneratinginteractive3dworldin072seconds/004277e4-cc32-4a24-bc57-88909bd31122_model.json b/wonderturbogeneratinginteractive3dworldin072seconds/004277e4-cc32-4a24-bc57-88909bd31122_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ae0919301a6f4cbad8df04d9cd6c43ebcf87d50e --- /dev/null +++ b/wonderturbogeneratinginteractive3dworldin072seconds/004277e4-cc32-4a24-bc57-88909bd31122_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f192155b8a2e22c6dd4dd4b19016d7482d16b3cdd1fdc53ab7492d4d1f5eae2c +size 107571 diff --git a/wonderturbogeneratinginteractive3dworldin072seconds/004277e4-cc32-4a24-bc57-88909bd31122_origin.pdf b/wonderturbogeneratinginteractive3dworldin072seconds/004277e4-cc32-4a24-bc57-88909bd31122_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..153db2ca63cdf8c23b03a04ad7ea8f7180d1438b --- /dev/null +++ b/wonderturbogeneratinginteractive3dworldin072seconds/004277e4-cc32-4a24-bc57-88909bd31122_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad33ec053e8cd0cae4435b9eac6a796ccba82faaff2977649d251514c317ad03 +size 14522416 diff --git a/wonderturbogeneratinginteractive3dworldin072seconds/full.md b/wonderturbogeneratinginteractive3dworldin072seconds/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dac754077c0f69692c36e303993e3f12486611d9 --- /dev/null +++ b/wonderturbogeneratinginteractive3dworldin072seconds/full.md @@ -0,0 +1,351 @@ +# WonderTurbo: Generating Interactive 3D World in 0.72 Seconds + +Chaojun Ni $^{1,2*}$ Xiaofeng Wang $^{2*}$ Zheng Zhu $^{2*†}$ Weijie Wang $^{2,3*}$ Haoyun Li $^{2,4}$ Guosheng Zhao $^{2,4}$ Jie Li $^{2}$ Wenkang Qin $^{2}$ Guan Huang $^{2}$ Wenjun Mei $^{1\dagger}$ Peking University $^{2}$ GigaAI $^{3}$ Zhejiang University Institute of Automation, Chinese Academy of Sciences + +![](images/c7c80b12d72bd6918e97abe1c59d6258a0d9691e1dff4ec939142eb73b41512d.jpg) +Figure 1. Beginning with a single image, users can freely adjust the viewpoint and interactively control the generation of a 3D scene, each interaction requiring only 0.72 seconds. + +# Abstract + +Interactive 3D generation is gaining momentum and capturing extensive attention for its potential to create immersive virtual experiences. However, a critical challenge in cur + +rent 3D generation technologies lies in achieving real-time interactivity. To address this issue, we introduce WonderTurbo, the first real-time interactive 3D scene generation framework capable of generating novel perspectives of 3D scenes within 0.72 seconds. Specifically, WonderTurbo accelerates both geometric and appearance modeling in 3D scene generation. In terms of geometry, we propose StepSplat, an innovative method that constructs efficient 3D + +geometric representations through dynamic updates, each taking only 0.26 seconds. Additionally, we design QuickDepth, a lightweight depth completion module that provides consistent depth input for StepSplat, further enhancing geometric accuracy. For appearance modeling, we develop FastPaint, a 2-steps diffusion model tailored for instant inpainting, which focuses on maintaining spatial appearance consistency. Experimental results demonstrate that WonderTurbo achieves a remarkable $15 \times$ speedup compared to baseline methods, while preserving excellent spatial consistency and delivering high-quality output. + +# 1. Introduction + +The online generation of 3D from a single image [69, 70], which involves instant creation and updating of 3D scenes, has attracted significant attention. Compared with offline generation methods [11, 22, 31, 33, 37, 71, 78], which generate a fixed 3D scene based on user texts and input images, online generation allows users to interactively create and edit 3D content, enhancing creation efficiency and flexibility. + +Despite these advantages, existing online 3D scene generation methods [69, 70] still face significant challenges, primarily due to the low inference efficiency. This efficiency bottleneck largely stems from the time-consuming processes of optimizing geometric details and generating or refining new view appearances. Specifically, 3D scene representation methods like 3D Gaussian Splittings (3DGS) [24] require iterative training to update new geometries, while appearance refinement is based on diffusion-based image inpainting models [40] that require extensive inference steps, further increasing computational overhead. For example, the current fastest online 3D generation approach, WonderWorld [69], takes nearly 10 seconds to update a single 3D view, which falls short of real-time performance expectations. Although some works [34, 43, 48, 49, 57, 77, 77] on generating novel views from a single image improve in speed, these methods only support generating views within small viewpoint changes. + +In this paper, we present WonderTurbo, a novel framework designed for real-time interactive 3D scene generation. To address the critical challenges of inference efficiency, we optimize both geometric representation and appearance modeling. For geometric optimization, we introduce StepSplat, a scalable method that accelerates 3D scene expansion in 0.26 seconds. Unlike conventional 3DGS methods [24, 33, 60, 65, 74] that rely on iterative training to update 3D representations, StepSplat leverages insights from recent feed-forward approaches [7, 9, 55, 59] to perform direct inference 3DGS. Moreover, StepSplat extends the feed-forward paradigm to interactive 3D geometric representation, ensuring consistency across dynamic viewpoint + +changes. This is achieved through the maintenance of a feature memory module, which adaptively constructs cost volumes as the viewpoint changes. Meanwhile, to further enhance depth coherence, we incorporate a lightweight depth completion module, QuickDepth, to provide a consistent depth prior for StepSplat to construct the cost volume. On the appearance front, we propose FastPaint, a highly efficient method for real-time appearance refinement. In contrast to traditional diffusion-based inpainting methods [40], which require dozens of inference steps to refine appearance modeling, FastPaint achieves comparable results with only 2 inference steps while preserving spatial appearance consistency. + +We present the results of various camera setups, including a panoramic camera path and two casual walking camera paths as shown in Fig. 1. The results demonstrate that WonderTurbo can accurately generate 3D scenes based on user-provided text while maintaining high consistency. Furthermore, our model achieves leading performance in CLIP-based metrics [38, 50, 52] and user study win rates, while significantly boosting speed, achieving a $15\times$ acceleration. + +The main contributions of this paper are summarized as follows: + +- We present WonderTurbo, the first real-time (inference cost: 0.72 seconds) 3D scene generation method that allows users to interactively create diverse and cohesively connected scenes. +- For geometric efficiency optimization, the proposed StepSplat extends feed-forward paradigm to interactive 3D geometric representation, accelerating the expansion of 3D scenes within 0.26 seconds. Besides, QuickDepth is introduced to ensure depth consistency during viewpoint changes. For appearance modeling, we present FastPaint for image inpainting with only 2-steps inference. +- We perform comprehensive experiments to validate that WonderTurbo, while achieving a $15 \times$ acceleration, surpasses other methods in generating high-quality 3D scenes, both in geometry and appearance. + +# 2. Related Work + +# 2.1. Offline 3D Scene Generation from Single Image + +Offline 3D scene generation from a single image [2, 11, 16, 22, 26, 31, 37, 64, 71] has been explored by various methods, which generally involve generating multiple views or panoramas of a scene and subsequently transforming them into a 3D representation. Approaches such as Text2Room [22] and LucidDreamer [11] begin with a single input image and a user's textual description, generate multiple images of the scene, and then employ 3D optimization to refine the scene and create a more accurate and consistent 3D representation. Meanwhile, Wonderland [27] predicts 3DGS in a feed-forward manner by + +constructing 3D reconstruction models based on a video diffusion model's latent space. In contrast, methods such as GenEx [31], Pano2Room [37], and Dreamscene360 [76] synthesize coherent panoramas by leveraging pretrained text-to-panorama diffusion models [17, 67], which are then elevated to 3D, ultimately producing explorable 3D worlds. However, these methods typically operate offline, preventing user interaction during the generation process. Furthermore, once the scene is generated, modifying or adjusting the layout or content becomes challenging. + +# 2.2. Online 3D Scene Generation from Single Image + +Some researches have focused on the online interactive generation of 3D scenes, which only requires an image to generate a new 3D scene and specify its content. WonderJourney [70] employs an LLM [4, 5, 21, 62] to generate scene descriptions, a text-driven pipeline for coherent 3D scene generation, and a vision-language model (VLM) [1, 25, 75] for result verification, while allowing users to adjust text and control the generation of new 3D scenes. However, this process takes several minutes, making it unsuitable for interactive use. WonderWorld [69] reduces reconstruction time using Fast Layered Gaussian Surfels (FLAGS) and generates geometrically consistent scenes with guided diffusion-based depth estimation, but still requires about 10 seconds per scene. In contrast, WonderTurbo achieves the generation of diverse scenes within 0.72 seconds through accelerations both in geometric and appearance modeling, meeting the needs for real-time interaction. + +# 2.3. 3D Scene Representations + +3DGS [24] has attracted significant attention, for its high efficiency and photorealistic rendering. However, a major limitation of traditional 3DGS-based methods [33, 53, 61, 65] is the need for per-scene optimization, which can be time-consuming. 2D Gaussian Splatting [23] addresses this by projecting 3DGS onto a 2D plane, allowing faster rendering while maintaining geometric accuracy. Additionally, WonderWorld [69] introduces FLAGS, which employs a multi-layered representation and geometry-based initialization to reduce the need for per-scene optimization. + +However, these methods still require significant computational time. Therefore, recent work [7, 9, 10, 45, 55, 59, 66] has explored using feed-forward networks to predict 3D geometry from images. PixelSplat [7] predicts 3DGS distributions by learning from paired images and MVSplat [9] leverages multi-view correspondence information through the construction of a cost volume. However, these methods are not well-suited for generating interactive 3D scenes. Specifically, they struggle to handle scenarios with gradually increasing views effectively and rely on unsupervised depth estimation, which results in poor generalization. + +Recent explorations [54, 55, 59] have attempted to ad + +dress these challenges. FreeSplat [55] reconstructs long sequence inputs by constructing an adaptive cost volume between adjacent views and aggregating features through multi-scale structures. However, it lacks depth supervision and does not address the requirement for handling a gradually increasing number of views. DepthSplat [59] integrates monocular depth estimation priors [63] into the feedforward process, but its generalization capability remains limited, and it does not meet the needs of gradually increasing views. + +# 3. Method + +# 3.1. Overall Framework of WonderTurbo + +Interactive 3D scene generation [69, 70] is constrained by computational efficiency due to the time-consuming geometry and appearance modeling. WonderWorld [69] introduces FLAGS to accelerate geometric modeling. However, it still requires hundreds of iterations to optimize the geometry representation, and its appearance modeling relies on a pretrained diffusion model [8, 35, 36, 40] that needs dozens of inference steps for inpainting. In contrast, WonderTurbo achieves real-time interactive 3D scene generation by accelerating both geometry and appearance modeling. Specifically, we propose StepSplat for geometric modeling acceleration, which directly infers 3DGS in 0.26 seconds. Within this framework, QuickDepth completes missing depth information in 0.24 seconds. For appearance modeling acceleration, we introduce FastPaint, which completes image inpainting in 0.22 seconds. + +We present the pipeline of WonderTurbo in Fig. 2. At the $i$ -th iteration, given a user-specified location, FastPaint leverages the rendered image $I_{render}^{i}$ from the current 3D scene and a user-provided textual description to generate a new scene appearance $I_{target}^{i}$ . Subsequently, QuickDepth generates depth maps $D_{target}^{i}$ using the rendered depth map $D_{render}^{i}$ and the newly generated appearance $I_{target}^{i}$ , ensuring that the geometry of the newly generated scene aligns with the existing 3D scene. Finally, StepSplat takes the depth map $D_{target}^{i}$ and the new scene appearance $I_{target}^{i}$ as inputs, incrementally fusing $G_{local}^{i}$ into the global representation $G_{global}^{i}$ . In the following sections, we delve into the details of StepSplat, QuickDepth and FastPaint. + +# 3.2.StepSplat + +To accelerate the modeling of appearance, we introduce StepSplat. As shown in Fig. 3, StepSplat takes as input the pose $P^i$ , image $I_{target}^i$ , and corresponding depth $D_{target}^i$ from QuickDepth, and first uses backbone to extract both the matching features $F_m^i$ and image features $F_e^i$ . Then, it queries the feature memory for nearby views' matching features to construct cost volumes. Then, cost volumes are concatenated with $F_e^i$ to predict Gaussian parameters. Mean + +![](images/a103c5f47297034ed048da15390b0fd4841fd7bf1c52f5fa495bcf200d3538fa.jpg) +Figure 2. The pipeline of WonderTurbo. As the user moves the real-time rendering camera and inputs the text, the rendered image and depth map are then processed by FastPaint and QuickDepth to generate coherent geometry and appearance. Finally, StepSplat performs incremental fusion based on the outputs of FastPaint and QuickDepth. + +![](images/5ee0f5c1d729c9abe5427515c84981eb14cb2ca0d7ac0139202924b3227a4017.jpg) +Figure 3. The structure of StepSplat. + +while, we leverage consistent input depth from QuickDepth as a geometric priority to construct the cost volume, which ensures the accuracy of Gaussian centers. Finally, an incremental fusion strategy merges newly generated $G_{local}^{i}$ from the current view into the $G_{global}^{i}$ , ensuring continuous and consistent 3D representation. + +Feature Memory. We introduce the feature memory to store the matching features of previous views, which are used to build the late cost volume. Given the input images $I_{target}^{i}$ and $P^{i}$ , we first feed them into backbone to extract image features $F_{e}^{i}$ and matching features $F_{m}^{i}$ . Then, we construct a tuple $(P^{i}, F_{m}^{i})$ that is updated into the feature memory. To accelerate inference, we adapt RepVGG [14] as the backbone. + +Depth Guided Cost Volume. For constructing the cost volume for the current view, we adaptively select $N_v$ neighboring views from the feature memory and use the input depth from QuickDepth for the depth candidates of the cost volume. To achieve this, we first compute the distance between + +$P^i$ and all stored poses $\{P^n\}_{n = 1}^{i - 1}$ in the Feature Memory: + +$$ +d \left(P ^ {n}, P ^ {i}\right) = \| P ^ {n} - P ^ {i} \| _ {2}, \tag {1} +$$ + +where $\| \cdot \|_2$ represents the L2 norm. From these distances, we select the $N_v$ closest poses to the current pose and extract their corresponding matching features from Memory as $\{(P^{t_n}, F_m^{t_n})\}_{n=1}^{N_v}$ , where each $t_n$ corresponds to one of the $N_v$ closest poses. + +For ensuring consistent 3D representation, inspired by Multi-View Stereo [18, 42, 54, 56], we use $D_{target}^{i}$ to guide the construction of the cost volume. Specifically, $N_{d}$ depth candidates $\{d_s\}_{s=1}^{N_d}$ are uniformly sampled from $R$ as: + +$$ +R = \left\{d \mid (1 - a) \cdot D _ {\text {t a r g e t}} ^ {i} \leq d \leq (1 + a) \cdot D _ {\text {t a r g e t}} ^ {i} \right\}, \tag {2} +$$ + +where $a$ is the offset value used to adjust the depth candidate range. Then, each neighboring view's matching feature $F_{m}^{t_{n}}$ is warped to the candidate depth $d_{s}$ planes of the current view using the plane-sweep stereo algorithm [12], the feature warping is formulated as: + +$$ +F _ {d _ {s}} ^ {t _ {n} \rightarrow i} = \mathcal {W} \left(F _ {m} ^ {t _ {n}}, P ^ {i}, P ^ {t _ {n}}, d _ {s}\right), \tag {3} +$$ + +where $\mathcal{W}$ denotes the differentiable warping operation. We then compute the normalized dot-product correlation between the current view's feature $F_{m}^{i}$ and each warped neighboring feature $F_{ds}^{i_n\rightarrow i}$ , and average the correlation maps from all neighboring views: + +$$ +S ^ {i} = \left[ \frac {1}{N _ {v}} \sum_ {n = 1} ^ {N _ {v}} F _ {m} ^ {i} \cdot F _ {d _ {s}} ^ {t _ {n} \rightarrow i} \right] _ {s = 1} ^ {N _ {d}}, \tag {4} +$$ + +where $N_{d}$ is the number of candidate depths, and the correlation maps from each depth are stacked to form the cost + +volume $S^i$ . Meanwhile, an additional 2D U-Net [41] is used to further refine and upsample the cost volume. We normalize the cost volume $S^i$ and perform a weighted average of all depth candidates to obtain the predicted depth map $\hat{d}$ : + +$$ +\hat {d} = \operatorname {s o f t m a x} \left(S ^ {i}\right) \cdot d. \tag {5} +$$ + +After obtaining the depth predictions, the depth values are unprojected as the center of the 3DGS. The cost volume and image feature are then decoded to obtain other Gaussian parameters, similar to MVSplat [9]. + +Incremental Fusion. To reduce the redundancy of Gaussians, we achieve the incremental fusion by updating $G_{local}^{i}$ to $G_{global}^{i+1}$ using a depth constraint. Specifically, given $G_{local}^{i}$ with a 2D coordinate $[x_{local}, y_{local}]$ and depth $d_{local}$ , we project all global Gaussians $\{\mu_{global}^{j}\}_{j=1}^{K}$ from $G_{global}^{i}$ onto the current pixel coordinate system using the camera projection matrix $\mathbf{P}^{i}$ : + +$$ +\left[ x _ {j} ^ {g}, y _ {j} ^ {g}, d _ {j} ^ {g} \right] ^ {\top} = \mathbf {P} ^ {i} \mu_ {\text {g l o b a l}} ^ {j}, \tag {6} +$$ + +We then construct a candidate set of global Gaussians projected to the same discrete pixel location via: + +$$ +\mathcal {S} _ {\text {g l o b a l}} = \left\{j \mid \left\lfloor x _ {j} ^ {g} \right\rfloor = x _ {\text {l o c a l}} \wedge \left\lfloor y _ {j} ^ {g} \right\rfloor = y _ {\text {l o c a l}} \right\}. \tag {7} +$$ + +To enforce geometric continuity, we prune conflicting Gaussians from $S_{global}$ that violate the depth consistency constraint. The Gaussians to be pruned are those in $\mathcal{C}$ , defined as: + +$$ +\mathcal {C} = \left\{j \in \mathcal {S} _ {\text {g l o b a l}} \mid \left| d _ {\text {l o c a l}} - d _ {j} ^ {g} \right| < \delta \cdot d _ {\text {l o c a l}} \right\}, \tag {8} +$$ + +where $\delta$ controls the depth tolerance. The global model is then updated by selectively merging the valid local Gaussians, which are those not included in $\mathcal{C}$ , into the existing global model, as shown by: + +$$ +G _ {g l o b a l} ^ {i + 1} \leftarrow G _ {g l o b a l} ^ {i} \cup \left(\mathrm {G} _ {l o c a l} ^ {i} \backslash \mathcal {C}\right). \tag {9} +$$ + +Training of StepSplat. Traditional 3DGS feed-forward methods [7, 9, 29, 55] struggle to meet the demands of interactive 3D scene generation. This is partly due to the limited diversity of datasets, which focus on specific scenes such as autonomous driving [6, 47] or indoor environments [13, 46]. Additionally, there is a significant gap between the viewpoint variations in these datasets and the requirements of interactive 3D scene generation. To address this challenge, we create a dataset utilizing 3D generation models [11, 22, 37, 69, 70] with simulated viewpoint changes for the purpose of training StepSplat, which is detailed in Section 3.5. During training, we randomly select an image sequence and feed these images into the model one by one to generate a global Gaussian representation. This representation is then used to render images from novel viewpoints, with the RGB images serving as supervision. + +![](images/582d455187c8504f220d6452e934984f30d0f9d6918d11cdaa06c72b434dfb52.jpg) +Figure 4. The process of constructing the interactive 3D generation dataset. + +# 3.3. QuickDepth + +In the field of depth completion, existing methods [19, 28-30, 63] have made notable advancements. However, these methods are generally designed for sparse depth completion and face challenges in completing depth for regions that lack any depth information, a critical requirement for interactive 3D scene generation. To address this, WonderWorld [69] introduces a training-free guided depth diffusion method, but it requires over 3 seconds per depth map. Invisible Stitch [15] trains a depth completion model through teacher distillation and self-training due to the lack of ground-truth data. However, its training data are limited, leading to a decline in performance in some scenarios. We introduce QuickDepth, a lightweight depth completion model trained on our constructed dataset for interactive 3D scene generation, with strong generalization capabilities, performing well across a wide range of scenarios. + +To adapt QuickDepth to interactive 3D scene generation, we construct a dataset comprising diverse scenes, including indoor and outdoor environments, and scenes from comics and artworks (detailed in Section 3.5). Instead of using random masks or projections to simulate the mask in interactive 3D scene generation [15], we design a series of camera trajectories that are more aligned with interactive 3D scene generation. Specifically, we design camera poses $\{T_1,\dots ,T_n\}$ and obtain frames $\{I_1,\dots ,I_n\}$ along with their corresponding depth maps $\{D_{1},\ldots ,D_{n}\}$ from the dataset. We then utilize the geometric relationships between adjacent frames. More precisely, for each frame $I_{j}$ $(j\in [1,n])$ , we project the depth map of the previous frame $D_{j - 1}$ into the coordinate system of $I_{j}$ using the relative pose $T_{j - 1\rightarrow j}$ . This process produces incomplete depth $D_{j - 1\rightarrow j}^{\prime}$ and binary validity mask $M_{j - 1\rightarrow j}$ , where invalid pixels indicate regions that require depth completion. + +During training, we construct inputs by either masking the target frame's ground truth depth $D_{j}$ entirely or selecting a warped depth-mask pair $(D_{j - 1\rightarrow j}^{\prime},M_{j - 1\rightarrow j})$ . Meanwhile, QuickDepth is initialized with a light pre-trained depth estimation model [3] and takes as input the target + +frame's RGB image, the incomplete depth map and the binary mask. It then predicts the full depth, where the prediction is supervised by the target's ground truth depth. + +# 3.4. FastPaint + +In 3D scene generation, image inpainting [32, 51, 68, 72] is crucial for modeling 3D appearance. Some methods, like Pano2Room [37], generate panoramic images from a single input but struggle to place content at user-specified locations. Others [69, 70] use Stable Diffusion-based fin-tuned inpainting models [40]. However, during fine-tuning, the inpainting regions differ significantly from those in 3D scene generation, requiring a separate model to verify generated content [70]. Moreover, these diffusion models need multiple inference steps. Therefore, we propose FastPaint, which reduces inference to 2 steps and enhances pretrained models' inpainting via distillation and fine-tuning. + +Specifically, our approach synergistically leverages the advantages of ODE trajectory preservation and reformulation [39] to perform knowledge distillation [20] on the pretrained model [40], reducing the number of inference steps required while maintaining the quality of appearance modeling. To address the issue where inpainting regions in 3D scene generation differ from those in fine-tuning, a dataset is constructed for training FastPaint. Especially, we designed camera poses to simulate the interactive 3D generation process. By obtaining depth maps and images and using projections to acquire masks, we ensure the dataset aligns with the specific requirements of the inpainting task in this context. The construction of this dataset shares similarities with the methodology used in StepSplat and QuickDepth, particularly in simulating camera trajectories to help the models adapt to interactive 3D scene generation. + +# 3.5. Interactive 3D Generation Dataset + +Interactive 3D generation from a single image should enable diverse input styles. However, real-world data is limited to specific scenes, mainly in autonomous driving [6, 47] or indoor environments [13, 46]. This limitation results in poor generalization for current 3D generation methods [11, 37, 76]. Meanwhile, some methods [69, 70] directly use pretrained models to construct the pipeline, which may not be tailored for 3D scene generation, resulting in the need for a Vision-Language Model (VLM) [1, 25, 75] to verify if the content matches the scene style or text. + +To overcome this limitation, we build a dataset based on current 3D scene generation methods and train all our modules using this dataset. Specifically, we employ multiple 3D scene generation methods [11, 22, 37, 44, 69, 70, 73, 76] to create 3D scenes that each method excels at. Meanwhile, a VLM model [70] is used to verify whether the generated data conforms to the defined scene. Finally, the dataset contains over 6 million frames rendered through + +simulated interactive trajectories, including rotational paths, linear movements, and hybrid trajectories. The dataset primarily covers four categories: indoor environments $(32\%)$ , urban landscapes $(28\%)$ , natural terrains $(25\%)$ , and stylized artistic scenes $(15\%)$ . When training StepSplat, we impose certain restrictions on the distances between adjacent frames as inputs for StepSplat. This is to avoid using frames that are too close together, ensuring better alignment with the practical application of 3D interactive generation. For FastPaint and QuickDepth, the depth of two adjacent frames is used to project and obtain the corresponding mask. + +# 4. Experiments + +In this section, we present our experimental setup, including baselines and implementation details. Subsequently, both quantitative and qualitative results are provided to demonstrate the superiority of WonderTurbo. Finally, we conduct ablation studies to validate the efficiency of each module. + +# 4.1. Experiment Setup + +Baselines. We select representative 3D generation methods, encompassing both offline and online approaches. The offline methods include LucidDreamer [11] and Text2Room [22], which generate 3D scenes by producing multi-view images, as well as Pano2Room [37] and DreamScene360 [76], which directly generate panoramic images. For online 3D scene generation, we evaluate WonderJourney [70] and WonderWorld [69]. + +Evaluation Metrics. To evaluate the quality of 3D scene generation, following WonderWorld [69], we used CLIP [38] scores (CS), CLIP [38] consistency (CC), CLIP-IQA+ [52] (CIQA), Q-Align [58], and CLIP aesthetic [38] scores (CA) as metrics. Additionally, a user study is conducted to gather subjective feedback on visual quality. More details are provided in the supplementary materials. + +Implementation Details. To ensure comprehensive evaluation, we use input images from LucidDreamer [11], WonderJourney [70], and WonderWorld [69]. We generate 32 scenes by creating 8 scenes for each of 4 test cases and conduct evaluations with a fixed panoramic camera. For efficiency, we compare the time taken to generate scenes of the same size within the camera's view. + +# 4.2. Main Results + +Generation Speed. The time cost is crucial for interactive 3D scene generation. Despite utilizing FLAGS for acceleration, WonderWorld [69], the fastest among the compared methods, still requires more than 10 seconds as shown in Tab. 1. LucidDreamer [11] and Text2Room [22] involve generating multiple views for each new scene, substantially increasing appearance modeling time. Meanwhile, Pano2Room [37] and DreamScene360 [76] have inherent delays in panoramic image generation and require + +![](images/e4f6b04d8ea12625723e17944a59fa2b76e04ecab377e5e0426a7331da53de35.jpg) +WonderTurbo WonderWorld WonderJourney Pano2Room DreamScene360 Text2Room LucidDreamer Inference Inference Inference Inference Inference +cost: 43.70s + +![](images/5ccf403d1a8c995a058a382695d5f698311a7fc86009a7cfa8d13378e9af8e4e.jpg) + +![](images/d7075ed465c91185e6000b42ee19e1272c61f566606128f2716325af2db6e366.jpg) + +![](images/7c31cedb154fd338712cfcdd8770965c245114a4419a3935b8ba8d8ede80ccdf.jpg) + +![](images/8be6f480f9b05caaa24899f393ca5da659c9ce4909692774114f92737c2130bb.jpg) + +![](images/57253bee92b2b9cf0c92dc04c8880f8f692eceb8a6257e5fa0be9703b145d0a2.jpg) + +![](images/ed592b9cc1454fcdf2d494c2f95cb1f7976fc365c00c1855a78a604e98ce7973.jpg) +Prompt: Roads, Temples, Cathedrals, Snow mountains +Figure 5. Qualitative comparisons of using a fixed panoramic camera path. + +![](images/45c772288201d5bff930f379913711ea2333b468ec7b474a04e46f0af5cf5636.jpg) + +![](images/6b02b68420092b37ba89dee7e052b7cbf9a1e228fc6921ead926da327403023b.jpg) + +![](images/093919e769e950501993a2c932783ca61ffc062e737abc28499ca8995ad7226f.jpg) + +![](images/b11132e9d74a2e34d634f52b5747197a48da7a3eb4a6797404aff855500e071d.jpg) + +![](images/accbf9140566a46f4628042ee1bc6490554eb1e6261c5bb0d3cd8e3d063f08e4.jpg) + +![](images/385598723a9da2a2c662af73075957bf461c08c8fc53194cdd40ef84987bcaa9.jpg) + +![](images/ca059727f7ead3ea55bf5b65affa9e77a5d79c5e6faa9a4e25105e876aedee8b.jpg) +Prompt: Gardens, Trees, Offices, Cafe + +per-scene optimization, significantly constraining their efficiency. Notably, WonderTurbo excels in both geometry and appearance modeling, accelerating overall time by $15 \times$ . + +Quantitative Results. In Tab. 2, we compare WonderTurbo with various 3D scene generation methods [11, 22, 37, 69, 70, 76]. The experimental results show that online 3D scene generation methods outperform offline approaches by better + +meeting user textual requirements, achieving higher CLIP scores and improved CLIP consistency. WonderWorld [69] surpasses all other baseline methods, leading across all metrics. However, even with a $15 \times$ acceleration, WonderTurbo maintains competitive performance across all metrics compared to WonderWorld [69]. Additionally, since WonderTurbo is fine-tuned specifically for interactive 3D gener + +Table 1. Time comparison of scene generation, including geometry and appearance modeling, on an H20 GPU. We compare the time to generate scenes of the same size within the camera's view. + +
MethodGeometry (s)Appearance (s)Total (s)
OfflineLucidDreamer [11]35.388.3243.70
Text2Room [22]34.237.3241.55
Pano2Room [37]27.911.4729.38
DreamScene360 [76]44.291.4545.74
OnlineWonderJourney [70]78.121.4579.57
WonderWorld [69]6.624.4311.05
WonderTurbo0.500.220.72
+ +Table 2. Evaluation of novel view renderings for all methods + +
MethodCS↑CC↑CIQA↑Q-Align↑CA↑
OfflineLucidDreamer [11]27.720.92130.60233.54396.8231
Text2Room [22]24.500.90350.49102.67326.5324
Pano2Room [37]25.670.86520.35342.13425.0367
DreamScene360 [76]24.500.84354.69732.46206.9846
OnlineWonderJourney [70]27.630.96520.47533.52767.0134
WonderWorld [69]28.140.96540.67643.78237.2121
WonderTurbo28.650.97320.68123.72537.3243
+ +Table 3. Comparing the win rates of WonderTurbo in rendering novel views. + +
MethodWin RateMethodWin Rate
vs. LucidDreamer [11]96.32%vs. Text2Room [22]98.47%
vs. Pano2Room [37]94.26%vs. DreamScene360 [76]96.23%
vs. WonderJourney [70]96.54%vs. WonderWorld [69]69.43%
+ +ation tasks, improvements are observed in CLIP scores, CLIP consistency, CLIP-IQA+ and CLIP aesthetic. + +User Study. Additionally, we conduct a user study to evaluate the quality of 3D scenes generated by various methods. As shown in Tab. 3, the results indicate that WonderTurbo achieves comparable performance to WonderWorld [69] with a lower scene generation time cost and significantly outperforms all other methods in terms of user preference. + +Qualitative Results. As shown in Fig. 5, we present a qualitative comparison between WonderTurbo and several baseline methods using the same settings. Notably, WonderTurbo delivers competitive scene generation quality while significantly reducing generation time. In contrast, DreamScene360 [76] and Pamo2Room [37] struggle with noticeable geometric distortions and lack aesthetic appeal due to limited generalization capabilities. Meanwhile, LucidDreamer [11] and Text2Room [22] fail to place content correctly, with some prompt details not materializing. The results from WonderTurbo and WonderWorld [69] are closely matched, both demonstrating strong performance. + +Table 4. Ablation study results on different geometry models. + +
CS↑CC↑CIQA↑Q-Align↑CA↑
WonderTurbo w/ FreeSplat27.650.95420.64603.15436.6235
WonderTurbo w/ DepthSplat27.320.96750.66203.21456.7432
WonderTurbo w/ StepSplat28.650.97320.68123.72537.3243
+ +Table 5. Ablation study results on novel view renderings. + +
CS↑CC↑CIQA↑Q-Align↑CA↑
Ours w/o depth guided27.720.95320.63593.43617.1734
Ours w/o incremental fusion27.870.96540.64593.54317.2734
Ours w/o FastPaint27.820.96830.65743.71467.2136
WonderTurbo28.650.97320.68123.72537.3243
+ +# 4.3. Ablation Study + +Geometry Modeling. We compare different geometry modeling methods, including FreeSplat [55] and DepthSplat [59], all fine-tuned with the same settings. As shown in Tab. 4, FreeSplat [55] and DepthSplat [29] underperform compared to StepSplat, especially in Q-Align and CA scores, due to their reliance on unsupervised depth estimation. In contrast, StepSplat uses consistent depth maps to guide cost volume, enabling adaptive 3D scene generation. + +StepSplat. We conduct ablation experiments on StepSplat to illustrate the effect of the depth guided cost volume and the incremental infusion. As shown in Tab. 5, The depth-guided cost volume plays a key role in providing accurate depth information for geometry modeling, affecting image quality. Meanwhile, the incremental fusion also contributes to the overall performance, as it helps reduce redundant Gaussians and avoid issues like floating points. + +FastPaint. We compare FastPaint with the pretrained inpainting model [40]. As shown in Tab. 5, FastPaint enhances the capability of 3D appearance modeling, with improvements across various metrics. + +# 5. Discussion and Conclusion + +Despite progress in 3D scene generation from a single image, efficiency remains a challenge due to time-consuming geometry optimization and viewpoint refinement. To address this, we propose WonderTurbo, an efficient framework for real-time interactive 3D scene generation that accelerates geometry optimization and appearance modeling. For accelerating geometry modeling, we introduce StepSplat, which expands 3D scenes within 0.26 seconds while maintaining visual quality, and QuickDepth, which provides consistent depth priors for cost volume construction. For appearance modeling, FastPaint is proposed to achieve this with only 2 inference steps while ensuring spatial consistency. Experiments show that WonderTurbo accurately generates 3D scenes from text, outperforming CLIP-based metrics and user study win rates, with a $15 \times$ speedup. + +# References + +[1] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 3, 6 +[2] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. Advances in Neural Information Processing Systems, 35:25102-25116, 2022. 2 +[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 5 +[4] Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, et al. Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954, 2024. 3 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 5, 6 +[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19457-19467, 2024. 2, 3, 5 +[8] Junwen Chen, Heyang Jiang, Yanbin Wang, Keming Wu, Ji Li, Chao Zhang, Keiji Yanai, Dong Chen, and Yuhui Yuan. Prismlayers: Open data for high-quality multilayer transparent image generative models. arXiv preprint arXiv:2505.22523, 2025. 3 +[9] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2, 3, 5 +[10] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. Advances in Neural Information Processing Systems, 37:107064-107086, 2025. 3 +[11] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2, 5, 6, 7, 8 +[12] Robert T Collins. A space-sweep approach to true multi-image matching. In Proceedings CVPR IEEE computer so + +ciety conference on computer vision and pattern recognition, pages 358-363. IEEE, 1996. 4 +[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 5, 6 +[14] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13733-13742, 2021. 4 +[15] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 5 +[16] Dave Epstein, Ben Poole, Ben Mildenhall, Alexei A Efros, and Aleksander Holynski. Disentangled 3d scene generation with layout learning. arXiv preprint arXiv:2402.16936, 2024.2 +[17] Mengyang Feng, Jinlin Liu, Miaomiao Cui, and Xuansong Xie. Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models. arXiv preprint arXiv:2311.13141, 2023. 3 +[18] Ziyue Feng, Liang Yang, Longlong Jing, Haiyan Wang, YingLi Tian, and Bing Li. Disentangling object motion and occlusion for unsupervised multi-frame monocular depth. In European Conference on Computer Vision, pages 228-244. Springer, 2022. 4 +[19] Eric Ghysels, Arthur Sinko, and Rossen Valkanov. Midas regressions: Further results and new directions. *Econometric reviews*, 26(1):53–90, 2007. 5 +[20] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 6 +[21] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 3 +[22] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2, 5, 6, 7, 8 +[23] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1-11, 2024. 3 +[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3 +[25] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 3, 6 + +[26] Haoran Li, Haolin Shi, Wenli Zhang, Wenjun Wu, Yong Liao, Lin Wang, Lik-hang Lee, and Peng Yuan Zhou. Dreamscene: 3d gaussian-based text-to-3d scene generation via formation pattern sampling. In European Conference on Computer Vision, pages 214–230. Springer, 2024. 2 +[27] Hanwen Liang, Junli Cao, Vedit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos N. Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image, 2024. 2 +[28] Lina Liu, Xibin Song, Xiaoyang Lyu, Junwei Dao, Mengmeng Wang, Yong Liu, and Liangjun Zhang. Fcfr-net: Feature fusion based coarse-to-fine residual learning for depth completion. In Proceedings of the AAAI conference on artificial intelligence, pages 2136-2144, 2021. 5 +[29] Zhiheng Liu, Ka Leong Cheng, Qiuyu Wang, Shuzhe Wang, Hao Ouyang, Bin Tan, Kai Zhu, Yujun Shen, Qifeng Chen, and Ping Luo. Depthlab: From partial to complete. arXiv preprint arXiv:2412.18153, 2024. 5, 8 +[30] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. Infusion: Inpainting 3d gaussians via learning depth completion from diffusion prior. arXiv preprint arXiv:2404.11613, 2024.5 +[31] Taiming Lu, Tianmin Shu, Junfei Xiao, Luoxin Ye, Jiahao Wang, Cheng Peng, Chen Wei, Daniel Khashabi, Rama Chellappa, Alan Yuille, et al. Genex: Generating an explorable world. arXiv preprint arXiv:2412.09624, 2024. 2, 3 +[32] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14211-14219, 2024. 6 +[33] Chaojun Ni, Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Wenkang Qin, Guan Huang, Chen Liu, Yuyin Chen, Yida Wang, Xueyang Zhang, et al. Recondreamer: Crafting world models for driving scene reconstruction via online restoration. arXiv preprint arXiv:2411.19548, 2024. 2, 3 +[34] Simon Niklaus, Long Mai, Jimei Yang, and Feng Liu. 3dken burns effect from a single image. ACM Transactions on Graphics (ToG), 38(6):1-15, 2019. 2 +[35] Yuyang Peng, Shishi Xiao, Keming Wu, Qisheng Liao, Bohan Chen, Kevin Lin, Danqing Huang, Ji Li, and Yuhui Yuan. Bizgen: Advancing article-level visual text rendering for infographics generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 23615-23624, 2025. 3 +[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3 +[37] Guo Pu, Yiming Zhao, and Zhouhui Lian. Pano2room: Novel view synthesis from a single indoor panorama. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2, 3, 5, 6, 7, 8 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 6 +[39] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. Advances in Neural Information Processing Systems, 37:117340-117362, 2025. 6 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 6, 8 +[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5 +[42] Shuwei Shao, Zhongcai Pei, Weihai Chen, Xingming Wu, Zhong Liu, and Zhengguo Li. Smudlp: Self-teaching multiframe unsupervised endoscopic depth estimation with learnable patchmatch. arXiv preprint arXiv:2205.15034, 2022. 4 +[43] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. 2 +[44] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 6 +[45] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 3 +[46] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5, 6 +[47] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5, 6 +[48] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. 2 +[49] Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 302–317, 2018. 2 + +[50] Bingchao Wang, Zhiwei Ning, Jianyu Ding, Xuanang Gao, Yin Li, Dongsheng Jiang, Jie Yang, and Wei Liu. Fix-clip: Dual-branch hierarchical contrastive learning via synthetic captions for better understanding of long text. arXiv preprint arXiv:2507.10095, 2025. 2 +[51] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. In European Conference on Computer Vision, pages 153-168. Springer, 2024. 6 +[52] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, pages 2555-2563, 2023. 2, 6 +[53] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 3 +[54] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xu Chi, Yun Ye, Ziwei Chen, and Xingang Wang. Crafting monocular cues and velocity guidance for self-supervised multi-frame depth learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2689-2697, 2023. 3, 4 +[55] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. Advances in Neural Information Processing Systems, 37:107326-107349, 2025. 2, 3, 5, 8 +[56] Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1164-1174, 2021. 4 +[57] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7467-7477, 2020. 2 +[58] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 6 +[59] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2, 3, 8 +[60] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. In European Conference on Computer Vision, pages 156-173. Springer, 2024. 2 +[61] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 3 +[62] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, + +Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 3 +[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3, 5 +[64] Xiuyu Yang, Yunze Man, Junkun Chen, and Yu-Xiong Wang. Scenecraft: Layout-guided 3d scene generation. Advances in Neural Information Processing Systems, 37: 82060-82084, 2025. 2 +[65] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20331-20341, 2024. 2, 3 +[66] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 3 +[67] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. Advances in Neural Information Processing Systems, 37:1304-1332, 2025. 3 +[68] Hang Yu, Ruilin Li, Shaorong Xie, and Jiayan Qiu. Shadowenlightened image outpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7850-7860, 2024. 6 +[69] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 2, 3, 5, 6, 7, 8 +[70] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 2, 3, 5, 6, 7, 8 +[71] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 30(12):7749-7762, 2024. 2 +[72] Shaofeng Zhang, Jina Huang, Qiang Zhou, Zhibin Wang, Fan Wang, Jiebo Luo, and Junchi Yan. Continuous-multiple image outpainting in one-step via positional query and a diffusion-based approach. arXiv preprint arXiv:2401.15652, 2024. 6 +[73] Songchun Zhang, Yibo Zhang, Quan Zheng, Rui Ma, Wei Hua, Hujun Bao, Weiwei Xu, and Changqing Zou. 3d-scenedreamer: Text-driven 3d-consistent scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10170-10180, 2024. 6 +[74] Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, + +Wenjun Mei, and Xingang Wang. Drivedreamer4d: World models are effective data machines for 4d driving scene representation. arXiv preprint arXiv:2410.13571, 2024. 2 +[75] Taijin Zhao, Heqian Qiu, Yu Dai, Lanxiao Wang, Hefei Mei, Fanman Meng, Qingbo Wu, and Hongliang Li. Vlm-guided explicit-implicit complementary novel class semantic learning for few-shot object detection. Expert Systems with Applications, 256:124926, 2024. 3, 6 +[76] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2024. 3, 6, 7, 8 +[77] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2 +[78] Teng Zhou, Xiaoyu Zhang, and Yongchuan Tang. Panolama: Generating endless and coherent panoramas with next-token-prediction llms. arXiv preprint arXiv:2411.15867, 2024. 2 \ No newline at end of file diff --git a/wonderturbogeneratinginteractive3dworldin072seconds/images.zip b/wonderturbogeneratinginteractive3dworldin072seconds/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..65ce417e440f18619238d3548183c3336f4ee4bf --- /dev/null +++ b/wonderturbogeneratinginteractive3dworldin072seconds/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9eb510343b834cf4702903dd303b0b7b6382d6950b1ef2d05ce9849f2c07d44b +size 1006465 diff --git a/wonderturbogeneratinginteractive3dworldin072seconds/layout.json b/wonderturbogeneratinginteractive3dworldin072seconds/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..92d562a2a842144cf5b27f0fa8e09a51df392722 --- /dev/null +++ b/wonderturbogeneratinginteractive3dworldin072seconds/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1a70b2aed34a7542bf4e374fbdb991d85df5f917b25dead425f4c09d17ed2d7 +size 436003 diff --git a/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/bd010426-5cc6-465c-8059-3ed79f4fec42_content_list.json b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/bd010426-5cc6-465c-8059-3ed79f4fec42_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..748f6ef455e6e55c609f89a2ce01d18b843e5f27 --- /dev/null +++ b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/bd010426-5cc6-465c-8059-3ed79f4fec42_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:352c657b19aecba69f47fefbca4a59256f0b0f8c770b2edcbeaee407978e340d +size 78567 diff --git a/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/bd010426-5cc6-465c-8059-3ed79f4fec42_model.json b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/bd010426-5cc6-465c-8059-3ed79f4fec42_model.json new file mode 100644 index 0000000000000000000000000000000000000000..94e708ff40157d938fd20c1a131d18afa83ca022 --- /dev/null +++ b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/bd010426-5cc6-465c-8059-3ed79f4fec42_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20f58d187f9cfc4d950053cc762ea09053f2ec28cbad2a6fa1962691fe772481 +size 97974 diff --git a/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/bd010426-5cc6-465c-8059-3ed79f4fec42_origin.pdf b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/bd010426-5cc6-465c-8059-3ed79f4fec42_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1671f467d3b9d91ce5cd5b2704e91179547d6f90 --- /dev/null +++ b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/bd010426-5cc6-465c-8059-3ed79f4fec42_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a304c89428fcbb180ba3b0f399dbf965236f26e0d3b56251dc7954cd3711869c +size 1797982 diff --git a/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/full.md b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0d3f6e1b17c60b32d31166175ab34ea81c2b773a --- /dev/null +++ b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/full.md @@ -0,0 +1,325 @@ +# World4Drive: End-to-End Autonomous Driving via Intention-aware Physical Latent World Model + +Yupeng Zheng $^{1,2,3*}$ , Pengxuan Yang $^{1,3,4*}$ , Zebin Xing $^{1,4*}$ , Qichao Zhang $^{1,4\dagger}$ , Yuhang Zheng $^{5}$ , Yinfeng Gao $^{1}$ , Pengfei Li $^{6}$ , Teng Zhang $^{4}$ , Zhongpu Xia $^{1}$ , Peng Jia $^{3}$ , XianPeng Lang $^{3}$ , and Dongbin Zhao $^{1,4}$ + +$^{1}$ Institute of Automation, Chinese Academy of Sciences, $^{2}$ Peng Cheng Laboratory, $^{3}$ Li Auto, $^{4}$ School of Artificial Intelligence, UCAS, $^{5}$ National University of Singapore, $^{6}$ Tsinghua University + +# Abstract + +End-to-end autonomous driving directly generates planning trajectories from raw sensor data, yet it typically relies on costly perception supervision to extract scene information. A critical research challenge arises: constructing an informative driving world model to enable perception annotation-free, end-to-end planning via self-supervised learning. In this paper, we present World4Drive, an end-to-end autonomous driving framework that employs vision foundation models to build latent world models for generating and evaluating multi-modal planning trajectories. Specifically, World4Drive first extracts scene features, including driving intention and world latent representations enriched with spatial-semantic priors provided by vision foundation models. It then generates multi-modal planning trajectories based on current scene features and driving intentions and predicts multiple intention-driven future states within the latent space. Finally, it introduces a world model selector module to evaluate and select the best trajectory. We achieve perception annotation-free, end-to-end planning through self-supervised alignment between actual future observations and predicted observations reconstructed from the latent space. World4Drive achieves state-of-the-art performance without manual perception annotations on both the open-loop nuScenes and closed-loop NavSim benchmarks, demonstrating an $18.0\%$ relative reduction in L2 error, $46.7\%$ lower collision rate, and $3.75\times$ faster training convergence. Codes will be accessed at https://github.com/ucaszyp/World4Drive. + +# 1. Introduction + +End-to-end autonomous driving integrates perception and planning into a unified, fully differentiable network. Given the complexity of the physical world and the uncertainty in + +![](images/dfbe8aca62d42bb654d61834b1546f8b9160b7ecef8d06010866f75ccd79d10d.jpg) +Figure 1. Our proposed World4Drive demonstrates superior convergence efficiency and performance compared to PerAct on the nuScenes dataset. As shown in the figure, where the x-axis represents training epochs (same iterations) and the y-axis shows normalized performance (calculated as the ratio of our minimum L2 error to the L2 error at each epoch). World4Drive achieves equivalent performance in $3.75 \times$ fewer training epochs and ultimately delivers a $1.18 \times$ improvement in peak performance. + +planning intentions [3], modeling multi-modal motion planning based on a holistic understanding of physical scenes (i.e., understanding spatial, semantic, and temporal information) is a critical challenge in the field. + +To enhance scene understanding, existing end-to-end approaches have explored diverse scene representations, including BEV-centric [15, 42], vector-based [17, 45], and sparse-centric representation [36]. Some works [19, 30, 35] leverage multi-modal large language models to enhance scene comprehension capabilities. Additionally, methods like VADv2 [3] and Hydra-MDP [22] model driving intentions through probabilistic planning. However, these approaches typically require perception annotations such as 3D bounding boxes and HD maps, which limits their scalability. Recently, VaVAM [1] leverages the learned representations of an auto-regressive video model to generate the driving trajectory directly. LAW [21] proposes a latent + +world model that constructs uni-modal latent features from raw images and acquires scene representations through temporal self-supervised learning, reducing dependence on perception annotations. However, extracting single-modal latent features from images struggles to capture the spatial-semantic information of the physical world and multi-modal driving intentions, resulting in slow training convergence and suboptimal performance, as shown in Fig. 1. + +To address these critical issues, we introduce World4Drive, an end-to-end framework that integrates multi-modal driving intentions with a latent world model to enable rational planning. This is achieved by subconsciously simulating how the physical world evolves under different driving intentions, closely mirroring the decision-making process of human drivers. + +Given multi-view images and a trajectory vocabulary input, World4Drive extracts the driving intentions and world latent representations through its driving world encoding module. Specifically, the driving world encoding module incorporates two key components: a physical latent encoder and an intention encoder. The physical latent encoder consists of a context encoder that leverages spatial and semantic priors from a metric depth estimation model and a vision-language model and a temporal module that aggregates temporal information to construct world latent representations enriched with physical scene context. Concurrently, the intention encoder extracts multi-modal driving intention features from a predefined trajectory vocabulary, enabling comprehensive representation of possible driving behaviors. Subsequently, World4Drive predicts future latent representations across multi-modal driving intentions and proposes a world model selector to select the most plausible one for self-supervised alignment training with actual world latent representations extracted from future frames. During inference, we fully leverage World4Drive's latent world model to evaluate and rank multi-modal trajectory candidates, enabling robust decision-making that guides the autonomous vehicle's planning process in complex driving scenarios. + +World4Drive achieves state-of-the-art (SOTA) end-to-end planning performance without requiring perception annotations on the challenging nuScenes [2] and NavSim [6] benchmarks and is comparable to advanced perception-based models. Compared to LAW [21], a previous SOTA unsupervised method, our method significantly reduces average planning displacement error by $18.0\%$ (from 0.61m to 0.50m) and average collision rate by $46.7\%$ (from $0.30\%$ to $0.16\%$ ). More remarkably, our approach achieves more than three times faster convergence speed by appropriately incorporating spatial-semantic priors from vision foundation models. + +Our key contributions are summarized as follows: + +Inspired by human driver decision processes, we propose + +an intention-aware latent world model that innovatively uses the world model to generate and evaluate multimodal trajectories under different intentions. + +- To enhance the world model's understanding of the physical world without perception annotations, we design a novel driving world encoding module that leverages prior knowledge of vision foundation models to extract physical latent representations of the driving environment. +- Our method achieves SOTA planning performance without perception annotations on open-loop nuScenes and closed-loop NavSim benchmarks while significantly accelerating convergence speed. + +# 2. Related Works + +# 2.1. End-to-End Autonomous Driving + +In recent years, with advancements in perception technologies using BEV-centric [23, 44], vector-based [25], and sparse-centric [27] scene representations, vision-based end-to-end autonomous driving has garnered increasing attention. Models such as UniAD [15], VAD [17], and SparseDrive [36] have explored those diverse representations, establishing end-to-end architectures including perception, prediction, and planning. Methods like [8, 50] leverages wolrd models to produce or evaluate trajectories, while some methods [16, 42, 52] implement parallelized end-to-end structures. To consider the intention uncertainty in planning, VADv2 [3] and Hydra-MDP [22] model driving intentions through probabilistic planning. DiffusionDrive and GoalFlow [26, 43] explores end-to-end methodologies based on diffusion models [11]. Furthermore, with the evolution of large language models (LLMs), several approaches [18, 30, 35, 53] enhance scene information through language models. For instance, VLP [32] incorporates linguistic understanding into scene information via contrastive learning. DriveVLM [38] constructs a dual slow-fast system that integrates vision-language model capabilities into the decision space. TOKEN [37] utilizes large language models to enhance object-level perception, improving planning capabilities in long-tail scenarios. These methods typically require extensive and costly perception or QAs annotations, which limits their scalability. + +# 2.2. Autonomous Driving World Models + +World models in autonomous driving aim to predict scene evolution following various actions. These include image-based video generation, 3D world models based on representations such as point clouds and occupancy grids, and latent feature-based future world generation. Image-based video generation encompasses driving video approaches using diffusion models like DriveDreamer [40, 48], Vista [7], and Drive-WM [41], as well as driving video generation methods based on autoregressive models such as Drive- + +![](images/78d334076db6b86404952762eb2c12ad222301eec6cc1c2dd391fd52acd6c83e.jpg) +Figure 2. We propose World4Drive, a novel approach that constructs an intention-aware latent world model to generate, evaluate, and rank multi-modal trajectories under multi-modal driving intentions. + +World [31] and GAIA[12]. 3D world models include point cloud-based world models [47] and occupancy-based world models [9, 20, 49, 51]. These construct models in 3D space to better capture dynamic changes in 3D scenes. Recently, approaches like VaVAM [1] and LAW [21] have employed video generation techniques to learn scene representations through self-supervised learning, eliminating the dependency on perception annotations. Specifically, LAW proposes latent world models that predict a single future scene latent feature through self-supervision learning, achieves SOTA end-to-end planning performance. However, constructing single-modal latent features from raw images often struggles to capture the spatial-semantic scene information and the uncertainty of multi-modal driving intentions, resulting in suboptimal performance. + +# 3. Method + +# 3.1. Overview + +The overall pipeline of World4Drive is illustrated in Fig. 2. World4Drive comprises two key modules: 1) Driving World Encoding (Sec. 3.2), which extracts driving intention and physical world latent representations from RGB images and trajectory vocabulary, and 2) Intention-aware World Model (Sec. 3.3), which predicts the latent representation of the future world according to multi-modal driving intentions and scores multi-modal planning trajectories via the world model selector. The two key modules are tightly coupled, enabling autonomous driving vehicles to imagine the future world under various intentions while achieving vision-based end-to-end planning without requiring perception annotations. + +# 3.2. Driving World Encoding + +In the Driving World Encoding module, we introduce an intention encoder that takes vocabulary as input to extract driving intentions and a physical latent encoder that utilizes a vision-language model and a metric depth estimation model to extract world latent representations that are aware of spatial, semantic and temporal context. + +# 3.2.1. Intention Encoder + +Given a randomly initialized ego query $Q_{ego}$ and a trajectory vocabulary $\mathcal{V} \in \mathbb{R}^{N \times S \times 2}$ input [3], we first obtain the intention point $P_I \in \mathbb{R}^{3 \times K \times 2}$ by adopting the k-means clustering algorithm on the endpoints of $\mathcal{V}$ . Among them, $N$ represents the number of trajectories in the trajectory vocabulary, 3 represents the three commands type(e.g., left, right, straight), $K$ represents the number of intentions for each command type, and $S$ represents the number of waypoints in each trajectory. Then we obtain the intention query $Q_I$ with sinusoidal position encoding. Finally, we utilize a self-attention layer to obtain the intention-aware multimodal planning query $Q_{plan}$ . Formally, + +$$ +Q _ {\text {p l a n}} = \text {S e l f A t t e n t i o n} \left(\left(Q _ {e g o} + Q _ {I}\right)\right). \tag {1} +$$ + +By default, $N$ is set to 8192, and $K$ is set to 6. + +# 3.2.2. Physical World Latent Encoding + +We introduce the Physical World Latent Encoding module to extract the world latent representations with a holistic understanding (i.e., spatial and semantic perception capabilities) of the 3D physical world. The module consists of a context encoder (see Fig. 3 for details) to incorporate spatial and semantic prior and a temporal aggregation module to enhance temporal context. + +Context Encoder. Given a frame of multi-view images $I_{t} \in \mathbb{R}^{M \times H \times W \times 3}$ at timestep $t$ as input, we first extract + +![](images/823490775a2b9c75fdc6c5e1116a374194fb5af274014c08dbb7541b326582b4.jpg) +Figure 3. Detailed pipeline of context encoder. It consists of a 3D spatial encoding module and a semantic understanding module, achieving a holistic understanding of the physical world. + +the corresponding image features $F_{t}\in \mathbb{R}^{M\times h\times w\times D}$ with an image backbone, where $D$ is the feature dimension and $M$ represents the number of camera views. Previous work, LAW [21], directly extracts camera features as world latent representation, lacking spatial and semantic understanding of driving scenarios. To address this issue, we introduce spatial-semantic priors with open-vocabulary semantic supervision and 3D geometry-aware positional encoding. + +Semantic Understanding. We utilize the vision-language model Grouded-SAM [34] to produce pseudo semantic labels. Given the prompt for the object of interest, we obtain 2D bounding boxes and the corresponding semantic mask $S_{t} \in \mathbb{R}^{H \times W \times C}$ via the Grouded-SAM model. Formally, + +$$ +S _ {t} = \operatorname {G r o u n d e d S A M} \left(F _ {t}\right), \tag {2} +$$ + +we only keep labels with high confidence to reduce incorrect labeling. Finally, we leverage cross-entropy loss $\mathcal{L}_{sem}$ to enhance the semantic understanding of the latent representations. + +3D Spatial Encoding. 3D Spatial Encoding aims to provide the model with accurate positional information in the physical world. Previous work, PETR [28], utilizes a postprojection approach by generating 3D meshgrids to provide different 3D positional encodings for each pixel. Inspired by this concept, we provide the scale-aware depth for each pixel to represent 3D space, offering accurate spatial understanding for end-to-end planning. Specifically, we employ a metric depth model [13, 46] to estimate multi-view depth maps $D_{t} \in \mathbb{R}^{M \times h \times w}$ . In contrast to PETR, we adopt a forward projection approach, obtaining the 3D position in the ego coordinate system $p = \{x,y,z\}$ of each pixel $(u,v)$ through depth maps and the camera intrinsic matrix. Thus, we can generate 3D position maps $P_{t} \in \mathbb{R}^{M \times h \times w \times 3}$ . Subsequently, we encode these 3D positions using sinu + +soidal positional encoding and obtain corresponding positional embedding $E_{t} \in \mathbb{R}^{M \times h \times w \times D}$ through a learnable MLP. Formally, + +$$ +E _ {t} = \operatorname {M L P} \left(\operatorname {S P E} \left(P _ {t}\right)\right), \tag {3} +$$ + +where $\mathrm{SPE}(\cdot)$ is the sinusoidal position encoding. Finally, by adding positional embedding $E_{t}$ to the image features $F_{t}$ , we obtain the semantic-spatial-aware visual feature $\hat{F}_t$ . + +Temporal Aggregation. Different from previous work [21] that uses randomly initialized queries to obtain latent representation. We employ a temporal aggregation module to obtain latent representation enriched with the temporal context. In particular, we preserve the visual feature $\hat{F}_{t - 1}$ at the prior timestamp $t - 1$ . To enhance the temporal information in the world latent representation, we aggregate historical information into the current visual features through cross-attention mechanisms to obtain world latent representations $L_{t}\in \mathbb{R}^{M\times h\times w\times D}$ . Formally, + +$$ +L _ {t} = \text {C r o s s A t t e n t i o n} \left(\hat {F} _ {t}, \hat {F} _ {t - 1}\right). \tag {4} +$$ + +The proposed Physical World Latent Encoder enriches the world latent representations with spatial, semantic, and temporal information, providing a holistic understanding of the dynamic driving environment, which is crucial for imagining the future world. + +# 3.3. Planning with Intention-aware World Model + +In this section, we propose the intention-aware world model to predict the latent representation of the future world according to multi-modal driving intentions (Sec. 3.3.1) and score multi-modal planning trajectories via the world model selector (Sec. 3.3.2). + +# 3.3.1. Intention-aware World Model Dreamer + +Action Encoding. Given the intention-aware multi-modal planning query $Q_{plan}$ , we first employ a cross-attention layer to aggregate scene context into $Q_{plan}$ . Then, we obtain the multi-modal trajectories $T = \{T^{1},\dots,T^{K}\} \in \mathbb{R}^{K\times S\times 2}$ with an MLP layer. Formally, + +$$ +T = \operatorname {M L P} \left(\operatorname {C r o s s A t t e n t i o n} \left(Q _ {\text {p l a n}}, L _ {t}\right)\right). \tag {5} +$$ + +Finally, we employ an action encoder (an MLP layer) to acquire intention-aware action tokens $A \in \mathbb{R}^{K \times D}$ , where $K$ is the number of intentions. + +Intention-aware World Model Prediction. Our objective is to predict future world latents $L_{t + n} = \{L_{t + n}^{1},\dots,L_{t + n}^{K}\}$ following actions corresponding to each driving intention, where $n$ is the timestamp interval. We concatenate action tokens $A$ and world latent $L$ along the dimension channel. Different from the previous work, we randomly initialize a learnable query $Q_{future}$ and adopt multi-layer cross-attention as the predictor. Formally, + +$$ +L _ {t + n} = \text {C r o s s A t t e n t i o n} \left(Q _ {\text {f u t u r e}}, \operatorname {C o n c a t} (A, L)\right). \tag {6} +$$ + +![](images/6a78badf25abfaffa76838cf09157c3f12953fd0b268c14f0daa2345d51355bd.jpg) +Figure 4. Detailed pipeline of World Model Selector. By computing and comparing feature distances between predicted latents and actual latent, the model selects the most plausible latent and its corresponding trajectory as output. The selector employs reconstruction loss between the selected latent and the actual latent to self-supervised learn scene representations while simultaneously training the world model scoring network using focal loss between predicted scores and the index of the selected latent. + +By default, we set $n = 3$ , the ablation study of timestamp interval $n$ can be seen in the supplementary material. + +# 3.3.2. World Model Selector + +We propose a World Model Selector module that evaluates trajectories under $K$ different intentions through the latent world model and selects the reasonable trajectory from among them. The detailed architecture is illustrated in Fig. 4. In particular, given the predicted intention-aware future latent $L_{t + n}$ and the actual future latent $\hat{L_{t + n}}$ , we compute the feature distance between the predicted latent representation and the actual latent representation for each modality. The modality with the minimum distance is selected as the final selected modality (assume the index of this modality is $j$ ), the corresponding latent distance is used as the reconstruction loss $\mathcal{L}_{\text{recon}}$ for optimization, and the corresponding trajectory $T^{j}$ is selected as the final planning trajectory. Simultaneously, we employ a classification network as the ScoreNet, $\mathcal{C}$ , to predict scores $\mathbb{S} = \{\mathbb{S}^1,\dots \mathbb{S}^K\}$ corresponding to the $K$ modalities. Formally, + +$$ +\mathbb {S} = \operatorname {S o f t m a x} \left(\mathcal {C} \left(L _ {t + n}\right)\right) \tag {7} +$$ + +We utilize a focal loss between scores $\mathbb{S}$ and the selected modality index $j$ to optimize the world model scoring network. + +Notably, 1) during inference, we directly select the trajectory corresponding to the world model with the highest score as the final trajectory, and 2) we employ the MSE loss to compute the latent distance. We do ablation on other losses in the supplementary material. + +# 3.4. Training Loss + +Following previous works, we apply the $L_{1}$ loss $\mathcal{L}_{\text{traj}}$ to guide the final planning trajectory $T^{j}$ with the expert trajectory $\hat{T}$ . World4Drive is end-to-end trainable. Therefore, + +the final loss for end-to-end training is + +$$ +\mathcal {L} = \alpha \mathcal {L} _ {\text {s e m}} + \beta \mathcal {L} _ {\text {r e c o n}} + \gamma \mathcal {L} _ {\text {s c o r e}} + \eta \mathcal {L} _ {\text {t r a j}}, \tag {8} +$$ + +we set $\alpha = 0.2, \beta = 0.2, \gamma = 0.5, \eta = 1.0$ by default. + +# 4. Experiments + +# 4.1. Benchmarks + +Open-loop nuScenes Benchmark. The nuScenes [2] benchmark is an open-loop evaluation framework developed for the real-world nuScenes dataset. The nuScenes dataset comprises 1000 driving videos captured across diverse environments. In line with previous methodologies [15, 17], we employ displacement error (L2) and collision rate (CR) as evaluation metrics for the predicted trajectories, which are sampled at $2\mathrm{Hz}$ over a 3-second horizon. + +Closed-loop NavSim Benchmark. The NavSim [6] benchmark is built upon the OpenScene [5] dataset, encompassing 1192 training scenarios and 136 test scenarios, with a total of over 100,000 keyframes. In alignment with the officially provided baseline, we interpolate the predicted trajectories (sampled at $2\mathrm{Hz}$ over a 4-second horizon) using an LQR controller. The model performance is evaluated through closed-loop PDM scores (PDMS), which are calculated based on five key factors: no at-fault collision (NC), drivable area compliance (DAC), time-to-collision (TTC), comfort (Comf.), and ego progress (EP). + +# 4.2. Implementation Details + +nuScenes Benchmark. Following the VAD-Tiny [17] configuration, we employ ResNet-50 [10] as the image backbone, processing 6 surround-view images with a resolution of $360 \times 640$ . The driving commands align with previous works [15], comprising three types: left, right, and straight. For each command, we predict 6 planning trajectories and select the one with the highest score corresponding to the world model as the final planning trajectory. We train the model for 12 epochs on 8 NVIDIA 3090 GPUs with a total batch size of 8 and an initial learning rate of 5e-5. + +NavSim Benchmark. Consistent with the NavSim benchmark, our closed-loop model takes a concatenated image formed by stitching front, front-left, and front-right camera views as input, which is then resized to $256 \times 1024$ . We employ ResNet-34 [10] to extract image features. We train the model for 60 epochs on 8 NVIDIA 3090 GPUs with a total batch size of 64. Since LAW [21] is not open-sourced, we reimplement and evaluate it under settings identical to ours. For vision foundation models, we utilize the giant model from Metric3D v2 [13] for depth estimation and Grounded-SAM [34] for semantic segmentation. + +# 4.3. Main Results + +As demonstrated in Tab. 1, we compare our proposed framework with several SOTA methods. Methods highlighted + +Table 1. End-to-end planning results on nuScenes benchmark [2] + +
MethodL2 (m) ↓Collision Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
ST-P3 [14]1.332.112.902.110.230.621.270.71
OccNet [39]1.292.132.992.130.210.591.370.72
UniAD [15]0.480.961.651.030.050.170.710.31
VAD [17]0.410.701.050.720.070.180.430.23
PPAD [4]0.310.560.870.580.080.120.380.19
GenAD [50]0.280.490.780.520.080.140.340.19
\( LAW^* \)[21] (Perception-based)0.240.460.760.490.080.100.390.19
BEV-Planner [24]0.300.520.830.550.100.371.300.59
\( LAW^* \)[21] (Perception-free)0.260.571.010.610.140.210.540.30
World4Drive (Ours)0.230.470.810.500.020.120.330.16
+ +* LAW [21] adopt Swin-Tiny [29] as image backbone while other methods adopt ResNet-50 [10] as image backbone. + +Table 2. End-to-end planning results on NavSim benchmark [2] + +
MethodInputNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
UniAD [15]C97.891.992.9100.078.883.4
PARA-Drive [36]C97.992.493.099.879.384.0
LTF [33]C97.492.892.4100.079.083.8
Transfuser [33]C & L97.792.892.8100.079.284.0
VADv2 [3]C & L97.289.191.6100.076.080.9
Hydra-MDP [22]C & L97.991.792.9100.077.683.0
DiffusionDrive [26]C & L98.296.294.7100.082.288.1
Ego-MLPE93.077.383.6100.062.865.6
LAW (Perception-free) [21]C97.293.391.9100.078.883.8
World4Drive (Ours)C97.494.392.8100.079.985.1
+ +In the Input column, C represents camera modality, L represents lidar modality, and E represents ego status. + +with blue background in the table require manual perception annotations, whereas methods highlighted with red background do not require manual perception annotations for training and inference. World4Drive achieves SOTA performance among perception annotation-free approaches, demonstrating an $18.0\%$ reduction in L2 error and a $46.7\%$ reduction in collision rate compared to strong baselines. Furthermore, World4Drive achieves the lowest collision rate among all methods. Compared to LAW, the SOTA perception-based approach, our method shows only a modest increase of less than $2\%$ in L2 error while significantly improving safety metrics. + +As demonstrated in Tab. 2, World4Drive also achieves competitive performance in closed-loop metric PDMS. Compared to the baseline, our approach demonstrates significant improvements in Time-to-Collision (TTC) and Drivable Area Compliance (DAC) metrics. These metrics specifically evaluate an autonomous vehicle's spatial awareness and understanding of the drivable area. The result indicates that incorporating the vision foundation model priors substantially enhances the model's comprehensive under + +standing of the physical world. Moreover, our closed-loop metrics surpass those of other methods requiring perception annotations, with the exception of DiffusionDrive [26]. + +# 4.4. Ablation Study + +In this section, we conduct several ablation studies to explore the effectiveness, robustness, and stability of our proposed World4Drive. All of the ablation experiments are conducted on the nuScenes [2] benchmark. All L2 errors and collision rates are averaged over a 3-second prediction horizon. + +# 4.4.1. Effectiveness of Each Component + +In this section, we assess the effectiveness of each component in our method. Detailed results are demonstrated in Tab. 3. Row 1 demonstrates the result of our baseline, LAW, which only has a single-modal world model. Comparing row 1 and row 2, we observe that incorporating vehicle intention significantly reduces L2 error and collision rate. Further, the comparison between row 1 and row 4 demonstrates substantial planning performance improve- + +Table 3. Ablation study of each proposed component + +
IDPhysical Latent EncoderIntention-aware WML2Collision
DepthSemanticWMIntentions
1✓0.610.30
2✓✓0.550.25
3✓✓✓0.510.29
4✓✓✓0.490.26
5✓✓✓0.610.36
6✓✓✓✓0.500.16
+ +Table 4. Performance under different driving conditions. + +
IDConditionWorld4Drive (Ours)LAW [21]
L2 (m) ↓Collision (%) ↓L2 (m) ↓Collision (%) ↓
1All0.500.160.610.30
2Day0.470.160.560.26
3Night0.760.080.670.22
4Sunny0.500.180.580.29
5Rainy0.490.050.540.16
+ +ments when integrating priors from vision foundation models and vision language models, highlighting the critical importance of comprehensive physical world understanding. + +To explore the contribution of different perceptual components, we conducted a more detailed analysis. Comparing rows 2 and 3, we find that introducing spatial priors enhances positional awareness, thereby improving trajectory-fitting capabilities. Similarly, a comparison between rows 3 and 6 reveals that semantic priors significantly reduce collision rates, suggesting better obstacle understanding. + +Finally, we investigate the necessity of combining intentions with world modeling. The comparison between rows 4 and 6 demonstrates that adding intention modeling substantially improves planning quality, which is intuitive as intentions provide multiple planning possibilities, enabling the model to select safer trajectories. However, the comparison between rows 5 and 6 reveals that intentions alone, without world modeling, actually lead to degraded planning performance. This confirms the crucial role of the world model in evaluating and ranking multi-modal intentions. + +# 4.4.2. Performance In Different Driving Conditions + +In this section, we analyze planning performance across diverse driving conditions, including varying weather conditions, illumination settings, and driving maneuvers. Following the official nuScenes scene descriptions, we categorize weather as sunny or rainy, illumination as day or night, and driving maneuvers as left, straight driving, or right. + +Tab. 4 presents a comparative analysis of our method against the baseline, LAW [21], across different weather and lighting conditions. Our approach consistently out + +performs LAW across almost all environmental scenarios. Notably, compared to LAW, in challenging nighttime and rainy conditions, our method reduces collision rates by $63.7\%$ and $68.8\%$ , respectively. This significant improvement can be attributed to the integration of priors from vision foundation models, which enables our system to comprehend higher-dimensional spatial and semantic information about the physical environment. Consequently, our approach demonstrates greater robustness to photometric inconsistencies inherent in nighttime and rainy weather conditions, which typically impede the temporal self-supervised training of latent world models in baseline methods. + +Tab. 5 demonstrates the comparison of planning performance under different driving maneuvers between our method and LAW. Compared to LAW, our method generates significantly safer planning trajectories across a variety of driving maneuvers. + +The superior planning performance under diverse driving conditions demonstrates the effectiveness and robustness of our approach. + +Table 5. Performance under driving maneuvers. + +
ModelL2 (m) ↓Collision (%) ↓
LeftRightStraightAllLeftRightStraightAll
LAW0.670.710.580.610.540.230.290.30
World4Drive0.630.690.480.500.400.200.140.16
+ +# 4.4.3. Scability of World4Drive + +To explore the scalability of our approach, we conduct experiments by varying both the size of the hidden dimension $D$ and the image backbone. As shown in Tab. 6, comparing rows 1, 4, and 5, we scale the image backbone from ResNet34 to ResNet50 and ResNet101, while comparing rows 3, 4, and 5, we scale the size of the hidden dimension from 125 to 256 and 384. The ablation results demonstrate that World4Drive exhibits excellent scalability for both image backbone and hidden dimension, which is different from previous methods [16]. In previous methods, increasing hidden dimensions typically yields greater benefits than scaling up the image backbone. Our method shows comparable improvements with both scaling strategies. It is natural because the extracted latent representations are directly utilized for planning tasks, and both caling strategies can effectively incorporate additional scene information that directly benefits vehicle planning. + +# 4.5. Qualitative Results + +In this section, we present the visualization of World4Drive on nuScenes benchmark. The qualitative result is shown in Fig. 5. The upper portion of the visualization demonstrates that World4Drive does safer planning during turning maneuvers compared to LAW. The lower portion shows that + +![](images/1dbdd3099e0d0134a5fcfaae7ef170cb0ff5fec7c081c8068279324dbea21d4c.jpg) + +![](images/6004e734dff1b7ab5e59967865334eea7046308031a1eebbaa12bdfb139cb00c.jpg) + +![](images/da7bcff5c3522d3e5192174353788247445ed210b45e78de6c4f5c361659747a.jpg) +Figure 5. Visualization of World4Drive. We render the ground truth annotations as the perception results. + +Table 6. Ablation study of scalability + +
IDBackboneDimensionL2 (m) ↓Collision (%) ↓
1ResNet-342560.520.25
2ResNet-501280.550.27
3ResNet-502560.500.16
4ResNet-503840.490.10
5ResNet-1012560.470.14
+ +the world model selector effectively selects the most reasonable trajectory from multi-modal planning intentions across diverse scenarios. Additional visualizations and failure case analyses are provided in the supplementary material. + +# 5. Conclusion + +In this paper, we present World4Drive, an intention-aware physical latent world model. World4Drive proposes a novel framework that incorporates driving intentions with a la + +tent world model, innovatively leveraging a latent world model to generate, evaluate, and select multi-modal trajectories. Specifically, World4Drive proposes a physical world latent encoding module, incorporating the spatial and semantic priors from vision foundation models and aggregating temporal information. Extensive experiments demonstrate World4Drive's profound and comprehensive understanding of the physical world, as well as the effectiveness of tightly coupling driving intentions with the latent world model. + +# 6. Acknowledgement + +This work is supported by the National Key Research and Development Program of China under Grant 2022YFA1004000, in part by Beijing Natural Science Foundation under Grant L253007 and 4242052, and the National Natural Science Foundation of China (NSFC) under Grant 62173325. + +# References + +[1] Florent Bartoccioni, Elias Ramzi, Victor Besnier, Shashanka Venkataramanan, Tuan-Hung Vu, Yihong Xu, Loick Chambon, Spyros Gidaris, Serkan Odabas, David Hurych, et al. Vavim and vavam: Autonomous driving through video generative modeling. arXiv preprint arXiv:2502.15672, 2025.1, 3 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giangcarlo Baldan, and Oscar Beijbom. Nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 5, 6 +[3] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 3, 6 +[4] Zhili Chen, Maosheng Ye, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Ppad: Iterative interactions of prediction and planning for end-to-end autonomous driving. In European Conference on Computer Vision, 2024. 6 +[5] OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving. https://github.com/OpenDriveLab/OpenScene, 2023.5 +[6] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, et al. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. Advances in Neural Information Processing Systems, 37:28706-28719, 2024. 2, 5 +[7] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. In Advances in Neural Information Processing Systems, 2024. 2 +[8] Yinfeng Gao, Qichao Zhang, Da-Wei Ding, and Dongbin Zhao. Dream to drive with predictive individual world model. IEEE Transactions on Intelligent Vehicles, 9(12): 8224-8238, 2024. 2 +[9] Songen Gu, Wei Yin, Bu Jin, Xiaoyang Guo, Junming Wang, Haodong Li, Qian Zhang, and Xiaoxiao Long. Dome: Taming diffusion model into high-fidelity controllable occupancy world model. arXiv preprint arXiv:2410.10429, 2024. 3 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 5, 6 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2 +[12] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 3 + +[13] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10579 - 10596, 2024. 4, 5 +[14] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, 2022. 6 +[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2023. 1, 2, 5, 6 +[16] Xiaosong Jia, Junqi You, Zhiyuan Zhang, and Junchi Yan. *Drivetransformer: Unified transformer for scalable end-to-end autonomous driving.* In *The Thirteenth International Conference on Learning Representations*, 2025. 2, 7 +[17] Bo Jiang, Shaoyu Chen, Qing Xu, Benchcheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 2, 5, 6 +[18] Bu Jin, Xinyu Liu, Yupeng Zheng, Pengfei Li, Hao Zhao, Tong Zhang, Yuhang Zheng, Guyue Zhou, and Jingjing Liu. Adapt: Action-aware driving caption transformer. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7554-7561. IEEE, 2023. 2 +[19] Bu Jin, Yupeng Zheng, Pengfei Li, Weize Li, Yuhang Zheng, Sujie Hu, Xinyu Liu, Jinwei Zhu, Zhijie Yan, Haiyang Sun, et al. Tod3cap: Towards 3d dense captioning in outdoor scenes. In European Conference on Computer Vision. Springer, 2024. 1 +[20] Xiang Li, Pengfei Li, Yupeng Zheng, Wei Sun, Yan Wang, and Yilun Chen. Semi-supervised vision-centric 3d occupancy world model for autonomous driving. arXiv preprint arXiv:2502.07309, 2025. 3 +[21] Yingyan Li, Lue Fan, et al. Enhancing end-to-end autonomous driving with latent world model. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 3, 4, 5, 6, 7 +[22] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziyue Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multitarget hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 1, 2, 6 +[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: learning bird's-eye-view representation from lidar-camera via spatiotemporal transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):2020-2036, 2024. 2 +[24] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6 +[25] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. In International Conference on Learning Representations, 2023. 2 +[26] Benchcheng Liao, Shaoyu Chen, et al. Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2, 6 +[27] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 2 +[28] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, 2022. 4 +[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6 +[30] Ana-Maria Marcu, Long Chen, Jan Hünermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, et al. Lingoqa: Visual question answering for autonomous driving. In European Conference on Computer Vision. Springer, 2024. 1, 2 +[31] Chen Min, Dawei Zhao, Liang Xiao, Jian Zhao, Xinli Xu, Zheng Zhu, Lei Jin, Jianshu Li, Yulan Guo, Junliang Xing, et al. Driveworld: 4d pre-trained scene understanding via world models for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[32] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[33] Aditya Prakash, Kashyap Chitta, et al. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6 +[34] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 4, 5 +[35] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beibwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In European Conference on Computer Vision. Springer, 2024. 1, 2 +[36] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Haoran Wu, and Sifa Zheng. Sparsedrive: End-to-end + +autonomous driving via sparse scene representation. In IEEE International Conference on Robotics and Automation, 2025. 1, 2, 6 +[37] Ran Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. arXiv preprint arXiv:2407.00959, 2024. 2 +[38] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. In The Conference on Robot Learning, 2024. 2 +[39] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 6 +[40] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jia-gang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In European Conference on Computer Vision. Springer, 2024. 2 +[41] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[42] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2 +[43] Zebin Xing, Xingyu Zhang, Yang Hu, Bo Jiang, Tong He, Qian Zhang, Xiaoxiao Long, and Wei Yin. Goalflow: Goaldriven flow matching for multimodal trajectories generation in end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 2 +[44] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[45] Pengxuan Yang, Yupeng Zheng, Qichao Zhang, Kefei Zhu, Zebin Xing, Qiao Lin, Yun-Fu Liu, Zhiguo Su, and Dongbin Zhao. Uncad: Towards safe end-to-end autonomous driving via online map uncertainty. In IEEE International Conference on Robotics and Automation, 2025. 1 +[46] Wei Yin, Chi Zhang, et al. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 4 +[47] Lunjun Zhang, Yuwen Xiong, Ze Yang, Sergio Casas, Rui Hu, and Raquel Urtasun. Copilot4d: Learning unsupervised world models for autonomous driving via discrete diffusion. In The Twelfth International Conference on Learning Representations, 2024. 3 + +[48] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: LIm-enhanced world models for diverse driving video generation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2 +[49] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In European Conference on Computer Vision, 2024. 3 +[50] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In European Conference on Computer Vision. Springer, 2024. 2, 6 +[51] Yupeng Zheng, Xiang Li, Pengfei Li, Yuhang Zheng, Bu Jin, Chengliang Zhong, Xiaoxiao Long, Hao Zhao, and Qichao Zhang. Monoocc: Digging into monocular semantic occupancy prediction. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 18398-18405. IEEE, 2024. 3 +[52] Yupeng Zheng, Zhongpu Xia, Qichao Zhang, Teng Zhang, Ben Lu, Xiaochuang Huo, Chao Han, Yixian Li, Mengjie Yu, Bu Jin, et al. Preliminary investigation into data scaling laws for imitation learning-based end-to-end autonomous driving. arXiv preprint arXiv:2412.02689, 2024. 2 +[53] Yupeng Zheng, Zebin Xing, Qichao Zhang, Bu Jin, Pengfei Li, Yuhang Zheng, Zhongpu Xia, Kun Zhan, Xianpeng Lang, Yaran Chen, et al. Planagent: A multi-modal large language agent for closed-loop vehicle motion planning. arXiv preprint arXiv:2406.01587, 2024. 2 \ No newline at end of file diff --git a/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/images.zip b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..35b7b9aac13e06abf0b85a1923f2ddff1b45f43a --- /dev/null +++ b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f0c7576fd5f18557f6ddf05d981ee508c366806d29794bdd931fae0f9dfa627 +size 626077 diff --git a/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/layout.json b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e147f602baa63aa0d8833f40489a30bed7674eef --- /dev/null +++ b/world4driveendtoendautonomousdrivingviaintentionawarephysicallatentworldmodel/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:891f6de32f14001971047e775dd1d6a9ce5f67c5ec4765d37ef1cf9b80f9d06b +size 390978 diff --git a/worldscoreaunifiedevaluationbenchmarkforworldgeneration/2e63d834-5044-4966-887d-e7d4112f0b43_content_list.json b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/2e63d834-5044-4966-887d-e7d4112f0b43_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4cb5a8fbc18eb7b4a17a89c310b024c0cc6f61dd --- /dev/null +++ b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/2e63d834-5044-4966-887d-e7d4112f0b43_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:896e6757f36db7660f251c2079785c3967084959d6cd548b0bff7a37d0110ba2 +size 87832 diff --git a/worldscoreaunifiedevaluationbenchmarkforworldgeneration/2e63d834-5044-4966-887d-e7d4112f0b43_model.json b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/2e63d834-5044-4966-887d-e7d4112f0b43_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c2591ddc6c3c5ba06d5f1edd4aa615499d917711 --- /dev/null +++ b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/2e63d834-5044-4966-887d-e7d4112f0b43_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3048d317ff132fbca339a8b17189741958b229a441b7adb49b42e93f727cea6 +size 114820 diff --git a/worldscoreaunifiedevaluationbenchmarkforworldgeneration/2e63d834-5044-4966-887d-e7d4112f0b43_origin.pdf b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/2e63d834-5044-4966-887d-e7d4112f0b43_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..20a03b3a84812b5bcd1e5e1b21cc7c0d906747a0 --- /dev/null +++ b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/2e63d834-5044-4966-887d-e7d4112f0b43_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:017b7509f9256c43cf3682e31ad257b92ede0944f91be5006904da2f02dc5a12 +size 5205674 diff --git a/worldscoreaunifiedevaluationbenchmarkforworldgeneration/full.md b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8729bb1a17a966ca7550ee6bc8a69615d873f7c9 --- /dev/null +++ b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/full.md @@ -0,0 +1,345 @@ +# WorldScore: A Unified Evaluation Benchmark for World Generation + +Haoyi Duan* Hong-Xing Yu* Sirui Chen Li Fei-Fei Jiajun Wu Stanford University + +# Abstract + +We introduce the WorldScore benchmark, the first unified benchmark for world generation. We decompose world generation into a sequence of next-scene generation tasks with explicit camera trajectory-based layout specifications, enabling unified evaluation of diverse approaches from 3D and 4D scene generation to video generation models. The WorldScore benchmark encompasses a curated dataset of 3,000 test examples that span diverse worlds: static and dynamic, indoor and outdoor, photorealistic and stylized. The WorldScore metric evaluates generated worlds through three key aspects: controllability, quality, and dynamics. Through extensive evaluation of 20 representative models, including both open-source and closed-source ones, we reveal key insights and challenges for each category of models. Our dataset, evaluation code, and leaderboard can be found at https://haoyi-duan.github.io/WorldScore/. + +# 1. Introduction + +Recent advances in visual generation have sparked growing interest in world generation—the creation of large-scale, diverse worlds with various scenes, which finds wide applications in entertainment, education, simulation, and embodied AI. The rapid progress in video generation [1, 6, 10, 88], 3D scene generation [11, 16, 90, 91], and 4D scene generation [3, 85, 89] has shown generating high-quality individual scenes, demonstrating the potential of these models as world generation systems. However, as the concept of world generation expands, users demand to generate more comprehensive worlds that seamlessly integrate multiple varied scenes with detailed spatial layout controls rather than disconnected individual environments. + +Achieving this vision requires a unified evaluation benchmark that systematically assesses different types of world generation models across large-scale, diverse worlds, which is currently absent. Existing benchmarks mainly focus on video generation [15, 45, 46, 48, 92] and evaluate only indi + +![](images/24d7d3dea01533aa5d18c85aa3ff5ec442bedbb02a8edf773b98273cb41f6257.jpg) +Figure 1. While existing video benchmarks like VBenchmark [26] rate Models A and B similarly based on single-scene video quality, our WorldScore benchmark differentiates their world generation capabilities by identifying that Model B fails to generate a new scene or follow the instructed camera movement. In https://haoyi-duan.github.io/WorldScore/, we show the videos to explain our WorldScore metrics. + +vidual scene generation. For example, VBench [26] primarily evaluates text-to-video (T2V) tasks using curated prompts without explicit spatial layout control, restricting their evaluations to single scenes (Figure 1). Moreover, despite the promising potential of 3D and 4D scene generation methods for world generation, current benchmarks lack essential components such as camera specifications and reference images, making them incompatible with many state-of-the-art 3D/4D scene generation methods that require an image or a camera trajectory as inputs [11, 16, 39, 90, 91]. + +We introduce WorldScore, a unified benchmark for world generation. Our key design is to decompose world generation into a sequence of next-scene generation tasks, where each step is characterized by a triplet of (current scene, next scene, layout). For unified evaluation across different methods, we provide both an image and a text prompt for a current scene, as well as both camera matrices and a textual description for a layout specifi + +
Benchmark# ExamplesMulti-SceneUnifiedLong Seq.Image Cond.Multi-StyleCamera Ctrl.3D Consist.
TC-Bench [15]150XXX✓XXX
EvalCrafter [45]700XXXXXXX
FETV [46]619XXXXXXX
VBench [26]800XXXXXXX
T2V-CompBench [71]700XXXXXXX
Meng et al. [48]160XXXXXXX
Wang et al. [78]423XX✓XXXX
ChronoMagic-Bench [92]1649XXXXXXX
WorldModelBench [40]350XXX✓XXX
WorldScore (Ours)3000✓✓✓✓✓✓✓
+ +Table 1. Comparison of Benchmarks. Our WorldScore benchmark is designed to evaluate various world generation approaches including 3D, 4D, I2V and T2V models. It is designed to generate multiple scenes with varying sequence lengths. Our benchmark also features multiple visual styles, accurate camera control evaluation, and 3D consistency evaluation, all of which are important factors in world generation yet currently missing in existing benchmarks. + +cation. This design allows our WorldScore benchmark to evaluate various approaches including 3D, 4D, text-to-video, and image-to-video models on large-scale world generation. All methods are evaluated on a common output format, i.e., rendered or generated videos, to enable direct comparison of generation across different types of approaches. + +Our evaluation metric, WorldScore, is computed by aggregating three key aspects: controllability, which measures the adherence of the generated worlds w.r.t. control inputs; quality, which measures the fidelity and consistency; dynamics, which measures how much the generated worlds exhibit accurate and stable motions. Each of these aspects comprises a few distinct metrics, leading to a total of 10 metrics that contribute to computing the WorldScore. + +To enable a comprehensive assessment, we curate a diverse dataset covering both static and dynamic world generation scenarios across different visual domains. For static worlds, we include 5 categories of indoor scenes and 5 categories of outdoor scenes with varying sequence lengths. For dynamic worlds, we include 5 distinct types of dynamics such as rigid motion and fluid motion. Additionally, each example in our dataset has a corresponding stylized counterpart sampled from a rich set of candidate styles, allowing the evaluation of various visual domains. In total, our dataset comprises 3000 high-quality test examples that span indoor/outdoor environments and photorealistic/stylized visual domains. + +We conduct extensive experiments by evaluating 20 diverse models, including 6 image-to-video models (with 2 leading closed-source models), 7 text-to-video models, 6 3D scene generation models, and a 4D generation model. In summary, our contributions are fourfold: + +- We propose the first world generation benchmark, WorldScore, which allows unified evaluation across various approaches including 3D, 4D, I2V, and T2V models. +- We curate a high-quality, diverse dataset for our benchmark evaluation. Our dataset covers diverse static and + +dynamic scenes across various categories with multiple visual styles. + +- We introduce the WorldScore metrics, which aggregate critical aspects in world generation model performances, including controllability, quality, and dynamics. +- Through the comprehensive evaluation of 18 open-source and 2 closed-source models, we reveal key insights and challenges in current world generation approaches, providing valuable guidance for future research. + +# 2. Related Work + +Video generation benchmarks. The progress of both open-source [1, 10, 84, 88] and closed-source [2, 6, 20, 58] video generation models has stimulated the proposal of numerous benchmarks [15, 26, 45, 46, 48, 92]. However, most existing benchmarks, such as VBench [26] and WorldModelBench [40], focus on evaluating video generation models based on single-scene video quality without layout control and multi-scene generation. Furthermore, their designs are incompatible with 3D/4D scene generation methods that require camera specification. In contrast, our WorldScore benchmark is designed to focus on evaluating world generation approaches with multi-scene generation tasks, and it is designed to accommodate 3D, 4D, I2V and T2V models. We show a detailed comparison in Table 1. + +Video generation models. Recent advances in image generation, including VAEs [36], GANs [5, 18, 30-33, 49], VQ approaches [13, 73], and Diffusion models [23, 52, 68, 70], have fueled explorations in video generation [25, 47, 65, 76, 77]. The advent of Sora [6] has further demonstrated the potential of video models as world generation models [29, 48, 83]. While most recent models focus on text-to-video (T2V) generation [9, 10, 14, 41], developments in image-to-video (I2V) [1, 84, 86, 88, 97] have also been significant. In our WorldScore benchmark, we evaluate both T2V and I2V models as world generation approaches, thanks + +![](images/cfe78322cb62963ce3cdf5ae1c26443de8f112be7693ce16a19f418af2ec1c52.jpg) + +![](images/b5c5798b81a33576c1630a2981ca935e9fd155c1d6d6252dac4aed9a8988cda6.jpg) + +![](images/b682e63a8587cb754178e0aac10f3db58c127ad8ab5659ae581f2f92b494c9cc.jpg) +Figure 2. Overview of the WorldScore benchmark design. Top left: World generation is decomposed into a sequence of next-scene generation tasks, where each step follows a structured world specification defining both spatial layout and semantic content. Bottom left: The unified world specification is used to instruct different types of models, including video generation and 3D/4D generation models. Bottom right: All models output videos for evaluation. Top right: Output videos are evaluated using the WorldScore metrics, which assess three fundamental aspects including controllability, quality, and dynamics. + +to our unified design that accommodates both image and text conditioning strategies. + +3D scene generation. Besides video models, our WorldScore benchmark also includes 3D and 4D generation methods. Recent 3D scene generation models rely mainly on generative diffusion models [16, 90], which formulate generating scenes in a sequential manner using supervision from 2D image outpainting models. These methods [11, 12, 24, 91] project the synthesized 2D scene extensions into a 3D representation by leveraging depth estimation models [4, 34, 37, 87]. + +To incorporate dynamics, 4D generation methods [39, 43, 56, 66, 95, 96, 96] further integrate multi-view and video diffusion priors. Due to the difficulty of scene-level generation, most of existing methods focus on object-level generation. Nevertheless, we include 4D-fy [3] in our benchmark due to its open-source accessibility. + +# 3. The WorldScore Benchmark + +Design overview. Our goal is to establish an evaluation benchmark for world generation that unifies different methodological approaches. Our WorldScore benchmark introduces three key components: (1) a standardized world specification, (2) a carefully curated dataset, and (3) multifaceted metrics. We show an overview in Figure 2. We decompose world generation into a sequence of next-scene generation tasks, where each step is defined by a world specification encompassing both spatial layout and semantic content (top left of Figure 2). This world specification enables us to instruct different types of models ranging from 3D/4D scene generation to video generation approaches. The + +generated outputs, standardized as videos (bottom right of Figure 2), are then evaluated using the WorldScore metrics (top right of Figure 2) that assess three critical aspects: controllability, quality, and dynamics. This unified evaluation approach ensures fair comparison across different methodological paradigms. + +# 3.1. World Specification + +**Formulation.** We decompose the world generation task into a sequence of next-scene generation tasks, where each step is specified by a triplet of $(\mathcal{C}, \mathcal{N}, \mathcal{L})$ , where $\mathcal{C} = \{\mathbf{I}, \mathcal{P}\}$ denotes the current scene given by a scene image $\mathbf{I}$ and a text prompt $\mathcal{P}$ , $\mathcal{N}$ denotes the next-scene text prompt, and $\mathcal{L} = \{\mathcal{T}, \mathcal{V}\}$ denotes the layout given by a camera trajectory $\mathcal{T} = (\mathbf{C}_1, \mathbf{C}_2, \dots, \mathbf{C}_N)$ where $\mathbf{C}_i$ denotes a camera matrix and a text prompt of camera movement $\mathcal{V}$ . Then, a world generation model is instructed to generate a video: + +$$ +\mathbf {V} = g _ {\text {w o r l d}} \left(w _ {\text {p r o c}} (\mathcal {C}, \mathcal {N}, \mathcal {L})\right), \tag {1} +$$ + +where $\mathbf{V}$ denotes a video, $g_{\mathrm{world}}$ denotes the world generation model, and $w_{\mathrm{proc}}$ denotes a model-specific pre-processing which we detail in Supp. A. + +Static and dynamic worlds. We explicitly disentangle the evaluation of dynamics aspect from the controllability and quality aspects due to their distinct natures. To this end, we have two types of tasks: + +Static world generation: We instruct a model to generate varying-length scene sequences for controllability and quality assessment. Here, the next-scene text prompt $\mathcal{N}$ describes the new scene contents, and the layout $\mathcal{L}$ describes large camera movements. + +![](images/cfb7022b204834861300852d47468109877dbf41a36e80ebcf055c25d0f1f276.jpg) + +![](images/120e3a3e70740f6146204ab5e1400c1c7700c3678b1a40dd48bd4e940b46cb63.jpg) + +![](images/f5d45f95da9c00436a3345658b861cc036780148c3754bb9d823244cfd919b0c.jpg) + +![](images/f7497b54c8419c4474e2e5b54c0ce3b45f73f96298f3974e027dc7dabe5dc902.jpg) + +![](images/acf5cdd550e2e022a4b9c154fbd372265359e58951762fecb65cbcd1b1fc618c.jpg) + +![](images/4f5d7bb29d2d19a7a43d07e8882e46cdf73cb7f4ac8c7c91cbbcb148c603f453.jpg) + +![](images/a4a2daec86ab759690afd81eb6ae1141c9ed025c257855f324065fd378362222.jpg) + +![](images/4a08925f253d8c240519450a1fd9196b06cc6d4b95da3c6261c37f09504152de.jpg) + +![](images/7e5ccc9fc9b7979fbc5d44d766f1ed89a983cb880306c5f5ecbdb9f6b5ce1243.jpg) + +![](images/3bb4ee55a2bcffb9a9fcb8829df8204d81241e480bba1aae518bf565e669414d.jpg) + +![](images/4e860816b1686f0afa39baaa83b30786f5df78d430e73fb3edfb8609ed2d88b9.jpg) +Figure 3. Showcasing of the current scene images. Top two rows: Static world generation examples are categorized into indoor (first row) and outdoor (second row) scenes, each containing 5 categories. Bottom row: Dynamic world generation examples are divided into 5 motion types. Each dynamic example comes with an annotation of motion mask that indicates where the motion should happen. + +![](images/b40506ce114fc20afdab640f289fb649d484c6052a7c50cdcd32a073d1251ea4.jpg) + +![](images/0ea19fd5dba828ef2cf6ccb58f097d490aedf2745f50565a119c1abc4316e987.jpg) + +![](images/0be4925c910f3b2827e0f10e4811173adb4b154cb6f840d4d5fe15578a71ef4c.jpg) + +![](images/a1b6ea66795afee83751f0af04f3b2d0d2db7be946356b67157a65e549b8fd5d.jpg) + +![](images/0e1b2d45e9ac382cc3c3f640ecf676f99fccdf79957d67c8367a8e42c5a73e60.jpg) +Figure 4. Curation on the current scene $\mathcal{C}$ . Top: Photorealistic worlds. Bottom: Stylized counterparts. + +Dynamic world generation: We instruct a model to generate in-scene motion for dynamics assessment. Here, the next-scene text prompt $\mathcal{N}$ describes the same scene content as $\mathcal{C}$ but with dynamics changes, e.g., an animal moving. The layout $\mathcal{L}$ explicitly specifies a fixed camera position without any camera motion. + +# 3.2. Dataset Curation + +Our dataset consists of 3000 examples (world specifications), including 2000 for static world generation and 1000 for dynamic world generation. We show a detailed statistics in Table S4 in the supplementary material. + +Curation on current scene $\mathcal{C}$ . The current scene $\mathcal{C} = \{\mathbf{I},\mathcal{P}\}$ is given by an image $\mathbf{I}$ and its text prompt $\mathcal{P}$ . We show an illustration of our curation process in Figure 4. + +For static world generation, we define 10 categories of scenes including 5 indoor and 5 outdoor scene types. Then, we source images from open-source scene datasets [8, 38, 42, 57, 62, 67, 69, 74, 98] and supplement with an online source, Unsplash [7]. We apply a very rigorous filtering strategy to ensure high quality and high diversity (Supp. B.1), leading to approximately 5000 images $\mathbf{I}$ in photorealistic style (they are either real photos or physically-based rendered images). Then, we query a Vision-Language Model (VLM), GPT-4o [51], to generate captions $\mathcal{P}$ for these images and do a 10-way classification to put each of them into a category. + +![](images/34efd36a488ac090903292d58b9ecee9a41187800b74ae696a9b3245702b9ff3.jpg) +Figure 5. Curation on layouts $\mathcal{L}$ . Left: Camera paths $\mathcal{T}$ and text $\mathcal{Y}$ . Right: A move-right example. + +Finally, we further filter each category by keeping the first 100 highest-quality images, leading to 1000 images $\mathbf{I}$ and their corresponding prompts $\mathcal{P}$ . + +Then, we create a stylized counterpart for each example in the photorealistic domain. For each example, we randomly pick a style from a set of 7 style candidates, and create a new text prompt $\mathcal{P}$ by adding the style text to the prompt of the photorealistic example (Supp. B.2). Then, we leverage a commercial style-controlled text-to-image generation model [55] to generate the stylized counterpart image I. We show examples in the top two rows in Figure 3. + +For dynamic world generation, we define 5 categories of motion types and source Unsplash to manually curate 100 images for each of the category. We follow a similar process as in the static world generation examples to create text prompts and stylized counterpart, eventually leading to a total of 1000 examples. We show examples in the bottom row in Figure 3. + +Curation on next-scene text prompts $\mathcal{N}$ . Each world generation consists of a sequence of next-scene generation tasks. The next-scene text prompt $\mathcal{N}$ can have varying lengths. In particular, we consider two cases: (1) a small world where $\mathcal{N}$ consists of only one new scene, and (2) a large world where $\mathcal{N}$ consists of three new scenes. + +To generate coherent and contextually relevant scene sequences, we adopt an auto-regressive scene description generation process [90], that is, we instruct an LLM to generate the next-scene text prompt that should be different from all current scene text prompts. For example, for a small world, + +$$ +\mathcal {N} = \operatorname {L L M} (\mathcal {J}, \mathcal {P}), \tag {2} +$$ + +where the LLM takes two inputs: (1) the task specification $\mathcal{J} =$ "Generate a scene description different from the past scenes." $^1$ , and (2) a collection of past and current scene descriptions. For a large world which consists of 4 scenes, we repeat this process for 3 times, so that $\mathcal{N} = \mathcal{N}_1 + \mathcal{N}_2 +$ $\mathcal{N}_3$ consists of three individual next-scene prompts. In our + +generation, $20\%$ of our static world generation examples are large worlds, and the others are small worlds. + +Curation on layouts $\mathcal{L}$ . A layout $\mathcal{L} = \{\mathcal{T},\mathcal{V}\}$ is given by a camera trajectory $\mathcal{T} = (\mathbf{C}_1,\mathbf{C}_2,\dots ,\mathbf{C}_N)$ and a text prompt of camera movement $\mathcal{V}$ . We curate a set of 8 camera movements (left of Figure 5) which are widely used in movie industry. This design achieves two objectives: Firstly, it covers all spatial directions; secondly, it facilitates text-to-video models to take the instruction $\mathcal{V}$ as most of them are trained on movie clips that often contain these camera movement descriptions. These movements include both intra-scene movements, such as moving into a scene, as well as inter-scene transitions, such as pulling out the camera. For each static scene generation example, we randomly assign a layout $\mathcal{L}$ to a next-scene generation task. We show an example in the right of Figure 5. When the assigned layout is intra-scene, we perform a replacement of $\mathcal{N}$ with $\mathcal{P}$ . + +We leave details of our dataset curation in Supp. B. + +# 3.3. The WorldScore Metrics + +Our WorldScore metrics include two overall scores: WorldScore-Static which measures only the static world generation capability, and WorldScore-Dynamic which measures dynamic world generation capability in addition to static worlds. They are defined as the aggregation of several individual metrics in the three key aspects: controllability, quality, and dynamics. We briefly introduce each individual metric in the following, and we leave details in Supp. C. + +Controllability. We have three metrics. + +Camera controllability: To evaluate how the models adhere to the instructed layout $\mathcal{L} = \{\mathcal{T},\mathcal{Y}\}$ , we compute camera errors as follows: + +$$ +e _ {\text {c a m e r a}} = \sqrt {e _ {\theta} \cdot e _ {t}}, \tag {3} +$$ + +where $e_{\theta}$ and $e_t$ are scale-invariant rotation and translation errors with respect to the ground truth trajectory $\mathcal{T}$ , respectively. We compute camera errors across all the frames of the generated video $\mathbf{V}$ . We leave more details in Supp. C.1. Object controllability: We evaluate whether the objects specified in the next-scene prompt $\mathcal{N}$ appear in the generated next scene. To this end, we measure the success rate of object detection. Specifically, we leverage a state-of-the-art open-set object detection model [44]. We extract one or two individual object descriptions from the text prompt $\mathcal{N}$ . We compute the success rate by matching the detected objects with the object descriptions. This provides a quantitative measure of how well the generated foreground objects adheres to the world specification. + +Content alignment: Besides the objects (which typically occupies approximately only $\frac{1}{4}$ of the text prompt length), we also assess whether the generated scenes are aligned with the entire text $\mathcal{N}$ using CLIPScore [22]. + +![](images/3e6f9eb47e7e5c57a2015a2132a5972c7fe3edc8a21e8c6303da4bccb4c0acf6.jpg) +Figure 6. Typical examples. Top: 3D consistency. The bad example on the right-hand-side has a sudden change in geometry rather than smooth transition. Middle: Photometric consistency. The bad example exhibits severe texture shift in the mountain grassland. Bottom: Motion accuracy. In the good example, the octopus moves while the jellyfish remains static. For bad example on the right, the jellyfish moves while the octopus remains static. A full version of all metrics is in Figure S3 and Figure S4 in supplementary material. In https://haoyi-duan.github.io/WorldScore/, we show videos to explain our WorldScore metrics. + +# Quality. We have four metrics. + +3D consistency: We evaluate the 3D consistency in the static world videos. This metric focuses on how the geometry of a scene remains stable across frames, regardless of slight changes in visual textures. To this end, we use DROID-SLAM [72], a standard SLAM method, to estimate dense pixel-wise depth for each frame, and then we compute the reprojection error between a pair of co-visible pixels in consecutive frames. Since DROID-SLAM is designed to be robust against appearance changes, this metric measures geometric inconsistency. We show an example in Figure 6, and we leave more details in Supp. C.2. + +Photometric consistency: While 3D consistency exclusively focuses on geometry, photometric consistency focuses on appearance (e.g., textures). Many video generation models struggle with maintaining consistent object textures, leading to appearance inconsistency issues such as texture flickering. Existing consistency metrics, such as those with CLIP or DINO features [26, 27], focus on categorical identity but fail to capture fine-grained texture changes. For example, the mountain in the middle row of Figure 6 remains a mountain (i.e., the same geometry and semantic class) across frames, but the texture (grass) has been shifted and distorted over time. This cannot be captured by CLIP/DINO features. + +To detect photometric artifacts, our photometric consistency metric estimates the optical flow between consecutive frames and computes the Average End-Point Error (AEPE). This metric effectively identifies unstable visual appearance, + +as shown in Figure 6. We leave more details in Supp. C.3. + +Style consistency: We evaluate the style consistency by computing the differences (F-norm) between the Gram matrices [17] of the first frame and the last frame of a single next-scene generation task. + +Subjective quality: We use automatic metrics to evaluate the human perceptual quality of the generated scenes. There exists some automatic image assessment metrics [82] and aesthetic metrics [75], and thus we consider ensemble them. To find a combination that best fits human perception, we perform a human study of 400 participants, enumerate different metric combinations, and we pick the combination (CLIP-IQA+ [75] with CLIP Aesthetic [63]) that best matches human preference. We leave more details in Supp. C.4. + +# Dynamics. We have three metrics. + +Motion accuracy: Accurate motion placement is essential in dynamics generation. For example, if a prompt specifies that a car should move while nearby pedestrians remain still, the model should animate the car, not the pedestrians. To quantify this, we introduce motion accuracy, which measures whether the motion specified in the next-scene prompt $\mathcal{N}$ occurs in the designated regions. As shown in the bottom row of Figure 6, the score is calculated by comparing optical flow within the intended region with the flow outside the region. We need to consider the outside flow as it cancels out the global motion caused by unintended camera movements. + +Motion magnitude: We measure a world generation model's + +
ModelsWorldScoreControllabilityQualityDynamics
-Static-DynamicCameraCtrlObjectCtrlContentAlign3DConsistPhotoConsistStyleConsistSubjectiveQualMotionAccMotionMagMotionSmooth
Gen-3 [58]60.7157.5829.4762.9250.4968.3187.0962.8263.8554.5327.4868.87
Hailuo [20]57.5556.3622.3969.5673.5367.1862.8254.9152.4463.4627.2070.07
DynamiCrafter [84]52.0947.1925.1547.3625.0072.9060.9578.8554.4041.1139.2526.92
VideoCrafter1-T2V [9]47.1043.5421.6150.4460.7864.8651.3638.0542.6311.7675.0018.87
VideoCrafter1-I2V [9]50.4747.6425.4624.2535.2774.4273.8965.1754.8555.6325.0042.49
VideoCrafter2 [9]52.5747.4928.9239.0772.4665.1461.8543.7956.7447.1230.4029.39
T2V-Turbo [41]45.6540.2027.8030.6869.1438.7234.8449.6568.7434.8740.097.48
EasyAnimate [86]52.8551.6526.7254.5050.7667.2947.3573.0550.3175.0031.1640.32
Allegro [97]55.3151.9724.8457.4751.4870.5069.8965.6047.4154.3940.2837.81
Vchitect-2.0 [14]42.2838.4726.5549.5465.7541.5342.3025.6944.5833.5933.8121.31
LTX-Video [19]55.4456.5425.0653.4139.7378.4188.9253.5049.0876.2229.9571.09
CogVideoX-T2V [88]54.1848.7940.2251.0568.1268.8164.2042.1944.6725.0047.3136.28
CogVideoX-I2V [88]62.1559.1238.2740.0736.7386.2188.1283.2262.4469.5626.4260.15
SceneScape [16]50.7335.5184.9947.4428.6476.5462.8821.8532.750.000.000.00
Text2Room [24]62.1043.4794.0138.9350.7988.7188.3637.2336.690.000.000.00
LucidDreamer [11]70.4049.2888.9341.1875.0090.3790.2048.1058.990.000.000.00
WonderJourney [90]63.7544.6384.6037.1035.5480.6079.0362.8266.560.000.000.00
InvisibleStitch [12]61.1242.7893.2036.5129.5388.5189.1932.3758.500.000.000.00
WonderWorld [91]72.6950.8892.9851.7671.2586.8785.5670.5749.810.000.000.00
4D-fy [3]27.9832.1069.9255.090.8535.471.5932.040.8922.2222.8880.06
+ +Table 2. WorldScore evaluation of 20 world generation models. Top: Close-source video models. Middle: Open-source video models. Bottom two rows: 3D and 4D models. Abbreviations: Ctrl=Controllability, Align=Alignment, Consist=Consistency, Photo=Photometric, Qual=Quality, Acc=Accuracy, Mag=Magnitude, Smooth=Smoothness. + +ability to create large motions by estimating the optical flow between the consecutive frames of the generated video. + +Motion smoothness: Temporal jittering is a common failure mode in dynamic world generation. We utilize a standard video frame interpolation model [93] to generate smooth interpolation as ground truth to evaluate the temporal smoothness of generated videos V. We leave details in Supp. C.7. + +Score normalization and aggregation. After computing individual evaluation metrics, we apply a linear normalization and mapping process based on empirical bounds (Supp. C.8) to ensure that the final scores fall within the range between zero to one, and then we scale it by 100. Then, we compute the arithmetic mean of the dimension scores within control and quality aspects to obtain our WorldScore-Static. Additionally, we further incorporate three dynamics dimension scores into the aggregation, resulting in WorldScore-Dynamic. For 3D scene generation models that do not support dynamic tasks, we assign 0 to each dynamics metric. + +# 4. Results + +Validation. We validate our metrics by human study. Our results suggest that WorldScore's metrics align with human preference, and WorldScore is robust to different video resolutions and aspect ratios. We leave details in Supp. D. + +Models. We evaluate 20 available world generation models on our WorldScore benchmark. We assess 13 video generation models, including two leading commercial closed-source I2V models—Gen-3 [58] and Hailuo [20], along with 7 well-known open-source I2V models: DynamiCrafter [84], VideoCrafter1-I2V [9], VideoCrafter2 [10], EasyAnimate [86], CogVideoX-I2V [88], LTX-Video [19] and Allegro [97], and 4 open-source T2V models: VideoCrafter1-T2V, T2v-Turbo [41], Vchitect-2.0 [14], and CogVideoXT2V. Additionally, we evaluate six well-known 3D scene generation models: SceneScape [16], Text2Room [24], LucidDreamer [11], WonderJourney [90], InvisibleStitch [12], and WonderWorld [91]. Moreover, we include an open-source 4D generation model, 4D-fy [3]. We leave details of these models in Table S1 in supplementary material. + +# 4.1. Observations and Challenges + +We show the WorldScore benchmark results in Table 2. We identify key challenges in world generation: + +3D models excel in static world generation. From the WorldScore-Static results, we observe that 3D scene generation models generally perform better, e.g., WonderWorld [91] (72.69) and LucidDreamer [11] (70.40) are the top-2, much better than the best video model CogVideoX-I2V [88] (62.15). This is because 3D models inherently have + +![](images/db7953578c4718a2f751da023956b1035dfc1f7ebf61338207e713d5bd53b9dc.jpg) +Figure 7. WorldScore-Static across different subdomains. + +high camera controllability and, thus, better content alignment due to the larger space they can create, as well as high 3D and photometric consistency. However, they do not allow for the generation of dynamic worlds. When extended to 4D for dynamics, 4D-fy [3] does not perform well, likely due to the intrinsic difficulty in 4D scene generation. + +Video models lack camera controllability. Even CogVideoX-T2V [88], the best video generation model in camera controllability (40.22), scored much lower than any 3D/4D generation model. This is the main challenge for video generation models to achieve good static world generation. Recent work in injecting camera conditioning [21, 81] might be a promising solution. + +The best open-source video models are as good as closed-source video models. Comparing CogVideoX-I2V [88], with Gen-3 and Hailuo [20], we observe that CogVideoX-I2V scored even higher than both closed-source models in both WorldScore-Static (62.15) and WorldScore-Dynamic + +(59.12). However, CogVideoX-I2V is not better than them in every aspect. For instance, we observe that CogVideoX-I2V is better at camera controllability yet worse at object controllability and content alignment. + +Trade-offs in motion smoothness and magnitude. Comparing motion smoothness and motion magnitude metrics for each method, we observe that larger motion often comes at the cost of lower smoothness, revealing current challenge for video models in maintaining both significant motion and natural transitions. + +Larger motion does not necessarily mean more accurate motion placement. The correlation between the motion magnitude and accuracy is weak. This implies that models that can produce large motion do not guarantee correct motion placement to follow instructions. Instead, they could hallucinate unintended camera motion or irrelevant motion. More robust motion modeling may be needed to balance the three dynamics metrics. + +Video models are weak in long sequence generation and in outdoor scenes. We further evaluate model performance across different subdomains, and we show WorldScore-Static results in Figure 7. We observe that video generation models struggle significantly with long-sequence (large world generation) tasks. In addition, video models are significantly weaker than 3D models in outdoor scenes, while the gap is smaller in indoor scenes. + +T2V models are easier to steer than I2V models. Compare T2V models to I2V models, e.g., CogVideoX-T2V and CogVideoX-I2V, we observe that T2V models generally have higher scores in the controllability aspect and larger motion magnitude, while I2V models have higher scores in quality aspect. Through empirical examination, we find that this is because T2V models are willing to generate larger camera motion, while I2V models tend to stick to the input image viewpoint. This reveals a challenging in controlling I2V models to generate new scene contents. We leave further visualizations in Supp. E. + +# 5. Conclusion + +The WorldScore benchmark reveals current limitations in world generation approaches. For 3D models, while they excel in static world generation, extending them to 4D representations and incorporating dynamics remains challenging. For video models, the main challenges include controllability, long-sequence generation, and generating outdoor scenes. These insights point to directions for future research: bridging the gap between 3D and 4D representations, developing more robust controllability mechanisms, and designing architectures capable of handling extended scene sequences. We believe the WorldScore benchmark will serve as a valuable tool for measuring progress toward more capable and versatile world generation systems. + +Acknowledgments. This work is in part supported by ONR YIP N00014-24-1-2117, ONR MURI N00014-22-1-2740, NSF RI #2211258 and #2338203, and the Okawa Foundation. We thank Mohamed El Banani and Christoph Lassner for their helpful discussion. + +# References + +[1] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 1, 2 +[2] Luma AI. Luma dream machine: New freedoms of imagination, 2025. https://lumalabs.ai/dream-machine, Accessed: 2025-02-24. 2 +[3] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996–8006, 2024. 1, 3, 7, 8, S2 +[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3 +[5] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2 +[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 2 +[7] Supported by SQUARESPACE. Unsplash, 2013. https://unsplash.com, Accessed: 2025-02-23. 4, S1 +[8] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 4, S1, S4 +[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation, 2023. 2, 7, S2 +[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models, 2024. 1, 2, 7, S2 +[11] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 1, 3, 7, S2 +[12] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 3, 7, S2 + +[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2 +[14] Weichen Fan, Chenyang Si, Junhao Song, Zhenyu Yang, Yinan He, Long Zhuo, Ziqi Huang, Ziyue Dong, Jingwen He, Dongwei Pan, et al. Vchitect-2.0: Parallel transformer for scaling up video diffusion models. arXiv preprint arXiv:2501.08453, 2025. 2, 7, S1 +[15] Weixi Feng, Jiachen Li, Michael Saxon, Tsu-jui Fu, Wenhu Chen, and William Yang Wang. Tc-bench: Benchmarking temporal compositionality in text-to-video and image-to-video generation. arXiv preprint arXiv:2406.08656, 2024. 1, 2 +[16] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7, S2 +[17] Leon A Gatys. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. 6 +[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[19] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024.7, S2 +[20] HailuoAI. Hailuo, 2024. https://hailuoai(video/, Accessed: 2025-02-24. 2, 7, 8, S1, S2 +[21] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 8 +[22] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[24] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 3, 7, S2 +[25] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[26] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 1, 2, 6 + +[27] Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, et al. Vbench++: Comprehensive and versatile benchmark suite for video generative models. arXiv preprint arXiv:2411.13503, 2024. 6 +[28] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17307-17316, 2023. S1 +[29] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 2 +[30] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2 +[31] Tero Karras. A style-based generator architecture for generative adversarial networks. arXiv preprint arXiv:1812.04948, 2019. +[32] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. +[33] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in neural information processing systems, 34:852-863, 2021. 2 +[34] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 3 +[35] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. S5 +[36] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[37] Katrin Lasinger, René Ranftl, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. arXiv preprint arXiv:1907.01341, 2019. 3 +[38] Hoang-An Le, Thomas Mensink, Partha Das, Sezer Karaoglu, and Theo Gevers. Eden: Multimodal synthetic dataset of enclosed garden scenes. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1579-1589, 2021. 4, S1, S4 +[39] Yao-Chih Lee, Yi-Ting Chen, Andrew Wang, Ting-Hsuan Liao, Brandon Y Feng, and Jia-Bin Huang. Vividdream: Generating 3d scene with ambient dynamics. arXiv preprint arXiv:2405.20334, 2024. 1, 3 +[40] Dacheng Li, Yunhao Fang, Yukang Chen, Shuo Yang, Shiyi Cao, Justin Wong, Michael Luo, Xiaolong Wang, Hongxu + +Yin, Joseph E Gonzalez, et al. Worldmodelbench: Judging video generation models as world models. arXiv preprint arXiv:2502.20694, 2025. 2 +[41] Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Yang Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. arXiv preprint arXiv:2405.18750, 2024. 2, 7, S1, S2 +[42] Mengtian Li, Yu-Xiong Wang, and Deva Ramanan. Towards streaming perception. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 473-488. Springer, 2020. 4, S4 +[43] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576-8588, 2024. 3 +[44] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2025. 5 +[45] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 1, 2 +[46] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[47] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. arXiv preprint arXiv:2303.08320, 2023. 2 +[48] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024. 1, 2 +[49] Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2 +[50] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. S6 +[51] OpenAI. Hello gpt-4o, 2024. https://openai.com/index/hello-gpt-4o/, Accessed: 2025-02-23. 4, S1 +[52] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2 + +[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. S1 +[54] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. S5 +[55] Recraft. Recraft image generation and editing api, 2025. https://www.recraft.ai/docs, Accessed: 2025-02-25.5,S1 +[56] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3 +[57] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 4, S1, S4 +[58] Runway. Introducing gen-3 alpha: A new frontier for video generation, 2024. https://runwayml.com/research/introducing-gen-3-alpha, Accessed: 2025-02-24. 2,7, S1, S2 +[59] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. S1 +[60] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. S4 +[61] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. S4 +[62] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 4, S1, S4 +[63] Christoph Schuhmann. Clip+ mlp aesthetic score predictor. Clip+ mlp aesthetic score predictor, 2022. 6, S1, S5, S7 +[64] SDXL. 106 styles for stable diffusion xl model, 2023. S1 +[65] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[66] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3 + +[67] Ivan Skorokhodov, Grigoriei Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14144-14153, 2021. 4, S1, S4 +[68] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[69] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 4, S1, S4 +[70] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[71] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. arXiv preprint arXiv:2407.14505, 2024. 2 +[72] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 6, S4 +[73] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2 +[74] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 4, S1, S4 +[75] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 6, S1, S5, S7 +[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2 +[77] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, pages 1-20, 2024. 2 +[78] Yiping Wang, Xuehai He, Kuan Wang, Luyao Ma, Jianwei Yang, Shuohang Wang, Simon Shaolei Du, and Yelong Shen. Is your world simulator a good story presenter? a consecutive events-based benchmark for future long video generation. arXiv preprint arXiv:2412.16211, 2024. 2 +[79] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In European Conference on Computer Vision, pages 36-54. Springer, 2024. S5 +[80] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to + +structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. S6 +[81] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 8 +[82] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 6, S5 +[83] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 2 +[84] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. 2023. 2, 7, S2 +[85] Dejia Xu, Hanwen Liang, Neel P Bhatt, Hezhen Hu, Hanxue Liang, Konstantinos N Plataniotis, and Zhangyang Wang. Comp4d: Lm-guided compositional 4d scene generation. arXiv preprint arXiv:2403.16993, 2024. 1 +[86] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture. arXiv preprint arXiv:2405.18991, 2024. 2, 7, S1, S2 +[87] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3 +[88] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 7, 8, S2 +[89] Heng Yu, Chaoyang Wang, Peiye Zhuang, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Laszlo A Jeni, Sergey Tulyakov, and Hsin-Ying Lee. 4real: Towards photorealistic 4d scene generation via video diffusion models. arXiv preprint arXiv:2406.07472, 2024. 1 +[90] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 3, 5, 7, S1, S2 +[91] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1, 3, 7, S1, S2 +[92] Shenghai Yuan, Jinfa Huang, Yongqi Xu, Yaoyang Liu, Shaofeng Zhang, Yujun Shi, Ruijie Zhu, Xinhua Cheng, Jiebo + +Luo, and Li Yuan. Chronomagic-bench: A benchmark for metamorphic evaluation of text-to-time-lapse video generation. arXiv preprint arXiv:2406.18522, 2024. 1, 2 +[93] Guozhen Zhang, Chunxu Liu, Yutao Cui, Xiaotong Zhao, Kai Ma, and Limin Wang. Vfimamba: Video frame interpolation with state space models. arXiv preprint arXiv:2407.02315, 2024. 7, S5, S6 +[94] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. S6 +[95] Yuyang Zhao, Zhiwen Yan, Enze Xie, Lanqing Hong, Zhenguo Li, and Gim Hee Lee. Animate124: Animating one image to 4d dynamic scene. arXiv preprint arXiv:2311.14603, 2023. 3 +[96] Yufeng Zheng, Xueting Li, Koki Nagano, Sifei Liu, Otmar Hilliges, and Shalini De Mello. A unified approach for text- and image-guided 4d scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7300–7309, 2024. 3 +[97] Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. Allegro: Open the black box of commercial-level video generation model. arXiv preprint arXiv:2410.15458, 2024. 2, 7, S2 +[98] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 4, S1, S4 \ No newline at end of file diff --git a/worldscoreaunifiedevaluationbenchmarkforworldgeneration/images.zip b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db08db46c94d4877a0e9fb8f8249eeafacb2e035 --- /dev/null +++ b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12c25f8f091cc2c1b89740c8f9dc642272c843452e323c25d19488b30e53a7fe +size 900547 diff --git a/worldscoreaunifiedevaluationbenchmarkforworldgeneration/layout.json b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..70f033027a6d9366b2bd68ae00ac962b9b10e746 --- /dev/null +++ b/worldscoreaunifiedevaluationbenchmarkforworldgeneration/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ae664b84342d32528641cf642fc7f11a98d73c26ce3481efdd4f2b1efcf7021 +size 429412 diff --git a/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/0bfd0212-1def-41aa-be95-53d5f33cfe97_content_list.json b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/0bfd0212-1def-41aa-be95-53d5f33cfe97_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..781e2b03d9f44d05b281e9c9dd551d296f6721bd --- /dev/null +++ b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/0bfd0212-1def-41aa-be95-53d5f33cfe97_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86ea2d49a046c79372faa9d00a11332da626386d4366fffb541bb5ad70579271 +size 78361 diff --git a/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/0bfd0212-1def-41aa-be95-53d5f33cfe97_model.json b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/0bfd0212-1def-41aa-be95-53d5f33cfe97_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b67185adfd4f0dc1224e313c1cf035ca3863cf7c --- /dev/null +++ b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/0bfd0212-1def-41aa-be95-53d5f33cfe97_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76895c08898d7f83409ab5a063ed967c900aa3ae0adcc8ba2c24495a9f02074f +size 100462 diff --git a/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/0bfd0212-1def-41aa-be95-53d5f33cfe97_origin.pdf b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/0bfd0212-1def-41aa-be95-53d5f33cfe97_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..547df31560253624fc5c73c93532103e6d62cb24 --- /dev/null +++ b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/0bfd0212-1def-41aa-be95-53d5f33cfe97_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb1ab25a6a733bd10a66164d34860b0a713d3034a4df6314a589ad4fb405e337 +size 1652269 diff --git a/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/full.md b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f13c709952f9d5f55638d30c96bf62a7b8c741c7 --- /dev/null +++ b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/full.md @@ -0,0 +1,335 @@ +# $\mathbf{X}^2$ -Gaussian: 4D Radiative Gaussian Splatting for Continuous-time Tomographic Reconstruction + +Weihao Yu $^{1}$ Yuanhao Cai $^{2}$ Ruyi Zha $^{3}$ Zhiwen Fan $^{4}$ Chenxin Li $^{1}$ Yixuan Yuan $^{1*}$ + +1The Chinese University of Hong Kong 2Johns Hopkins University + +3The Australian National University 4University of Texas at Austin + +![](images/86a06b78efa64b69e3f23733b198a9b2b9288f903e13002e9cb4392d608ce363.jpg) +Figure 1. Dynamic reconstruction results of the proposed $\mathbf{X}^2$ -Gaussian on public DIR Dataset [9]. The red dashed line is the reference line for diaphragm movement, and blue dashed box shows some tissue deformation. Our method demonstrates superior capability in continuous-time reconstruction, significantly outperforming existing approaches. + +# Abstract + +Four-dimensional computed tomography (4D CT) reconstruction is crucial for capturing dynamic anatomical changes but faces inherent limitations from conventional phase-binning workflows. Current methods discretize temporal resolution into fixed phases with respiratory gating devices, introducing motion misalignment and restricting clinical practicality. In this paper, We propose $X^2$ -Gaussian, a novel framework that enables continuous-time 4D-CT reconstruction by integrating dynamic radiative Gaussian splatting with self-supervised respiratory motion learning. Our approach models anatomical dynamics through a spatiotemporal encoder-decoder architecture that predicts time-varying Gaussian deformations, eliminating phase discretization. To remove dependency on external gating devices, we introduce a physiology-driven periodic consistency loss that learns patient-specific breathing cycles directly from projections via differentiable optimization. Extensive experiments demonstrate state-of-the + +art performance, achieving a 9.93 dB PSNR gain over traditional methods and 2.25 dB improvement against prior Gaussian splatting techniques. By unifying continuous motion modeling with hardware-free period learning, $X^2$ -Gaussian advances high-fidelity 4D CT reconstruction for dynamic clinical imaging. Code will be publicly available at: https://github.com/CUHK-AIM-Group/X2-Gaussian. + +# 1. Introduction + +Four-dimensional computed tomography (4D CT) has become a cornerstone in dynamic medical imaging, especially for respiratory motion management in clinical applications such as image-guided radiotherapy (IGRT) [12, 37]. By capturing both spatial and temporal information of the chest cavity during breathing cycles, 4D CT enables clinicians to monitor and assess respiratory-induced tumor motion and other dynamic anatomical changes during treatment [3, 13, 48]. + +Traditional 4D CT reconstruction follows a phase- + +binning workflow. It first divides the projections into discrete respiratory phases using external gating devices that require direct patient contact, followed by independent reconstruction of each phase to obtain a sequence of 3D volumes. Within this framework, 3D reconstruction methods such as Feldkamp-David-Kress (FDK) algorithm [41], or total variation minimization [46, 47], can be directly applied to 4D CT reconstruction. Due to the limited number of projections available per phase, the reconstructed CT images frequently exhibit significant streak artifacts, which degrade the visibility of fine tissue structures. To address this issue, several researchers [5, 11, 32, 40, 61] have proposed methods for extracting patient-specific motion patterns to compensate for respiratory motion across different phases. Meanwhile, other studies [22, 24, 25, 30, 62] have explored the use of Convolutional Neural Networks (CNNs) to restore details in artifact-contaminated images. + +Recent advances in Neural Radiance Fields (NeRF) [36] have introduced improved methods for CT reconstruction [7, 58]. These approaches enable high-fidelity 3D reconstruction from sparse views, thereby mitigating the projection undersampling issues caused by phase partitioning. The emergence of 3D Gaussian splatting (3DGS) [28] has further facilitated the development of more efficient and higher-quality methods [6, 59]. Despite these progress, the reconstruction of 4D CT still suffers from two challenges rooted in the traditional phase-binning paradigm. Firstly, previous methods simulate 4D imaging through a series of disjoint 3D reconstructions at predefined phases, failing to model the continuous spatiotemporal evolution of anatomy. This discretization introduces temporal inconsistencies, limits resolution to a few static snapshots per cycle, and produces artifacts when interpolating between phases. Secondly, they heavily relies on external respiratory gating devices, not only introducing additional hardware dependencies and potential measurement errors that can compromise reconstruction accuracy, but also imposing physical constraints and discomfort on patients during the scanning process. + +To overcome these limitations, we propose $\mathbf{X}^2$ -Gaussian, a novel framework that achieves genuine 4D CT reconstruction by directly modeling continuous anatomical motion. Firstly, unlike previous approaches that perform sequential 3D reconstructions, our method introduces a dynamic Gaussian motion model that explicitly captures the continuous deformation of anatomical structures over time by extending radiative Gaussian splatting [59] into the temporal domain. Specifically, we design a spatiotemporal encoder that projects Gaussian properties onto multi-resolution feature planes, effectively capturing both local anatomical relationships and global motion patterns. The encoded features are then processed by a lightweight multi-head decoder network that predicts deformation parameters for each Gaussian at any queried timestamp, enabling true 4D reconstruction without discrete phase binning. Secondly, we intro + +duce a self-supervised respiratory motion learning method to eliminate the requirement of external gating devices. By leveraging the quasi-periodic nature of respiratory motion, our approach learns to estimate the breathing period directly from the projection data through a novel physiology-driven periodic consistency mechanism that enforces temporal coherence across respiratory cycles. This approach fundamentally differs from traditional phase-based methods by transforming the discrete phase assignments into learnable continuous parameters, enabling our model to automatically discover and adapt to patient-specific breathing patterns. + +As shown in Fig. 1, $\mathbf{X}^2$ -Gaussian exhibits superior reconstruction performance compared to existing state-of-the-art methods, establishing a new benchmark in 4D CT reconstruction. Our contributions can be summarized as follows: + +- We present $\mathrm{X}^2$ -Gaussian, the first method to directly reconstruct time-continuous 4D-CT volumes from projections, which bypasses phase binning entirely, enabling motion analysis at arbitrary temporal resolutions. +- We extend the static radiative Gaussian splattering into the temporal domain. To our knowledge, this is the first attempt to explore the potential of Gaussian splattering in dynamic tomographic reconstruction. +- We introduce a novel self-supervised respiratory motion learning module that jointly estimates the respiratory cycle and enforces periodic consistency, eliminating reliance on external gating devices. +- Extensive experiments demonstrate that our method significantly improves reconstruction quality, reduces streak artifacts, and accurately models respiratory motion, while also showing potential for automatic extraction of various clinical parameters. + +# 2. Related Work + +# 2.1. CT Reconstruction + +Traditional 3D computed tomography reconstruction methods mainly include two categories: analytical algorithms [15, 54] and iterative algorithms [1, 34, 43, 45]. Analytical methods estimate the radiodensity by solving Radon transformation and its inverse version. Iterative algorithms are based on optimization over iterations. In recent years, deep learning based models [2, 18, 26, 31, 33] like CNNs have been employed to learn a brute-force mapping from X-ray projections to CT slices. Another technical route is to employ the 3D rendering algorithms such as neural radiance fields (NeRF) [36] and 3D Gaussian Splitting (3DGS) [28] to solve the CT reconstruction problem in a self-supervised manner, i.e. using only 2D X-rays for training. Based on these algorithms, when coping with 4D CTs, researchers typically segment the projections into ten discrete respiratory phases for sequential 3D reconstruction. This approach not only necessitates external devices for phase measurement during scanning but also impedes accurate modeling of the continuous motion of anatomical structures. Concur + +![](images/fa33956d80040fb8344ee71042d2223330febcc9ded2cc08c3d14bdd4394a487.jpg) +Figure 2. Framework of our $\mathbf{X}^2$ -Gaussian, which consists of two innovative components: (1) Dynamic Gaussian motion modeling for continuous-time reconstruction; (2) Self-Supervised respiratory motion learning for estimating breathing cycle autonomously. + +rent work [17] also employs dynamic Gaussian splatting. However, they merely establish ten timestamps corresponding to ten phases, thereby maintaining a discrete representation. In contrast, this paper is dedicated to achieving truly continuous-time 4D CT reconstruction. + +# 2.2. Gaussian Splatting + +3D Gaussian splatting [28] (3DGS) is firstly proposed for view synthesis. It uses millions of 3D Gaussian point clouds to represent scenes or objects. In the past two years, 3DGS has achieved great progress in scene modeling [50, 53, 55, 60], SLAM [35, 52, 56], 3D Generation [39, 51], medical imaging [6, 59], etc. For instance, Cai et al. design the first 3DGS-based method, X-GS [6], for X-ray projection rendering. Later work $\mathbf{R}^2\mathrm{GS}$ [59] rectifies 3DGS pipeline to enable the direct CT reconstruction. Nonetheless, these algorithms show limitations in reconstructing dynamic CT volumes. Our goal is to cope with this problem. + +# 3. Preliminaries + +Radiative Gaussian Splitting [59] represents 3D CT using a collection of Gaussian kernels $\mathbb{G} = \{G_i\}_{i=1}^K$ , each characterized by its central position $\pmb{\mu}_i \in \mathbb{R}^3$ , covariance matrix $\pmb{\Sigma}_i \in \mathbb{R}^{3 \times 3}$ , and isotropic density $\rho_i$ : + +$$ +G _ {i} (\boldsymbol {x} | \rho_ {i}, \boldsymbol {\mu} _ {i}, \boldsymbol {\Sigma} _ {i}) = \rho_ {i} \cdot \exp \left(- \frac {1}{2} (\boldsymbol {x} - \boldsymbol {\mu} _ {i}) ^ {T} \boldsymbol {\Sigma} _ {i} ^ {- 1} (\boldsymbol {x} - \boldsymbol {\mu} _ {i})\right). \tag {1} +$$ + +The covariance matrix can be decomposed as: $\boldsymbol{\Sigma}_i = \boldsymbol{R}_i\boldsymbol{S}_i\boldsymbol{S}_i^T\boldsymbol{R}_i^T$ , where $\boldsymbol{R}_i \in \mathbb{R}^{3 \times 3}$ is the rotation matrix and $\boldsymbol{S}_i \in \mathbb{R}^{3 \times 3}$ is the scaling matrix. Then the total density at position $\boldsymbol{x}$ is computed as the sum of all contributed Gaussian kernels: + +$$ +\sigma (\boldsymbol {x}) = \sum_ {i = 1} ^ {N} G _ {i} (x | \rho_ {i}, \boldsymbol {\mu} _ {i}, \boldsymbol {\Sigma} _ {i}). \tag {2} +$$ + +For 2D image rendering, the attenuation of X-ray through a medium follows the Beer-Lambert Law [27]: + +$$ +I (\boldsymbol {r}) = \log I _ {0} - \log I ^ {\prime} (\boldsymbol {r}) = \int \sigma (\boldsymbol {r} (t)) d t, \tag {3} +$$ + +where $I_0$ is the initial X-ray intensity, $\boldsymbol{r}(t) = \boldsymbol{o} + t\boldsymbol{d} \in \mathbb{R}^3$ represents a ray path, and $\sigma(\boldsymbol{x})$ denotes the isotropic density at position $\boldsymbol{x} \in \mathbb{R}^3$ . Thus, the final pixel value is obtained by integrating the density field along each ray path + +$$ +I _ {r} (\boldsymbol {r}) = \sum_ {i = 1} ^ {N} \int G _ {i} (\boldsymbol {r} (t) | \rho_ {i}, \boldsymbol {\mu} _ {i}, \boldsymbol {\Sigma} _ {i}) d t, \tag {4} +$$ + +where $I_r(\boldsymbol{r})$ is the rendered pixel value. + +# 4. Methods + +# 4.1. Overview + +Given a sequence of X-ray projections $\{I_j\}_{j=1}^N$ acquired at timestamps $\{t_j\}_{j=1}^N$ and view matrices $\{M_j\}_{j=1}^N$ , our goal + +is to learn a continuous representation of the dynamic CT volume that can be queried at arbitrary timestamps, thereby overcoming the inherent limitations of discrete phase binning. To accomplish this, as shown in Fig. 2, our method seamlessly integrates dynamic Gaussian motion modeling with a self-supervised respiratory motion learning scheme into a unified, end-to-end differentiable framework. Specifically, raw Gaussian parameters are initialized from $\{I_j\}_{j=1}^N$ and $\{M_j\}_{j=1}^N$ . Given a timestamp $t_j$ , dynamic Gaussian motion modeling module predicts the deformation of each parameter, allowing continuous-time reconstruction. Additionally, we model the respiratory cycle as a learnable parameter and sample another timestamp accordingly. Through carefully designed periodic consistency loss, we mine the real breathing period in a self-supervised way. + +# 4.2. Dynamic Gaussian Motion Modeling + +To achieve continuous 4D CT reconstruction, we introduce a deformation field that models the anatomical dynamics. At the core of our method is a time-dependent deformation field $\mathcal{D}(\pmb{\mu}_i,t)$ that predicts the deformation parameters $\Delta G_{i}$ for each Gaussian at time $t$ . The deformed Gaussians $G_{i}^{\prime}$ can be computed as: + +$$ +G _ {i} ^ {\prime} = G _ {i} + \Delta G _ {i} = \left(\boldsymbol {\mu} _ {i} + \Delta \boldsymbol {\mu} _ {i}, \boldsymbol {R} _ {i} + \Delta \boldsymbol {R} _ {i}, \boldsymbol {S} _ {i} + \Delta \boldsymbol {S} _ {i}, \rho_ {i}\right), \tag {5} +$$ + +where $\Delta \mu_{i},\Delta R_{i}$ , and $\Delta S_{i}$ are the deformation offsets for position, rotation, and scaling, respectively. Our deformation field $\mathcal{D}$ is implemented as a composition of two components: $\mathcal{D} = \mathcal{F}\circ \mathcal{E}$ , where $\mathcal{E}$ is a spatiotemporal encoder and $\mathcal{F}$ is a deformation-aware decoder. + +Decomposed Spatio-Temporal Encoding. To encode the spatiotemporal features of Gaussian primitives, a straightforward approach would be to employ neural networks to directly parameterize $\mathcal{E}$ . But such a method may lead to low rendering speed and potential overfitting issues, especially given the sparse projection data in 4D CT reconstruction. Inspired by recent advances in dynamic scene reconstruction [8, 14, 50], we adopt a decomposed approach that factorizes the 4D feature space into a set of multi-resolution K-Planes [16], which reduces memory requirements while preserving the ability to model complex spatiotemporal patterns in respiratory motion. + +Specifically, given a Gaussian center $\pmb{\mu} = (x,y,z)$ and timestamp $t$ , we project 4D coordinates $\pmb{v} = (x,y,z,t)$ onto six orthogonal feature planes: three spatial planes $\mathcal{P}xy$ , $\mathcal{P}xz$ , $\mathcal{P}yz$ and three temporal planes $\mathcal{P}xt$ , $\mathcal{P}yt$ , $\mathcal{P}zt$ . Each plane $\mathcal{P} \in \mathbb{R}^{d \times lM \times lM}$ stores learnable features of dimension $d$ at multiple resolutions $l \in 1,\dots,L$ , where $M$ is the basic resolution, enabling simultaneous modeling of fine local motion and global respiratory patterns. The encoded feature $\pmb{f}_e$ is computed through bilinear interpolation across multi-resolution planes: + +$$ +f _ {e} = \oplus_ {l} \otimes_ {(a, b)} \psi (\mathcal {P} _ {a b} ^ {l} (\boldsymbol {v})), \tag {6} +$$ + +![](images/6e4ef2a91328743f6313e2fd2ccafb576357d7f3cd0282dacaca573ce2eee08d.jpg) +t=0.1s + +![](images/7bd7462551371a74e3ff111ed68f0bb513e7653b511fa4e8e09d044b4b3494b3.jpg) +$t = 3.1s$ + +![](images/cc3ed8a6dc2808484f3fee07ea5d615b44252726bcd79b3d6bd36bbd842e8285.jpg) +t=6.1s + +![](images/40508320e2f34a895cbfbfaa6b333f0302fb41d8fd7e65092766359b59186d13.jpg) +$t = 1.7s$ +Figure 3. Periodic display of respiratory motion ( $T = 3s$ ). A specific anatomical structure (framed by boxes of the same color) at time $t$ has the same position at time $t + nT$ . + +![](images/89e895c0da513ea47ffdc44a4963d099f57ba2ba4b1178395ab46e7a9692fb4b.jpg) +$t = 4.7s$ + +![](images/0c7108a94f387b26693e380e1b096bf5c7dee482e783d5d1073115a5ef2c7941.jpg) +$t = 7.7s$ + +where $\psi$ denotes bilinear interpolation, $\oplus$ represents feature concatenation, $\otimes$ is Hadamard product, and $(a,b) \in \{(x,y),(x,z),(y,z),(x,t),(y,t),(z,t)\}$ . Then $\pmb{f}_{e}$ is further merged through a tiny feature fusion network $\phi_h$ (i.e. one layer of MLP) as $\pmb{f}_h = \phi_h(\pmb{f}_e)$ . + +Deformation-Aware Gaussian Decoding. Once the spatiotemporal features are encoded, we employ a lightweight multi-head decoder network $\mathcal{F}$ to predict the deformation parameters for each Gaussian: + +$$ +\Delta \boldsymbol {\mu}, \Delta \boldsymbol {R}, \Delta \boldsymbol {S} = \mathcal {F} _ {\boldsymbol {\mu}} (\boldsymbol {f} _ {h}), \mathcal {F} _ {R} (\boldsymbol {f} _ {h}), \mathcal {F} _ {S} (\boldsymbol {f} _ {h}). \tag {7} +$$ + +Such decoupled design allows specialized learning of different motion characteristics: position shifts for translational movements, rotation for orientation changes, and scaling for volumetric expansion/contraction. Then the deformed Gaussian parameters at timestamp $t$ can be calculated according to Eq. 5. In this way, our dynamic Gaussian motion modeling not only allows independently fine-tune different aspects of motion but also facilitates continuous interpolation across time, yielding smooth temporal transitions in the reconstructed CT volume. + +# 4.3. Self-Supervised Respiratory Motion Learning + +To eliminate the need for external respiratory gating devices while accurately capturing breathing patterns, we introduce a self-supervised approach that directly learns respiratory motion from projection data. Our method leverages the inherently periodic nature of human respiration to establish temporal coherence across respiratory cycles. + +Physiology-Driven Periodic Consistency Loss. Respiratory motion exhibits an inherently cyclic pattern, with anatomical structures returning to approximately the same position after each breathing cycle [20]. This physiological characteristic serves as a powerful prior to constrain + +![](images/30c50cc44bfd23c243da0b0d3604d8578628bc4b5203121397e8249c58e61f70.jpg) +Figure 4. Convergence behavior of the learnable period $\hat{T}$ . Without Bounded Cycle Shifts, $\hat{T}$ undergoes wide-ranging oscillations approaching half the true period. Without Log-Space Parameterization, the optimization curve exhibits large oscillations. With both techniques implemented, $\hat{T}$ converges stably and accurately to the correct breathing cycle. + +the reconstruction process. As illustrated in Fig. 3, a given anatomical position at time $t$ should match its state at time $t + nT$ , where $T$ represents the respiratory period and $n$ is an integer. To explicitly encode this periodicity, we enforce a consistency constraint on the rendered images: + +$$ +I (t) = I (t + n T). \tag {8} +$$ + +In practice, we define a periodic consistency loss: + +$$ +\mathcal {L} _ {p c} = \mathcal {L} _ {1} (I (t), I (t + n T)) + \lambda_ {1} \mathcal {L} _ {s s i m} (I (t), I (t + n T)), \tag {9} +$$ + +which encourages the reconstructed images at times $t$ and $t + nT$ to be similar. Here, $\mathcal{L}_1$ and $\mathcal{L}_{ssim}$ are L1 loss and D-SSIM loss [49], respectively. This constraint effectively reduces the temporal degrees of freedom in our model by enforcing cyclic coherence, helping to mitigate artifacts and improve reconstruction quality, especially in regions with significant respiratory-induced motion. + +Differentiable Cycle-Length Optimization. In realistic scenarios, the true respiratory cycle $T$ is not available a priori. Hence, we treat it as a learnable parameter $\hat{T}$ within our framework. Instead of being provided externally, $\hat{T}$ is optimized directly from the projection data by backpropagating the periodic consistency loss. This allows the network to automatically discover the breathing period in a self-supervised manner. To ensure numerical stability and avoid harmonic artifacts, we implement two critical designs: + +- Bounded Cycle Shifts: We restrict the integer $n$ in our periodic consistency loss to $n \in \{-1, 1\}$ , focusing only on adjacent respiratory cycles. This restriction is critical for avoiding potential ambiguities in period estimation. When using larger values of $n$ , the optimization might converge to period estimates that are multiples or divisors of the true period. For example, if the true period $T$ is 3 + +seconds and our model learns $\hat{T} = 4$ seconds, then with $n = 6$ , we would enforce consistency between times $t$ and $t + 24$ seconds, which coincidentally satisfies periodicity (as 24 is divisible by the true period of 3). By limiting $n$ to adjacent cycles, we ensure the model learns the fundamental period rather than its harmonics. + +- Log-Space Parameterization: We represent $\hat{T} = \exp(\hat{\tau})$ where $\hat{\tau} \in \mathbb{R}$ is an unbounded learnable variable. This ensures positivity and provides smoother gradient updates compared to direct period estimation. This logarithmic parameterization ensures $T$ remains positive, improves numerical stability by preventing extremely small period values, and creates a more uniform gradient landscape for optimization. + +As shown in Fig. 4, these two technical designs are critical for accurate and stable period estimation. Without bounded cycle shifts, the learned period $\hat{T}$ oscillates with large amplitude approaching sub-harmonics (i.e. $T / 2$ ) of the true respiratory period, as the periodic consistency loss can be satisfied by most common multiples of sub-harmonics. Direct optimization in linear space leads to pronounced oscillations in the learning trajectory of $\hat{T}$ . With both techniques implemented, $\hat{T}$ converges stably and accurately to the correct breathing cycle. In this way, we reformulate Eq. 9 as + +$$ +\begin{array}{l} \mathcal {L} _ {p c} = \mathcal {L} _ {1} \left(I (t), I (t + n \exp (\hat {\tau}))\right) \\ + \lambda_ {1} \mathcal {L} _ {s s i m} \left(I (t), I (t + n \exp (\hat {\tau}))\right), \tag {10} \\ \end{array} +$$ + +where $n \in \{-1, 1\}$ . Then the optimal period $T^{*}$ can be learned via + +$$ +\tau^ {*} = \underset {\hat {\tau}} {\arg \min } \mathcal {L} _ {p c}, \quad T ^ {*} = \exp \left(\tau^ {*}\right). \tag {11} +$$ + +Through this self-supervised optimization approach, our model automatically discovers patient-specific breathing patterns directly from projection data without requiring external gating devices, simplifying clinical workflow while improving reconstruction accuracy. + +# 4.4. Optimization + +Loss Function. We optimize our framework by employing a compound loss function. Similar to $\mathcal{L}_{pc}$ , we use L1 loss and D-SSIM loss to supervise the rendered X-ray projections as $\mathcal{L}_{render} = \mathcal{L}_1 + \lambda_2\mathcal{L}_{ssim}$ . Following [59], we integrate a 3D total variation (TV) regularization term [42] $\mathcal{L}_{TV}^{3D}$ to promote spatial homogeneity in the CT volume. We also apply a grid-based TV loss [8, 16, 50] $\mathcal{L}_{TV}^{4D}$ to the multi-resolution k-plane grids used during spatiotemporal encoding. The overall loss function is then defined as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {r e n d e r}} + \alpha \mathcal {L} _ {p c} + \beta \mathcal {L} _ {T V} ^ {3 D} + \gamma \mathcal {L} _ {T V} ^ {4 D}, \tag {12} +$$ + +where $\alpha, \beta$ , and $\gamma$ are weights that control the relative influence of the periodic consistency and regularization terms. + +Table 1. Comparison of our ${\mathrm{X}}^{2}$ -Gaussian with different methods on the DIR dataset. + +
MethodPatient1Patient2Patient3Patient4Patient5Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
FDK [41]34.470.83625.050.62434.230.82628.050.70925.250.63829.410.727
IntraTomo [57]40.040.96530.620.88933.550.88833.000.91032.80.93534.000.917
NeRF [36]40.850.96432.870.91733.430.89733.660.92234.290.95535.020.931
TensoRF [10]33.210.90730.320.86433.470.88133.640.81332.400.92832.610.898
NAF [58]38.210.94531.730.87534.110.90033.950.91131.740.92733.950.912
SAX-NeRF [7]37.210.96131.530.93836.710.92934.300.94433.140.94734.580.942
3D-GS [28]34.190.84722.960.71332.530.84026.320.79329.890.81229.180.801
X-GS [6]38.000.90325.320.73933.540.85428.690.80728.770.79330.860.819
R²-GS [59]40.510.96633.750.92139.660.95636.450.93835.090.93737.090.943
Ours44.60.97835.320.93543.220.97237.180.94236.360.94739.340.955
+ +Table 2. Comparison of our $\mathrm{X}^2$ -Gaussian with different methods on the 4DLung and SPARE datasets. + +
Method4DLungSPARE
PSNRSSIMPSNRSSIM
FDK [41]27.030.61114.250.359
IntraTomo [57]34.280.93927.290.871
TensoRF [10]34.550.93726.910.857
NAF [58]34.940.93628.440.893
X-GS [6]29.620.70518.200.442
R²-GS [59]37.310.95231.120.908
Ours38.610.95732.240.922
+ +Progressive Training Procedure. During training, we first train a static 3D radiative Gaussian splatting model [59] for 5000 iterations. This warm-up phase ensures that the model effectively captures the underlying anatomical structures from the projection data. After the warm-up period, we extend the framework to its full 4D form. The Gaussian parameters, spatiotemporal encoder/decoder, and the learnable respiratory period parameter $\hat{\tau}$ are jointly optimized using the combined loss $\mathcal{L}_{total}$ . This progressive training strategy enables the model to build on a robust 3D reconstruction before incorporating temporal dynamics, resulting in stable convergence and high-quality dynamic reconstruction. + +# 5. Experiments + +# 5.1. Dataset and Implementation Details + +We conducted experiments on 4D CT scans from 13 patients across three public datasets: 5 patients from DIR dataset [9], 5 from 4DLung dataset [23], and 3 from SPARE dataset [44]. Each patient's 4D CT consists of 10 3D CTs from different phases. We used the tomographic toolbox TIGRE [4] to simulate clinically significant one-minute 4D CT sampling. The respiratory cycle was configured at 3 + +seconds, with the corresponding phase determined based on sampling time to obtain X-ray projections. For each patient, 300 projections were sampled, which is substantially fewer than the several thousand projections currently required in clinical settings. + +Our $\mathrm{X}^2$ -Gaussian was implemented by PyTorch [38] and CUDA [19] and trained with the Adam optimizer [29] for 30K iterations on an RTX 4090 GPU. Learning rates for position, density, scale, and rotation are initially set at 2e-4, 1e-2, 5e-3, and 1e-3, respectively, and decay exponentially to $10\%$ of their initial values. The initial learning rates for the spatio-temporal encoder, decoder, and learnable period are set at 2e-3, 2e-4, and 2e-4, respectively, and similarly decay exponentially to $10\%$ of their initial values. $\hat{\tau}$ was initialized to 1.0296 ( $\hat{T} = 2.8$ ). $\lambda_1$ and $\lambda_2$ in $\mathcal{L}_{pc}$ and $\mathcal{L}_{render}$ were 0.25. $\alpha$ , $\beta$ , and $\gamma$ in Eq. 12 were set to 1.0, 0.05, and 0.001, respectively. During testing, We used PSNR and SSIM to evaluate the volumetric reconstruction performance. $\mathrm{X}^2$ -Gaussian predicted 103D CTs corresponding to the time of each phase, with PSNR calculated on the entire 3D volume and SSIM computed as the average of 2D slices in axial, coronal, and sagittal directions. + +# 5.2. Results + +Tab. 1 and Tab. 2 illustrate the quantitative results of our $\mathbf{X}^2$ -Gaussian and SOTA 3D reconstruction methods which follow the phase-bining workflow, including traditional methods (FDK [41]), NeRF-based methods (IntraTomo [57], NeRF [36], TensoRF [10], NAF [58], SAX-NeRF [7]), and GS-based methods (3D-GS [28], X-GS [6], $\mathbb{R}^2$ -GS [59]). As can be seen in Tab. 1, our method significantly outperforms other approaches in reconstruction quality. Specifically, compared to the traditional FDK method, our approach demonstrates a 9.93 dB improvement in PSNR, achieving approximately a $34\%$ enhancement. When compared to state-of-the-art methods, our approach surpasses the NeRF-based method SAN-NeRF by 4.76 dB and the GS-based method R2-GS by 2.25 dB. Similar results can be observed in Tab. 2, demonstrating the superiority of our + +![](images/c59cfe0447c2fceb635767c64e73b8b3dc6fd65f27772bb7cecf9781f565417c.jpg) +Figure 5. Qualitative comparison of reconstruction results across coronal, sagittal, and axial planes. Our method shows superior performance in modeling dynamic regions (e.g. diaphragmatic motion and airway deformation) while preserving finer anatomical details compared to existing approaches. + +# method. + +Fig. 5 shows the quantitative comparison of reconstruction results between our method and existing approaches. Examination of the coronal and sagittal planes shows that our method distinctly captures diaphragmatic motion with remarkable fidelity, which can be attributed to the powerful continuous-time reconstruction capability of $\mathbf{X}^2$ -Gaussian. Similarly, on the axial plane, $\mathbf{X}^2$ -Gaussian successfully reconstructs the deformed airways. Additionally, $\mathbf{X}^2$ -Gaussian preserves fine anatomical details that competing approaches fail to recover, underscoring its effectiveness for high-fidelity volumetric reconstruction. + +# 5.3. Ablation study + +Period Optimization Tab. 3 demonstrates the effectiveness of our $\mathrm{X}^2$ -Gaussian for respiratory cycle estimation and examines how various optimization techniques influence estimation precision. Our approach achieves exceptional accuracy with an average error of just 5.2 milliseconds—approximately one-thousandth of a typical human respiratory cycle. This precision stems from two key technical contributions: Log-Space Parameterization and Bounded Cycle Shifts. Without Log-Space Parameteriza- + +tion, we observe oscillatory convergence behavior that compromises accuracy. More dramatically, when Bounded Cycle Shifts are omitted, the optimization incorrectly converges to harmonic frequencies rather than the fundamental cycle, resulting in a 40-fold increase in estimation error. These findings highlight the critical importance of our optimization framework in achieving reliable respiratory cycle estimation. + +Component Analysis We conducted ablation experiments on DIR dataset to validate the effect of key components in $\mathbf{X}^2$ -Gaussian, including the dynamic gaussian motion modeling (DGMM) and self-supervised respiratory motion learning (SSRML). Tab. 4 reports the results. As we can see, DGMM extends the static 3D radiative Gaussian splatting model to temporal domain, enabling continuous-time reconstruction and achieving improved 4D reconstruction results. Building upon this foundation, SSRML leverages the periodicity of respiratory motion to directly learn breathing patterns. Remarkably, this approach not only successfully captures specific respiratory cycles but also further enhances reconstruction quality by $0.78~\mathrm{dB}$ , demonstrating its significant contribution to improving temporal coherence + +![](images/e6c6768203eebbef2e05895d53e33d9aaa032133b9192ad02758c89098ace5a6.jpg) +(a) Reconstruction results with different view numbers + +![](images/a96052b457673820d742fb51484fa0ea2846fb41a9cc7736893234a333b948ce.jpg) +(b) Tidal volume during respiratory cycle +Figure 6. (a) Reconstruction results of $\mathrm{X}^2$ -Gaussian using different numbers of projections. (b) Temporal variations of lung volume in 4D CT reconstructed by $\mathrm{X}^2$ -Gaussian. + +and physiological motion plausibility. + +Table 3. Results of respiratory cycle estimation and different optimization techniques used on DIR dataset. + +
MethodPSNRSSIMEst. error of T (ms)
Ours39.340.9555.2
- Log-sp. param.39.320.95412.0
- B. cyc. shifts39.280.954216.8
- Both39.230.953914.0
+ +Hyperparameter Analysis We further analyzed the impact of different weights $\alpha$ of periodic consistency loss $\mathcal{L}_{pc}$ in Tab. 4. The optimal performance is achieved when periodic consistency loss and rendering loss are equally weighted (i.e. $\alpha = 1.0$ ), as this balance enables the model to simultaneously preserve visual fidelity while enforcing physiologically plausible temporal dynamics. When the weighting is either too high or too low, this equilibrium is disrupted, leading to performance degradation due to either over-constraining the periodic structure at the expense of reconstruction accuracy or prioritizing visual appearance without sufficient temporal coherence. + +# 5.4. Discussion + +Projection Numbers Fig. 6 (a) demonstrates the reconstruction results of $\mathrm{X}^2$ -Gaussian using different numbers of projections. As can be observed, the reconstruction quality gradually improves with an increasing number of available projections. Surprisingly, when compared with Tab. 1, we found that even when trained with only $100\mathrm{X}$ -ray images, our method still achieves better reconstruction results than the current SOTA method $\mathbb{R}^2$ -GS using $300\mathrm{X}$ -rays (37.41 dB vs. 37.09 dB). This clearly demonstrates the powerful capability of our approach. + +Respiratory Motion Quantification We densely sampled our $\mathrm{X}^2$ -Gaussian reconstructed 4D CT within 9 seconds, resulting in 180 3D CT volumes. With automated segmentation algorithm [21], we extracted lung masks and + +Table 4. Ablation studies on components and hyperparameters. DGMM denotes dynamic gaussian motion modeling in Sec. 4.2, and SSRML is self-supervised respiratory motion learning in Sec. 4.3. $\alpha$ is the weight of periodic consistency loss in Eq. 12. + +
MethodPSNRSSIM
Baseline37.090.943
+ DGMM38.560.947
+ DGMM + SSRML39.340.955
α = 0.138.860.952
α = 0.539.140.954
α = 1.039.340.955
α = 2.038.410.949
+ +calculated the volumetric changes of the lungs over time, as displayed in Fig. 6 (b). The pulmonary volume dynamics exhibit a periodic sinusoidal pattern, which precisely correlates with the subject's respiratory cycle, demonstrating that our method successfully models respiratory dynamics while achieving truly temporally continuous reconstruction. Furthermore, clinically relevant parameters can be quantitatively extracted from the volume-time curve: Tidal Volume (TV) is $370\mathrm{ml}$ , Minute Ventilation (MV) is $7.4\mathrm{L / min}$ , I:E Ratio is 1:1.9, etc. These automatically extracted clinical parameters demonstrate the potential of $\mathbf{X}^2$ -Gaussian in radiomic-feature-guided treatment personalization. + +# 6. Conclusion + +This paper presents $\mathbf{X}^2$ -Gaussian, a continuous-time 4D CT reconstruction framework that leverages dynamic radiative Gaussian splatting to capture smooth anatomical motion. Our method bypasses the limitations of phase binning and external gating by integrating dynamic Gaussian motion modeling with a self-supervised respiratory motion module. Experimental results on clinical datasets demonstrate notable improvements in reconstruction fidelity and artifact suppression. This work bridges the gap between discrete-phase reconstruction and true 4D dynamic imaging, offering practical benefits for radiotherapy planning through improved motion analysis and patient comfort. + +# Acknowledgement + +This work was supported by the Hong Kong Research Grants Council (RGC) General Research Fund 14204321, 14220622 and Hainan Province Clinical Medical Center. + +# References + +[1] Anders H Andersen and Avinash C Kak. Simultaneous algebraic reconstruction technique (sart): a superior implementation of the art algorithm. Ultrason. Imaging, 6(1):81-94, 1984. 2 +[2] Rushil Anirudh, Hyojin Kim, Jayaraman J Thiagarajan, K Aditya Mohan, Kyle Champley, and Timo Bremer. Lose the views: Limited angle ct reconstruction via implicit sinogram completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6343-6352, 2018. 2 +[3] Pia Baumann, Jan Nyman, Morten Hoyer, Berit Wennberg, Giovanna Gagliardi, Ingmar Lax, Ninni Drugge, Lars Ekberg, Signe Friesland, Karl-Axel Johansson, et al. Outcome in a prospective phase ii trial of medically inoperable stage i non-small-cell lung cancer patients treated with stereotactic body radiotherapy. J. Clin. Oncol., 27(20):3290-3296, 2009. 1 +[4] Ander Biguri, Manjit Dosanjh, Steven Hancock, and Manuchehr Soleimani. Tigre: a matlab-gpu toolbox for cbct image reconstruction. Biomed. Phys. Eng. Express, 2(5): 055010, 2016. 6 +[5] Marcus Brehm, Pascal Paysan, Markus Oelhafen, and Marc Kachelrieß. Artifact-resistant motion estimation with a patient-specific artifact model for motion-compensated cone-beam ct. Med. Phys., 40(10):101913, 2013. 2 +[6] Yuanhao Cai, Yixun Liang, Jiahao Wang, Angtian Wang, Yulun Zhang, Xiaokang Yang, Zongwei Zhou, and Alan Yuille. Radiative gaussian splattering for efficient x-ray novel view synthesis. In Eur. Conf. Comput. Vis., pages 283-299. Springer, 2024. 2, 3, 6 +[7] Yuanhao Cai, Jiahao Wang, Alan Yuille, Zongwei Zhou, and Angtian Wang. Structure-aware sparse-view x-ray 3d reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11174-11183, 2024. 2, 6 +[8] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 130-141, 2023. 4, 5 +[9] Richard Castillo, Edward Castillo, Rudy Guerra, Valen E Johnson, Travis McPhail, Amit K Garg, and Thomas Guerrero. A framework for evaluation of deformable image registration spatial accuracy using large landmark point sets. Phys. Med. Biol., 54(7):1849, 2009. 1, 6 +[10] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Eur. Conf. Comput. Vis., pages 333-350. Springer, 2022. 6 +[11] Mingqing Chen, Kunlin Cao, Yefeng Zheng, and R Alfredo C Siochi. Motion-compensated mega-voltage cone beam ct using the deformation derived directly from 2d projection images. IEEE Trans. Med. Imag., 32(8):1365-1375, 2012. 2 +[12] Joanne N Davis, Clinton Medbery, Sanjeev Sharma, John Pablo, Frank Kimsey, David Perry, Alexander Muacevic, and + +Anand Mahadevan. Stereotactic body radiotherapy for centrally located early-stage non-small cell lung cancer or lung metastases from the rssearch® patient registry. *Radiat. Oncol.*, 10:1-10, 2015. 1 +[13] Achilles J Fakiris, Ronald C McGarry, Constantin T Yiannoutsos, Lech Papiez, Mark Williams, Mark A Henderson, and Robert Timmerman. Stereotactic body radiation therapy for early-stage non-small-cell lung carcinoma: four-year results of a prospective phase ii study. Int. J. Radiat. Oncol. Biol. Phys., 75(3):677-682, 2009. 1 +[14] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia, pages 1-9, 2022. 4 +[15] Lee A Feldkamp, Lloyd C Davis, and James W Kress. Practical cone-beam algorithm. Journal of the Optical Society of America A, 1(6):612-619, 1984. 2 +[16] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12479-12488, 2023. 4, 5 +[17] Yabo Fu, Hao Zhang, Weixing Cai, Huiqiao Xie, Licheng Kuo, Laura Cervino, Jean Moran, Xiang Li, and Tianfang Li. Spatiotemporal gaussian optimization for 4d cone beam ct reconstruction from sparse projections. arXiv preprint arXiv:2501.04140, 2025. 3 +[18] Muhammad Usman Ghani and W Clem Karl. Deep learning-based sinogram completion for low-dose ct. In IVMSP, pages 1-5. IEEE, 2018. 2 +[19] Design Guide. Cuda c programming guide. NVIDIA, July, 29:31, 2013. 6 +[20] Emma J Harris, Naomi R Miller, Jeffrey C Bamber, J Richard N Symonds-Tayler, and Philip M Evans. Speckle tracking in a phantom and feature-based tracking in liver in the presence of respiratory motion using 4d ultrasound. Phys. Med. Biol., 55(12):3363, 2010. 4 +[21] Johannes Hofmanninger, Forian Prayer, Jeanny Pan, Sebastian Röhrich, Helmut Prosch, and Georg Langs. Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem. Eur. Radiol. Exp., 4:1-13, 2020. 8 +[22] Dianlin Hu, Yikun Zhang, Jin Liu, Yi Zhang, Jean Louis Coattrieux, and Yang Chen. Prior: Prior-regularized iterative optimization reconstruction for 4d cbct. IEEE J. Biomed. Health Inform., 26(11):5551-5562, 2022. 2 +[23] Geoffrey D Hugo, Elisabeth Weiss, William C Sleeman, Salim Balik, Paul J Keall, Jun Lu, and Jeffrey F Williamson. Data from 4d lung imaging of nsclc patients. 2016. 6 +[24] Zhuoran Jiang, Yingxuan Chen, Yawei Zhang, Yun Ge, Fang-Fang Yin, and Lei Ren. Augmentation of cbct reconstructed from under-sampled projections using deep learning. IEEE Trans. Med. Imag., 38(11):2705-2715, 2019. 2 +[25] Zhuoran Jiang, Zeyu Zhang, Yushi Chang, Yun Ge, Fang-Fang Yin, and Lei Ren. Enhancement of 4-d cone-beam computed tomography (4d-cbct) using a dual-encoder convolutional neural network (decnn). IEEE Trans. Radiat. Plasma Med. Sci., 6(2):222-230, 2021. 2 +[26] Kyong Hwan Jin, Michael T McCann, Emmanuel Froustey, and Michael Unser. Deep convolutional neural network for + +inverse problems in imaging. IEEE Trans. Image Process., 26(9):4509-4522, 2017. 2 +[27] Avinash C Kak and Malcolm Slaney. Principles of computerized tomographic imaging. SIAM, 2001. 3 +[28] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 6 +[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[30] Anish Lahiri, Gabriel Maliakal, Marc L Klasky, Jeffrey A Fessler, and Saiprasad Ravishankar. Sparse-view cone beam ct reconstruction using data-consistent supervised and adversarial learning from scarce training data. IEEE Trans. Comput. Imaging, 9:13-28, 2023. 2 +[31] Suhyeon Lee, Hyungjin Chung, Minyoung Park, Jonghyuk Park, Wi-Sun Ryu, and Jong Chul Ye. Improving 3d imaging with pre-trained perpendicular 2d diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10710-10720, 2023. 2 +[32] T Li, Eduard Schreibmann, Y Yang, and L Xing. Motion correction for improved target localization with on-board cone-beam computed tomography. Phys. Med. Biol., 51(2):253, 2005. 2 +[33] Yiqun Lin, Zhongjin Luo, Wei Zhao, and Xiaomeng Li. Learning deep intensity field for extremely sparse-view cbct reconstruction. In MICCAI, pages 13-23. Springer, 2023. 2 +[34] Stephen H Manglos, George M Gagne, Andrzej Krol, F Deaver Thomas, and Rammohan Narayanaswamy. Transmission maximum-likelihood reconstruction with ordered subsets for cone beam ct. Phys. Med. Biol., 40(7):1225, 1995. 2 +[35] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splattering slam. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18039-18048, 2024. 3 +[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Commun. ACM., 65(1):99-106, 2021. 2, 6 +[37] Hiroshi Onishi, Hiroki Shirato, Yasushi Nagata, Masahiro Hiraoka, Masaharu Fujino, Kotaro Gomi, Katsuyuki Karasawa, Kazushige Hayakawa, Yuzuru Niibe, Yoshihiro Takai, et al. Stereotactic body radiotherapy (sbrt) for operable stage i non-small-cell lung cancer: can sbrt be comparable to surgery? Int. J. Radiat. Oncol. Biol. Phys., 81(5):1352-1358, 2011. 1 +[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Adv. Neural Inform. Process. Syst., 32, 2019. 6 +[39] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3 +[40] Simon Rit, Jasper Nijkamp, Marcel van Herk, and Jan-Jakob Sonke. Comparative study of respiratory motion correction techniques in cone-beam computed tomography. *Radiat. Oncol.*, 100(3):356-359, 2011. 2 + +[41] Thomas Rodet, Frédéric Noo, and Michel Defrise. The cone-beam algorithm of feldkamp, davis, and kress preserves oblique line integrals. Med. Phys., 31(7):1972-1975, 2004. 2, 6 +[42] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Phys. D: Nonlinear Phenom., 60(1-4):259-268, 1992. 5 +[43] Ken Sauer and Charles Bouman. A local update strategy for iterative reconstruction from projections. IEEE Trans. Signal Process., 41(2):534-548, 1993. 2 +[44] Chun-Chien Shieh, Yesenia Gonzalez, Bin Li, Xun Jia, Simon Rit, Cyril Mory, Matthew Riblett, Geoffrey Hugo, Yawei Zhang, Zhuoran Jiang, et al. Spare: Sparse-view reconstruction challenge for 4d cone-beam ct from a 1-min scan. Med. Phys., 46(9):3799-3811, 2019. 6 +[45] Emil Y Sidky and Xiaochuan Pan. Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization. Phys. Med. Biol., 53(17):4777, 2008. 2 +[46] T Solberg, J Wang, W Mao, X Zhang, and L Xing. Enhancement of 4d cone-beam computed tomography through constraint optimization. In ICCR, 2010. 2 +[47] Jiayu Song, Qing H Liu, G Allan Johnson, and Cristian T Badea. Sparseness prior based iterative image reconstruction for retrospectively gated cardiac micro-ct. Med. Phys., 34 (11):4476-4483, 2007. 2 +[48] Jan-Jakob Sonke, Lambert Zijp, Peter Remeijer, and Marcel Van Herk. Respiratory correlated cone beam ct. Med. Phys., 32(4):1176-1186, 2005. 1 +[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 5 +[50] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In IEEE Conf. Comput. Vis. Pattern Recog., pages 20310-20320, 2024. 3, 4, 5 +[51] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. In Eur. Conf. Comput. Vis., pages 1-20. Springer, 2024. 3 +[52] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19595-19604, 2024. 3 +[53] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3 +[54] Lifeng Yu, Yu Zou, Emil Y Sidky, Charles A Pelizzari, Peter Munro, and Xiaochuan Pan. Region of interest reconstruction from truncated data in circular cone-beam ct. IEEE transactions on medical imaging, 25(7):869-881, 2006. 2 +[55] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19447-19456, 2024. 3 + +[56] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 3 +[57] Guangming Zang, Ramzi Idoughi, Rui Li, Peter Wonka, and Wolfgang Heidrich. Intratomo: self-supervised learning-based tomography via sinogram synthesis and prediction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1960-1970, 2021. 6 +[58] Ruyi Zha, Yanhao Zhang, and Hongdong Li. Naf: neural attenuation fields for sparse-view cbct reconstruction. In MIC-CAI, pages 442–452. Springer, 2022. 2, 6 +[59] Ruyi Zha, Tao Jun Lin, Yuanhao Cai, Jiwen Cao, Yanhao Zhang, and Hongdong Li. R2-gaussian: Rectifying radiative gaussian splatting for tomographic reconstruction. arXiv preprint arXiv:2405.20693, 2024. 2, 3, 5, 6 +[60] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. In Eur. Conf. Comput. Vis., pages 341-359. Springer, 2024. 3 +[61] Shaohua Zhi, Marc Kachelrieß, and Xuanqin Mou. High-quality initial image-guided 4d cbct reconstruction. Med. Phys., 47(5):2099-2115, 2020. 2 +[62] Shaohua Zhi, Marc Kachelrieß, Fei Pan, and Xuanqin Mou. Cyn-net: A convolutional neural network specialized for 4d cbct images refinement. IEEE Trans. Med. Imag., 40(11): 3054-3064, 2021. 2 \ No newline at end of file diff --git a/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/images.zip b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2403456784967f2f17fd35f18133e3c6d200f4fb --- /dev/null +++ b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09952085d2585997295e4072cf398710d3113334afadb4a9d4e5192c980b9cb6 +size 716687 diff --git a/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/layout.json b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5938c59ca8b761478ddd42a6398f14e463ffc4c0 --- /dev/null +++ b/x2gaussian4dradiativegaussiansplattingforcontinuoustimetomographicreconstruction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5e25d633c75c9594955c1f8659d9f193123c233c38495592c22860d9149619b +size 458957 diff --git a/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/73b0573d-4884-4d06-9b66-e9a473055724_content_list.json b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/73b0573d-4884-4d06-9b66-e9a473055724_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b9031ec24e1792a9d22fd395cc859e188cf354c8 --- /dev/null +++ b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/73b0573d-4884-4d06-9b66-e9a473055724_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11cfe7a501975f6bfe6734b38bf6747b4836022d19f8b1acab9e81efd872f883 +size 92993 diff --git a/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/73b0573d-4884-4d06-9b66-e9a473055724_model.json b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/73b0573d-4884-4d06-9b66-e9a473055724_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3d422d2f4375853e4ccbbe27eac1dc0a7da8b61a --- /dev/null +++ b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/73b0573d-4884-4d06-9b66-e9a473055724_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7591588014695f483f8052af16d17876fa06f6c21bc7f4adfbf85388ff6ff929 +size 121084 diff --git a/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/73b0573d-4884-4d06-9b66-e9a473055724_origin.pdf b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/73b0573d-4884-4d06-9b66-e9a473055724_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..43e8dc115c623036869af5e770eacfeeccdbe588 --- /dev/null +++ b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/73b0573d-4884-4d06-9b66-e9a473055724_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64db9cfad27adc27a1ffc56f546bafa4a31d622460abb1571bfd3c3131d97965 +size 8942762 diff --git a/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/full.md b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dd5a7217a1a457ef564408d8fa8709c93172e1a6 --- /dev/null +++ b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/full.md @@ -0,0 +1,369 @@ +# X2I: Seamless Integration of Multimodal Understanding into Diffusion Transformer via Attention Distillation + +Jian Ma $\boxtimes$ OPPO AI Center majian2@oppo.com + +Qirong Peng* OPPO AI Center pengjirong@oppo.com + +Xu Guo† Tsinghua University guo-x24@mails.tsinghua.edu.cn + +Chen Chen $\boxtimes$ OPPO AI Center chenchen4@oppo.com + +Haonan Lu OPPO AI Center n luhaonan@oppo.com + +Zhenyu Yang OPPO AI Center yangzhenyu@oppo.com + +# Abstract + +Text-to-image (T2I) models are well known for their ability to produce highly realistic images, while multimodal large language models (MLLMs) are renowned for their proficiency in understanding and integrating multiple modalities. However, currently there is no straightforward and efficient framework to transfer the multimodal comprehension abilities of MLLMs to T2I models to enable them to understand multimodal inputs. In this paper, we propose the X2I framework, which endows Diffusion Transformer (DiT) models with the capability to comprehend various modalities, including multilingual text, screenshot documents, images, videos, and audio. X2I is trained on a 100K English corpus in 160 GPU hours. Building on the DiT teacher model, we adopt an innovative distillation method to extract the inference capabilities of the teacher model and design a lightweight AlignNet structure to serve as an intermediate bridge. Compared to the teacher model, X2I shows a decrease in performance degradation of less than $1\%$ while gaining various multimodal understanding abilities, including multilingual to image, image to image, image-text to image, video to image, audio to image, and utilizing creative fusion to enhance imagery. Furthermore, it is applicable for LoRA training in the context of image-text to image generation, filling a void in the industry in this area. We further design a simple LightControl to enhance the fidelity of instructional image editing. Finally, extensive experiments demonstrate the effectiveness, efficiency, multifunctionality, and transferability of our X2I. The open-source code and checkpoints for X2I can be found at the following link: https://github.com/OPPO-Mente-Lab/X2I. + +# 1. Introduction + +The recently open-sourced T2I models[50, 54, 61, 62, 66], such as Flux.1[32], have ushered in a new era of AI art due to their ability to generate realistic, photograde images that are both interesting and creative. The framework of T2I models has evolved from early GAN-based models[16, 27], auto-regressive models[60, 85], and UNet-based diffusion models[54] to DiT models[8, 17, 71]. To enhance both controllability and practicality in visual generation, numerous implicit[41, 46, 48] and explicit[34, 47, 84, 88] instruction-based editing models leveraging reference images have been developed quickly. Additionally, several unified image instruction editing models[22, 68, 86] have appeared. They all share a critical characteristic: the requirement of collecting comprehensive instruction-based editing datasets and committing to costly training resources. + +The text encoder in T2I is critical for generating semantically relevant images, as seen from early models such as CLIP[57] to later models such as T5[58] and advanced LLMs[13, 72]. The main goal is to improve the semantic understanding of text without considering the impact of other modalities on visual output. In contrast, VLMs, including early approaches like Flamingo[2] and BLIP2[35], along with newer models like LlamaVision[12], QwenVL[3], and InternVL[10], have effectively integrated pre-trained visual encoders with LLMs for enhanced text and visual comprehension. Is there a more streamlined way to transfer VLM or MLLM capabilities to DiT to enable diverse modal inputs? The industry is primarily tackling this challenge through two strategies. One method, similar to PEA-Diffusion[45], utilizes feature distillation by aligning multilingual LLM encoders to the original dimensions using MLP. However, it requires substantial resources of 1600 GPU hours and large-scale image-text datasets. This distillation relies on the U-Net[63] block's + +![](images/0cce55654716856506be5bd5fe659f5e2a3b6ddef14874826eeb8089b277f1dc.jpg) +multilingual to image(t) image variants(i) + +![](images/c129c7583671837002e7efae56eee2a85705f8c7cecd605877e3933d2bf61112.jpg) + +![](images/ba71ebe9ce5febe155abf4a59eb27489eb99dd9b5f55dd8924486175958d7517.jpg) +The is wearing a and riding a on the , the is also wearing a token mixing $(\mathrm{i} + \mathrm{i} + \mathrm{i} + \mathrm{i} + \mathrm{i})$ + +![](images/a77f28784f4d74bc9fa7484b031dbba17b3f806933d24ba9681859a4353b28c7.jpg) +image:Sunset Background text:Wearing scarf and hat video:Sora-generated1-minute video creativefusion(i+ii+V) +image: Sanxingdui artifact text: Science fiction novel parlay audio: Mechanical operation creative fusion(1+1) + +![](images/caadf8e3e400e87df48501ffca8fc8786bd324ad5e04c9fa9d1b66cf40b71069.jpg) + +![](images/cd39e26b0c7a2f00a2c4405f0a4177724e16a2a2dca30005aac6018d3d7c1b91.jpg) +image: Ink wash painting text: Artistic conception poetry audio: Moonlight Sonata for Piano creative fusion(i+tt+a) + +![](images/e6c2de20bb57effd9f3d81ea84df7448d7762385edc53cecf7a145ac5f7f53e9.jpg) + +![](images/e0d81bee7ec58260ca338d879a324cbe1505caf8909270c15dc859d65ff79d67.jpg) +audio1: Morning insects +chirping and birds singing +audio2: Train whining +audio to image +music1: Mariage d'Amou music2: Turkish March music to image +Figure 1. The primary applications of X2I include multilingual text-to-image, image-to-image, image-text-to-image, video-to-image, audio-to-image, and various multimodal combinations for image generation. The red boxes embedded in the images indicate the input images, text, video, or audio. The red box text below the images represents the abbreviated application names, with the initial letters of different modalities in parentheses. The descriptions below some of the images provide a brief overview of the input audio or video. + +![](images/707c024ee5f635562d6368219978136fa3ddca1dbdc69db737ddb33ff8e0aeee.jpg) + +![](images/83e1adb8b64c93a180d9301100b5736f278e430d2345643794eb91eccf6f41b1.jpg) +video1: Blurred video with cards and flowers +video2: A person wearing red clothes is skiing +video to image + +![](images/3df05041dbf0ece2dda12cf10f918d9b1150404f7c1d838dc3ff59fea5cf2559.jpg) + +![](images/2e88cd1f294b9ae70497f4737c82700434ef9d6d1b613acbb5d9627ef624c03d.jpg) + +outputs and restricts support for multimodal input. Another method is the VLMs-based replacement of T2I[43], but it is costly, requiring extensive training or fine-tuning of T2I or MLLMs, necessitating substantial data quality and quantity[22, 68]. Alternatively, a solution akin to GlueGen[55] could be employed to align the features of a single-modal or multimodal encoder with the latent space of existing T2I models. However, aligning solely through the encoder still results in suboptimal overall performance. + +In this work, we introduce an efficient framework, X2I, along with a simple LightControl module to facilitate the enhancement of image transition from weak to strong fidelity. Utilizing a DiT-based image generation model as a teacher, the student model mirrors the entire architecture of the teacher model but substitutes its text encoder with MLLMs. Both models are capable of processing the same linguistic input. To optimize our training data without gathering images, we opt to distill the teacher model's inference abilities. In T2I models, this necessitates performing distillation during the reverse denoising phase. Furthermore, we efficiently transfer the visual generation capabilities of the teacher model to the student model by ingeniously design + +ing the AlignNet structure and distilling the attention. As a result, after alignment, the student model acquires the multimodal comprehension abilities of MLLMs during image generation. Moreover, we formulate a template input format intended for the student model to effectively interpret instructions from different modalities. Since MLLMs generally have proficiency in understanding the global semantic features of visual content, we develop LightControl to derive structured data from reference images. This enhances visual fidelity and facilitates precise and controllable image editing. + +After training the X2I, we conduct various experiments to compare the capabilities of T2I and instructional-based image editing. Additionally, we conduct subjective experiments for image generation with various multimodal mixed inputs, including multilingual, image (I2I), imagetext (IT2I), video (V2I), and audio (A2I). The results demonstrate that X2I acquires robust multimodal understanding capabilities with less than $1\%$ performance degradation in the T2I generation. Furthermore, for more personalized tasks like I2I or IT2I, we conduct experiments involving the training of traditional LoRA[23] to fulfill in + +dustry requirements, addressing a training gap in LoRA for these purposes. Moreover, to enhance the adaptability of the framework, only the AlignNet parameters within the model are updated, enabling the aligned student model to accommodate a variety of downstream tasks, such as LoRA, ControlNet[70], IP-Adapter[84], different fine-tuning models, and compression models. Notably, some objective metrics for tasks involving ControlNet and IP-Adapter show improvements over the teacher model. This improvement is due to X2I's robust support for image conditional information input, which facilitates further extraction of image features. We also perform experimental analyses on training efficiency, attaining $98.2\%$ of the teacher model's performance with 48K training samples. + +In summary, our key contributions are as follows. + +- We introduce a novel method for transferring the understanding capabilities of MLLMs to DiT models, achieving fast fitting speeds with only a small amount of monolingual text data through the carefully designed AlignNet and Attention Distillation. +- We design LightControl to enable the generation of image instruction editing from weak fidelity to high fidelity. +- The architecture itself possesses strong modularity, leading to performance improvements in certain IT2I tasks. It also has the ability to train LoRA in IT2I tasks. +- X2I is the first image generation model that supports audio understanding in addition to text and visual understanding. + +# 2. Related Work + +Diffusion Models and Text Encoder. Diffusion models have shown immense potential in the field of T2I. Earlier works[17, 62, 65, 82] use U-Net as the backbone of the diffusion model, which, to some extent, impacts the scalability of the models. Inspired by DiT[52], many works[8, 17, 32, 36] replace the U-Net backbone with transformers, greatly improving the quality of the generated images. In order to improve the understanding of the diffusion model's prompts, Stable Diffusion (SD) used CLIP to encode text information, while Flux.1 and eDiff-I[5] utilize information encoded by both T5 and CLIP to guide image generation. Recently, works like LuminaT2X[19], Colors[71], and LiDiT[44] directly use pre-trained decoder-only LLMs as the prompt encoder, which enhances the prompt-following ability in image generation. However, these methods can only encode the text prompt, lacking the ability to understand information from images, videos, audio, etc. + +MLLMs as Text Encoders for T2I. VLMs[3, 10, 35, 39, 42, 74, 75, 78] have made significant progress in visual language understanding tasks. Furthermore, models such as AnyGPT[87], NextGPT[76], X-LLM[7], and MiniCPMo[24] have further developed the capabilities in audio understanding. In articles that fuse multimodal models with + +image generation models, MUMU[6] and ML-MGIE[18] inject multimodal information into the image generation model using VLM but also change the weights of the generation model during training, lacking plug-and-play capabilities. KOSMOS-G[51] and Easyref[91] achieve alignment while preserving the weights of the original SD[16] and SDXL[54] models. KOSMOS-G trains a VLM and AlignerNet from scratch, whereas Easyref trains the features and projector of the final layer output of the VLM. These methods consume significant computational resources and are limited to a single task, failing to fully leverage the zero-shot capabilities of VLM. Furthermore, the aforementioned methods focus on alignment with U-Net-based models like SDXL. Currently, there is a lack of an efficient alignment method for DiT models, such as Flux.1 and SD3. Work in the related field, such as HunyuanVideo[30], requires an additional CLIP text encoder for alignment. Qwen2VLFlux[43] not only retains the structure of T5 but also alters the weights of Flux.1 itself, significantly reducing the model's plug-and-play and zero-shot capabilities. + +Distillation of DiT. Knowledge distillation (KD) has been widely applied in diffusion models[28, 89, 90]. SSKD[73] involves the direct distillation of critical attention mechanism information from teacher to student, which can considerably reduce the performance gap between both. EFFICIENT-VDIT[14] utilizes a multi-step consistency distillation technique to accelerate DiT sampling. PEA-Diffusion integrates multilingual CLIP with SDXL by employing the L2-norm distance, thus allowing the T2I model to support multilingual capabilities. However, this method is specifically tailored for the U-Net backbone and demands significant computational resources. + +# 3. Methodology + +# 3.1. Preliminary + +The diffusion process is performed in the latent space, where a transformer denoiser $\epsilon_{\theta}$ is employed to predict noise $\epsilon$ with the current timestep $t$ , noisy latent $x$ and generation conditions, $c$ and $c_{p}$ , where $c_{p} = \tau_{\theta}(y)$ is produced by encoding the text prompts $y$ with a pre-trained CLIP text encoder $\tau_{\theta}$ , and $c = \varepsilon_{\theta}(y)$ is produced by encoding the text prompts $y$ with a pre-trained T5 text encoder $\varepsilon_{\theta}$ . To enhance the capability of feature fusion, MM-DiT[17] uses adaptive layer normalization (AdaLN) to improve the adaptability of the model when processing different input conditions. There are two fully connected layers plus a SiLU activation function forming the feature extraction and normalization networks $\zeta_{\theta}$ and $\delta_{\theta}$ , which respectively extract features for $x$ and $c$ . These features are used as the normalization and scaling parameters for the multi-head self + +attention, as shown below: + +$$ +\begin{array}{l} x, \alpha_ {1}, \beta_ {2}, \gamma_ {2}, \alpha_ {2} = \zeta_ {\theta} (x, t, c _ {p}) \\ c, \alpha_ {1 c}, \beta_ {2 c}, \gamma_ {2 c}, \alpha_ {2 c} = \delta_ {\theta} (c, t, c _ {p}), \\ \end{array} +$$ + +then, concatenate $x$ and $c$ directly with $\varphi(\cdot)$ and use the self-attention to achieve feature interaction and fusion. The output is then split according to the corresponding indices, as shown below: + +$$ +x _ {A}, c _ {A} = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {0}}}\right) \cdot V, \tag {2} +$$ + +where $Q = W_{Q} \cdot \varphi(x, c)$ , $K = W_{K} \cdot \varphi(x, c)$ , $V = W_{V} \cdot \varphi(x, c)$ , $x_{A}, c_{A}$ represents the feature value output on the image and text sides obtained after computing attention and truncating according to the respective index values. $W_{Q}, W_{K}, W_{V}$ are learnable projection matrices. + +Both the text and image sides undergo the following process, and only the image-side processing is described below for the convenience of presentation. After the residual module, the regressed scaling parameter $\alpha_{1}$ is used to scale the weights of the image-side input, + +$$ +x _ {L N} = \mu \left(x + \alpha_ {1} * x _ {A}\right), \tag {3} +$$ + +where $\mu$ represents Layer Normalization (LN). Then, normalization is performed utilizing adaptive parameters $\beta_{2},\gamma_{2}$ + +$$ +x _ {F F} = \nu \left(x _ {L N} * \left(1 + \gamma_ {2}\right) + \beta_ {2}\right), \tag {4} +$$ + +where $\nu$ represents FeedForward (FF). Finally, the output of the current block is obtained by multiplying the regression scaling coefficients and the residuals + +$$ +x _ {O} = \left(x + \alpha_ {1} * x _ {A}\right) + x _ {F F} * \alpha_ {2}. \tag {5} +$$ + +# 3.2. Model Overview + +The framework consists of two parts, as shown in Fig. 2. The initial part involves X2I training exclusively on a text corpus, with AlignNet as the trainable parameters. Distribution alignment between student and teacher is attained via attention distillation. The subsequent part incorporates the LightControl module, forming X2I Enhanced Training, thereby improving the accuracy of reference images in image instruction editing tasks. + +# 3.3. AlignNet with MLLM's Hidden State + +Many approaches use MLLM as text encoders for T2I or feature extractors for other modules, typically utilizing only the last[51] or penultimate[79] hidden states. However, according to Oscar Skean's analysis[69], the quality of hidden states in the intermediate layers of transformer-based LLMs is often superior to that of the final and initial layers. Recognizing that employing a static intermediate layer's + +![](images/e842b54d288ac2787b4f07f2d6744407813abcb1fd3a27c007955ef4405f055b.jpg) +Figure 2. Overview diagram of the X2I. In the deep red box on the left represents the X2I training, where only the text data needs to be input for training, with trainable parameters limited to AlignNet and the distillation location situated within the attention output of MM-DiT. The light green box at the bottom right represents X2I enhanced training after the X2I training, AlignNet is fixed, and the trainable parameters limited to LightControl. + +representation might be suboptimal for adapting to diverse MLLM models, we propose utilizing representations from every layer of MLLMs as input to AlignNet. + +To integrate hidden states from all layers of MLLMs, we design a simple CNN to capture spatial and channel-wise relationships between layers. Convolutional operations allow the extraction of more complex patterns and improve the rationality of weight allocation, enabling spatial adaptability and establishing inter-layer dependencies. By adjusting the kernel size, we can also create cross-layer receptive field correlations, achieving better multi-scale fusion. + +In MLLM, define a feature extractor that projects text $y$ into a middle feature $H \in \mathbb{H}^{b \times m \times s \times z}$ , where $m$ represents all layers of MLLM's hidden states, each layer has a hidden size of $z$ and a sequence length of $s$ . Next, we define a CNN mapping network $\Psi$ to enable self-learning of weights across all layers' hidden states, resulting in the aggregation of hidden states from $m$ layers into 1 layer of weighted hidden states as the following formula, + +$$ +y _ {p}, y = \Phi (\Psi (m, 1, k, p) (H)), \tag {6} +$$ + +where the kernel size is $k$ , and the padding is denoted as $p$ . $\Phi$ is a simple MLP network that maps the feature learned by the $\Psi$ with weight information to the corresponding dimension of $c$ and $c_{p}$ , thereby obtaining new features $y$ and $y_{p}$ . + +# 3.4. Attention Distillation + +KD can be divided into two main categories: logits distillation and feature distillation. In MM-DiT, intermediate layer features encompass low-level information such as pixel values, shape, and gradient information, as well as high-level information such as color, theme, and lighting details. In this paper, the student model is conditioned to inject multimodal feature information. Knowledge transfer occurs layer by layer through the concatenation of the information injected at each level. Therefore, for deeply encoded information that logits cannot express, we opt for intermediate feature distillation between layers to achieve alignment. + +Further, each MM-DiT consists of AdaLN as Eq. (1), self-attention layer as Eq. (2), LN as Eq. (3), and FF layers as Eq. (4). In addition to regressing the $\gamma$ and $\beta$ after the LN, AdaLN also regresses an $\alpha$ before the end of each residual module. Since biases here directly affect the distribution of features, we choose to perform feature distillation in the position of the attention output $x_{A}, c_{A}$ in order to minimize unnecessary regression and maintain consistency in the distribution affected by normalization parameters. This provides an important conclusion in exploring direct feature distillation in the MM-DiT feature layers as well as the single-DiT block of Flux.1. + +Regarding the choice of attention distillation loss, we adopt reverse KL (RKL) from MiniLLM[20]. KL performs well in traditional tasks due to the smaller output space and fewer modes in conventional classification tasks. However, for LLMs, the output space is more complex with more modalities; RKL can prevent the student model from overfitting to the low-probability outputs of the teacher model, and the same applies to the MM-DiT structure. For the Flux.1 structure, the target loss for $l$ layers of MM-DiT is defined as follows: + +$$ +\mathcal {L} (\theta) = \mathrm {K L} [ \eta_ {l} | | v _ {l} ] = \left[ - \mathbb {E} \log \frac {v _ {l} (A _ {t} | \epsilon , c)}{\eta_ {l} (A _ {s} | \epsilon , y)} \right], \tag {7} +$$ + +where $\eta, \upsilon$ specifically represent the attention output mappings in each layer of the student and teacher models in MM-DiT, respectively. $A_{t}$ denotes the attention output $x_{A}, c_{A}$ mentioned above; similarly, $A_{s}$ corresponds to the output of the attention output in the student model. $\epsilon$ represents noise input because we perform distillation during the inference process, which is the reverse denoising process. + +# 3.5. Training Stages + +For most pre-trained MLLMs, features from different modalities are mapped to a common feature space, where similar information from different modalities is brought closer together. This implies that different modality features share a common semantic space. Based on this consensus, we opt to perform alignment by inputting pure text modality while simultaneously aligning the inferential capabilities + +of the teacher model, that is, aligning capabilities during the process of pure noise denoising. The Appendix Sec. 8 Fig. 7 provides a more detailed analysis of the changes in the distribution of different modality spaces before and after alignment, which effectively explains how X2I gained the multimodal understanding capabilities of MLLM simply through textual semantic alignment. + +Since MLLMs tend to understand images more from a global semantic perspective, they often lack the crucial fine-grained understanding needed for image generation. To address this issue, we design a simple LightControl in parallel with the MM-DiT structure named X2I enhanced training. This pathway is a convolutional module composed of ResNet blocks, similar to those mentioned in ControlNeXt[53]. We initialize a ResNet network with 19 layers, mirroring those used in MM-DiT. We define the ResNet module $\Gamma$ to extract the conditional controls; similar to ControlNet, the output of each ResNet module is: + +$$ +y _ {c} ^ {l} = \Gamma^ {l} (c _ {p}, c _ {i}; \Theta), \tag {8} +$$ + +where $y_{c}^{l}$ is the feature output of $l$ the layer of the ResNet, $c_{p}$ is the corresponding conditional text input, $c_{i}$ is the conditional reference image input, and $\Theta$ is the corresponding trainable parameter. The final output here is also similar to ControlNet, where $y_{c}$ and $x_{O}$ are added feature-wise. + +# 4. Experiments + +# 4.1. Datasets and Implementation + +We randomly select 100K image-text pairs from the Laion2B [67], and utilize InternVL to generate new captions from the images for training X2I. Furthermore, based on the T2I-Adapter's[49] sketch-guided model and SDXL, we construct a training dataset of 400k images containing five styles: Monet, Baroque, Cartoon, Pixar, and Van Gogh. For the text encoder in the student model, we utilize open-source models from the InternVL series, Qwen2.5-VL-7B[4], and MiniCPM-o, for the DiT structure, we use Flux.1. To optimize performance, inference ability is distilled by focusing on the denoising process between steps 0 and 1. The X2I is trained on 8 A100 GPUs for 24K steps. X2I training requires over 80GB of GPU memory. We optimize this by separating training and inference, and overlapping communication, achieving a global batch size of 6. Acceleration techniques are detailed in the appendix Sec. 9. + +# 4.2. Tasks and Benchmarks + +T2I and Multilingual T2I. Evaluation Datasets. We randomly select five sub-evaluation datasets from T2I-CompBench++[25], each containing 1k images. These include: complex compositions bench, non-spatial bench, attributes binding bench related to color and texture categories, numeracy bench. GenAI-Bench[33], EvalMuse[21], + +and Multilingual-General[45]. Metrics includes four main categories: image-text matching score indicators like ClipScore (CS) with ViT-bigG[11] and FGA-BLIP2 (FB)[21] Score. Human-consistency feedback score indicators include PickScore (PS)[29], HPSv2[77], and ImageReward (IR)[81]. Fine-grained score indicators have BLIP-VQA (BV)[25] and UniDet[25]. Model scoring indicators reflecting complex textual semantic alignment such as GPT-4o and VQAScore (VS)[38], with GPT-4o instructions specified by the T2I-CompBench++ for complex compositions evaluation. Furthermore, the IR score ranges from $98.2\%$ probability within [-2, 2], and the FB score ranges from [1, 5]. We use eight normalized metrics to determine the relative gap compared to the teacher model, defining a Performance Ratio (PR) metric as a performance percentage indicator. Baseline model for universal T2I capability comparison is the Flux.1 teacher model, used to assess reduction in student model's basic T2I capabilities. Qwen2VL-Flux has similar functions, but MLLM does not support text input and relies solely on the source model's T5. Thus, it is excluded from comparison. For the comparison of multilingual T2I capabilities, we evaluate both PEA-Diffusion and Sana[80]. + +IT2I and I2I. Evaluation Datasets. We chose DreamBench[64] and we generate six images for each text prompt[34]. Metrics. We report the average DINO, CLIP-I, and CLIP-T scores based on all pairs of real and generated images. Baseline model. The comparison model includes methods based on different frameworks such as SD, Imagen, Phi-3, SD3, and Flux.1. In the appendix Sec. 12, we also compare the differences in Qwen2VL-Flux's capabilities in image understanding and image editing based on objective metrics. + +Image stylization. Evaluation Datasets consists of 50 images from MSCOCO[37] and 50 photorealistic images generated by DiffArtist[26]. Baseline model is also selected DiffArtist. + +ControlNet. Evaluation Datasets is sourced from MultiGen-20M[56] and includes four tasks: A model based on canny edge detection, evaluated using the F1 score. Models based on hed, evaluated using Structure Similarity Index Measure (SSIM). A model based on depth information is evaluated using RMSE. + +# 4.3. Quantitative Results + +Universal T2I Generation. As shown in Tab. 1, the one that performs closest to the teacher model is the complex compositions evaluation set. Additionally, on some subevaluation sets and certain metrics, the performance of X2I even surpass that of the Flux.1. The overall average PR across the five evaluation sets is $99.21\%$ of that of Flux.1, with such subtle differences being virtually indistinguishable subjectively. Appendix Sec. 10 Tab. 4 displays the re + +sults from the GenAI-Bench and EvalMuse evaluation sets, with performance scores also exceeding $99\%$ . This further demonstrates that X2I's results in T2I generation are nearly indistinguishable from those of the teacher model Flux.1. + +Multilingual T2I Generation Tab. 5 includes three multilingual evaluation datasets from Multilingual-General, as well as four additional translations in German, Portuguese, Spanish, and French. While the performance in English remains consistent with a one percentage point difference similar to before. For more details, please refer to Sec. 11. + +IT2I Generation Tab. 2 shows the performance of different methods on the DreamBench. Overall, X2I has a slight advantage in DINO scores. CLIP-I shows a two-point difference compared to KOSMOS-G and there is a considerable gap in CLIP-T compared to recent methods like OmniControl and IP-Adapter based on Flux.1, indicating that X2I still has room for improvement in its ability to faithfully follow the semantics of images. + +# 4.4. Qualitative Results + +X2I training achieves outstanding results through feeding pure text. Fig. 1 illustrates the effects of some of these functions. T2I possesses the capability to support over 29 natural languages, covering the major language families worldwide; I2I includes image variations and the fusion of concepts from multiple images. IT2I encompasses single image instruction editing, style transfer, image stylization, instruction editing for multiple images, and the result of text token mixing with multiple images. It integrates multiple visual concepts to create unique images, extracts specific concepts from a set of input images, and synthesizes these concepts into new images, allowing precise control over the visual details of the images and avoiding common vagueness and uncertainty in text descriptions. Examples include changing clothes in the combination of clothes and people, performing group photos in person combinations, and creative conceptual combinations with three or more images. + +A2I enables auditory information to be visualized, providing a way to "see sound". This aspect is not limited to regular audio information but also processing various audio types, including natural water flow sounds, animal calls, iconic IP character voices, and the emotions that different music pieces convey, translating these into corresponding visual images. V2I can inductively summarize single frames from videos as images and generate high-definition images from low-resolution videos. + +Additionally, more creative combinations can be achieved. In the middle of Fig. 1, the last three examples in the second row illustrate this: the first input example is a 1-minute video generated by Sora[40], a sunset background image along with the text instruction "wearing a hat and scarf", the generated image successfully maintains the original video's semantic information while adhering well + +
T2I++ CompBenchMethodsCSPSHPSv2IRFBGPT-4oVSBVUniDetPR
ComplexFlux.10.42790.22970.20550.83543.69900.91810.87990.5402-99.77%
X2I0.42340.22830.20960.81243.67230.91420.87790.5417-
Non-spatialFlux.10.44180.27510.24721.30133.59630.89620.90750.6762-99.12%
X2I0.43630.27060.24621.27863.54490.88210.89960.6571-
ColorFlux.10.47330.28720.24171.41794.22680.70600.90770.6642-99.72%
X2I0.46910.28810.24211.38724.18450.71130.90140.6643-
TextureFlux.10.45720.25890.24111.19944.06010.75880.88070.5281-99.18%
X2I0.45040.25700.24011.12874.00550.74470.88090.5172-
NumeracyFlux.10.47300.26640.24311.34593.48520.78600.78410.45280.611299.36%
X2I0.47130.26700.24851.27233.41400.76800.77980.44640.6206
+ +Table 1. Objective performance on the T2I-CompBench++ evaluation dataset. + +
MethodsBaseDINOCLIP-ICLIP-T
Textual Inversion[83]0.5690.7800.255
DreamBooth[64]0.6680.8030.305
Custom Diffusion[31]0.6430.7900.305
BLIP-Diffusion[34]SD[62]0.5940.7790.300
Subject-Diffusion[47]0.7110.7870.293
IP-Adapter[84]0.6670.8130.289
KOSMOS-G[51]0.6940.8470.287
SuTI[9]Imagen[65]0.7410.8190.304
OmniGen[51]Phi-3[1]-0.8010.315
UNIC-Adapter[15]SD3[17]0.8160.8410.306
IP-Adapter[84]0.7680.8030.322
OminiControl[70]Flux.1[32]0.7400.7680.329
X2I0.8170.8260.304
+ +Table 2. Performance of different methods on DreamBench. + +to both the text and image instructions. The second input combines an image of the Sanxingdui bronze sacred tree, a sci-fi novel text fragment, and a sound of mechanical operation, resulting in an output that blends all the semantic information, adding a mysterious touch. The third input is a Chinese ink painting, a classical poem, and Beethoven's Moonlight Sonata, resulting in a final image that perfectly integrates the three modalities into an evocative artwork. For more interesting effect displays, please refer to the appendix Sec. 17. + +X2I enhanced Training. The purpose of constructing LightControl is to enhance the fidelity of image instruction editing. Essentially, it is not limited to a specific downstream task. In this paper, relevant experimental validation is only conducted in the stylization domain. For more comparative results, please refer to the appendix Sec. 16. + +IT2I LoRA. For highly personalized IT2I image generation tasks tailored to the user, we can leverage the X2I to train a LoRA. In this paper, we train X2I using reference images in the style of abstract art line drawings. More comparative results, please refer to the appendix Sec. 15. + +Plug-and-Play. In this paper, we have validated the prominent variant models and related downstream models in the Flux.1 community. Specifically, we compare objective metrics for tasks involving ControlNet and IP-Adapter. Additionally, incorporating the original image as input in + +the X2I task resulted in improvements in both objective and subjective metrics. More results in the appendix Sec. 13. + +Reasoning ability and multi-turn dialogue generation ability. MLLMs themselves possess certain reasoning abilities and multi-turn dialogue capabilities. X2I also inherits some of these related abilities. More results in the appendix Sec. 6. + +# 4.5. X2I Training Convergence + +Benefitting from the design of AlignNet and attention distillation, the training convergence speed of X2I is remarkably fast. Here, we have calculated the performance metrics including CS, FB Score, PS, IR, and VS, and used the average of these five objective metrics as the Performance Ratio. Additionally, we have also calculated the variation of the SSIM metric. As shown in Fig. 3 with Qwen2.5-VL-7B, the convergence speed in the early stages of training exhibits an almost linear growth, achieving a Performance Ratio of $98.2\%$ with only 8K steps of training. Appendix Sec. 7 further demonstrates the experimental results of InternVL[10] with different capacity adaptations for X2I. + +![](images/3daa81b6a5174e3c64b16544ee405d1a893396d7b00d1cce11fae05400221723.jpg) +Figure 3. X2I training performance vs. training steps. + +# 4.6. Ablation Study + +In this paper, we design three types of ablation experiments: AlignNet structure with the number of feature layers extracted by MLLMs, the feature distillation position in the + +
ModelsCSPSIRFBSSIMPR
FLUX.10.45170.25781.38913.7789-100%
A10.43130.23391.11493.49460.508995.40%
A30.43480.24071.21983.5440.513196.62%
A3_ada0.44290.24471.22783.60550.513197.36%
A3_t5_ada0.43970.24521.27573.60160.503697.57%
A29_ada0.44630.24781.29543.67890.527198.40%
A29_t5_ada0.43610.24211.19883.57120.508496.73%
A29_CNN0.45080.25181.37993.67480.518999.11%
Block0.1483-0.0220-2.28001.08540.047545.65%
FF0.29950.1027-0.70192.23260.462369.58%
LN0.43920.24951.30043.65570.512898.16%
Oneside0.44700.24611.30303.63120.507798.13%
KL0.45280.25321.40393.69880.515299.26%
JS0.44900.25371.37953.69720.520299.26%
RKL0.45430.25771.40433.75680.520899.70%
+ +Table 3. Three ablation results of X2I. + +MM-DiT structure, and different target losses for feature distillation. The evaluation dataset uses the Multilingual-General. The evaluation metrics are CS, FB Score, PS and IR, and the SSIM metric for image variants. Additionally, the overall PR of these four normalized metrics. The MLLMs chosen is Qwen2.5-VL. + +AlignNet Structure Ablation. Tab. 3 upper portion show AlignNet structure ablation. "A1" indicates extracting only the last layer features of Qwen2.5-VL while the AlignNet structure consists of a simple MLP, similar to PEA-Diffusion, performing only simple dimensional mapping. The overall performance of the relevant metrics for T2I generation is $95.4\%$ of the teacher, showing a relatively large decline. "A3" extracts features from the first and the last two layers of Qwen2.5-VL, where the first layer features are static token-mapped vector features trained by MLLMs, containing the most original token information, but lacking contextual information. Therefore, combining with the last two layers features offers complementary advantages. By directly taking the average pooling of the features, the performance improvement is about one percentage point compared to "A1". "A3_ada" further learns adaptive weights for each layer of features, improving by $0.74\%$ . "A3_t5_ada" increases the nonlinearity complexity of AlignNet by adding a custom T5 module for feature mapping, but the improvement is minimal. "A29_ada" selects features from all 29 layers of Qwen2.5-VL, leading to about a one percentage point improvement over "A3_ada". However, "A29_t5_ada" shows some metric decline, suggesting that deep nonlinear complex mapping is not required for such modal alignment learning. Furthermore, the scalar parameter learning fails to capture the synergy between low-level edge features and high-level semantic features, as well as issues of spatial insensitivity and poor dynamic adaptability. Therefore, "A29_CNN" ultimately adopts a CNN to further enhance the non-linear learning of each layer's feature weights, resulting in optimal performance, reducing the + +performance loss compared to the teacher to less than one percentage point. + +Distillation Position Ablation. The middle four lines of Tab. 3 show distillation position ablation. "Block" refers to direct distillation of each layer's MM-DiT and Single-DiT output positions in Flux.1, analogous to PEA-Diffusion's method of distilling each U-Net block output in SDXL. The experimental result shows a suboptimal performance of $45.65\%$ . "FF" denotes distillation at the output position of each block's FeedForward layer, as shown in Eq. (4), results indicate a significant improvement compared to "Block" distillation. This is likely due to the influence of the regression scaling factor $\alpha_{2}$ on distribution alignment in Eq. (5). LN denotes distillation at the second layer norm output position of each block, i.e., the output of Eq. (3), approaching optimal results but still possibly affected by the regression scaling factor $\alpha_{2}$ . "Oneside" indicates distillation solely at the conditional side output after MM-DiT's self-attention, aiming to explore whether the interaction information's purest form can be obtained after self-attention and token length segmentation. Experimental results remain about one percentage point below the optimal results. + +Feature Alignment Loss Ablation. We additionally compare three other target loss functions: KL, JS, and RKL divergence. Both KL divergence and JS divergence consistently exhibited slightly better performance compared to the MSE target loss like "A29_CNN" before. However, the best performance was achieved with RKL, which outperformed KL divergence by $0.44\%$ percentage points. + +# 5. Conclusion and Limitations + +In this paper, we introduce a novel framework X2I for aligning MLLMs and the DiT visual generation model. This framework can be fully trained with minimal training resources and minimal training data. Our extensive experimental results demonstrate that X2I maintains its general performance while extending additional functionalities, including multilingual image generation, image instruction editing, conceptual fusion of multiple images, image generation from videos or audio, and further exploration of creative combinations from different modalities to unleash imagery with X2I. Additionally, X2I supports LoRA training for multiple modalities, enabling seamless integration with various controllable downstream tasks in the open-source community. Furthermore, we have developed LightControl to explore more precise and controllable possibilities for visual generation. + +The limitations of this paper lie in the somewhat inferior precision and control of X2I in the unified direction of image instruction editing, as well as the failure to fully explore the logical reasoning, few-shot learning, and chain-of-thought capabilities of MLLMs. In the future, we will continue to delve deeper into this direction. + +# References + +[1] Marah Abdin and et al Jyoti Aneja, Hany Awadalla. Phi-3 technical report: A highly capable language model locally on your phone, 2024. 7 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning, 2022. 1 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 1, 3 +[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report, 2025. 5 +[5] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers, 2023. 3 +[6] William Berman and Alexander Peysakhovich. Mumu: Bootstrapping multimodal image generation from text-to-image data, 2024. 3 +[7] Feilong Chen, Minglun Han, Haozhi Zhao, Qingyang Zhang, Jing Shi, Shuang Xu, and Bo Xu. X-llm: Bootstrapping advanced large language models by treating multi-modalities as foreign languages, 2023. 3 +[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 1, 3 +[9] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W. Cohen. Subject-driven text-to-image generation via apprenticeship learning, 2023. 7 +[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 3, 7 +[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scal + +ing laws for contrastive language-image learning. arXiv preprint arXiv:2212.07143, 2022.6 +[12] Xiangxiang Chu, Jianlin Su, Bo Zhang, and Chunhua Shen. Visionllama: A unified llama backbone for vision tasks, 2024. 1 +[13] DeepSeek-AI. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model, 2024. 1 +[14] Hangliang Ding, Dacheng Li, Runlong Su, Peiyuan Zhang, Zhijie Deng, Ion Stoica, and Hao Zhang. Efficient-vdit: Efficient video diffusion transformers with attention tile, 2025. 3 +[15] Lunhao Duan, Shanshan Zhao, Wenjun Yan, Yinglun Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Mingming Gong, and Gui-Song Xia. Unic-adapter: Unified image-instruction adapter with multi-modal transformer for image generation, 2024. 7 +[16] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2021. 1, 3 +[17] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 1, 3, 7 +[18] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models, 2024. 3 +[19] Peng Gao, Le Zhuo, Dongyang Liu, Ruoyi Du, Xu Luo, Longtian Qiu, Yuhang Zhang, Chen Lin, Rongjie Huang, Shijie Geng, Renrui Zhang, Junlin Xi, Wenqi Shao, Zhengkai Jiang, Tianshuo Yang, Weicai Ye, He Tong, Jingwen He, Yu Qiao, and Hongsheng Li. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers, 2024. 3 +[20] Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models, 2024. 5 +[21] Shuhao Han, Haotian Fan, Jiachen Fu, Liang Li, Tao Li, Junhui Cui, Yunqiu Wang, Yang Tai, Jingwei Sun, Chunle Guo, and Chongyi Li. Evalmuse-40k: A reliable and fine-grained benchmark with comprehensive human annotations for text-to-image generation model evaluation, 2024. 5, 6 +[22] Zhen Han, Zeyinzi Jiang, Yulin Pan, Jingfeng Zhang, Chaojie Mao, Chenwei Xie, Yu Liu, and Jingren Zhou. Ace: All-round creator and editor following instructions via diffusion transformer, 2024. 1, 2 +[23] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 2 +[24] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, Xinrong Zhang, Zheng Leng Thai, Kaihuo Zhang, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. Minicpm: Unveiling the potential of small language models with scalable training strategies, 2024. 3 + +[25] Kaiyi Huang, Chengqi Duan, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench++: An enhanced and comprehensive benchmark for compositional text-to-image generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2025. 5, 6 +[26] Ruixiang Jiang and Changwen Chen. Diffartist: Towards aesthetic-aligned diffusion model control for training-free text-driven stylization, 2024. 6 +[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 1 +[28] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion, 2024. 3 +[29] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation, 2023. 6 +[30] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyuan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. Hunyuanvideo: A systematic framework for large video generative models, 2025. 3 +[31] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion, 2023. 7 +[32] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.1,3,7 +[33] Baiqi Li, Zhiqiu Lin, Deepak Pathak, Jiayao Li, Yixin Fei, Kewen Wu, Tiffany Ling, Xide Xia, Pengchuan Zhang, Graham Neubig, and Deva Ramanan. Genai-bench: Evaluating and improving compositional text-to-visual generation, 2024. 5 +[34] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 1, 6, 7 +[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 1, 3 +[36] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang, Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Ji-hong Zhang, Chao Zhang, Meng Chen, Jie Liu, Zheng Fang, + +Weiyan Wang, Jinbao Xue, Yangyu Tao, Jianchen Zhu, Kai Liu, Sihuan Lin, Yifu Sun, Yun Li, Dongdong Wang, Mingtao Chen, Zhichao Hu, Xiao Xiao, Yan Chen, Yuhong Liu, Wei Liu, Di Wang, Yong Yang, Jie Jiang, and Qinglin Lu. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding, 2024. 3 +[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6 +[38] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation, 2024. 6 +[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3 +[40] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, Lifang He, and Lichao Sun. Sora: A review on background, technology, limitations, and opportunities of large vision models, 2024. 6 +[41] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 1 +[42] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding, 2024. 3 +[43] Pengqi Lu. Qwen2vl-flux: Unifying image and text guidance for controllable image generation, 2024. 2, 3 +[44] Bingqi Ma, Zhuofan Zong, Guanglu Song, Hongsheng Li, and Yu Liu. Exploring the role of large language models in prompt encoding for diffusion models, 2024. 3 +[45] Jian Ma, Chen Chen, Qingsong Xie, and Haonan Lu. Peadiffusion: Parameter-efficient adapter with knowledge distillation in non-english text-to-image generation. arXiv preprint arXiv:2311.17086, 2023. 1, 6 +[46] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently. arXiv preprint arXiv:2303.17870, 2023. 1 +[47] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 1, 7 +[48] Jian Ma, Yonglin Deng, Chen Chen, Nanyang Du, Haonan Lu, and Zhenyu Yang. Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models, 2025. 1 +[49] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models, 2023. 5 +[50] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and + +Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[51] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models, 2024. 3, 4, 7 +[52] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. 3 +[53] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation, 2024. 5 +[54] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 1, 3 +[55] Can Qin, Ning Yu, Chen Xing, Shu Zhang, Zeyuan Chen, Stefano Ermon, Yun Fu, Caiming Xiong, and Ran Xu. Gluegen: Plug and play multi-modal encoders for x-to-image generation, 2023. 2 +[56] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, Stefano Ermon, Yun Fu, and Ran Xu. Unicontrol: A unified diffusion model for controllable visual generation in the wild, 2023. 6 +[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1 +[58] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 1 +[59] Samyam Rajbhandari, Jeff Rasley, Olatunj Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models, 2020. 1 +[60] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation, 2021. 1 +[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3, 7 +[63] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1 +[64] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine + +tuning text-to-image diffusion models for subject-driven generation, 2023. 6, 7 +[65] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 3, 7 +[66] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[67] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 5 +[68] Yichun Shi, Peng Wang, and Weilin Huang. Seededit: Align image re-generation to image editing, 2024. 1, 2 +[69] Oscar Skean, Md Rifat Arefin, Yann LeCun, and Ravid Shwartz-Ziv. Does representation matter? exploring intermediate layers in large language models, 2024. 4 +[70] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. *Omnicontrol: Minimal and universal control for diffusion transformer*, 2025. 3, 7 +[71] Kcolors Team. Kcolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 1, 3 +[72] Hugo Touvron, Louis Martin, and Kevin Stone. Llama 2: Open foundation and fine-tuned chat models, 2023. 1 +[73] Kai Wang, Fei Yang, and Joost van de Weijer. Attention distillation: self-supervised vision transformer students need more guidance, 2022. 3 +[74] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 3 +[75] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2024. 3 +[76] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm, 2024. 3 +[77] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023. 6 +[78] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue + +Wu, Bingxuan Wang, Zhenda Xie, Yu Wu, Kai Hu, Jiawei Wang, Yaofeng Sun, Yukun Li, Yishi Piao, Kang Guan, Aixin Liu, Xin Xie, Yuxiang You, Kai Dong, Xingkai Yu, Haowei Zhang, Liang Zhao, Yisong Wang, and Chong Ruan. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding, 2024. 3 +[79] Bin Xia, Yuechen Zhang, Jingyao Li, Chengyao Wang, Yitong Wang, Xinglong Wu, Bei Yu, and Jiaya Jia. Dreamomni: Unified image generation and editing, 2024. 4 +[80] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers, 2024. 6 +[81] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation, 2023. 6 +[82] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths, 2024. 3 +[83] Jianan Yang, Haobo Wang, Yanming Zhang, Ruixuan Xiao, Sai Wu, Gang Chen, and Junbo Zhao. Controllable textual inversion for personalized text-to-image generation, 2023. 7 +[84] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 3, 7 +[85] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation, 2022. 1 +[86] Qifan Yu, Wei Chow, Zhongqi Yue, Kaihang Pan, Yang Wu, Xiaoyang Wan, Juncheng Li, Siliang Tang, Hanwang Zhang, and Yueting Zhuang. Anyedit: Mastering unified high-quality image editing for any idea, 2024. 1 +[87] Jun Zhan, Junqi Dai, Jiasheng Ye, Yunhua Zhou, Dong Zhang, Zhigeng Liu, Xin Zhang, Ruibin Yuan, Ge Zhang, Linyang Li, Hang Yan, Jie Fu, Tao Gui, Tianxiang Sun, Yugang Jiang, and Xipeng Qiu. Anygpt: Unified multimodal IIm with discrete sequence modeling, 2024. 3 +[88] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1 +[89] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation: Improved latent consistency distillation by semi-linear consistency function with trajectory mapping, 2024. 3 +[90] Yuanzhi Zhu, Xingchao Liu, and Qiang Liu. Slimflow: Training smaller one-step diffusion models with rectified flow, 2024. 3 +[91] Zhuofan Zong, Dongzhi Jiang, Bingqi Ma, Guanglu Song, Hao Shao, Dazhong Shen, Yu Liu, and Hongsheng Li. Easyref: Omni-generalized group image reference for diffusion models via multimodal llm, 2024. 3 \ No newline at end of file diff --git a/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/images.zip b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2784883bb911f72e18c96f7a87f4238a8b62a143 --- /dev/null +++ b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd07bcbde5215fceb6948f82e4a4110ba5934aaa861f73c63da74ac022f18e9b +size 572982 diff --git a/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/layout.json b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dd8716176f4d36f44ee8dd5e90393b4f26caf78f --- /dev/null +++ b/x2iseamlessintegrationofmultimodalunderstandingintodiffusiontransformerviaattentiondistillation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b344d7afd661c559714d79affaa2fb5aa12e4fc155ff9f875a6c5762641fe88 +size 451792 diff --git a/xcaptureanopensourceportabledeviceformultisensorylearning/6b08175f-1aac-4b47-837d-556a41b9c21a_content_list.json b/xcaptureanopensourceportabledeviceformultisensorylearning/6b08175f-1aac-4b47-837d-556a41b9c21a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7f3ceb3f0de6b4f288b534c87e07d67e3e6acc2e --- /dev/null +++ b/xcaptureanopensourceportabledeviceformultisensorylearning/6b08175f-1aac-4b47-837d-556a41b9c21a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfcf9ccd7a24e7727a7d239ecc7509d2a9a50188ecc8f4d6327c25333e33b20a +size 81570 diff --git a/xcaptureanopensourceportabledeviceformultisensorylearning/6b08175f-1aac-4b47-837d-556a41b9c21a_model.json b/xcaptureanopensourceportabledeviceformultisensorylearning/6b08175f-1aac-4b47-837d-556a41b9c21a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8ae60beafa7d2619e4203808ea5cbd8486baebe1 --- /dev/null +++ b/xcaptureanopensourceportabledeviceformultisensorylearning/6b08175f-1aac-4b47-837d-556a41b9c21a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20cfa04fe700838947c99d68aa878698653c2d77d74c486ae7e01e2a810b9805 +size 99334 diff --git a/xcaptureanopensourceportabledeviceformultisensorylearning/6b08175f-1aac-4b47-837d-556a41b9c21a_origin.pdf b/xcaptureanopensourceportabledeviceformultisensorylearning/6b08175f-1aac-4b47-837d-556a41b9c21a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..03452b8e97020e6d92bddf81b9b145750dafb2ff --- /dev/null +++ b/xcaptureanopensourceportabledeviceformultisensorylearning/6b08175f-1aac-4b47-837d-556a41b9c21a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e85920a3be02246c3e732ac929f3a6e9525a6b387077cbdc7ab999ac1b3ce577 +size 14181867 diff --git a/xcaptureanopensourceportabledeviceformultisensorylearning/full.md b/xcaptureanopensourceportabledeviceformultisensorylearning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4b1c648adc7350b8df357e33427159007f1ce278 --- /dev/null +++ b/xcaptureanopensourceportabledeviceformultisensorylearning/full.md @@ -0,0 +1,289 @@ +# X-Capture: An Open-Source Portable Device for Multi-Sensory Learning + +Samuel Clarke + +Suzannah Wistreich + +Yanjie Ze + +Jiajun Wu + +Stanford University + +![](images/55f76900b091201231ae277c856f6a8e43e2853682c1a6bbb16d0d2ce736853c.jpg) +Figure 1. X-Capture for multi-sensory data capture. (Left) The user captures tactile data from a vase in a living room. (Right) The sensor readings for each modality from the same probed point on the vase, as well as a visualization of the hammer impulse and 3D pose vectors for the image and tactile captures, shown in blue and red, respectively. + +![](images/ad800ce50e283934355e7557cd0f81a979c67de08d8fbce9c4e20c097ba9b101.jpg) +RGB + +![](images/fa239026393483e372f988a84ee54b11c8c8a903bd2ad2056797058aecd504fe.jpg) +Spectrogram + +![](images/d9555c47faac2213e9c90ecdc507d24cd2e82cd8cd62fa80a7948436e495273f.jpg) +Depth + +![](images/1c61fc8c01f671145d087d4dc05d696da0fb5ece2857bbce17814948f6ff4ffb.jpg) +Hammer Impulse + +![](images/f7996ee30fc80e2fd14565f151e9e39c8c17984e123ee8acc108c331aa739477.jpg) +Tactile + +![](images/73e40815cbc9923da2fb62c43d5fd93ea60e5bb05bff0a3ade39fdd24c5287d1.jpg) +Capture Pose + +# Abstract + +Understanding objects through multiple sensory modalities is fundamental to human perception, enabling cross-sensory integration and richer comprehension. For AI and robotic systems to replicate this ability, access to diverse, high-quality multi-sensory data is critical. Existing datasets are often limited by their focus on controlled environments, simulated objects, or restricted modality pairings. We introduce X-Capture, an open-source, portable, and cost-effective device for real-world multi-sensory data collection, capable of capturing correlated RGBD images, tactile readings, and impact audio. With a build cost under $1,000, X-Capture democratizes the creation of multisensory datasets, requiring only consumer-grade tools for assembly. Using X-Capture, we curate a sample dataset of 3,600 total points on 600 everyday objects from diverse, real-world environments, offering both richness and variety1. Our experiments demonstrate the value of both the quantity and the sensory breadth of our data for both pretraining and fine-tuning multi-modal representations for + +object-centric tasks such as cross-sensory retrieval and reconstruction. X-Capture lays the groundwork for advancing human-like sensory representations in AI, emphasizing scalability, accessibility, and real-world applicability. + +# 1. Introduction + +As humans, we experience objects in our everyday environments through a combination of every sensory modality we possess. Each sensory modality provides us with unique information about the object, which can complement the information evident from other modalities. Touching what visually appears to be a ripe fruit can reveal that it is in fact not yet ripe or has a hidden bruise. Hearing a rigid drinking glass being tapped can disambiguate whether it is made of glass, crystal, or plastic. However, while each of these sensory modalities may complement each other, they are often highly correlated, as they each derive from the underlying physical properties of a given object [25]. Thus, we are able to intuitively relate different modalities and generate expectations of other modalities from experiencing only one [26]. In order for robots and agents to understand objects in the same way humans do, they must similarly possess such intuition about the relationships between objects' + +different sensory modalities. + +In this paper, we focus on the sensory modalities of vision (both RGB and depth), sound, and touch—modalities for which popular commercially available sensors exist. Numerous powerful models have been developed to relate these modalities within a shared latent representation for interesting downstream cross-sensory inference and generation tasks [19, 36, 44, 46]. Such models rely on large datasets of examples correlating sensory modalities with each other. Although many such datasets exist, they suffer from key deficiencies that hamper their usefulness in training representations to enhance the understanding of real-world, in-the-wild objects. First, many multi-sensory datasets focus on scenes rather than objects, reducing their relevance to applications where object understanding is a priority, such as robotic manipulation. Of those focusing on objects, many only include data collected from simulated objects, or from real objects exclusively within controlled environments. Both simulated and controlled real data can present a large domain gap relative to data from real in-the-wild scenarios. Additionally, those collected in controlled environments often require expensive rigs and equipment, a drawback to both accessibility and scalability. Finally, most object-centric multi-sensory datasets lack breadth in the sensory modalities they correlate, often linking only two sensory modalities, such as touch-vision or audio-vision. This hinders representations from forming direct rather than emergent alignment among more than two modalities. + +To address each of these deficiencies, we introduce X-Capture, a portable, low-cost device for capturing correlated RGBD images, tactile images, and impact audio samples from objects in the wild. Our device connects to and is powered by a laptop, with a user interface (UI) that visualizes data during the collection process. We ensure that each sensor takes independent readings (e.g., the touch sensor is not visible in camera images) through careful design of the device and UI, and we explicitly measure an input-output relationship for both touch and audio. These features address common shortcomings of many existing object-centric datasets and collection methods. We open-source the mechanical design and parts list of the device, with a total bill of materials of less than $1000 at time of writing. The device requires only a consumer-grade 3D printer and soldering iron to assemble. Figure 1 shows the X-Capture device in use, along with sensor readings for each modality from a probed sample object. As shown in Table 1, our device is more versatile, portable, and relatively cheaper compared with other data-capturing devices. + +We collect a sample dataset of 600 different objects with our device and show how our dataset can be used to fine-tune existing multi-sensory representations to improve their performance in object-centric benchmark tasks such as cross-sensory retrieval, reconstruction, and detection. + +
Device/SetupObject PropertiesPort.Cost
RGB 3D Audio Touch
UBC ACME [35]✓✓✓✓✘N/A*
RealImpact [6]✓✓✓✘✘$8,000
Obj.Folder Real [16]✓✓✓✓✘$11,000
TVL [11]✓✘✘✓✓$560
X-Capture (Ours)✓✓✓✓✓$1000
+ +Table 1. Comparing prior multi-sensory object-centric data capture devices for what object properties they are designed to capture, their portability ("Port"), and their estimated cost. Note that by "Audio" we are referring to impact sounds. The X-Capture device captures images, point clouds, impact audio, and tactile data from objects in a portable and affordable package. (*UBC ACME [35] used equipment which is no longer commercially available at time of writing.) + +# 2. Related Work + +Among large foundation models are many popular multisensory representation models, which attempt to acquire implicit knowledge of the physical world through their representations. Contrastive Language-Image Pretraining (CLIP) trained a representation between images and their text captions to relate images and text [36]. ImageBind used CLIP as a backbone, along with separate datasets linking images to audio, depth, and thermal data, to train disparate modality-specific encoders to share a single multisensory representation [19]. Though each dataset was used to train the model to link images with only one other modality, they showed evidence of emergent alignment between other modalities. A Unified Representation of Language, Images, and Point Clouds (ULIP) [44] trained a unified representation of text descriptions, images, and point clouds of objects, showing that the representation was made measurably stronger by aligning all three modalities simultaneously during training, rather than only two at a time as ImageBind had done. Later works used both real and simulated tactile data to bind tactile images to the CLIP representation space as well [4, 11, 46, 49]. + +Training these representations required a curated dataset linking two or more distinct sensory modalities. Many datasets link images with text through datasets of human-captioned images [12, 39, 40]. Other datasets link sensory signals from two or more modalities, the majority of which consist of multi-sensory data from simulated environments [13, 14], with some works showing how such data could be used to train models for downstream tasks in either simulated [24] or real [15, 17] embodied environments. However, the ObjectFolder Benchmark showed that models fine-tuned with their proposed dataset of 100 real objects generally performed significantly better on real-world embodied tasks than those trained only with simulated data [16]. Many multi-sensory datasets were col + +
DatasetObj.Correlated ModalitiesEnv.
RGB 3D Audio Touch
Feeling of Success [3]106✓✓X✓T
VisGel [31]195✓XX✓T
Touch and Go [45]3971✓XX✓W
SSVTP [28]N/A✓XX✓T
HCT [11]N/A✓XX✓W
Greatest Hits [34]N/A✓XX✓W
RealImpact [6]50✓✓✓XC
ObjectFolder 2.0 [15]1000✓✓✓✓S
ObjectFolder Real [16]100✓✓✓✓C
X-Capture (Ours)600✓✓✓✓W
+ +Table 2. Comparing publicly available multi-sensory datasets by their number of objects, their correlated sensory modalities, and the environments in which they are collected (C=Controlled, S=Simulation, T=Tabletop, and W=Wild). The X-Capture dataset includes the widest breadth of correlated sensory modalities of a dataset collected in the wild. + +lected using automated setups to scale up data collection [3, 6, 28, 31, 35], but similar to those of ObjectFolder Benchmark, such setups are often costly to replicate, collected in only controlled environments, and naturally impose some constraints on the types of objects which can be scanned. See Table 1 for a comparison of X-Capture to relevant data collection setups and devices, and Table 2 for a comparison of our dataset to prior multi-sensory datasets. We include additional details in Appendix B. + +Collecting multi-sensory data in situ can produce datasets with distributions matching expected test conditions of embodied agents more closely. The authors of the Touch and Go dataset press a GelSight sensor against objects in the wild while holding a webcam behind the sensor [45]. The authors of the Greatest Hits dataset [34] similarly took video and audio samples of a drumstick striking objects. But without measuring inputs explicitly, such as the contact location and the pressing or striking force, it is impractical to infer the essential input-output relationship between force and touch or impact and sound of an object. + +Hand-held sensorized devices can combine the best of both worlds, offering the inter-sensor measurement consistency of sophisticated setups with the portability needed to collect data in situ. Multiple works designed handheld devices for capturing correlated vision and tactile signals [10, 11]. The authors of [2] designed a handheld device for collecting multi-sensory tactile data to characterize thermal properties of objects. Numerous works have proposed sensorized handheld devices that approximated robot end-effectors [5, 42, 48]. Humans used these devices to collect demonstration data by performing tasks with everyday objects, which could be used to train a policy for a robot arm or mobile manipulator [21, 41]. Our device brings the advantages of handheld data collection devices into a uni + +![](images/8786b78ced80a2e8023201ce34303363b2bd0d9f1c2ad6899b62111514b6d478.jpg) +Figure 2. Exploded view of X-Capture. The device rigidly constrains all sensor assemblies into fixed relative poses on a compact chassis with an ergonomic grip. Wires and circuitry are not shown. + +fied framework for measuring vision, touch, and audio data from objects, using commercially available sensors popular in many embodied learning applications. + +# 3. The X-Capture Device + +The X-Capture device supports sensing correlated RGB, depth, tactile, and impact audio samples using a combination of six distinct sensors. It is compact and power-efficient, making it as convenient as possible to carry, only requiring control and power from a laptop. The chassis is constructed out of 3D printed PLA, with an ergonomic grip supporting an enclosure. The enclosure houses a USB hub that routes power and communications to all sensors and connects to a laptop with a single USB-C cable. Figure 2 and the supplementary video show how the device and sensor assemblies physically fit together. We detail the compositions of the sensor assemblies and how they are used below, organized by sensory modality. See Appendix C for additional details on the hardware, including internal photos, a bill of materials, and support for alternative sensors, and the supplementary video for a sequential breakdown of how the sensors and assemblies fit together. + +# 3.1. Touch + +For tactile sensing, X-Capture uses an assembly mounted on the front of the device consisting of a DIGIT sensor [30] attached to a load cell. Vision-based tactile sensors are popular for both benchmark datasets [16] and robotic applications [1, 28, 43]. We choose the DIGIT sensor for its commercial availability and relative durability [30]. However, the images the sensor collects during contact with an object are not only a function of the object's intrinsic geometry and stiffness at the point of contact, but also extrinsic factors such as the angle and the force at which the DIGIT sensor presses against the object. Thus, we design a novel assembly to explicitly measure this input-output relation- + +ship by annotating tactile images from DIGIT in real time with relative angle and calibrated normal force measurements using an accelerometer and load cell, respectively. The device additionally uses the accelerometer to subtract the gravitational force exerted by the DIGIT sensor at the current device angle, dynamically re-zeroing the pressing force. This method uses low-cost, commercially available sensors to obviate the need to collect thousands of training points with a $6,000 force sensor for estimating calibrated forces from the vision-based sensor [9, 23, 51]. While collecting tactile data, our device automatically takes a snapshot of the DIGIT's tactile image at a pressing force within a 0.5N range of different target levels. We use 10, 15, and 20N as target levels during data collection. + +# 3.2. Vision and Depth + +X-Capture captures RGBD images of objects with a RealSense D405 stereoscopic depth camera. The camera is compact and lightweight, with a 7cm minimum distance for measuring depth, lending itself well to close-up capture of objects. X-Capture snapshots both the RGBD image and the accelerometer state simultaneously, such that the user can ensure a consistent angle of approach between the RGBD image and the tactile reading of a given point. + +# 3.3. Audio + +While many objects independently make sounds and therefore have semantic audio properties, we focus on collecting impact sounds from objects, which can reveal important geometric or material properties. X-Capture collects audio samples from objects by striking them with an impact hammer and recording the resulting impact sound with a microphone situated behind the point of impact. The impact hammer is 3D printed from PLA, with a Thorlabs PK2JA2P1 piezo stack attached to a small steel rod at its tip. The piezo stack measures the impact force as it excites the object and produces the sound, in order to measure the input-output relation between the acting contact forces and resulting sound at each point. The base of the hammer fits into a 3D-printed elastomer base which functions as a spring. For each audio sample, the user pulls the hammer back until it adheres to an electromagnet mounted behind the hammer, storing potential energy in the elastomer base. After a pre-configured delay, the device automatically deactivates the electromagnet, releasing the hammer to strike the object in a silent and repeatable manner. The device records the impact sounds with a Dayton Audio EMM6 measurement microphone, which has a relatively uniform frequency response within the range of human-audible frequencies. + +The impact hammer and the microphone's outputs are recorded by a HiFiBerry DAC2 ADC Pro board attached to a Raspberry Pi Zero 2W. The HiFiBerry board ensures that the hammer and microphone signal recordings are timesynchronized with minimal noise. The board supports high + +sample rates as well as digitally controlled gains, such that our device can dynamically and repeatably adjust the volume gain of each recording according to the loudness of each object's impact sounds. This ensures that the audio signal is recorded at a maximal level without clipping. + +Finally, we design a custom PCB which provides precise power and input conditioning, as well as stable physical connections for all of the above electrical components. This novel assembly collects impact audio data at a comparably high fidelity to prior works [6, 16], while significantly optimizing power usage, size, and cost ( $130 vs.$ 2,000), to improve portability and affordability. + +# 3.4. User Interface and Workflow + +Similar to the view shown at the right of Figure 1, the UI visualizes all modalities of data in one screen to provide important feedback during data collection, allowing for retakes and comparisons to previous captures. For each example, the user captures readings from each modality at one specific point on the object. The user first positions X-Capture's RGBD camera such that a target point on the object is centered in the image, at a depth 8 to $13\mathrm{cm}$ away from the camera, as measured by depth image and displayed on a bar below the image. The UI displays the captured RGB and depth images, with the target point marked by small crosshairs superimposed at the center of both images. The user then presses the DIGIT sensor against the same point, using both the crosshairs on the RGB image and a display of the current angle of the device's accelerometer as a guide to ensure contacting the object at the same point and angle. During contact, a bar below the tactile image displays the current pressing force. The user pushes gradually up to 20N with snapshots automatically collected at 10, 15, and 20N. Finally, the user positions the impact hammer to hover over the target point, initiates a recording, pulls the hammer back to the electromagnet, then holds the device still while the electromagnet releases the hammer to strike the object after a delay. The UI displays the recorded spectrogram of the impact audio, as well as a time-domain graph of the impulse measured by the hammer, so the user can verify that the hammer made a single, clean impulse. See the supplementary video for a demonstration of this workflow. Note that our sensors allow constraints such as depth, angle, and impact quality to enforce some consistency in our sample dataset, but such constraints can be relaxed to optimize for higher throughput, if desired. + +# 4. The X-Capture Dataset + +To evaluate our device's data collection pipeline, we present a novel dataset of correlated vision, tactile, audio, and postprocessed point cloud data from a total of 3,600 points on 600 objects in real-world environments, collected in just under three weeks. Our multi-sensory dataset includes data from objects across a diverse class of materials, geometries, + +![](images/224088a7757d6bc13695ebd1d9fc5cc6faa583aaec477bb487659b092a76176b.jpg) +RGB +Figure 3. Example multi-sensory data points from the X-Capture dataset. Each row shows aligned multi-sensory data captured from a single point on an object in a distinct natural environment: (from left) RGB image centered on the point, depth image, impact audio spectrogram, and tactile image. (Objects, from top: Cat Brush, Insulated Steel Cup, Glass Storage Bowl, Computer Speaker, and Sheet Metal Container.) + +![](images/e99f363324a733184fe3e912880a83ee235814eb0e2a2b722c239300218d01ba.jpg) +Depth + +![](images/65546fc26b816faae100c9f262f52daab8eccc4d82f639bfba7e652cd262ccd8.jpg) +Audio + +![](images/225fdf483679d1fd5c13ff5a22e8da9ea6e3ec2ae29930df9f253ecca25f5c18.jpg) +Touch + +and functional uses. Aided by our device's flexible capabilities and portable design, we capture a wide breadth of objects encountered in everyday settings. We include further details of the objects and environments comprising our dataset in Appendix D.1 and a comprehensive comparison of our datasets to existing alternatives in Appendix B. + +# 4.1. Collection Procedure + +We collect data across 12 different environments— one indoor workspace with relatively controlled conditions, and 11 in-the-wild locations, ranging from everyday indoor settings to a dynamic outdoor area. We capture 325 diverse objects in the indoor workspace, a quiet office with consistent lighting, providing a distribution baseline. We probe the remaining 275 objects in 11 diverse natural environments, including a Bathroom, Home Office, Classroom, Workshop, Bedroom, Laundry Room, Living Room, two kitchens, and two outdoor patios. We capture an equal number of objects in each in-the-wild environment. + +For each object, we collect readings of each modality from six distinct points on the object. We choose point lo + +cations which cover the breadth of each object's surface and capture unique local features. Each of the six points has a corresponding RGBD image, impact audio recording, and tactile impressions captured at 10N, 15N, and 20N. We use our device and UI to manually register these sensory readings with each other at each point by following the procedure described in Section 3.4. This collection process takes about nine minutes per object, on average. + +# 4.2. Dataset Labeling and Post-Processing + +We provide a brief text description of each object, noting salient materials and the relevant state of the object (e.g., if a can is full or empty). We further postprocess our data, using the RGB and depth data to estimate point clouds of objects and using the noted recording gains and hammer signals to normalize the audio. We include details of both these post-processing steps in Appendix D.3. + +# 5. Experiments + +We validate the usefulness of the sample dataset we collect with the X-Capture device with three popular multi-sensory benchmark tasks: cross-sensory retrieval, image generation, and point cloud generation. We also demonstrate how we can use our data to train an audio-based object detector. Unless otherwise noted, we randomly split our dataset into 500 training and 100 test objects. See Appendix E.5 for additional details on training and testing procedures. + +# 5.1. Baselines + +We evaluate recent cross-sensory representation frameworks on our dataset. The ImageBind framework [19] provides encoders for the RGB, audio, and depth modalities, each pretrained on web-scale datasets pairing images with each modality, but the framework lacks encoders for point clouds or tactile readings. In order to cover all of the modalities we provide in our dataset, we also evaluate the performance of a combination of pretrained encoders from different sources, which each specialize in a specific modality or cross-sensory representation. We encode RGB images with CLIP's [36] pretrained ViT-L encoder. For tactile images, we use the ViT-L encoder publicly released with [11], which has been pretrained on the TVL dataset. We encode our audio recordings with the Audio Spectrogram Transformer (AST) [20], with publicly released weights from pretraining on ImageNet [8] and AudioSet [18]. Finally, for point clouds, we use the PointBERT model [50] with publicly released weights from pretraining on ULIP [44]. + +For both the ImageBind encoders and the ensemble of modality-specific encoders, we evaluate three different training configurations. We first evaluate the publicly-released out-of-the-box pretrained weights. Then we test fine-tuning all pretrained models on our dataset's training set, using two distinct formulations of the contrastive InfoNCE loss [33]. In the "Image Loss" formulation, we fine + +![](images/fd4ea19b551a486be02975cd8b547bdd6899115893530b11edbf79a67d6abc60.jpg) +Figure 4. Comparing test retrieval performance of our cross-modal encoders trained with varying quantities of objects, with each plot grouping results by the query modality used for retrieval. The right-most plot shows an average across all modality combinations. + +![](images/1a611602bf1810b0e4f8916aca5d36ebe55c0d290e78bb82e66e0c5a17867b77.jpg) + +![](images/5162cac531583737cd65258a6fc56b5f6a59dac4ee38e995fa021aa44bbafb9a.jpg) + +![](images/5a077246d89bc0c5bd1c0a47354750e1d8774e713650480de1cac1a6e97e4fdd.jpg) + +![](images/6f1624077f8cce976c92054216124cecf313fe3a626f8b2be34272876e65ca77.jpg) + +
ImageBind Top-5 Accuracy (%)RGB→Audio→Depth→Avg.
AudioDepthRGBDepthRGBAudio
Random Guess5.05.05.05.05.05.05.0
Out-of-the-Box Pretrained7.011.46.04.67.65.06.9
Fine-Tune w/ Image Loss46.063.847.628.264.025.445.8
Fine-Tune w/ Cross-Sens. Loss47.066.249.028.664.628.047.2
+ +Table 3. Cross-sensory retrieval top-5 accuracies of ImageBind trained with different strategies using our dataset. The top and bottom column headers denote the query and retrieved modalities, respectively. Out-of-the-Box weights do not generalize well to our data, whereas Fine-Tuning with a Cross-Sensory Loss outperforms other configurations, across all modalities. + +
ImageBind Top-1 Accuracy (%)RGB→Audio→Depth→Avg.
AudioDepthRGBDepthRGBAudio
Random Guess16.716.716.716.716.716.716.7
Out-of-the-Box Pretrained16.720.516.017.019.018.718.0
Fine-Tune w/ Image Loss21.034.320.822.745.221.227.5
Fine-Tune w/ Cross-Sens. Loss22.336.523.321.846.720.328.5
+ +Table 4. Contact localization top-1 accuracies of ImageBind trained with different strategies using our dataset. The top and bottom column headers denote the query and retrieved modalities, respectively. Out-of-the-Box weights generalize poorly to our data, and Fine-Tuning with a Cross-Sensory Loss achieves best average performance. + +tune according to a symmetric InfoNCE loss between images and each other available modality, the same technique used by ImageBind. In the "Cross-Sensory Loss" formulation, we fine-tune according to a symmetric InfoNCE loss between each pairing of all modalities, similar to the loss used among vision, language, and point cloud encoders in ULIP. For all configurations, we keep the encoders for the RGB image modality frozen during training. + +# 5.2. Cross-Sensory Retrieval + +Cross-sensory retrieval assesses models' abilities to connect multi-sensory information, akin to the human ability to intuitively link sight, sound, and touch, to make valuable inferences about unknown properties of objects. Consequently, it has become a common benchmark in cross-sensory learning [16, 19, 46]. We formulate the task as an inter-object classification task, testing each cross-sensory frameworks' performance as follows: given a randomly selected point from each of $N$ objects, where $N = 100$ for our test dataset, + +can the framework correctly associate a sensory reading from one modality with another from the same point? + +We show the top-5 accuracies for ImageBind in Table 3. Note that for 100 objects, the expected value of random selection would be $5\%$ under our test conditions. While the encoders have been pretrained on very diverse datasets, their out-of-the-box weights struggle on our object-centric data, especially with the audio modality. Our results also suggest that using a full cross-sensory loss comparing all modalities directly provides a generally stronger representation on our data. We show the results for the cross-sensory encoder ensemble in Table 7 of Appendix E.1. Interestingly, in these results we see less of a clear performance advantage to training with the Cross-Sensory Loss versus the Image Loss, as compared to the advantage we observe in our ImageBind experiments. For modality pairings the ensemble and ImageBind have in common, such as RGB $\rightarrow$ Audio, we see similar results to those of ImageBind. + +# 5.3. Cross-Sensory Contact Point Localization + +Contact point localization assesses models' abilities to differentiate between sensory signals coming from different points on the same object. We thus formulate contact point localization as a classification task similar to cross-sensory retrieval, except that instead of comparing different sensory signals from a single point from one object to points of other objects, we compare different sensory signals from a single point to those of another modality from all $M$ points on the same object, where $M = 6$ in our dataset. + +We show the top-1 accuracies for ImageBind in Table 4. In this task, the expected value of random selection is $\sim 16.7\%$ . Once again, ImageBind's pretrained weights struggle to outperform random chance on our object-centric data. However, this is clearly a difficult task even after finetuning. ImageBind seems to excel at differentiating object points with RGB and depth much more than it does with audio, perhaps because the impact sounds a single object produces at different points often have very minute differences. We see similar results for our cross-sensory encoder ensemble in Table 8 of Appendix E.2, where the association between RGB and point cloud is much stronger than associations between other modalities. + +For both this task and the retrieval task of Section 5.2, we show results of using arithmetic combinations of Image-Bind embeddings from two different modalities to retrieve from the third modality in Appendix E.3. + +# 5.4. Scaling Law + +We claim that the X-Capture device makes multi-sensory object-centric data collection more efficient than prior alternatives, allowing us to scale up the collection of a valuable dataset. Though we have already collected a sample dataset of 600 objects, we evaluate how a model's performance improves according to the quantity of our training objects we use during fine-tuning. The results in Figure 4 demonstrate that each modality continues to benefit from additional training examples, and the performance improvement seems far from plateauing at 500 objects, confirming the value of scaling up the X-Capture dataset even further. + +# 5.5. Using X-Capture Data as a Pretraining Set + +Multi-sensory object understanding is important in many different downstream applications, including robotic manipulation, anomaly detection, and generation. However, each potential application may involve unique data with a domain gap differentiating it from the data in existing large-scale datasets. Fortunately, such large-scale datasets can still be used to pretrain models, which can be fine-tuned on additional data for downstream applications. We evaluate whether data from X-Capture could be similarly helpful by using it as a pretraining set. + +We evaluate cross-sensory retrieval on the real object data from the ObjectFolder Benchmark [16]. Though ObjectFolder's subject matter is quite similar to ours, there is a significant domain gap in how each modality is recorded. Their RGB images are background-less renderings from 3D scans of objects rather than real images. Their tactile data is from a GelSight sensor, which has a different texture detail and lighting conditions than the DIGIT. Finally, they collect impact sounds in an acoustically-treated room, with objects suspended by string to reduce the contact-damping usually affecting objects' impact sounds in real-world settings. + +We evaluate the cross-sensory encoder ensemble's vision, audio, and tactile encoders trained with three different configurations using the cross-sensory loss. In "Our Pre-trained Only" configuration, we train encoders with only X-Capture data. In "Fine-Tuned Only", we train the encoders only with data from the 70 real training objects from ObjectFolder. And in "Our Pretrained + Fine-Tuned", we pretrain the encoders with our data, then fine-tune on the real training objects from ObjectFolder. We evaluate cross-sensory retrieval on the 30 real test objects of ObjectFolder Benchmark and show the top-5 accuracies in Table 5. The expected value of random selection is $16.7\%$ . We see that "Our Pretrained Only" model outperforms random chance, but still performs rather poorly in this zero + +
ObjectFolder-Real (30 Objects) Top-5 Accuracy (%)RGB→Audio→Tactile→Avg.
AudioTactileRGBTactileRGBAudio
Random Guess16.716.716.716.716.716.716.7
Our Pretrained Only33.325.026.724.220.823.325.6
Fine-Tuned Only27.533.332.533.332.537.532.8
Our Pretrained + Fine-Tuned45.033.340.830.032.535.036.1
+ +Table 5. Results of ablating both pretraining with our dataset and fine tuning with train objects from the ObjectFolder Real dataset [16] for downstream evaluation of cross-sensory retrieval on the real test objects from ObjectFolder. + +shot generalization, likely struggling to surmount the domain gap without fine-tuning on ObjectFolder's data. However, "Our Pretrained + Fine-Tuned" model outperforms the "Fine-Tune Only" model, suggesting that pretraining with the X-Capture dataset helps bridge the generalization gap in the low-data regime of the ObjectFolder Real training set. The improvement is evident in the audio modality, but not in the tactile modality, suggesting that the domain gap between different tactile sensors may be especially challenging. + +# 5.6. X-to-2D/3D Generation + +For humans, the sound or feel of an object can often conjure a mental image of the object. We thus evaluate the ability of our representations aligned to CLIP image features to provide a useful representation for both a CLIP-based 3D implicit function generator and CLIP-based image generator. We use our dataset to train our tactile and audio encoders to align to the outputs of CLIP's ViT-L encoder of the corresponding image of each point. For this experiment, we use a mean-squared error (MSE) rather than contrastive loss to prioritize alignment to CLIP over cross-sensory association and differentiation. After training alignment on our training objects, we use the signals encoded from our held-out test objects as input to pretrained Shap-E [27] to generate 3D implicit functions and to pretrained Stable Diffusion [38] to generate images. Both models have been pretrained by their authors to use CLIP ViT-L features encoded from text to generate 3D implicit functions and images, respectively. + +We show qualitative results for both image generation and 3D neural radiance field generation in Figure 5. The results vary substantially in quality, but they seem to provide some interesting insights into what features our encoders are able to glean from each respective sensory modality. We see that both shape and image generations from the image encoding tend to match the original object in both color and semantic features. Generations from audio encodings often successfully match the salient materials of the object, and occasionally some semantic geometric features. Generations from tactile encodings seem to excel at matching the geometric features local to the contact point, such as the local curvature, and also occasionally match in material properties such as hardness. These results lend further evidence of the complementary information which can be inferred from different sensory modalities when interacting with ob + +![](images/9cb58b62c8815029acf4b91237baedb75b7b8166a0b3258410c5fda8138db8b5.jpg) +Figure 5. Results of using Shap-E [27] to generate 3D neural radiance fields and Stable Diffusion [38] to generate images from outputs of multimodal encoders which have been trained on our data to align to CLIP features. The three left columns show the RGB images, audio spectrograms, and tactile images inputted to their respective encoders. The next three columns show the neural radiance fields generated from using the outputs of the encoders from these RGB, audio, and tactile inputs, respectively, as input to Shap-E. The last three columns similarly show the images generated from using the outputs of the RGB, audio, and tactile inputs, respectively, as input to Stable Diffusion. + +![](images/9ed5e08eeee5109bddf2a83f252bb3bc20cd55dd789a94fa6dbbe7dc0d9557e4.jpg) + +![](images/6d0847b2ad4bc93a3f41cd98d874ac9e8eaefbbb081e04152d223526905cf7ef.jpg) + +![](images/9e5c68d720bfe03141cdc0e1d2a11a4c80f5696f3ec094aeef378fd07236b354.jpg) +Figure 6. Still frame from prompting Detic [52] with our pretrained audio encoder's embedding of the real sound of the ceramic plate (at bottom, highlighted pink) being placed on the counter. Detic identifies the plate, as well as the ceramic mugs, drinking glass, and car, as likely sources of the sound. It successfully ignores appliances, cardboard tea boxes, and metal utensils. + +jects. We show additional examples of prompting generations with arithmetic combinations of embeddings from different modalities in Appendix E.3. + +# 5.7. Zero-Shot Audio-Based Object Detection + +Also similar to ImageBind [19], we use our dataset to train a CLIP-aligned audio embedding which can replace the input for the text-based Detic detection model [52]. We train our audio encoder on our dataset contrastively to ViT/B-32 CLIP features, then use embeddings from this encoder to prompt Detic's CLIP-based object detector. Though our dataset is collected with objects in the wild, all objects are captured in static configurations, whereas humans + +and robots often perceive objects through dynamic interactions. In order to test generalization to audio from such natural dynamic interactions, we use clips from the EPIC-KITCHENS-100 dataset [7] of humans using their kitchens naturally, and select clips where a human impacts an object against a table or another object and use the audio embedding of the impact sound to prompt the Detic model. We show one such result in Figure 6 and additional results in Figure 12 of Appendix E.4. We include the original video clips with audio in our supplementary video. The detector mostly selects either the correct item or items of similar materials and acoustic properties. ImageBind does not provide weights from this task for comparison. + +# 6. Conclusion and Limitations + +We introduced X-Capture, an open-source and low-cost device for collecting multi-sensory data in the wild. Using X-Capture, we collected a sample dataset of correlated RGB, depth, audio, and tactile readings of 3,600 points from 600 objects in natural environments, enabling direct benchmarking of cross-sensory encoding frameworks and loss functions on retrieval and contact localization tasks. Our results suggest that cross-sensory representations can be strengthened by learning from object-centric data correlating as many sensory modalities as possible, and that pretraining on this data yields valuable representations that can be fine-tuned to improve performance on other object-centric tasks. However, a limitation of X-Capture is that it captures objects in static configurations and environments, whereas humans and robots learn about objects interactively and dynamically while manipulating them. We hope our work inspires new, perhaps automated, collection efforts to further scale up multi-sensory learning from real objects. + +Acknowledgments. We thank Roger Clarke, Ryan Williams, Anirudh Jain, Mark Rau, and Fernando Lopez-Lezcano for advice with the hardware design, Klemen Kotar, Le Xue, Stephen Tian, and Weiyu Liu for valuable conceptual discussions, and Andrej Krevl and Matt Wright for their facility support. This work is in part supported by NSF CCRI #2120095 and RI #2338203 and ONR MURI N00014-22-1-2740. S. Clarke is supported by the Meta PhD Fellowship. + +# References + +[1] Bo Ai, Stephen Tian, Haochen Shi, Yixuan Wang, Cheston Tan, Yunzhu Li, and Jiajun Wu. Robopack: Learning tactile-informed dynamics models for dense packing. arXiv preprint arXiv:2407.01418, 2024. 3 +[2] Tapomayukh Bhattacharjee, Joshua Wade, Yash Chitalia, and Charles C Kemp. Data-driven thermal recognition of contact with people and objects. In 2016 IEEE Haptics Symposium (HAPTICS), pages 297-304. IEEE, 2016. 3 +[3] Roberto Calandra, Andrew Owens, Manu Upadhyaya, Wenzhen Yuan, Justin Lin, Edward H Adelson, and Sergey Levine. The feeling of success: Does touch sensing help predict grasp outcomes? In Conference on Robot Learning, pages 314-323. PMLR, 2017. 3, 2 +[4] Ning Cheng, Changhao Guan, Jing Gao, Weihao Wang, You Li, Fandong Meng, Jie Zhou, Bin Fang, Jinan Xu, and Wenjuan Han. Touch100k: A large-scale touch-language-vision dataset for touch-centric multimodal representation. arXiv preprint arXiv:2406.03813, 2024. 2 +[5] Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. In Proceedings of Robotics: Science and Systems (RSS), 2024. 3 +[6] Samuel Clarke, Ruohan Gao, Mason Wang, Mark Rau, Julia Xu, Jui-Hsien Wang, Doug L James, and Jiajun Wu. Realimpact: A dataset of impact sound fields for real objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1516-1525, 2023. 2, 3, 4, 6 +[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 8 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[9] Won Kyung Do, Bianca Jurewicz, and Monroe Kennedy. Densetact 2.0: Optical tactile sensor for shape and force reconstruction. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 12549-12555. IEEE, 2023. 4 + +[10] Yiming Dou, Fengyu Yang, Yi Liu, Antonio Loquercio, and Andrew Owens. Tactile-augmented radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26529-26539, 2024. 3 +[11] Letian Fu, Gaurav Datta, Huang Huang, William Chung-Ho Panitch, Jaimyn Drake, Joseph Ortiz, Mustafa Mukadam, Mike Lambeta, Roberto Calandra, and Ken Goldberg. A touch, vision, and language dataset for multimodal alignment. In Forty-first International Conference on Machine Learning, 2024. 2, 3, 5 +[12] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36, 2024. 2 +[13] Chuang Gan, Jeremy Schwartz, Seth Alter, Damian Mrowca, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, et al. Threadworld: A platform for interactive multi-modal physical simulation. arXiv preprint arXiv:2007.04954, 2020. 2 +[14] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations. arXiv preprint arXiv:2109.07991, 2021. 2 +[15] Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. Objectfolder 2.0: A multisensory object dataset for sim2real transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10598–10608, 2022. 2, 3 +[16] Ruohan Gao, Yiming Dou, Hao Li, Tanmay Agarwal, Jeannette Bohg, Yunzhu Li, Li Fei-Fei, and Jiajun Wu. The objectfolder benchmark: Multisensory learning with neural and real objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17276–17286, 2023. 2, 3, 4, 6, 7 +[17] Ruohan Gao, Hao Li, Gokul Dharan, Zhuzhu Wang, Chengshu Li, Fei Xia, Silvio Savarese, Li Fei-Fei, and Jiajun Wu. Sonicverse: A multisensory simulation platform for embodied household agents that see and hear. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 704–711. IEEE, 2023. 2 +[18] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 5 +[19] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 2, 5, 6, 8 +[20] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778, 2021.5 + +[21] Huy Ha, Yihuai Gao, Zipeng Fu, Jie Tan, and Shuran Song. Umi-on-legs: Making manipulation policies mobile with manipulation-centric whole-body controllers. In 8th Annual Conference on Robot Learning, 2024. 3 +[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 8 +[23] Erik Helmut, Luca Dziarski, Niklas Funk, Boris Belousov, and Jan Peters. Learning force distribution estimation for the gelsight mini optical tactile sensor based on finite element analysis. arXiv preprint arXiv:2411.03315, 2024. 4 +[24] Yining Hong, Zishuo Zheng, Peihao Chen, Yian Wang, Junyan Li, and Chuang Gan. Multiply: A multisensory object-centric embodied large language model in 3d world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26406-26416, 2024. 2 +[25] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis. arXiv preprint arXiv:2405.07987, 2024. 1 +[26] Roland S Johansson and J Randall Flanagan. Coding and use of tactile signals from the fingertips in object manipulation tasks. Nature Reviews Neuroscience, 10(5):345-359, 2009. 1 +[27] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 7, 8 +[28] Justin Kerr, Huang Huang, Albert Wilcox, Ryan Hoque, Jeffrey Ichnowski, Roberto Calandra, and Ken Goldberg. Self-supervised visuo-tactile pretraining to locate and follow garment features. In CVPR, 2023. 3, 2 +[29] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6 +[30] Mike Lambeta, Po-Wei Chou, Stephen Tian, Brian Yang, Benjamin Maloon, Victoria Rose Most, Dave Stroud, Raymond Santos, Ahmad Byagowski, Gregg Kammerer, et al. Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation. IEEE Robotics and Automation Letters, 5(3):3838-3845, 2020. 3, 2 +[31] Yunzhu Li, Jun-Yan Zhu, Russ Tedrake, and Antonio Torralba. Connecting touch and vision via cross-modal prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10609-10618, 2019. 3 +[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 8 +[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5 +[34] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H Adelson, and William T Freeman. Visually + +indicated sounds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2405-2413, 2016. 3, 2 +[35] Dinesh K Pai, Kees van den Doel, Doug L James, Jochen Lang, John E Lloyd, Joshua L Richmond, and Som H Yau. Scanning physical interaction behavior of 3d objects. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 87-96, 2001. 2, 3 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5 +[37] Mark Rau, Orchisama Das, and Elliot K Canfield-Dafilou. Improved carillon synthesis. In Proceedings of the 22nd international conference on digital audio effects, Birmingham, UK, pages 1-8, 2019. 6 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 7, 8 +[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 2 +[41] Nur Muhammad Mahi Shafiullah, Anant Rai, Haritheja Etukuru, Yiqian Liu, Ishan Misra, Soumith Chintala, and Lerrel Pinto. On bringing robots home. arXiv preprint arXiv:2311.16098, 2023. 3 +[42] Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters, 5(3):4978-4985, 2020. 3 +[43] Sudharshan Suresh, Haozhi Qi, Tingfan Wu, Taosha Fan, Luis Pineda, Mike Lambeta, Jitendra Malik, Mrinal Kalakrishnan, Roberto Calandra, Michael Kaess, et al. Neuralfeels with neural fields: Visuotactile perception for in-hand manipulation. Science Robotics, 9(96):eadl0628, 2024. 3 +[44] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 2, 5 +[45] Fengyu Yang, Chenyang Ma, Jiacheng Zhang, Jing Zhu, Wenzhen Yuan, and Andrew Owens. Touch and go: Learning from human-collected vision and touch. arXiv preprint arXiv:2211.12498, 2022. 3, 2 + +[46] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, et al. Binding touch to everything: Learning unified multimodal tactile representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26340-26353, 2024. 2, 6 +[47] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 6 +[48] Sarah Young, Dhiraj Gandhi, Shubham Tulsiani, Abhinav Gupta, Pieter Abbeel, and Lerrel Pinto. Visual imitation made easy. In Conference on Robot Learning, pages 1992-2005. PMLR, 2021. 3 +[49] Samson Yu, Kelvin Lin, Anxing Xiao, Jiafei Duan, and Harold Soh. Octopi: Object property reasoning with large tactile-language models. arXiv preprint arXiv:2405.02794, 2024. 2 +[50] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19313-19322, 2022. 5 +[51] Wenzhen Yuan, Siyuan Dong, and Edward H Adelson. Gelsight: High-resolution robot tactile sensors for estimating geometry and force. Sensors, 17(12):2762, 2017. 4 +[52] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European conference on computer vision, pages 350-368. Springer, 2022. 8 \ No newline at end of file diff --git a/xcaptureanopensourceportabledeviceformultisensorylearning/images.zip b/xcaptureanopensourceportabledeviceformultisensorylearning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..72f6fb9392aa67992d751ceeefc79761207eb629 --- /dev/null +++ b/xcaptureanopensourceportabledeviceformultisensorylearning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97ef9fa16bd96637541566754308b5ac167c6a63416024fb41432fe6f0f7e1cf +size 560924 diff --git a/xcaptureanopensourceportabledeviceformultisensorylearning/layout.json b/xcaptureanopensourceportabledeviceformultisensorylearning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bb5039525e44142db9e7d98fffc4f03b53681630 --- /dev/null +++ b/xcaptureanopensourceportabledeviceformultisensorylearning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7cc024cadd14c1222cfc34492544c7b36a0757bace85addaefad69c50089d851 +size 320242 diff --git a/xdancerexpressivemusictohumandancevideogeneration/893438c0-43d3-4ba2-a298-9c386258256a_content_list.json b/xdancerexpressivemusictohumandancevideogeneration/893438c0-43d3-4ba2-a298-9c386258256a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..083d09d00ffacdd7f66c8e91917f37cd850a328c --- /dev/null +++ b/xdancerexpressivemusictohumandancevideogeneration/893438c0-43d3-4ba2-a298-9c386258256a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dddd49f13806601d6a522456616a32737c531aa000b93e9dd6a90ca10f8965b +size 73840 diff --git a/xdancerexpressivemusictohumandancevideogeneration/893438c0-43d3-4ba2-a298-9c386258256a_model.json b/xdancerexpressivemusictohumandancevideogeneration/893438c0-43d3-4ba2-a298-9c386258256a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..19a4ae4913e6ee94a3c7fff15113bd9148619e45 --- /dev/null +++ b/xdancerexpressivemusictohumandancevideogeneration/893438c0-43d3-4ba2-a298-9c386258256a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db17537d0a5cd6d470fac0f5c81ad51c2d334fbcf1120d57fd488fde3f891094 +size 93706 diff --git a/xdancerexpressivemusictohumandancevideogeneration/893438c0-43d3-4ba2-a298-9c386258256a_origin.pdf b/xdancerexpressivemusictohumandancevideogeneration/893438c0-43d3-4ba2-a298-9c386258256a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..14b153b34d43f5a8545236ead21fc44c8ee09bd1 --- /dev/null +++ b/xdancerexpressivemusictohumandancevideogeneration/893438c0-43d3-4ba2-a298-9c386258256a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3fa34bb1415b4abcfe5e12870a4d72068e879d53b3a3e74715e92532d0139c9 +size 5665134 diff --git a/xdancerexpressivemusictohumandancevideogeneration/full.md b/xdancerexpressivemusictohumandancevideogeneration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1d09371869d7104a7c92b005a57b1f4610da9fa9 --- /dev/null +++ b/xdancerexpressivemusictohumandancevideogeneration/full.md @@ -0,0 +1,274 @@ +# X-Dancer: Expressive Music to Human Dance Video Generation + +Zeyuan Chen $^{1,2}$ Hongyi Xu $^{2}$ Guoxian Song $^{2}$ You Xie $^{2}$ Chenxu Zhang $^{2}$ Xin Chen $^{2}$ Chao Wang $^{2}$ Di Chang $^{2,3}$ Linjie Luo $^{2}$ $^{1}$ UC San Diego $^{2}$ ByteDance $^{3}$ University of Southern California + +![](images/92ecfc14529c8b0ea729fd6891a2168fbb6f5dfd172efb17e247355bc33a515f.jpg) +Figure 1. We present X-Dancer, a unified transformer-diffusion framework for zero-shot, music-driven human image animation from a single static image. Our approach accommodates diverse body forms and appearances, generating diverse and highly expressive synchronized full-body dance motions. It captures both fine-grained movements of the head and hands and large-scale actions such as body rotations and jumps, seamlessly translating them into vivid and lifelike dance videos. + +# Abstract + +We present X-Dancer, a novel zero-shot music-driven image animation pipeline that creates diverse and long-range lifelike human dance videos from a single static image. As its core, we introduce a unified transformer-diffusion framework, featuring an autoregressive transformer model that synthesizes extended and music-synchronized token sequences for 2D body, head and hands poses, which then guide a diffusion model to produce coherent and realistic dance video frames. Unlike traditional methods that primarily generate human motion in 3D, X-Dancer addresses data limitations and enhances scalability by modeling a + +wide spectrum of 2D dance motions, capturing their nuanced alignment with musical beats through readily available monocular videos. To achieve this, we first build a spatially compositional token representation from 2D human pose labels associated with keypoint confidences, encoding both large articulated body movements (e.g., upper and lower body) and fine-grained motions (e.g., head and hands). We then design a music-to-motion transformer model that autoregressively generates music-aligned dance pose token sequences, incorporating global attention to both musical style and prior motion context. Finally we leverage a diffusion backbone to animate the reference image with these synthesized pose tokens through AdaIN, + +forming a fully differentiable end-to-end framework. Experimental results demonstrate that X-Dancer is able to produce both diverse and characterized dance videos, substantially outperforming state-of-the-art methods in term of diversity, expressiveness and realism. See our project page for more results: zeyuan-chen.com/X-Dancer/ + +# 1. Introduction + +Dance is a universal form of self-expression and social communication, deeply embedded in human behavior and culture. With the rise of social media platforms like TikTok and YouTube Shorts, people increasingly share self-expressive dance videos online. However, creating expressive choreography typically demands practice and even professional training. From a computational aspect, generating realistic dance movements is challenging due to the freeform, personalized nature of dance and its alignment with musical rhythm and structure. In this work, we tackle the challenge of creating continuous, expressive and lifelike dance videos from a single static image, driven solely by a music track. + +This study addresses two key challenges in music-driven human image animation: (1) generating smooth, diverse full-body movements at finer scales that capture the complex, non-linear synchronization with music inputs, and (2) translating these generated body movements into high-fidelity video outputs that maintain visual consistency with the reference image and ensure temporal smoothness. Prior approaches [34, 39] mainly focus on computationally generating 3D human poses, such as SMPL skeletons [28], from music inputs, utilizing diffusion- or GPT-based frameworks. While these methods excel in producing high-quality, clean motions, they are constrained by limited training datasets (primarily the multi-view AIST dataset [23, 40]), which lack diversity, and contain 3D body poses only (excluding head and hands motions). Expanding these datasets with widely available 2D monocular dance videos would require 3D human pose estimation, which is often error-prone and risks degraded motion quality and consistency. Moreover, we target at photorealistic dance video generation rather than 3D skeleton or mesh animations. + +With recent advances in diffusion models, numerous works have leveraged their generative capabilities to synthesize visually compelling videos by animating a reference image with motion signals such as 2D skeletons [6, 16, 17], dense poses [49], and depth maps [60]. Unlike motion transfer settings that derive motion from a driving video, our goal is to generate motion signals that are consistent with the reference body shape and aligned with the input musical beats. Recently, a few studies [37, 48] have attempted to synthesize visual outputs end-to-end from audio inputs using diffusion models. While these methods have advanced in realism and dynamic quality, they still struggle to capture + +long-range motion and audio context due to high computational demands. Moreover, these frameworks have primarily been validated on audio-driven talking heads, leaving uncertain whether they can accommodate the complexities of full-body kinematics and rapid motion transitions. + +To this end, we propose X-Dancer, a unified framework that integrates an autoregressive transformer model for generating extended dance sequences attuned to input music, coupled with a diffusion model to produce high-resolution, lifelike videos. In contrast to prior methods focused on 3D music-to-dance motion generation, our approach models 2D human motion, leveraging widely accessible dance videos where 2D pose estimation is both reliable and scalable. For effective autoregressive motion generation, we develop a multi-part tokenization scheme for per-frame 2D whole-body poses, incorporating detected keypoint confidences to capture multi-scale human motions with motion blur and various occlusions. Thereafter we train a cross-modal transformer that auto-regressively predicts future N-frame pose tokens, paired with per-frame music features extracted with Jukebox [10] and Librosa [21]. Our design enables the model to capture a broader diversity of expressive, music-aligned movements with enhanced scalability in both model complexity and data scale. We then leverage a T2I diffusion model with temporal layers to animate the reference image by implicitly translating the generated confidence-aware motion tokens into spatial guidance via a trainable motion token decoder. Specifically, it integrates motion tokens through AdaIN [18] along upsampling from a learnable feature map into multi-scale spatial feature guidance. By co-learning pose translation with temporal and appearance reference modules [5, 13], the diffusion backbone interprets motion tokens with temporal context and body shape reference, leading to better shape-disentangled pose control, smoother and more robust visual outputs even under low-confidence or jittering pose sequences. This design establishes an end-to-end transformer-diffusion learning framework, merging the transformer's strengths in long-context understanding and generation with the diffusion model's capability in high-quality visual synthesis. + +To the best of our knowledge, X-Dancer represents the first music-driven human image animation framework. Trained on a large curated music-dance video dataset, our method excels at generating diverse, expressive and detailed whole-body dance videos attuned to input music, adaptable to both realistic and stylized human images across various body types. We extensively evaluate our model on challenging benchmarks, and X-Dancer outperforms state-of-the-art baselines both quantitatively and qualitatively. Additionally, we highlight its scalability and customization capabilities, showcasing scalability across varying model and data scales, as well as fine-tuning to characterized choreography. We summarize our contributions as follows, + +- A novel transformer-diffusion based music-to-dance human image animation, achieving state-of-the-art performance in terms of motion diversity, expressiveness, music alignment and video quality. +- A cross-modal transformer model that captures long-range synchronized dance movements with music features, employing a multi-scale tokenization scheme of whole-body 2D human poses with keypoint confidence. +- A diffusion-based human image animation model that interprets temporal pose tokens and translates them into consistent high-resolution video outputs. +- Demonstration of captivating zero-shot music-driven human image animations, along with characterized choreography generation. + +# 2. Related Work + +# 2.1. Music to Dance Generation + +Significant progress has been made in realistic human motion generation [8, 29, 34, 39, 55, 59] from various inputs, such as speech [2, 12, 27, 32, 54] or text [19, 31, 36, 57]. However, the task of music-to-dance generation [1, 22, 25, 26, 30, 34, 39, 46, 50, 53, 61] presents unique challenges: (1) ensuring the generated dance rhythmically aligns with the music, and (2) producing intricate motions with diverse styles and speeds. There has been several 3D human pose datasets [22-24, 30] proposed. The AIST++ dataset [23] is a 3D music pose dataset containing 1,408 dance motions tailored to various music genres. FineDance [24], a 3D dataset focusing on fine-grained hand motions, includes 14 hours of dance data. Models like Bailando [34] and EDGE [39] have leveraged AIST++ for training. Bailando [34] pioneered using a VQ-VAE for 3D pose encoding, followed by an Actor-Critic GPT model to generate body poses conditioned on music. EDGE [39] applied a diffusion framework to predict human poses from a noisy sequence, conditioning on music and employing long-form sampling for extended dance generation. DabFusion [50] explores image-to-video generation conditioned on music. Despite these advancements, existing approaches trained on 3D datasets often generate dance poses with limited diversity. X-Dancer overcomes these data constraints by leveraging a broad spectrum of 2D dance motions, aligning complex movements with musical beats, and utilizing widely available video data to enhance scalability. + +# 2.2. Human Image Animation + +With the advancements in diffusion models [4, 13, 15, 43, 47], generating high-quality human videos has become feasible. The introduction of ControlNet [45, 56] and PoseGuider [16, 17] has further empowered prior methods [6, 7, 16, 49, 60] to create realistic dance videos from a reference image using motion signals such as 2D skele + +tons, dense poses, and depth maps. However, generating expressive, temporally smooth 2D motion sequences that maintain consistency with reference body shapes and align well with music remains an open challenge. Several methods [20, 37, 48] adopt an end-to-end pipeline for generating realistic talking head videos from audio by injecting audio features directly into the network via cross-attention layers. While these methods excel in capturing micro-expressions and lip movements, they fall short in handling highly dynamic and articulated dance movements. X-Dancer implicitly incorporates generated motion tokens into a diffusion backbone to animate the reference image, enabling high-quality and shape-consistent dance video generation. + +# 3. Method + +Given a single portrait as the reference image $I_{R}$ and a conditioning music sequence represented as $M_{t}$ , our objective is to generate a sequence of dance frames $I_{M_t}$ , where $t = 1, \dots, T$ denotes the frame index. The generated sequence $I_{M_t}$ seeks to maintain the appearance and background context depicted in $I_{R}$ while present expressive dance movements in harmony with the provided musical rhythm and beats. As illustrated in Fig. 2, our model is trained in two stages: transformer-based 2D human dance motion generation (Section. 3.2), and diffusion-based video synthesis (Section. 3.3) from the generated motion sequence. + +# 3.1. Data Representations + +Video Generation. To generate human dance videos, we employ Latent Diffusion Models [33] that synthesize samples in the image latent space, facilitated by a pretrained image auto-encoder. During training, latent features of images are progressively corrupted with Gaussian noise $\epsilon$ , following the Denoising Diffusion Probabilistic Model (DDPM) framework [14, 35]. A UNet-based denoising framework, enhanced with spatial and temporal attention blocks, is then trained to learn the reverse denoising process. + +We apply human-centric cropping to the training videos of half and full-body dances, yielding a unified resolution of $896 \times 512$ . Rather than modeling intricate pixel-wise movements directly on music, we first establish the correlation between music and human dance movement, which subsequently guides the final visual synthesis. + +Motion Modeling. In contrast to prior methods that generate 3D human motions, we represent diverse dance motions as 2D pose sequences. Compared to its 3D counterparts, 2D human dance motions are widely accessible from large collections of monocular videos, eliminating the need for complex multiview capture setups or labor-intensive 3D animations. Furthermore, the 2D pose detection is significantly more reliable and robust. To enhance the realism and expressiveness, we model not only large body articulations but also finer details on head movements and hand gestures. + +![](images/17cc3dd5544c71c8ef001efa06c24f82ed653733db16b957070c2e547c965924.jpg) +Figure 2. Overview of X-Dancer. We propose a cross-conditional transformer model to autoregressively generate 2D human poses synchronized with input music, followed by a diffusion model to produce high-fidelity videos from a single reference image $I_R$ . First, we develop a multi-part compositional tokenization for 2D poses, encoding and quantizing each body part independently with keypoint confidence. A shared decoder merges these tokens into a whole-body, confidence-aware pose. Next, a GPT-based transformer autoregressively predicts future pose tokens with causal attention, conditioned on past poses and aligned music embeddings, as well as global music styles and prior motion context. A motion decoder is trained to generate multi-scale spatial pose guidance upsampled from a learned feature map, integrating the generated motion tokens within a temporal window (16 frames) via AdaIN. By co-training the motion decoder and temporal modules, our diffusion model is capable of synthesizing temporally smooth and high-fidelity video frames, while maintaining consistent appearance with the reference image with a trained reference net. + +Notably, we incorporate keypoint confidence into our pose representation, allowing the model to account for motion blur and occlusions. Each per-frame whole-body pose with $P$ joints is thus represented as $p \in \mathcal{R}^{P \times 3}$ , where the last dimension encodes the keypoint confidence. + +Music Embedding. Inspired from [39], we utilize the pretrained Jukebox model [10] to extract rich musical features, supplemented by rhythmic information with one-hot encoding of music beats using an audio processing toolbox Librosa [21]. We resample and synchronize these extracted embeddings—denoted as $F_{1:T}^{J}$ for Jukebox and $F_{1:T}^{L}$ for Librosa—to the video frame rate, ensuring per-frame alignment between the music and visual elements. + +# 3.2. Transformer-Based Music to Dance Motion + +Given a collection of monocular, single-person, music-synchronized dance videos with paired 2D whole-body pose detections, we aim to model the intricate, non-linear correlation between skeletal dance motion and musical features. To achieve this, we first introduce a compositional, confidence-aware VQ-VAE, which captures diverse and nuanced human dance poses across different body parts. Next, we leverage a GPT-like transformer to autoregressively predict future pose tokens, modeling temporal motion token transitions in synchronization with music embeddings. + +Compositional Confidence-Aware Human Pose Tok + +enization. Our approach builds on the standard VQ-VAE framework, trained in a self-supervised manner. Given whole-body 2D poses with associated keypoint confidences $p$ , a 1D convolutional encoder maps $p$ into a latent pose embedding $z_{e}(p) = E(p)$ , which is then quantized by mapping to its nearest representation $z_{q}(p)$ within a learnable codebook, and finally the decoder $D$ reconstructs the pose $\hat{p}$ from $z_{q}(p)$ . However, as prior studies [34, 44, 52] suggest, the dependencies between spatial keypoints are complex, and a vanilla VQ-VAE often struggles to capture subtle pose details, such as finger movements and head tilts, due to information loss during quantization and multi-frequency nature of pose variations across different body parts. + +To improve expressive coverage, we train independent 2D pose encoders $E^{j}$ and learn $B = 5$ separate codebooks $Z_{q}^{j}$ for upper and lower half bodies, left and right hands, and head respectively, allowing the model to spatially decompose 2D whole-body pose variations across different frequencies. With pose partition, distinct body-part codes can be flexibly combined, enhancing the range of expressiveness represented within each individual codebook. To capture part-wise spatial correlations and ensure information flow across body parts, we concatenate the quantized pose latents and feed them into a shared decoder. The resulting embedding is then mapped to reconstructed keypoint coordinates through separate projection heads, enabling joint re + +construction while preserving nuanced part-specific details. + +We train encoder and decoder simultaneously with the compositional codebooks with the following loss function: + +$$ +\mathcal {L} _ {\mathrm {V Q}} = \sum_ {j = 1} ^ {B} \| \hat {p} ^ {j} - p ^ {j} \| _ {2} + \beta \sum_ {j = 1} ^ {B} \| \operatorname {s g} \left[ z _ {e} ^ {j} (p) \right] - z _ {q} ^ {j} (p) \| _ {2}, \tag {1} +$$ + +where $\mathrm{sg}$ is a stop gradient operation, and the second term is a commitment loss with a trade-off $\beta$ . We utilize exponential moving average (EMA) [42] to update codebook vectors and remove the embedding loss $\sum_{j=1}^{B} \|z_e^j(p) - \mathrm{sg}[z_q^j(p)]\|_2$ for more stable training. + +Cross-Conditional Autoregressive Motion Modeling. After training the compositional quantized multi-part codebooks, 2D human poses detected in our training videos can be represented as sequences of codebook tokens via encoding and quantization. Specifically, each detected pose is mapped to a sequence of corresponding codebook indices, structured as one-hot vectors indicating the nearest codebook entry for each element. We denote this as $C_{1:T}^{j} = ((c_{1,1}^{j},\ldots ,c_{K,1}^{j}),\ldots ,(c_{1,T}^{j},\ldots ,c_{K,T}^{j}))$ , where $K$ is the number of tokens per body part $j$ in each frame. + +With this quantized motion representation, we design a temporal autoregressive model to predict the multinomial distributions of possible subsequent poses for each body part, conditioned on both the input music embeddings $F_{1:T}^{J}$ and $F_{1:T}^{L}$ and the preceding motion tokens. Our motion generation transformer is conditioned on the music inputs in two ways. First, we combine the Jukebox and Librosa music embeddings to form starting tokens, which serves as a global music context, such as styles and genres, that informs the entire motion sequence generation. Second, inspired by [34], we concatenate the frame-wise projected music embeddings with the corresponding motion tokens as inputs to the transformer model, ensuring precise synchronization between the motion and music features. This dual conditioning allows our model to produce both globally coherent and locally synchronized dance movements. We use $T = 64$ frames for our autoregressive model training. To handle extended motion sequence generation with consistent motion styles and smooth transitions, we additionally incorporate cross-segment motion context into the transformer model. Specifically, we uniformly sample a subset of 8 frames from the previous motion segment as the motion context, and append them after the global music context inputs. We denote the combined context as $F_{g}$ . + +Since we model body parts independently, maintaining coherence in the assembled whole-body poses is essential to avoid asynchronous movements (e.g., upper and lower body moving in different directions). To address this, we leverage mutual information across multi-part motions, designing our model with cross-conditioning across body parts. Specifically, we employ a GPT model to estimate the joint distributions of $C_{1:T}^{j}$ as follows, + +$$ +\phi (C _ {1: T} ^ {1: B} | F _ {g}) = \prod_ {t = 1} ^ {T} \prod_ {j = 1} ^ {B} \prod_ {k = 1} ^ {K} +$$ + +$$ +\phi \left(c _ {k, t} ^ {j} \mid c _ {1: K, < t} ^ {1: B}, c _ {1: K, t} ^ {< j}, c _ {< k, t} ^ {j}, F _ {\leq t} ^ {J}, F _ {\leq t} ^ {L}, F _ {g}\right) \tag {2} +$$ + +We structure the cross-conditioning between body parts in two ways: (1) the current motion token is conditioned on all preceding motion information from all body parts, ensuring inter-part temporal coherence; (2) by ordering body parts as upper and lower body, head, and hands, we build the hierarchical dependencies from primary components (upper/lower body) to finer, high-frequency movements (head and hands). Since each body part's pose is represented with a small set of tokens, we empirically observe that causal attention is sufficient to model the next-token distribution within each part. This modeling strategy preserves motion coherence of each body part as a whole, producing expressive and plausible dance movements. + +Our GPT model is optimized through supervised training using a cross-entropy loss on the next-token probabilities. Notably, because our pose tokenization includes associated keypoint confidences, the transformer also learns to model temporal variations in confidence, such as those caused by motion blur and occlusions, enabling it to capture more realistic motion distributions observed in videos. + +# 3.3. Diffusion-Based Dance Video Synthesis + +We employ a diffusion model to synthesize high-resolution, realistic human dance videos, conditioned on the generated motions from our trained transformer together with a given reference image. To achieve this, we leverage a pretrained T2I diffusion backbone [33] and incorporate additional temporal modules [13] for temporal consistency. For transferring the reference image context, a reference network, as a trainable copy of the backbone UNet, extracts reference features of identity appearance and background which are cross-queried by the self-attention within the backbone UNet. Motion control is achieved through an addition module, often configured in ControlNet [56] or light-weight PoseGuider [16, 17], which translates the motion conditions into 2D spatial guidance additive to the UNet features. + +To incorporate the generated motion tokens into human image animation, one approach is to utilize the trained VQ-VAE decoder $D$ to decode pose tokens into keypoint coordinates, which are then visualized as 2D skeleton maps for motion conditioning. These skeleton maps can provide motion guidance to the final diffusion synthesis through a PoseGuider module. While effective to some extent, this approach introduces a non-differentiable process due to the skeleton map visualization, which impedes end-to-end training and often results in the loss of keypoint confidence information. Additionally, since the temporal module is trained on real video data where pose sequences are typically smooth, it may struggle with jittery or inconsistent + +![](images/61a48a52d282df0db3423d00b806f72550a8d98b64ae4b9a2954722fbbe6a11c.jpg) + +![](images/f7d0fd535c4f216d3f4fb7677732341009263279d72ad5d0194bf5d5f13160ff.jpg) +Reference + +![](images/8c6c3453ff02a75b806554f29c99b25a055119b8c30974ad16451679d82893e5.jpg) + +![](images/64bdafc93c040c4e1750ef71343c992664a07810b3fb8f978c7c5e8d5f2b6190.jpg) +X-Dancer + +![](images/ff62170b30adf095537d41a597c041322398f425b0354335276d7637c0662142.jpg) + +![](images/0ef59d0fbbfa49075ecb1629ff77b307349ef47b3f92ad7114448efc8ddcf1e1.jpg) +Hallo +Figure 3. Qualitative Comparisons. Among all the methods, X-Dancer achieves the most expressive and high-fidelity human dance video synthesis, maintaining the highest consistency with both the reference human characteristics and the background scene. + +![](images/f424ffc937018dcfacd18dcd45cb9589caebfec52052bb93d96d279f7129dc01.jpg) + +![](images/c992601744a67e38c104dc71d0ec12e9d1c9a1caa631cbf22fca06f1b68f8699.jpg) +Bailando + Pose Guider + +![](images/d6dbbf06679427225a4fb4d1126a28e9619c66c599f141c7567bb7130eb2d9fe.jpg) + +![](images/2bcbf1765b66ed2c6df514c547027456260c452fa9a50a3b9a4ede3f365bbb9c.jpg) +EDGE + Pose Guider + +motions generated by the transformer model at inference. + +In place of explicit translation of motion tokens into 2D skeleton maps as pose guider conditions, we introduce a trainable motion decoder that implicitly and directly translates the 1D pose tokens into 2D spatial guidance, added towards the intermediate features of the denoising UNet. Starting from a learnable 2D feature map, the decoder injects the 1D pose token sequences including the keypoint confidences with AdaIN layers [18], progressively upsampling this feature map into multiple scales aligned with the resolutions of the denoising UNet features. The motion decoder is trained alongside the temporal module within a 16-frame sequence window, effectively decoding token sequences into continuous pose guidance that integrates temporal context from adjacent frames. Moreover, by incorporating reference image context during training, we observe empirically that the decoded pose guidance retains minimal identity and body shape information compared to explicit 2D poses, enabling generated pose tokens to adapt seamlessly to subjects with varied body shapes and appearances. + +# 4. Experiments + +# 4.1. Implementation Details + +Dataset. Our model is trained on a curated in-house visual-audio dataset of 76,818 monocular dance recordings in indoor and outdoor settings, averaging 15 seconds per clip. Each video is cropped to $896 \times 512$ resolution and resampled at 30 fps, covering half- to full-body dances across diverse music and performer characteristics. Details on the dataset are provided in the supplementary paper. + +Training. We train our full pipeline in three stages on 8 A100 GPUs using the AdamW optimizer [51]. First, we train a multi-part pose VQ-VAE to encode and quantize 60 joint coordinates and keypoint confidences, into 5-part pose tokens. Each body part's pose uses 6 tokens, each with a unique 512-entry codebook of 6D embeddings. This VQ- + +VAE is trained for 40k steps with a batch size of 2048 at a learning rate of $2 \times 10^{-4}$ . Next, we train an autoregressive model for pose token prediction over 300k steps, initialized with pretrained GPT-2 weights, using a batch size of 24 and a learning rate of $1 \times 10^{-4}$ . The model operates on 64-frame pose sequences with a 2224-token context window. Lastly, we fine-tune the denoising UNet from SD1.5 [33] and ReferenceNet with two randomly selected frames per video, followed by co-training the motion token decoder and temporal module [13] with diffusion losses on consecutive 16 frames. The diffusion stage is trained for 90k steps at a learning rate of $1 \times 10^{-5}$ and a batch size of 16. + +Inference. We initiate our autoregressive dance motion generation from pose tokens encoded from the reference image pose. Extended dance sequences are then generated in 64-frame sliding segments with a 12-frame overlap, while 8 uniformly sampled frames from last segment serve as global motion context. We synthesize all video frames simultaneously with the full pose token sequence, applying prompt traveling [38] to improve temporal smoothness. + +# 4.2. Evaluations and Comparisons + +To the best of our knowledge, no existing approach addresses music-driven 2D video generation from a single human image. We adapted and combined established models to create two baselines for comparisons. First, we adapted the audio-driven, diffusion-based portrait animation model Hallo [48], retraining it for our task by substituting its audio features with our music embeddings to animate human images via cross-attention modules. For the second baseline, we utilize 3D music-to-dance generation models like Bailando [34] and EDGE [39] for motion synthesis, projecting their outputs into 2D pose maps, which are then fed into a diffusion model with a pose guider for controlled image animation. For fair comparison, we also train our motion transformer (stage 2, X-Dancer-AIST) on the AIST dataset [40], consistent with Bailando and EDGE, but using detected 2D + +Table 1. Quantitative comparison on motion generation. +Table 2. Quantitative comparison on visual synthesis. + +
MetricsAIST++ Dataset / In-House DatasetMetricsIn-House Dataset
FVD ↓DIV ↑BAS ↑FVD ↓FID-VID ↓ID-SIM ↑
Ground Truth509.58/129.7534.10/29.670.24/0.22Hallo [48]609.0876.990.4870
Hallo [48]548.81/249.1228.66/28.980.16/0.20Bailando [34] + PG583.26100.020.3392
Bailando [34]621.22/534.0222.34/24.050.19/0.19EDGE [39] + PG613.8193.730.3034
EDGE [39]639.46/303.3624.87/27.290.26/0.24Our motion + PG735.0572.710.4894
X-Dancer-AIST620.73/309.1825.25/24.310.26/0.22X-Dancer-AIST549.3874.060.3652
X-Dancer531.52/238.2225.61/28.080.23/0.21X-Dancer507.0661.940.5317
+ +Table 3. Ablation on our transformer model designs. + +
FVD ↓DIV ↑BAS ↑
w/o global music context265.7327.040.2142
w/o global motion context247.5426.420.2154
sub-dataset + GPT-medium402.6324.400.2112
sub-dataset + GPT-small332.9324.580.2046
X-Dancer238.2228.080.2182
+ +poses instead of 3D skeletons. However, since AIST has limited diversity in identities and appearances, we use our diffusion model—trained on our curated dataset—for comparison on synthesized videos. We conduct separate evaluations of all models on the test split of AIST (40 videos) as well as our curated music-video dataset (230 videos). + +Quantitative Evaluation. We numerically compare X-Dancer with baselines in terms of quality of both motion generation and video synthesis. Specifically, we calculate the Fréchet Video Distance (FVD) [11, 41] between generated dance motions and ground-truth training sequences for assessment of motion fidelity. For motion diversity, we compute the distributional spread of generated pose features (DIV) [23, 34], To numerically evaluate the alignment between the music and generated dancing poses, we follow [23, 34] to measure the Beat Align Score (BAS) by the average temporal distance between each music beat and its closest dancing beat. These evaluations are conducted in 2D pose space, where we detect poses over synthesized videos from the retrained Hallo model, and project the 3D skeleton motion of Bailando and EDGE into 2D via orthographic projection, using camera parameters that best align with the human bounding box in the reference image. + +As shown in Tab. 1, our method surpasses all baselines in terms of motion fidelity (FVD) and music beat alignment (BAS), while achieving the second-best diversity (DIV). Notably, Bailando and EDGE is trained on professional dances (AIST++ [23]), whereas our dataset comprises videos of everyday individuals, reflected in lower BAS even for ground-truth videos. Our method trained on AIST (X-Dancer-AIST) achieves the highest BAS whereas our model trained on our curated videos still significantly outperforms + +Table 4. Ablation on VQ-VAE and video diffusion. PG and MD denotes PoseGuider and Motion Decoder respectively. + +
Pose L1 FullBody/Head/Hands ↓
Single-Part VQVAE0.83 / 0.64 / 0.52
Multi-Part VQVAE0.50 / 0.40 / 0.42
Video PSNR ↑ / LPSIS ↓ / FVD ↓
GT Pose + PG19.465 / 0.197 / 294.52
Single-Part + MD17.079 / 0.258 / 350.73
Multi-Part + PG18.836 / 0.207 / 384.16
Multi-Part + MD19.148 / 0.207 / 295.87
+ +Bailando and Hallo for beat alignment. Hallo entangles motion generation and video synthesis, and achieves a higher DIV score than X-Dancer primarily due to its extremely noisy video outputs which results in jittering and chaotic skeleton motions following pose projection. + +For evaluation of video synthesis fidelity, we measure the FVD and FID-VID [3] score between the ground-truth and generated dance videos. Additionally, we assess identity preservation using the ArcFace score [9], which measures the cosine similarity of face identity features (ID-SIM). All metrics are evaluated on our test video dataset. As an extra baseline, we replace the motion token decoder in our full pipeline with a pose guider (Our motion + PG). As shown in Tab. 2, our method achieves the highest visual quality and identity preservation, which we attribute to our disentangled design of motion generation and video synthesis (compared to Hallo) and the use of an implicit motion token decoder rather than an explicit pose guider. + +Qualitative Comparisons. We present visual comparisons between our method and the baselines in Fig. 3. For more dynamic comparisons and user study, please refer to our supplementary material. The modified Hallo [48] represents an end-to-end diffusion pipeline that directly synthesizes the final video without intermediate motion generation. However, it exhibits noticeable artifacts, particularly in large articulated body motions, and often fails to preserve the human body's articulation topology. Bilando [34] and EDGE [39] generate body motion in 3D space without con + +![](images/1eb907a536244cb94803f7e0dc2184cf90548436f9e4e7891e914a383f8233e2.jpg) +Figure 4. Our Multi-part VQ-VAE captures intricate poses such as hand gesture (left), arm bending, head tilting and leg lifting(right) whereas single-part pose tokenization fails to fully replicate them. + +sidering the scene context or the human shape in the reference image. Despite post-processing for skeleton alignment in the pose guider input, these methods still struggle with significant identity and shape distortions, often producing unnatural interactions with the background scene. Furthermore, they do not model head and hand motions, leading to more rigid and less expressive dance movements compared to X-Dancer. We also note that our model trained on monocular dance videos is efficient and versatile, easily adapting to specific dances. This is not easily achievable with EDGE or Bailando, which require extensive effort in crafting 3D skeleton dances (supplementary paper Sec.3) + +# 4.3. Ablation Study + +We conduct ablation studies by systematically removing individual design components from our full training pipeline. + +Multi-part VQ-VAE. As shown in Tab. 4, our compositional multi-part VQ-VAE significantly reduces pose reconstruction error compared to single-part whole-body pose tokenization. The multi-part VQ-VAE better captures fine-scale pose variations, while single-part tokenization often loses high-frequency details, resulting in more rigid and less expressive motion generation. To further assess its impact, we train a video generation diffusion model using single-part pose tokenization. As illustrated in Figure 4, the single-part representation struggles to control fine-grained motions like hip swaying, leg lifting and head tilting, leading to perceptually lower-quality videos. This is further supported by its inferior video reconstruction metrics (PSNR and LPIPS [58]) in Tab. 4. + +Motion Generation. Next, we assess the impact of global music and motion context on motion generation. As presented in Tab. 3, both contexts contribute to producing more consistent, plausible, and music-synchronized motions. We further analyze the benefits of 2D motion modeling and transformer-based autoregressive generation by scaling both model parameters and dataset size. Across all metrics, we observe significant performance gains (Tab. 3) as the number of training parameters increases from 117M (GPT-small) to 345M (GPT-medium) and data scale from 10k to 100k videos, underscoring the scalability potential of + +monocular dance video data with our architecture, shedding light on further performance improvements as they scale. + +Pose-Guided Video Synthesis. In Tab. 2, we compare our motion token decoder (X-Dancer) against a pose guider using explicitly decoded skeleton map (Our motion + PG). While our transformer-generated motion exhibits jittering, our motion token decoder significantly reduces jitter and enhances temporal consistency by leveraging temporal motion context. Additionally, it demonstrates superior identity and body shape preservation compared to the pose guider. + +In Tab.4, we evaluate the effectiveness of our motion token decoder (Multi-Part + MD) against the pose guider (Multi-Part + PG) in self-driven video synthesis. Our motion decoder directly infers body pose coordinates and confidences from pose tokens, achieving lower reconstruction errors (PSNR and LPIPS) compared to the pose guider, which relies on explicitly decoded pose maps. However, due to motion blurriness in fast dance motions and fine-grained structural variations in small pixel regions like the face and hands, visual rendering artifacts remain present in both methods. Notably, these artifacts stem primarily from the diffusion model's limitations rather than errors in pose reconstruction, as evidenced by the similar video error observed in GT Pose+PG (Tab. 4). + +# 5. Conclusion + +We present X-Dancer, a novel framework that unites an autoregressive transformer with a diffusion model to generate high-quality, music-driven human dance videos from a single reference image. Unlike prior works, X-Dancer models and generates dance movements in 2D space, harnessing widely accessible 2D poses from monocular dance videos to capture diverse, expressive whole-body motions. Our method achieves state-of-the-art results in video quality, motion diversity and expressiveness, providing a scalable and adaptable solution for creating vivid, music-aligned dance videos across various human forms and styles. + +Limitations and Future Work. Our model is trained solely on curated real-human daily dance videos, which can be noisy and lack the motion precision found in professional dancer videos. Consequently, out-of-domain human images may lead to rendering artifacts, and the generated dance motions may occasionally lack music alignment. More failure cases are present in the supplementary paper. While we designed our pipeline to be end-to-end trainable and scalable, we currently implement it in stages due to memory limitations. Future work will explore large-scale, multi-machine training to further enhance performance and efficiency. + +Ethics Statement. Our work aims to improve human image animation from a technical perspective and is not intended for malicious use like fake videos. Therefore, synthesized videos should clearly indicate their artificial nature. + +# References + +[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM TOG, 2023. 3 +[2] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM TOG, 2022. 3 +[3] Yogesh Balaji, Martin Renqiang Min, Bing Bai, Rama Chellappa, and Hans Peter Graf. Conditional gan with discriminative filter generation for text-to-video synthesis. In IJCAI, 2019. 7 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465, 2023. 2 +[6] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In ICML, 2023. 2, 3 +[7] Di Chang, Hongyi Xu, You Xie, Yipeng Gao, Zhengfei Kuang, Shengqu Cai, Chenxu Zhang, Guoxian Song, Chao Wang, Yichun Shi, et al. X-dyna: Expressive dynamic human image animation. arXiv preprint arXiv:2501.10021, 2025.3 +[8] Ling-Hao Chen, Jiawei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In ICCV, 2023. 3 +[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 7 +[10] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020. 2, 4 +[11] Songwei Ge, Aniruddha Mahapatra, Gaurav Parmar, JunYan Zhu, and Jia-Bin Huang. On the content bias in fréchet video distance. In CVPR, 2024. 7 +[12] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In CVPR, 2019. 3 +[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 3, 5, 6 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020. 3 +[15] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35, 2022. 3 + +[16] Li Hu. Animate anyone: Consistent and controllable imaged-to-video synthesis for character animation. In CVPR, 2024, 2, 3, 5 +[17] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. 2, 3, 5 +[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 2, 6 +[19] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NeurIPS, 2023. 3 +[20] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 3 +[21] Yanghua Jin, Jiakai Zhang, Minjun Li, Yingtao Tian, Huachun Zhu, and Zhihao Fang. Towards the automatic anime characters creation with generative adversarial networks. arXiv preprint arXiv:1708.05509, 2017. 2, 4 +[22] Nhat Le, Thang Pham, Tuong Do, Erman Tijputra, Quang D Tran, and Anh Nguyen. Music-driven group choreography. In CVPR, 2023. 3 +[23] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In ICCV, 2021. 2, 3, 7 +[24] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In ICCV, 2023. 3 +[25] Ronghui Li, Yuqin Dai, Yachao Zhang, Jun Li, Jian Yang, Jie Guo, and Xiu Li. Exploring multi-modal control in music-driven dance generation. In ICASSP, 2024. 3 +[26] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In CVPR, 2024. 3 +[27] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In CVPR, 2022. 3 +[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851–866. 2023. 2 +[29] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In ECCV, 2022. 3 +[30] Zhenye Luo, Min Ren, Xuecai Hu, Yongzhen Huang, and Li Yao. Popdg: Popular 3d dance generation with popdanceset. In CVPR, 2024. 3 +[31] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In ECCV, 2022. 3 + +[32] Shenhan Qian, Zhi Tu, Yihao Zhi, Wen Liu, and Shenghua Gao. Speech drives templates: Co-speech gesture synthesis with learned templates. In ICCV, 2021. 3 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 5, 6 +[34] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In CVPR, 2022. 2, 3, 4, 5, 6, 7 +[35] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[36] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 3 +[37] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2, 3 +[38] Jonathan Tseng, Rodrigo Castellon, and C. Karen Liu. Edge: Editable dance generation from music, 2022. 6 +[39] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In CVPR, 2023. 2, 3, 4, 6, 7 +[40] Shuhei Tsuchida, Satoru Fukayama, Masahiro Hamasaki, and Masataka Goto. Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In ISMIR, 2019. 2, 6 +[41] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7 +[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 5 +[43] Yilin Wang, Zeyuan Chen, Liangjun Zhong, Zheng Ding, and Zhuowen Tu. Dolfin: Diffusion layout transformers without autoencoder. In ECCV, 2024. 3 +[44] Yuan Wang, Zhao Wang, Junhao Gong, Di Huang, Tong He, Wanli Ouyang, Jile Jiao, Xuetao Feng, Qi Dou, Shixiang Tang, and Dan Xu. Holistic-motion2d: Scalable whole-body human motion generation in 2d space. arXiv preprint arXiv.: 2024. 4 +[45] Yilin Wang, Haiyang Xu, Xiang Zhang, Zeyuan Chen, Zhizhou Sha, Zirui Wang, and Zhuowen Tu. Omnicontrol-net: Dual-stage integration for conditional image generation. In CVPR Workshops, 2024. 3 +[46] Shuang Wu, Shijian Lu, and Li Cheng. Music-to-dance generation with optimal transport. arXiv preprint arXiv:2112.01806, 2021. 3 +[47] Haiyang Xu, Yu Lei, Zeyuan Chen, Xiang Zhang, Yue Zhao, Yilin Wang, and Zhuowen Tu. Bayesian diffusion models for 3d shape reconstruction. In CVPR, 2024. 3 +[48] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao + +Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 6, 7 +[49] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In CVPR, 2024. 2, 3 +[50] Wang Xuanchen, Wang Heng, Liu Dongnan, and Weidong Cai. Dance any beat: Blending beats with visuals in dance video generation. In WACV, 2025. 3 +[51] Zhewei Yao, Amir Gholami, Sheng Shen, Mustafa Mustafa, Kurt Keutzer, and Michael Mahoney. Adalessian: An adaptive second order optimizer for machine learning. In AAAI, 2021. 6 +[52] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 4 +[53] Canyu Zhang, Youbao Tang, Ning Zhang, Ruei-Sung Lin, Mei Han, Jing Xiao, and Song Wang. Bidirectional autoregressive diffusion model for dance generation. In CVPR, 2024. 3 +[54] Chenxu Zhang, Chao Wang, Yifan Zhao, Shuo Cheng, Linjie Luo, and Xiaohu Guo. Dr2: Disentangled recurrent representation learning for data-efficient speech video synthesis. In WACV, 2024. 3 +[55] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 3 +[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 5 +[57] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 3 +[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 8 +[59] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In ICCV, 2023. 3 +[60] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 2, 3 +[61] Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2dance: Dancenet for music-driven dance generation. ACM TOMM, 2022. 3 \ No newline at end of file diff --git a/xdancerexpressivemusictohumandancevideogeneration/images.zip b/xdancerexpressivemusictohumandancevideogeneration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a0856bd35453b4f5ac7adf2e413ca171b94bf8d --- /dev/null +++ b/xdancerexpressivemusictohumandancevideogeneration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02dd01dae370e1fbd9998765cbbfe2bb2b3ad46422f74f0c3b5fff19d46d0fa5 +size 576477 diff --git a/xdancerexpressivemusictohumandancevideogeneration/layout.json b/xdancerexpressivemusictohumandancevideogeneration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3b1501c7b6e6ff3a4ebaf8c1fec8e9b2f43804e3 --- /dev/null +++ b/xdancerexpressivemusictohumandancevideogeneration/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9c03665ad4192907ef137fa5f06341b076458531784bf491127fe2b0d0b798a +size 337029 diff --git a/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/425c73c4-35d6-4f59-b075-35ebadf4b8b2_content_list.json b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/425c73c4-35d6-4f59-b075-35ebadf4b8b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cef8ef599c97a0825ffcff3e380ac16cb22cd6be --- /dev/null +++ b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/425c73c4-35d6-4f59-b075-35ebadf4b8b2_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:913b48276376d6c0596c1797246f1ad06b75d426aaaf06739753f61ab3186583 +size 87213 diff --git a/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/425c73c4-35d6-4f59-b075-35ebadf4b8b2_model.json b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/425c73c4-35d6-4f59-b075-35ebadf4b8b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..296e9cedee122193f3fc3fee51a700e4d81203d6 --- /dev/null +++ b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/425c73c4-35d6-4f59-b075-35ebadf4b8b2_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f05f30b5eaa19d21ded2717ce7efc375b70ab7cb78a435eaebe8adcf4978a5d +size 113927 diff --git a/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/425c73c4-35d6-4f59-b075-35ebadf4b8b2_origin.pdf b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/425c73c4-35d6-4f59-b075-35ebadf4b8b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..648746e26edbe48341ada274abd0eb866e5272a1 --- /dev/null +++ b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/425c73c4-35d6-4f59-b075-35ebadf4b8b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6b1ab0e083663f61147389f1a5bbc4b56652dcbe96494ad7c03b72e0b44ad0d +size 1471329 diff --git a/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/full.md b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2eedd44dc7efc1ba41d8e1642807860b96dfe865 --- /dev/null +++ b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/full.md @@ -0,0 +1,417 @@ +# X-Fusion: Introducing New Modality to Frozen Large Language Models + +Sicheng Mo $^{1}$ Thao Nguyen $^{2}$ Xun Huang $^{3}$ Siddharth Srinivasan Iyer $^{3}$ Yijun Li $^{3}$ Yuchen Liu $^{3}$ Abhishek Tandon $^{3}$ Eli Shechtman $^{3}$ Krishna Kumar Singh $^{3}$ Yong Jae Lee $^{2}$ Bolei Zhou $^{1}$ Yuheng Li $^{3}$ $^{1}$ University of California, Los Angeles $^{2}$ University of Wisconsin-Madison $^{3}$ Adobe Research + +https://sichengmo.github.io/XFusion/ + +# Abstract + +We propose X-Fusion, a framework that extends pretrained Large Language Models (LLMs) for multimodal tasks while preserving their language capabilities. X-Fusion employs a dual-tower design with modality-specific weights, keeping the LLM's parameters frozen while integrating vision-specific information for both understanding and generation. Our experiments demonstrate that X-Fusion consistently outperforms alternative architectures on both image-to-text and text-to-image tasks. We find that incorporating understanding-focused data improves generation quality, reducing image data noise enhances overall performance, and feature alignment accelerates convergence for smaller models but has minimal impact on larger ones. Our findings provide valuable insights into building efficient unified multimodal models. + +# 1. Introduction + +Large Language Models (LLMs) [1-13] have not only achieved unprecedented capabilities for language processing tasks (e.g., conversational AI [14-20]), but also emerged as foundation tools to solve multiple language-related challenges (e.g., coding [21-23]). However, as humans, we do not communicate solely through text, but also extend to other modalities, such as vision. For example, instead of simply saying "This dog is cute", we might show a photo of the dog to enhance the message: "[image] is cute" (Fig. 1). Thus, a truly versatile AI model should not only understand, reason, and generate textual output, but also must have abilities to understand, reason, and generate visual information. Moreover, these models should be unified to process and generate language and vision simultaneously, creating a more comprehensive interactive experience. + +To achieve a unified model, some approaches focus on training unified vision-language models entirely from scratch using next-token prediction loss [24-27]. A recent alternative, Transfusion [28], adopts a domain-specific strategy by combining next-token prediction loss for lan + +![](images/f22faea1af5c7c01f85c55b7c9e55424094bb00448a231969d5f7ebbe9309bd9.jpg) +Figure 1. We introduce X-Fusion - a novel framework that adapts pretrained LLMs (e.g., LLaMA) to new modalities (e.g., vision) while retaining their language capabilities and world knowledge. + +guage with diffusion loss for images. This hybrid architecture has significantly advanced performance, demonstrating greater promise than purely autoregressive approaches like Chameleon [26]. However, training such models from scratch demands immense computational resources (e.g., [28] trained on 2T tokens) and necessitates full retraining for each new modality. Given these shortcomings, another prominent research direction explores how to reuse powerful pretrained LLMs and introduce vision abilities to them [29-31], offering a more practical and efficient way for unified multimodal model training. + +Research on adapting LLMs [14-18] with image understanding has shown promising results through "visual instruction tuning" [32-35]. These models typically fine-tune the LLM to align the text feature space with pretrained vision encoders (e.g., CLIP [36]), thus, might degrade original language capabilities [37, 38]. Unlike image understanding, image generation poses greater difficulties, as it demands output capabilities in a new feature space. A large body of work [29, 30, 39-43] tackled this issue by leveraging pretrained image generation models (e.g., Stable Diffusion [44]). However, as these frameworks are not + +![](images/0e50d047078a33d06306a2682e544729fd1c1f0ab380d9db66e0d86ae1317de7.jpg) +The image features a classic bicycle with a large round clock on its front wheel, its face is white with black numbers and hands, indicating a time of 10:10. The bicycle appears to be stationary and the background is plain, allowing the bicycle and clock to be the focal point. The clock and bicycle are set against a white backdrop, emphasizing their details and shapes. + +>Input Image: +![](images/a59e456229124b43eec6253a7e1d5f9d8c00673531e63d59717d97e16343e5cd.jpg) +A French Bulldog stands on a bicycle, donning a black leather jacket, white-rimmed goggles, and a colorful bandana on its head. The canine is positioned as if ready to ride, with the bicycle's handlebars and part of the frame visible beneath it. The backdrop features a blue surface, suggesting the photo was taken outdoors, possibly at a casual gathering or event. + +>Input Image: +Figure 2. Captions generated by X-Fusion demonstrate high details and strong visual alignment with the image inputs. +![](images/198fbeed811b217f1919061fbd12465f9d8649dcac485468920dbf2b623cec15.jpg) +Two women stand close together, smiling for the camera, sharing a vibrant pink umbrella to shield them from the rain. The woman on the left is clad in a black coat and a pink scarf, while the woman on the right wears a white puffy coat. They appear cheerful and dressed for a day out in the city, with a backdrop featuring a historic building and other pedestrians navigating the wet streets. + +>Input Image: +![](images/3d3cf60956b47097a0e1c4bba8fb00f38af30f7dd75d1f7e2dd68d1538556391.jpg) +Curious feline with a striped coat is peeking from behind a beige cabinet door, its eyes wide and alert. The setting is indoors, with a muted, warm-colored wall and a textured rug underfoot. The cat's fur is well-groomed, and its expression suggests a sense of playfulness and interest in the surrounding environment. + +unified, this approach creates several limitations: limited cross-modal reasoning, restricted in-context learning, and increased error accumulation [45]. Most critically, these approaches typically require fine-tuning the LLM backbone, degrading inherited text generation ability [37, 38]. This raises fundamental research question: Is there a better way to introduce new modalities to pretrained LLMs? + +Motivated by these observations, we propose X-Fusion, a new approach that addresses two challenges: (i) retaining the language abilities of the pre-trained LLM while (ii) adapting it with image generation capabilities. First, X-Fusion freezes all language weights, denoted as the text tower, thus preserving the inherent language abilities. Second, instead of fine-tuning the LLM, we introduce a vision tower with separate vision weights in each layer to help process visual information for the LLM (Fig 1). This approach aligns text and vision features not only at input or output level, but also at the intermediate processing level. It is worth noting that this architecture is flexible in terms of design — the vision and text towers can have asymmetric architectures. Moreover, the framework naturally extends to additional modalities (e.g., audio) by introducing dedicated modality-specific towers, ensuring efficient and scalable integration while keeping each modality independent. + +While architectural innovations are crucial for integrating visual capabilities into LLMs, nowadays, understanding the impact of training data is equally important. Thus, we conduct a comprehensive set of ablation studies from a data-centric perspective to build a scalable training strategy. First, we note that creating image understanding samples without introducing noise to the images is crucial for training diffusion-based unified models. This approach enhances both image generation and understanding abilities due to more semantic visual representation learned from clean images. Based on this key component, we further observed cross task synergy—including more image under + +standing data that enhances generation performance. We also investigate the effectiveness of aligning our vision features with additional pre-trained representations. Our findings suggest that this extra alignment loss term may help smaller models converge faster, but the benefit diminishes as the model size increases. + +In short, our contributions are: (i) X-Fusion - a novel framework that adapts pretrained LLMs to new modalities while retaining their language capabilities. (ii) A systematic study on training strategy, offering insights for optimizing multi-modal learning. (iii) Experimental results on both image-to-text and text-to-image tasks, validating the effectiveness of the proposed architecture. + +# 2. Related Work + +Large Multimodal Models. The development of artificial intelligence has historically followed separate, modality-specific paths. Large Language Models (LLMs) [1-13] exclusively processed text input and generated textual responses, while computer vision models specialized in visual content understanding (e.g., object detection [46]) or visual generation (e.g., StyleGAN [47]). The emergence of vision encoder models like CLIP [36] bridged text and image modalities, enabling two key advances: (1) Vision-language models that allow LLMs to "see" (e.g., LLaVA [32]); and (2) Conditional visual generation models that can process textual input (e.g., Stable Diffusion [44]). Current research frontier are focusing on integrating vision and language capabilities into unified models that can both process and generate multimodal content. There are three main approaches: (1) Merging LLMs with pretrained image generation models (e.g., DreamLLM [29], GILL [30]); (2) Training LMMs via next-token prediction (e.g., Chameleon [26]); or (3) Training LMMs using both diffusion and next-token prediction losses (e.g., Transfusion [28]). Following the third approach, which has achieved state-of-the-art results across + +![](images/fc1548dcde2d0661b90d2125ca582f892057f3d28a70e100fd3af2422d0ea7a8.jpg) +A sentient AI robot holding a flower, contemplating human emotions. + +![](images/77054584337e1dd5e9f4c62dfbe5a19c85cafea78622a1855869073830f76692.jpg) + +![](images/ae5dc597e2c3b32c3eb062767fd707038bf89b6d880ccddd359caad120f42779.jpg) +An astronaut riding a camel on Mars during a sunset. + +![](images/a375d555956952eb58b95c3b4af3d01b206a1c39cd9b5fa18684ce769e3801d4.jpg) +A medieval castle floating upside down over a still lake. + +![](images/cbfabc11e1b518b45a2150622550b8ebb97e6e42feed3d032fbd8e9904f7adea.jpg) + +![](images/d3d20450630609305467a38db9de68f72329421701e46017442ef56bec731f64.jpg) +A photo of a whimsical garden where oversized flowers sing in the breeze. +Figure 3. Images generated by X-Fusion demonstrate high visual quality and strong text alignment with the input prompts. + +![](images/212606f534ba2a777cf725a83b84b6edd307378541f09e44c447bf2d65cc1068.jpg) +A ship sailing across a desert made of stars, its sails catching cosmic winds. +A photo of a cunning fox in a suit and tie reading a newspaper in a city park. + +![](images/58861371900f1eecd3debe7006bb3d38003555e945a816a26a5446821cc889d6.jpg) +A photo of a red book and a yellow vase on a polished table by a sunny window. + +![](images/36a6ad612e48ae04e0cbb64d0a115793dc657b4d5e14905a347c245c1f86ac10.jpg) +A photo of a playful puppy chasing vibrant holographic bubbles in a park. + +![](images/a9173eb7e0cda503e542313ba300426bab0630cf41f9685bc100907f0a967198.jpg) +A futuristic warrior queen commanding an army of robotic. +A crystalline rose blooming on a frozen lake. + +modalities, we instead propose initializing from frozen LLMs rather than training from scratch, significantly reducing computational costs and retaining the LLMs knowledge. Leveraging Pretrained LLMs. Since the success of LLMs [1-13], researchers have discovered that pretrained LLMs can serve as effective baselines for various purposes (e.g., coding [21-23]). Beyond being adapted for domain-specific tasks, prior works have shown that these pretrained models can also be fine-tuned to acquire new abilities, paving the way for a cost-effective approach to include more modalities like image understanding (e.g., LLaVA [32], Mini-GPT-4 [34]), image generation (e.g., GILL [30], DreamLLM [29], SEED-X [40]), or both image understanding and generation (e.g., Show-o [48], Emu3 [27], MetaMorph [31]). However, this method is not without limitations—often, when fine-tuning the LLMs' backbone, it risks compromising the original knowledge. To alleviate this shortcoming, in this work, we propose a novel method to extend vision-related abilities for LLMs while keeping all the language layers frozen, thus maintaining its original language abilities untouched. A concurrent work is LMFusion [49], with a similar high-level approach. However, they use joint attention across text and vision tokens, which is less flexible than our proposed model design. + +Task-specific Weights. People have been exploring the use of specialized models tailored to different tasks in both the language [50-53] and vision domains [54-56]. The use of unshared parameters has also been explored in multi-modal settings. For instance, CLIP [36] and ImageBind [57] focus on representation learning, while other works [58- + +60] emphasize vision-conditioned language model pretraining. However, these multi-modal approaches predominantly focus on visual understanding, neglecting generation tasks. Importantly, they are typically trained from scratch, which is computationally expensive. A concurrent work, Playground-v3 [61], takes a different approach by building upon LLMs, but freezes them to perform generation tasks only, without addressing visual understanding. + +# 3. Preliminaries + +In this section, we provide a brief preliminary overview of the state-of-the-art recipe (proposed by [28]) for training unified models in a hybrid manner, incorporating next-token prediction for language and diffusion models for image. + +# 3.1. Language Modeling via Autoregression + +LLMs are typically trained using an autoregressive modeling objective, where the joint probability of a sequence of language tokens $\mathbf{x}^{\mathrm{txt}} = \{x_1^{\mathrm{txt}}, x_2^{\mathrm{txt}}, \ldots, x_N^{\mathrm{txt}}\}$ is factorized as a product of conditional probabilities: + +$$ +P _ {\theta} (\mathbf {x} ^ {\mathrm {t x t}} \mid c) = \prod_ {i = 1} ^ {N} P _ {\theta} \left(x _ {i} ^ {\mathrm {t x t}} \mid x _ {< i} ^ {\mathrm {t x t}}, c\right). +$$ + +Here, $c$ represents optional conditioning information, which could include extracted features from other modalities (e.g., image representations) or task-specific context. However, in the standard setting, LLMs are usually trained without any additional conditioning ( $c$ is absent), and the predictions depend solely on the preceding tokens $x_{Modeltext only MMLU (↑)text2img FID (↓)CLIP (↑)img2text BLIP(↑)LLaMA3.2-1B [17]32.2———Single Tower25.019.1022.6330.2Gated Tower32.224.5121.9114.5Dual Projection32.220.2222.4630.9Dual Tower (Ours)32.214.2022.8131.3 + +Table 1. Architecture design comparison. Our dual-tower approach surpasses other baselines in image generation tasks and delivers competitive performance in image understanding, while maintaining the original language capability. + +$\lambda_{\mathrm{DM}} = 1$ , then train with AdamW optimizer ( $\beta_{1} = 0.9$ , $\beta_{2} = 0.95$ ), a linear warm-up scheduler, a maximum learning rate of $lr = 1 \times e^{-4}$ , and DeepSpeed Stage 2 [65] distributed training on H100 GPUs. We train with batches of 0.8M tokens for 100k steps, processing a total of 0.08T tokens for most experiments. We extend the training to 200K steps on our X-Fusion model initialized from LLaMA-3.1-8B and report the evaluation results in Supplementary. + +Task. We evaluate X-Fusion's performance on: (i) image understanding and (ii) image generation, as these tasks reflect the connection between visual and textual modalities. + +- Image Generation: To assess image generation, we evaluate text-to-image performance. Specifically, we generate 30K images using randomly sampled prompts from the MS-COCO [66] and report image quality using FID [67]. +- Image Understanding: To assess image understanding, we evaluate image captioning (or image-to-text) performance. Specifically, we generate captions for 30K images from the MS-COCO dataset [66] and report caption quality using BLIP2-ITM [68]. We also considered additional metrics, such as CIDEr [69] and BertScore [70], however, these were either inappropriate for long captions or failed to capture meaningful changes during training. Further analysis can be found in Supplementary. + +Data. Otherwise stated, we sample 0.08T tokens/patches from an in-house licensed dataset. We create image-caption pairs by center-cropping and resizing images to $256 \times 256$ , and pairing them with detailed captions generated by the InternVL-2.0 26B model [35]. These pairs are then formatted for both image-to-text tasks (serving as understanding data), and text-to-image tasks (serving as generation data). + +# 5. Architecture Design Choices for Adding Vision Abilities + +To evaluate the effectiveness of our Dual Tower design, we compare it against three alternative transformer block variants (Fig. 4) that are designed for multimodal integration. Apart from the transformer blocks, all other components—including tokenizers, encoder, and decoder modules—are kept identical across configurations. + +![](images/60c171acdc163e07246318684f1f60daf660c72dfc50b521f5ae886d684be377.jpg) +Figure 5. Performance of image generation and understanding at various data ratios. Increasing visual understanding data improves visual generation performance. + +![](images/dbda73621eb621b3572630398430d5d35587976749a7bab8187e02efb6093167.jpg) + +![](images/69a0c78643922d54ee8d5d39d7c67eb7097e383d4c824654c52611057cda455d.jpg) + +![](images/d50a119680b792765a9f0bfb00733521ff36526a44b42b73c596378d0444941e.jpg) + +![](images/4d457a76d41f4a0078ba10612b6c24b2ccf0e962a60ed4ebde981077bb746a1b.jpg) +Figure 6. Performance of image generation and understanding at various noise limits in the image-to-text samples. Providing clear images for image-to-text samples enhances visual generation and understanding simultaneously. + +![](images/c905435547b7a5f958c7fd001d5299ea7b756db392363b7376486ab434a5c08a.jpg) + +- Single Tower: This simple baseline uses the original language model transformer block to process both inputs directly. Note that this is equivalent to Transfusion [28] but with pre-trained LLM as initialization (Fig. 4a). +- Gated Tower: Inspired by [71], we duplicate the language transformer block into a trainable "gated block" (Fig. 4b), with both blocks taking the same input sequence: + +$$ +\mathbf {H} ^ {\mathrm {t x t}} = \mathcal {L} ^ {\mathrm {t x t}} (\mathbf {E} ^ {\mathrm {i n}}), \quad \mathbf {H} ^ {\mathrm {g a t e}} = \mathcal {L} ^ {\mathrm {g a t e}} (\mathbf {E} ^ {\mathrm {i n}}). +$$ + +After that, the gated block output will be added to the language transformer block output, multiplied by a learnable value $\gamma$ which is initialized as 0: + +$$ +\mathbf {H} ^ {\text {o u t}} = \mathbf {H} ^ {\text {t x t}} + \gamma * \mathbf {H} ^ {\text {g a t e}}. +$$ + +- Dual Projection: We copy the language weights consisting of one self-attention and one MLP (Fig. 4c). However, the data flow is different from dual-tower. First of all, based on the modality of $\mathbf{e}_i$ , we apply separate projection matrices for query $(\mathbf{Q})$ , key $(\mathbf{K})$ , and value $(\mathbf{V})$ and then a joint attention is operated on all tokens: + +$$ +\mathbf {H} ^ {\mathrm {a t t n}} = \operatorname {a t t n} \left(\mathrm {Q K V} _ {\text {m o d a l i t y}} \left(\mathbf {e} _ {i}\right)\right), +$$ + +After that, the output feature from self-attention is passed through a modality-specific MLP: + +$$ +\mathbf {H} ^ {\text {o u t}} = \mathrm {M L P} _ {\text {m o d a l i t y}} (\mathbf {H} ^ {\text {a t t n}}). +$$ + +This variant is similar to concurrent work [49]. However, dual-tower offers more flexibility, as the vision and language layers can be designed differently. + +FLOPs. Let $N$ denote the number of text tokens and $M$ denote the number of image tokens. For all three alternative designs, the attention complexity is $O((N + M)^2)$ . In our dual-tower design, for a fair comparison, we choose not to use the X-Fuse operation. As a result, the complexity in the vision tower is $O(M \cdot (N + M))$ (since we do not need query tokens for text), and the complexity in the text tower is $O(N \cdot (N + M))$ (since we do not need query tokens for image). In total, this results in the same complexity of $O((N + M)^2)$ . + +Quantitative Results. Tab. 1 presents a comparison of different architectural designs, all are using pretrained LLaMA-3.2-1B [17]. Among them, the Dual Tower architecture achieves the best performance in both image generation and understanding tasks. In contrast, the Gated Layer architecture performs the weakest in both tasks, likely due to the limitations of simple addition operations. Among baselines, the Single-Tower model delivers decent performance across both tasks; however, our Dual Tower model achieves a $23\%$ lower FID while maintaining the same number of training parameters. More importantly, the Single-Tower model compromises the inherent knowledge of the original language model due to training on T2I and I2T tasks. To assess the model's general knowledge, we evalu + +ate it on the MMLU benchmark [72], a multiple-choice test with four answer options, using a 5-shot setting. Results show that the Single Tower model's performance drops to 25.0, which is equivalent to chance-level performance. + +Dual Tower and Dual Projection share a common insight: modality-specific operations. While their text generation capabilities are nearly same, Dual Tower outperforms in image generation. This superiority can be attributed to its flexibility: The vision layers in Dual Tower can generate new QKV representations for input text features, followed by attention operations between the image and the new text features. In contrast, Dual Projection is limited to using the original language model's text QKV matrices for attention computation. Also, it is worth noting that conceptually, Dual Tower offers greater flexibility: while our current implementation replicates the language layer as the vision layer, the vision layer does not have to be identical to the language layer, as long as it produces outputs with the same feature dimensions. For the rest of the paper, all experiments will use Dual Tower design. + +Text processing flexibility affects image generation but not image understanding performance. + +# 6. Effect of Data Ratios and Noise on Generation and Understanding Tasks + +We now turn to examine the effects of training data on X-Fusion's performance. In the following section, we address two key questions: (i) How does noise level affect performance in a joint training setting? (ii) Does multitask training provide mutual benefits between tasks? + +# 6.1. Effect of Noise Amount + +As a unified model, it should be able to perform both image denoising and text autoregressive tasks simultaneously on input data. However, the level of noise applied to images in image-to-text (I2T) samples remains a question. While diffusion-based image modeling benefits from noisy input for generation tasks, excessive noise can degrade visual quality and hinder image understanding. This issue was also noted by [28], where they proposed limiting the diffusion noise on I2T samples to a maximum of $t = 50\% T$ to reduce distortion for visual understanding while still can treat those noisy images as generation training data. + +We argue that while this approach helps, it may not be optimal. We hypothesize that training with clean images (i.e., without adding noise to I2T samples) can lead to a stronger vision tower for image understanding, ultimately improving generation quality, although this reduces the amount of denoising data available for generation. + +To validate our hypothesis, we conducted comprehensive experiments where we systematically varied the max noise + +level for image understanding (I2T) tasks: $0\%$ (clean image), $25\%$ , $50\%$ , $75\%$ , and $100\%$ (normal generation setting). Throughout these experiments, we maintained a 1:2 data ratio between understanding and generation tasks. Results are provided in Fig. 6. As can be seen, generally the more noisy the images are, the more image understanding performance is degraded. Our proposed strategy (providing clean image for understanding) consistently achieves the best results for image understanding (2nd and 3rd col.). Interestingly, using clean image for understanding also helps to boost performance of image generation! (1st col.). + +Using clean images for visual understanding improve performance for both tasks. + +# 6.2. Effect of Data Ratio + +Along with the noise addition strategy, task data ratio is also a major factor for training unified multimodal models. To investigate the synergy between visual understanding and generation from data's perspective, we kept the architecture unchanged and trained for 100k steps with a batch size of 0.8M tokens on a total of 0.08T tokens, varying the composition of text-to-image (T2I) and image-to-text (I2T) tasks. Specifically, we begin with the training data composed entirely of T2I tasks (100% T2I, 0% I2T, or denoted as 100/0 for short), then progressively decrease the proportion of T2I data while increasing I2T data (i.e., 66/33, 50/50, 33/66) until the dataset consists solely of I2T tasks (0/100). + +Fig. 5 illustrates the results. To investigate how training data from one task influences performance on other tasks, we plot performance metrics against the number of tokens observed for each specific task (e.g., first column, x-axis indicates how many image generation tokens are trained). In another words, we want to ask: When two models are trained on an identical number of image generation tokens, but differ in their exposure to image understanding tokens, how do their performance diverge?, and vice versa. + +We observe that incorporating image understanding data (I2T) improves generation quality (T2I). As the proportion of I2T data increases while keeping total T2I data fixed, generation performance consistently improves (1st and 2nd panels). In contrast, visual generation data (T2I) does not positively impact the understanding task (I2T) (third panel). Overall, there's an asymmetric relationship: understanding data benefits generation, but generation data does not enhance understanding. Based on our findings, we recommend a 66/33 (or 2:1) training ratio, which strikes a strong balance and optimizes performance across both tasks. + +$\boxed{ \begin{array}{r l} \end{array} }$ Incorporating image understanding data enhances image generation performance, while adding generation data does not impact image understanding tasks. + +![](images/308174714abb83dce0026d2172057f1723bb13df45e83b2d95ecfc6f6530b160.jpg) +>Input Image: + +# X-Fusion + +Image depicts a home office setup with a laptop open on a desk, displaying a webpage. To the right of the laptop, there's a desktop computer with a monitor showing a webpage. + +The desk is cluttered with various items, including a printer, a lamp, and a few books. A window with curtains is visible in the background, and items are scattered around the room. + +# X-Fusion (Pretrained DiT) + +Image showcases a home office setup with a laptop open on a desk, displaying the screen with a photo. + +To the right, there's a desktop computer with a prominent tower, a monitor, a keyboard, and a mouse. Behind the desk, a curtain with floral patterns is visible, suggesting a personal space. + +On the desk, there's a mouse and a mousepad, indicating a well-equipped workspace. + +![](images/25a8d6a40b3d1ca4472ac632dbf57fd516ba0fcb3a753893ff2e932f8a6e0933.jpg) +Pretrained DiT + +![](images/75d500458d2aeada9feb43e8de65bb74636f2c2e2e58ad11ba2b95163dcf171f.jpg) +X-Fusion +Figure 7. Qualitative comparison between pretrained DiT model, X-Fusion(DiT) and vanilla X-Fusion on image generation and understanding task. By initializing the vision tower from pretrained text-to-image diffusion model, X-Fusion(DiT) achieves stronger image generation capability and competitive performance on image understanding compared to vanilla X-Fusion. + +![](images/8e3d6e4f0e5ca2fe82537fe763d08bcf1201410fb273d96e3b509b2a40260a4e.jpg) +X-Fusion (Pretrained DiT) + +# 7. Extension of X-Fusion + +In this section, we introduce three extensions of X-Fusion, including X-Fuse layer, transferring from pretraind diffusion models, and finetune for downstream tasks. + +![](images/ab1ba4089464595b59233658dc4a9167db13fed2a2f16177dd73bb92fdbdeda1.jpg) +Figure 8. Ablation: X-Fuse layer. Our X-Fusion model with the X-Fuse layer outperforms the baseline X-Fusion model on image generation and understanding tasks. + +![](images/31d330cc8f39014a3ec7e5c9a11bc553d0648bf000ae7801b90a85013c29a4f0.jpg) + +# 7.1. Effectiveness of X-Fuse Layer + +So far, our study is conducted on a dual-tower architecture without the X-Fuse operation, maintaining the same FLOPs as other baseline designs. In Sec 5, we further propose the X-Fuse operation, which merges features from both towers to trade increased FLOPs for improved performance. We conduct an ablation study to evaluate this design, and the results are shown in Fig 8. As illustrated, applying the X-Fuse operation leads to improvements in both capabilities. + +# 7.2. Transfer from Pretrained Diffusion Model + +While X-Fusion successfully kept its language generation capability, its image generation capability still needs to be trained from scratch. One solution is to transfer the knowledge from existing image generation models. In our dual-tower design, each block in the language and vision tower processes the entire feature sequence independently, therefore allowing non-identical block designs in both towers. + +With this advantage, we could transfer the image generation capability from a large-scale pretrained diffusion model that uses diffusion transformers [64, 73]. + +We train a variation of X-Fusion using Llama3.1-8B as the language tower and an in-house pretrained text-to-image DiT model as its vision tower, notated as X-Fusion(Pretrained DiT), using the same training recipe in the previous section for 50K iterations. To align the feature dimension in both towers, we add linear projection layers within the X-Fuse layer. Figure 8 shows that this operation further enhances the model's capability. Figure 7 qualitatively compares the image generation and image understanding performance between the pre-trained DiT model, X-Fusion(Pretrained DiT), and vanilla X-Fusion-8B models. X-Fusion(Pretrained DiT) obtained stronger image generation capability and on-par image understanding performance compared to the vanilla X-Fusion. + +# 8. Conclusion + +This paper introduces X-Fusion, a novel framework for adapting pretrained Large Language Models to new modalities (e.g., vision) while retaining their original language capabilities. We propose a Dual Tower architecture in which language weights remain frozen, while visual features are processed via a trainable vision tower with separate weights. Alongside this novel architecture, we provide a systematically comprehensive set of ablation studies that offer valuable insights from a data perspective. Our findings reveal that: (i) incorporating understanding-focused data improves generation performance, (ii) reducing noise in image data enhances overall results, and (iii) feature alignment benefits primarily smaller models. We hope our paper will step forward building an unified Large Multimodal Models in a more efficient way. + +Acknowledgment: The work was supported in part by Amazon Research Award through the UCLA Science Hub, NSF IIS2404180, Microsoft Accelerate Foundation Models Research Program, and Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea government (MSIT) (No. 2022-000871, Development of AI Autonomy and Knowledge Enhancement for AI Agent Collaboration) and (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pre-training). + +# References + +[1] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 1, 2, 3 +[2] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. +[3] Tom B. Brown and et al. Language models are few-shot learners, 2020. +[4] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020. +[5] Jack W. Rae and et al. Scaling language models: Methods, analysis & insights from training gopher, 2022. +[6] Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. Finetuned language models are zero-shot learners, 2022. +[7] Jordan Hoffmann and et al. Training compute-optimal large language models, 2022. +[8] PaLM team. Palm: Scaling language modeling with pathways, 2022. +[9] PaLM 2 team. Palm 2 technical report, 2023. +[10] Albert Q. Jiang and et al. Mixtral of experts, 2024. +[11] Albert Q. Jiang and et al. Mistral 7b, 2023. +[12] DeepSeek-AI. Deepseek llm: Scaling open-source language models with longtermism, 2024. +[13] Gemma Team. Gemma: Open models based on gemini research and technology, 2024. 1, 2, 3 +[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, March 2023. 1 +[15] Llama 2 team. Llama 2: Open foundation and fine-tuned chat models, 2023. +[16] Phi-3 team. Phi-3 technical report: A highly capable language model locally on your phone, 2024. +[17] Abhimanyu Dubey and et al. The llama 3 herd of models. ArXiv, abs/2407.21783, 2024. 5, 6 +[18] Qwen team. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1 +[19] InternLM2 team. Internlm2 technical report, 2024. + +[20] DeepSeek-AI. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model, 2024. 1 +[21] Mark Chen and et al. Evaluating large language models trained on code, 2021. 1, 3 +[22] Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis, 2023. +[23] Raymond Li and et al. Starcoder: may the source be with you!, 2023. 1, 3 +[24] Lili Yu and et al. Scaling autoregressive multi-modal models: Pretraining and instruction tuning, 2023. 1 +[25] Gemini Team. Gemini: A family of highly capable multimodal models, 2024. +[26] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2 +[27] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 3 +[28] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. 2024. 1, 2, 3, 6, 7 +[29] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Dreamllm: Synergistic multimodal comprehension and creation, 2024. 1, 2, 3 +[30] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models. NeurIPS, 2023. 1, 2, 3 +[31] Shengbang Tong, David Fan, Jiachen Zhu, Yunyang Xiong, Xinlei Chen, Koustuv Sinha, Michael Rabbat, Yann LeCun, Saining Xie, and Zhuang Liu. Metamorph: Multimodal understanding and generation via instruction tuning. arXiv preprint arXiv:2412.14164, 2024. 1, 3 +[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 3 +[33] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. +[34] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 3 +[35] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 5 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021. 1, 2, 3, 4 +[37] Yuexiang Zhai, Shengbang Tong, Xiao Li, Mu Cai, Qing Qu, Yong Jae Lee, and Yi Ma. Investigating the catastrophic forgetting in multimodal large language model finetuning. In Conference on Parsimony and Learning (Proceedings Track), 2023. 1, 2 +[38] Didi Zhu, Zhongyi Sun, Zexi Li, Tao Shen, Ke Yan, Shouhong Ding, Chao Wu, and Kun Kuang. Model tailor: mitigating catastrophic forgetting in multi-modal large language models. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org, 2024. 1, 2 +[39] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. 2023. 1 +[40] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 3 +[41] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. +[42] Zineng Tang, Ziyi Yang, Mahmoud Khademi, Yang Liu, Chenguang Zhu, and Mohit Bansal. Codi-2: In-context, interleaved, and interactive any-to-any generation. 2023. +[43] Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Januspro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025. 1 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 2 +[45] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy, 2017. 2 +[46] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection, 2016. 2 +[47] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 2 +[48] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 3 +[49] Weijia Shi, Xiaochuang Han, Chunting Zhou, Weixin Liang, Xi Victoria Lin, Luke Zettlemoyer, and Lili Yu. Llamafusion: Adapting pretrained language models for multimodal generation. arXiv preprint arXiv:2412.15188, 2024. 3, 6 +[50] Noam M. Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. + +Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. ArXiv, abs/1701.06538, 2017. 3 +[51] Dmitry Lepikhin, HyoukJoong Lee, Yuzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam M. Shazeer, and Z. Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. ArXiv, abs/2006.16668, 2020. +[52] William Fedus, Barret Zoph, and Noam M. Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. ArXiv, abs/2101.03961, 2021. +[53] Nan Du and et al. Glam: Efficient scaling of language models with mixture-of-experts. 2021. 3 +[54] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Neural Information Processing Systems, 2021. 3 +[55] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Collaging class-specific gans for semantic image synthesis. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14398-14407, 2021. +[56] Jiapeng Zhu, Ceyuan Yang, Kecheng Zheng, Yinghao Xu, Zifan Shi, and Yujun Shen. Exploring sparse moe in gans for text-conditioned image synthesis. ArXiv, abs/2309.03904, 2023. 3 +[57] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind one embedding space to bind them all. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15180-15190, 2023. 3 +[58] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. ArXiv, abs/2208.10442, 2022. 3 +[59] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. ArXiv, abs/2111.02358, 2021. +[60] Sheng Shen, Zhewei Yao, Chunyuan Li, Trevor Darrell, Kurt Keutzer, and Yuxiong He. Scaling vision-language models with sparse mixture of experts. ArXiv, abs/2303.07226, 2023. 3 +[61] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. ArXiv, abs/2409.10695, 2024. 3 +[62] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. CVPR, 2022. 4 +[63] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Q. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russ Howes, Po-Yao (Bernie) Huang, Shang-Wen Li, Ishan Misra, Michael G. Rabbat, Vasu Sharma, Gabriel Synnaeve, Huijiao + +Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. *ArXiv*, abs/2304.07193, 2023. 4 +[64] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 5, 8 +[65] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '20, page 3505-3506, New York, NY, USA, 2020. Association for Computing Machinery. 5 +[66] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[67] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[68] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 5 +[69] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation, 2015. 5 +[70] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert, 2020. 5 +[71] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 6 +[72] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding, 2021. 7 +[73] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 8 \ No newline at end of file diff --git a/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/images.zip b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d68642276e86f8a717ad6ec0f8e0b0b5d5d170f4 --- /dev/null +++ b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5acac70cab31554bde42de3162c8673a4c1b9464a40671097e18c9ce8e18b0c +size 591250 diff --git a/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/layout.json b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..19f4871ab52a3242fa896235e7587f873f8100ed --- /dev/null +++ b/xfusionintroducingnewmodalitytofrozenlargelanguagemodels/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03fc5c04a87988e2eb79e2f1c91bd61ca0303b8a6d85887848ce69bfad926cb6 +size 464258 diff --git a/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/66c6a6ed-9847-4426-937e-5e7c6b19faf9_content_list.json b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/66c6a6ed-9847-4426-937e-5e7c6b19faf9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0e09f3dce97cdb8128c51ab70f2602b5cda678d3 --- /dev/null +++ b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/66c6a6ed-9847-4426-937e-5e7c6b19faf9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cb0d3236a8bb1accdb8466e894af3ae7c4f80ab6c52c25da845f960af7b2e14 +size 102050 diff --git a/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/66c6a6ed-9847-4426-937e-5e7c6b19faf9_model.json b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/66c6a6ed-9847-4426-937e-5e7c6b19faf9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a4600980cf85b212dbabae7692120beac5488a14 --- /dev/null +++ b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/66c6a6ed-9847-4426-937e-5e7c6b19faf9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00aed7b13dea2f2ca6631eca8f0597587200e7f310e081dd8f7a1335e03dc62b +size 127960 diff --git a/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/66c6a6ed-9847-4426-937e-5e7c6b19faf9_origin.pdf b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/66c6a6ed-9847-4426-937e-5e7c6b19faf9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e27b4b6adabc26302f36cddb800a91e3f269808 --- /dev/null +++ b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/66c6a6ed-9847-4426-937e-5e7c6b19faf9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07e32b4f6fc28a11812ed4ffe732b76faef00bcf22743f9bbcfd2d3b97ad91f3 +size 3805079 diff --git a/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/full.md b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ebae04f7039a3e3534b1b96e22de69e7effab74 --- /dev/null +++ b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/full.md @@ -0,0 +1,423 @@ +# X-Prompt: Generalizable Auto-Regressive Visual Learning with In-Context Prompting + +Zeyi Sun $^{1,2}$ , Ziyang Chu $^{2,3}$ , Pan Zhang $^{2}$ , Tong Wu $^{4}$ , Yuhang Zang $^{2}$ + +Xiaoyi Dong $^{2,5}$ , Yuanjun Xiong $^{7}$ , Dahua Lin $^{2,5,6}$ , Jiaqi Wang + +$^{1}$ Shanghai Jiao Tong University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ Tsinghua University + +4Stanford University 5The Chinese University of Hong Kong 6CPII under InnoHK 7Adobe + +# Abstract + +Recent advances in large language models have enabled task prompting for open-ended text generation. In the vision domain, a longstanding goal is developing models capable of general visual learning, encompassing tasks such as image generation, editing, low-level processing, and dense perception. Although recent efforts have aimed at building vision foundation models that support prompting, significant challenges remain, particularly in accurately comprehending visual prompts and addressing the ambiguity inherent in textual prompts. To address this, we introduce X-Prompt, a purely auto-regressive large vision-language model designed for generalizable visual learning via in-context prompting. X-Prompt can process visual and textual prompts as context, enabling precise task interpretation and accurate execution. A novel prompt-token fusion mechanism effectively extracts relevant task information from complex prompts while significantly reducing the token length. Additionally, a unified training strategy for text and image prediction enhances task awareness, enabling seamless adaptation to open-ended prompts. Extensive experiments demonstrate that X-Prompt effectively interprets in-context prompts and exhibits generalization across both in-domain and out-of-domain visual tasks, paving the way for future advancements in general visual learning. + +# 1. Introduction + +Accurately interpreting tasks from prompts and executing them reliably at inference time has long been a significant challenge in machine learning. Large language models (LLMs) [7, 14, 41, 62, 63] have demonstrated considerable capability in following textual prompts [7]. However, in the vision domain, many tasks present complexities that cannot be easily captured through textual descriptions alone. For instance, in image personalization tasks [48, 65], prompts based solely on language may fail to capture the nuanced + +style of lesser-known artists adequately. + +Recent research has explored incorporating visual prompting into vision foundation models [3, 18, 67, 68], primarily focusing on low-level image processing and dense perception tasks. Nevertheless, tasks requiring fine-grained control and precise alignment to visual exemplars, such as image editing [6, 85] and customization [48, 65], remain challenging when relying solely on visual prompts, due to the inherent demands for semantic visual understanding and reasoning. Therefore, it is essential to leverage both visual learning and linguistic priors to facilitate a deeper and more accurate comprehension of visual tasks. + +In order to leverage the strong generalization ability of LLMs, recent approaches [12, 50, 59, 72] have explored developing vision-language models (VLMs) for vision learning. One line of researches [20, 54, 55] incorporates an additional diffusion decoder for image generation while facing challenges to maintain visual fidelity, particularly in tasks demanding precise editing or fine-grained visual precision. Another line, exemplified by Chameleon [59] and concurrent studies [69, 71, 73-75], directly utilizes an LLM to predict VQ-VAE [64] or VAE [28] features in an autoregressive manner, effectively preserving visual details and integrating the reasoning abilities of LLMs. Although most of these works primarily focus on image generation and lack support for visual prompting, the promising potential of autoregressive VLMs provides extensive research opportunities for facilitating generalizable visual learning. + +The primary challenge of visual prompting is improving the ability of foundation models to reasoning the intent behind visual prompts. Since these prompts often consist of several images that implicitly convey the target task without explicit explanations, it is crucial for foundation models to effectively identify and describe the differences or changes between each pair of images. Another challenge in leveraging in-context prompts is the substantial context length required as a single image typically requires 1024-4096 tokens. In an image-to-image task, at least four images are necessary for context, leading to a prohibitive training con + +# Chameleon + +![](images/5ce7693624b9c53e0a72039e1409062f55d501891953d525ace6ae7594c97ca6.jpg) + +![](images/e961013d2c8a28ddca43ccb87860ffcad3076da767a19ba3470805dcb03c097e.jpg) +Provide sametransformationon this image f + +![](images/0cf94ed5fb8412754619d1862eee48a9080524a0348656c3b8c280de9dd8748b.jpg) +Input + +# In Domain Tasks + +![](images/dbcf02a952a18cff991dbfd9e9dee7d44089ed699004f794a4c6d71d34adf046.jpg) +In-context Prompt + +![](images/4d65e43facc2d0c5e3789e9132b181fe38b73cde6c84b5f806bba68b6ebcf8ec.jpg) + +![](images/860dd72c03a0f415bd39737f9355e521ddeb06c0c5ee34f101639d0627e180dc.jpg) +Input + +![](images/c8c593cc7d812a55df39a7b2a6594ecc8540a98cb0e06901d472cae9bbd9515e.jpg) +Result + +![](images/62e75e2848db9eea46596bdf7fddc4edcec8a8e184b0fbd42152cb72b9c7867b.jpg) + +![](images/f5b067e22605026e959a55629b0a3d7ca4f475a447e78047bdba6288da006376.jpg) + +![](images/0f4c2dd268efeba222abb68dd1396cc1401d00317a0a13ea5dba37e066265c0f.jpg) + +![](images/cea1631f22cd0140b0d318efff40db170d3975e153eb4c67dfef87c54e514f41.jpg) + +![](images/3404f6a7858bdd26bb867978d262d66a175c0e420c72d829e18d691710c72531.jpg) + +![](images/fc333883e9b03764a7c48ef20a42a0cf975b5471502bf90e458db6e317eb05a4.jpg) + +![](images/dea1c10fb4d78891621ed5da2aafc78daffce872a4f050b66935959d900f1712.jpg) + +![](images/3ab5a37b5485789e9671284063f62094aaaab4c72e0549bab1ef89285cc38669.jpg) +Result + +![](images/c882d0324c701c726627b51cc3db2f6b0d014def3ab0cae1a7e488f26d2b0a86.jpg) + +![](images/b8ab3877ec82a5ffe7c8887f332ad36dd41a0f40d260fa63484c8e7697aba1ef.jpg) +Input +Make the girl in dress cloth in + +![](images/c5f2194dc225020751e3aab4fc48bfcbc2929ad00526db932f86bb38dd5575b2.jpg) + +# Novel Tasks + +![](images/be760df2c598318c5ae807d0e1f13068be8497b76e932ae00f45954006e5ab4b.jpg) +In-context Prompt + +![](images/80d94c9b49d9e41025e0e617f600bce66f437b7906aa80c1758b70c6e5eb3d2e.jpg) + +![](images/99a3b1efc8fbb185cd6a55b8bec321e8181638942fdfc280fb44f6c7d089907c.jpg) + +![](images/dfa68e4a4d0325a79eeb5dc9720172bcade93925aebd179af5e41aad1973a6e7.jpg) + +![](images/ea65031f830dfbe73dab33dee207839fdd089106d68c441a7f2650732038566e.jpg) +Figure 1. X-Prompt can perform multi-modal generation based on in-content examples in a pure auto-regressive foundation model. After training on diverse in-domain tasks, X-Prompt can generalize to previous unseen tasks like image inpainting, colorization, one-shot segmentation and video frame prediction in an visual prompting manner. + +![](images/c51cb2603dbed05a916247c736f10910f1a71748f3b3147a103bc4a523f38ecf.jpg) + +![](images/cb335168083d56d6edc752d95034346a6085f6a9442c0b3435f20ee86e76985e.jpg) + +![](images/dea33580d0f2a282b94b07ec5804732dcdd03f69c82d9b5a60560f1e3059e4cb.jpg) + +text length. This limitation, also noted by [74], restricts support to only three images, making direct visual prompting impractical due to the excessive context length. + +Based on Chameleon [59], we present X-Prompt, an auto-regressive large vision-language model designed for generalizable visual learning tasks through contextual prompting. The framework seamlessly integrates visual and textual prompts as context, enabling precise task interpretation and context-aware execution. (1) X-Prompt introduces + +# Chameleon + +![](images/06f88a4ce23d963a3ae700c462ea3501d9b940cacad2fac9fa324c12dcd509df.jpg) + +![](images/f0aeaf6505ffe880112bbfbf32d712712efafe3c7db8f87c30590087ed7b8e94.jpg) +X-Prompt +Input + +What's the relationship difference between and ? + +![](images/39b7d01687d1ac2a8bf33cc9b788f130fa5085d40eba6ba665bbdfbc050070d5.jpg) + +![](images/0bcf253530fde5106cec8e204540cc89317b73cef498ceb1bae0542657500ef4.jpg) + +# Input + +A woman with long golden hair, waring a white toga + +![](images/1b0d2dca5be1137cd7ea22b17795ffcb5b4efa4297027f64e73b0ef3a6e92752.jpg) + +![](images/a8a769bc42aedc253760a4f25521fef48d3df9519d3b06bacf5885b5260a1169.jpg) + +![](images/3257db54b54dc78c70f436945d3155cfc22f2d2f95f441ff7a5487f953073e14.jpg) + +![](images/4f100dc23a063db2b20a195fd9523c32297c85af70d7c0fb9cfe02bbb75f5158.jpg) +Result + +![](images/0be552398b3ce0a18409dafabb6f09b65b231a0fb8ea8ec0ef931332dd4d1533.jpg) + +![](images/8eb23416ed4cea4cd616dac72f035c30166f3e4f318e20904268e7dd236ca924.jpg) + +![](images/18f46c58aaef173b190e605fed17e5444ddf35758455ebc6fa5f6ea6035edccb.jpg) + +![](images/fb356ef687bbc41539e053433866815e91f8f886af8eb7f2f861c1aa33a3166b.jpg) + +![](images/4934b81d1042601b707df178e5bd375ceebc3412c6217e7836a5b3f0f9456767.jpg) + +![](images/a75b620069549b90f730b742d9df61be75291e7cd55d38b5d9cc4728e2886e7b.jpg) +Add the flag of to the plane in + +![](images/e1e11add9d7284f91f5601ce30746807a42ee6dbbb2333f79d704cb4d22a0f8b.jpg) +Result + +![](images/e99ad33b2ae3c2568ed5eda5e11c035ecd80e2497c16ae91f7c88b58e4024b95.jpg) +In-context Prompt + +![](images/0c15ea7f1050ef985f1a4648b687f2ed5722790aca8329726b7d5e4de8e482ed.jpg) + +![](images/99f01babc4739de9c676cd1115ab3798450708b1f361e751618e3a3b96ffde68.jpg) +Input + +![](images/79d402fbef4e35c86894cf7ed92dfb6f7ddfff39c75fcbc67025c491d4e230e6.jpg) +Result + +![](images/f567bb139e43d0f1fc3e69ba557f2da22785ade65911fc3767eb0828c5dd8cf8.jpg) + +![](images/dc84d32693d422de2782ee9c00a0bf4cf2edcf15fc99a66a2cf48fb79d36e582.jpg) + +![](images/4aa436130d6b1ae607a3f6945f308ffe869c3a74e3afebd89074d4ddbeba04fb.jpg) + +![](images/b80fe94efdb6152165c9c1d878b6678a78149525bcd587aa036dd1d421752482.jpg) + +a novel approach that compresses the content of examples (comprising thousands of tokens) into a sequence of fixed-length compressed tokens. These tokens serve as contextual cues for reasoning on new images, significantly reducing the maximum context length required in training and inference. Moreover, the compressed tokens enhance the interpretation of target tasks, demonstrating improved generalizability to unseen tasks in experimental evaluations. (2) A unified training strategy for text and image prediction is + +![](images/8e740bed337bb6dc70c60a198d03b6d6bf699f2a9f93b89ea997e043b44be551.jpg) +Figure 2. Two stages of X-Prompt in both training and inference. In stage 1, in-context examples are compressed into X-Prompt tokens. In stage 2, these tokens (with their hidden states) serve as task guidance for the specified tasks. + +employed to enhance task awareness. By incorporating reversed image prediction tasks and text prediction tasks, such as generating descriptions of image differences, we explicitly encourage the model to infer the intent behind visual prompts, thereby boosting its overall performance and generalization capabilities. Visualization results in Fig. 1 illustrate the model's ability to perform diverse visual tasks through prompting. + +Our work makes three key contributions: + +- We introduce X-Prompt, a purely auto-regressive large vision-language model designed for generalizable visual learning through in-context prompting +- The X-Prompt effectively extracts useful information from visual prompts and encodes it into compressed tokens. Additionally, a unified training strategy for both image and text prediction tasks is introduced to further enhance X-Prompt's capability in interpreting prompts. +- We integrate image customization, editing, low-level vision, and dense perception tasks into a single framework, showcasing the effectiveness and generalization of visual prompting across seen and unseen tasks. + +# 2. Related Work + +Large Vision-Language Models (LVLMs). The emergence of large language models (LLMs) [1, 7, 10, 24, 41, 61] have made remarkable breakthroughs. The domain of research has increasingly turned its attention toward Large Vision-Language Models (LVLMs). Previous advances in this field focus on the integration of vision understanding capabilities with LLMs [2, 13, 30, 34, 40, 43, 56, 76, 77, 83]. Recent works start to focus on integrating vision generation abilities. One early line of these works [19, 20, 27, 54, 55, 72] compress visual features with LLMs into compressed embeddings and use diffusion decoder (like SDXL [42]) to generate visual contents. However, this suffers great information loss during LLM encoding process, leading to unsatisfying results in image editing tasks. Another line of work, pioneered by Chameleon [59], + +uses unified image tokens from VQ-VAE [15, 64] to unify perception and generation. In this work, we aim to fully unlock the potential of Chameleon for general image generation in a unified in-context learning paradigm for both seen and unseen tasks. + +Auto Regressive based Image Generation. While previous state of the art image generation dominant by diffusion models [16, 23, 32, 45, 47, 52, 57], recent works on auto-regressive for image generation have shown promising results [33, 36, 37, 53, 60, 69, 73, 84]. However, these models only show experiments results on text-to-image generation [31, 53, 58, 71, 73, 73], with limited research on other types of image generation tasks or lack of quantitative results [54, 80]. In this work, we not only enhance text-image alignment but also extend our exploration to diverse visual learning tasks, including image editing, controlled image generation, and perception tasks such as semantic segmentation and depth estimation. We demonstrate that auto-regressive models can achieve competitive results across these tasks in a unified framework. + +In-Context Prompting. GPT-3 [7] introduced the paradigm of in-context prompting, where diverse NLP tasks are reformed as text completion tasks, which significantly boosts performance on related tasks. In-context prompting has also been explored in the vision domain [3, 4, 67, 68], but these models are limited to vision-only tasks, lacking multi-modal versatility. Multimodal models like Emu1/2 [54, 55] demonstrate in-context prompting capabilities; however, their reliance on integration of an external diffusion model restricts their effectiveness in dense prediction tasks. In contrast, our approach uses a unified early fusion representation based on Chameleon [59], enabling a single model to generalize across wider of tasks with improved performance and enhanced generalizability. The works on in-context prompting in diffusion models [25, 26] have demonstrated high-quality image generation. However, they lack the reasoning capability required for tasks such as image editing or rely on task-specific fine-tuning. + +![](images/eed034bbd514ce000397d5c330614ea949facbb279b86d67a189d75a43246b3f.jpg) +Figure 3. Training data pair augmentation and list of training prototype tasks and subtasks. We introduce reverse task and difference description task through next text token prediction to improve the performance and generalizability. + +![](images/270780aa47a976f38641068b1e3041984d640763cf0df9df7f603819021082b8.jpg) + +![](images/dc3093b97aa0cb8edfba5ca4f2eb9a75ec8f501fd3c52a1bfbdcede550aa8561.jpg) + +![](images/f7140b2e047fa94763e9c449a42510f0aa38667ca757dfa652355300fc0168cb.jpg) + +![](images/1505225b7b53e03833860747384c4f2812e219f95e0b26ee7397d4c4135077b5.jpg) + +# 3. Method + +# 3.1. Visual Prompt Compression + +We introduce a context-aware compression mechanism to better model visual prompts based on Chameleon. As illustrated in Fig. 2. This mechanism defines three types of tokens: In-Context Example Tokens (IE), X-Prompt Tokens (XP), and TODO Tokens (TD). Given an in-context example as a prompt, our model embeds it into a feature space, producing In-Context Example Tokens $X_{IE} \in \mathbb{R}^{IE \times C}$ where $IE$ denotes the number of tokens and $C$ denotes the feature dimension. The model further includes learnable X-Prompt Tokens, represented as $X_{XP} \in \mathbb{R}^{S \times C}$ , where $S$ is the number of learnable tokens optimized during training. + +To encourage the model to store contextual information in $X_{XP}$ , we disconnect the relationship between $X_{IE}$ and the TODO Tokens $X_{TD}$ through attention masking, forcing the model to rely on X-Prompt Tokens for context representation. The model then generates the TODO Tokens $X_{TD} \in \mathbb{R}^{TD \times C}$ , with $TD$ as the target sequence length, by sequentially maximizing the conditional probability of each token in $X_{TD}$ given $X_{XP}$ and all previously generated tokens $X_{TD_{TypeMethodsDepth Est. RMSE↓Semantic Seg. mIoU↑Surface Normal Est. Mean Angle Error↓Lowlight Enhans. PSNR↑SSIM↑Deblur PSNR↑SSIM↑Derain PSNR↑SSIM↑NYUv2ADE20KNYU-Depth V2LOLGoProRain100HUnified Model (Continuous)Painter [67]0.28849.90×22.400.87220.120.70920.400.714InstructDiffusion [21]×29.71×××23.58-16.110.683OmniGen [74]0.480××13.380.39213.390.3219.110.132Unified Model (Discrete)Unified-IO [35]0.38725.71-××××××LVM [3]0.44311.7131.0610.350.41112.070.51212.120.483Lumina-mGPT [33]×20.8722.10××17.590.53613.310.322Ours0.27731.2119.1719.710.81021.040.76124.770.815 + +Table 1. Comparison of X-Prompt with task-specific and vision generalist baselines across six representative tasks, covering both high-level visual understanding and low-level image processing. $\times$ ’ indicates that the method is incapable of performing the task. The best and second best results are marked in green and red . + +where we use the cosine similarity of CLIP [44] text features as the distance function dist. This retrieved example serves as in-context guidance for the model, which then generates the edited output $I_{\mathrm{output}}$ based on both Instrcurrent and the retrieved pair (Instr_retrieved, $P_{\mathrm{retrieved}}$ ): + +$$ +I _ {\text {o u t p u t}} = \operatorname {M o d e l} \left(I _ {\text {i n p u t}}, \operatorname {I n s t r} _ {\text {c u r r e n t}}, \operatorname {I n s t r} _ {\text {r e t r i e v e d}}, P _ {\text {r e t r i e v e d}}\right). +$$ + +It is worth noticing that dist can be extended to other function beyond simply compute text feature similarity. This automated retrieval process reduces the need for manual intervention and can also be customized by users to achieve more precise, tailored image editing. RAIE is an optional choice, and we apply this method only when specifically mentioned in the experiment section. + +By leveraging in-context examples, RAIE enhances the consistency and accuracy of image editing tasks. RAIE is naturally suited for a unified auto-regressive model. However, it poses significant challenge for current state-of-the-art diffusion models, which rely solely on text encoders and lack comprehensive understanding capabilities. + +# 4. Experiments + +Though we train a unified model across different tasks, we report the data preparation process separately in each subsection for clarity. The complete training dataset comprises approximately 5 million data pairs, expanding to around 8 million pairs after task reversion and text prediction task augmentation (detailed in Appendix G). For all tasks that take an image with text as input and produce an image with text as output, we include an in-context example of the same task type. We set the batch size to 1024 and the context window size to 5120. The learning rate is set to 1e-4 with a cosine learning rate scheduler. Training is conducted on 128 NVIDIA A100-80G GPUs over 20,000 steps. X-Prompt token number is set to 32 with abalation study in Appendix D. + +# 4.1. Image Dense Prediction + +Settings. We use representative datasets for dense prediction tasks: NYU-v2 [51] for depth and surface normal estimation, ADE-20K [86] for semantic segmentation, and + +Rain-13K [17], LOL [70], and GoPro [39] for corresponding low-level vision tasks. Full training data details are available in the Appendix G. + +Results. We select previous vision generalist for comparison. Results are shown in Tab. 1, where our model can achieve competitive results on dense prediction task (semantic segmentation, geometric prediction) and low level vision task. Our model is the first to deliver promising results using a unified, discrete token approach. We provide qualitative results in Appendix J, where our model successfully achieve semantic segmentation, norm estimation and low-level vision tasks. Performance in low-level vision tasks that require high-fidelity output is restricted due to the inherent information loss in the VQ-VAE discretization process, as Chameleon [59] adopts $16\mathrm{x}$ compression rate. Upper bound analysis is detailed in Appendix K. By discretizing images and adopting a next-token prediction approach akin to that used in large language models, our method offers promising scalability for future advancements with better VQ tokenizer around horizon. + +# 4.2. Text-to-Image Generation + +Settings. To enhance Chameleon's [59] text-to-image generation capabilities, we utilize QWen2-VL [66] to rewrite dense captions for 500K high-quality images filtered from the LAION [49] dataset (detailed in Appendix H.4). For evaluation, we adopt the GenEval [22] benchmark. + +Results. As shown in Tab. 2, we significantly enhance Chameleon's original text-to-image generation capabilities. Leveraging the image dense description task, our model achieves competitive results compared to other auto-regressive models for text-to-image generation. This experiment highlights the effectiveness of unifying dense image description and image generation tasks, resulting in notable improvements, particularly in tests involving complex multi-object and color attributes. The ability to generate images that accurately follow text prompts is essential for our downstream applications like image-editing. Qualitative visualization and user preference study are available in Appendix A and Appendix F. + +
TypeModelSingle Obj.Two Obj.CountingColorsPositionColor Attri.Overall
DiffusionLDM [47]0.920.290.230.580.020.050.37
SD-1.5 [47]0.970.380.350.760.040.060.43
SD-2.1 [47]0.980.510.440.850.070.170.50
DALL-E 2 [46]0.940.660.490.770.100.190.52
Show-o [75]0.950.520.490.820.110.280.53
SDXL [42]0.980.740.390.850.150.230.55
DALLE 3 [5]0.960.870.470.830.430.450.67
Auto-regressiveLLamaGen [53]0.710.340.210.580.070.040.32
Emu3Gen [69]0.980.710.340.810.170.210.54
Chameleon [59]------0.39
X-Prompt0.970.690.280.710.140.150.49 (+0.10)
X-Prompt (+text pred.)0.980.730.330.850.260.280.57 (+0.18)
+ +Table 2. Evaluation of text-to-image generation ability on GenEval [22] benchmark. Unifying image dense description task through next text token prediction can significantly improve the text-image alignment of images generated by Chameleon [59]. + +![](images/967242589abd5ec544a06c3724f59e62339da10fb4ea8c1d004dd39b1250e757.jpg) +Figure 4. Qualitative Results on MagicBrush [81] testset comparing with MagicBrush results w/ and w/o context examples. + +# 4.3. Image Editing with RAIE + +Settings. As we introduced Retrieval-Augmented Image Editing (RAIE) in Sec. 3.3, we also prepare training data in the same way. We use publicly available UltraEdit [85] (500K) and MagicBrush [81] (8K) For the training. We use CLIP-B/32 [44] text encoder to encode the the edit instruction for each training sample and retrieval the most similar instruction feature as its neighbor sample (excluding the sample itself). During training, we input the neighboring sample as a task prompt context and prompt the model to predict the edited image. Besides generation task, we also use QWen2-VL [66] to describe the differences between the input image and the edited image. We add these difference description tasks into the training to help model gain better understanding of images and their variations. Preparing each editing pair with a similar editing pair for in-context generation is crucial to the success of RAIE, as we frequently observe similar editing pairs in both the Ultra-Edit and MagicBrush datasets. We provide detailed analysis of RAIE on MagicBrush in Appendix I. + +For evaluation, We use MagicBrush [81] benchmark, we also encode the edit instruction using CLIP-B/32 text encoder. We only use MagicBrush training set as reference database to perform Retrieval-Augmented Image Editing (RAIE) proposed in Sec. 3.3. The testing metrics and CLIP + +and DINO [9] models are consistent with [81, 85]. + +
TypeMethodsCLIPdir↑CLIPout↑CLIPimg↑DINO↑
ContinuousInstructPix2Pix [6]0.0810.2760.8520.750
MagicBrush [81]0.1060.2780.9330.899
UltraEdit [85]0.0930.2740.8990.848
DiscreteLumina-mGPT [33]0.0250.2530.8100.751
Ours + text prompt0.0670.2630.8230.785
Ours + RAIE0.0970.2790.8620.792
+ +Table 3. Image Editing Results. Comparison of different methods on the MagicBrush [81] testset. + +Results. As shown in Tab. 3, training solely with image editing pairs does not yield satisfactory results, as the model tends to replicate the original image rather than apply meaningful edits. Incorporating the additional difference description task through training model with next text token prediction encourages the model to identify distinctions between the input and edited images. This significantly improves the overall performance. Testing with an editing pair retrieved from the training set serving as an in-context example further enhances the quality of the edits, which proves the effectiveness of RAIE. Qualitative results are presented in Fig. 4 with user study in Appendix F. This experiment demonstrates the effectiveness of our in-context generation strategy for image editing tasks that require advanced semantic comprehension. + +
SettingsLow Light Enhancement LOLDerain Rain100HObject Addition InstructP2P [6]Object RemovalDepth Estimation NYU-v2 RMSE↓
PSNR↑SSIM↑PSNR↑SSIM↑CLIPdir↑CLIPout↑CLIPdir↑CLIPout↑
OmniGen [74] (in-context)8.9230.24313.140.4110.0540.2430.0310.233×
Full training19.710.77021.550.6330.1120.2830.1030.2650.279
No In-context9.1400.2537.9240.212-0.0310.2520.0230.2440.745
In-context w/o XP compression17.000.63318.100.5090.0920.2620.0690.2460.390
In-context w/ XP compression17.220.65318.910.5120.0920.2740.0730.2510.352
+ +Table 4. Results of in-context generation in novel task settings. "Full training" denotes for model trained with corresponding training set. While the other settings evaluate performance on tasks not encountered during training. + +
Context exampleInputw/ contextw/o contextContext exampleInputw/ contextw/o context
OmniGenperform depth estimationadd a castle
Ours
OmniGenperform image derainremove the sun
Ours
+ +Figure 5. Novel task in-context testing compared to OmniGen [74]. X-Prompt can achieve novel task generalization with a given example. While OmniGen [74] fall short in in-context generation (such as adapting to new color spectrum or preserve details when adding object to the image). + +
MethodImageInpaintingColorization
MSE↓LPIPS↓MSE↓LPIPS↓
Visual Prompting [4]0.980.650.670.40
Painter [67]1.120.840.940.72
Large Vision Model [3]0.620.500.510.46
OmniGen [74]0.910.720.930.91
Ours (w/o XP compression)0.520.440.580.44
Ours (w/ XP compression)0.450.340.490.37
+ +Table 5. Unseen Tasks Evaluation. X-Prompt achieve strong performance on previous unseen tasks of image inpainting and colorization compared to other in-context generation methods. + +# 4.4. In-Context Prompting on Novel Tasks + +Settings. Following the approach of GPT-3 [38], this experiment primarily investigates the generalizability of our model on novel tasks, given only a single example as context. We choose Low light enhancement and Image derain from low-level vision, object addition and object removal from Image Editing as novel task to perform this study. During training, we remove the training data for each novel tasks. For Image derain, we remove the training of generating derained image of Rain-13K. For low light enhancement, we remove the training of generating enhanced image of Mit-5K [8] and LOL [70]. For image editing task, we fil + +tered out the sample in Ultra-Edit [85] and MagicBrush [81] using LLama-3-Instruct-7B [14] by querying whether the instruction involves object addition or removal. Our test is perform on Rain100H [78] (Image derain), LOL-val (low light enhancement) and manually filtered 100 editing samples of Object Addition/Removal task from publicly available instructP2P [6] dataset. For depth estimation we test a new color palette that match the distance of the pixel on image to a different color spectrum. For low level vision, we report PSNR and SSIM as image evaluation metric. For Image Editing, we report CLIP [44] score consistent with [81, 85]. For depth estimation task, we report the RMSE on the previous unseen color spectrum. + +We report the quantitative results in Tab. 4. "Full training" refers to the model fully trained on corresponding training set. Both "In-context" and "No In-context" settings are trained with the corresponding training data completely removed. On novel task, in-context example can significantly improve results. In contrast, the model fails to perform novel tasks without the in-context example. We also ablate the effectiveness of token compression proposed in Sec. 3.1, where X-Prompt's attention mask force the model + +Style as Context +Figure 6. X-Prompt can support diversified context to achieve style personalization and action preservation. +![](images/fb8ec091570b6217e04dda1f71d1863e7ed3d0ed78a532a689f69ad19e734baf.jpg) +An old man with a straw hat and a Hawaiian shirt A woman dressed in a burgundy velet gown and gloves A girl with blonde hair, dressed in blue uniform + +compresses the in-context visual prompt implicitly. This achieves improved result (w/ or w/o XP compression). Detailed explanation of training and inference process are in Appendix B. We also compare our in-context generation capabilities with OmniGen [74], where we input three image follow the same prompt template in [74]. As shown in both Tab. 4 and Fig. 5, due to prohibitive context-length, OmniGen struggled to achieve generalized in-context generation. In depth estimation, it fails to generalize to unseen color spectrums. In image editing, OmniGen is unable to keep unchanged parts of the image consistent with the original, nor did it effectively follow contextual cues. For image deraining, the model struggled to interpret the context accurately, leading to unexpected results. In contrast, our model, leveraging unified text and image next-token prediction loss, demonstrates superior generalization to previously unseen tasks. + +For evaluation of unseen tasks compared to training set, we adopt the evaluation in image inpainting and colorization following the same setting as LVM [3], which are completely unseen domain compared to our training data. As reported in Tab. 5, X-Prompt achieve strong performance in completely unseen tasks compared to other visual prompting methods. Qualitative results are visualized in Fig. 1. The generalizability to previous unseen tasks proves the effectiveness of our vast tasks in-context prompting frame-work based on large multimodal pre-trained model Chameleon [59]. The effectiveness of X-Prompt compression is validated with improved performance. Details of X-Prompt process and efficiency analysis are in Appendix B and Appendix C. + +# 4.5. Other In-context Prompting Form + +Settings. In addition to training our model on existing datasets, we also create two small datasets for style + +personalization and action preservation to demonstrate X-Prompt's ability to extract diverse contextual information. For style-personalization, we use RB-Modulation [48] to generate image pairs based on style image and further filter low quality data with QWen2-VL [66] (detailed in Appendix H.2). For action preservation, we generate diversified human actions and use pose estimation model and ControlNet [82] to generate different person doing same action in the same pose. For each task, we generate 10K pairs for in-context generation. For style personalization, we give model an example transformation pair to prompt model perform similar transformation on an unseen image. For action and pose preservation, we give model two images of a person doing same action in similar pose and a new person description and prompt model to generate a new image. + +Results. We show qualitative results in Fig. 6. X-Prompt can extract both the high level semantics and low level details of the context example and perform successful transformation on new image or generation based on text prompt. This experiments demonstrate that X-Prompt can achieve in-context prompting in diversified multi-modal tasks. + +# 5. Discussion + +In this work, we propose empowering the autoregressive foundation model Chameleon [59] to achieve unified image generation through in-context prompting. We demonstrate its promising performance across tasks such as text-to-image generation, dense prediction, low-level vision, and image editing, and showcase its generalizability to previously unseen tasks when provided with in-context examples. We hope this work will pave way for this promising direction to achieve the "GPT-3 moment" in the unified multi-modal field in image generation. + +# Acknowledgments + +This work was supported by National Key R&D Program of China 2022ZD0161600, Shanghai Artificial Intelligence Laboratory, Hong Kong RGC TRS T41-603/20-R, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. Dahua Lin is a PI of CPII under the InnoHK. + +# References + +[1] Rohan Anil, Andrew M. Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Tachard Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Z. Chen, Eric Chu, J. Clark, Laurent El Shafey, Yanping Huang, Kathleen S. Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan A. Botha, James Bradbury, Siddhartha Brahma, Kevin Michael Brooks, Michele Catasta, Yongzhou Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, C Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, M. C. D'iaz, Nan Du, Ethan Dyer, Vladimir Feinberg, Fan Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, An Ren Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wen Hao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Mu-Li Li, Wei Li, Yaguang Li, Jun Yu Li, Hyeontaek Lim, Han Lin, Zhong-Zhong Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alexandra Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Marie Shelby, Ambrose Slone, Daniel Smilkov, David R. So, Daniela Sohn, Simon Tokumine, Dasha Valter, Vijay Vasudevan, Kiran Vodrahalli, Xuezhi Wang, Pidong Wang, Zirui Wang, Tao Wang, John Wieting, Yuhuai Wu, Ke Xu, Yunhan Xu, Lin Wu Xue, Pengcheng Yin, Jiahui Yu, Qiaoling Zhang Steven Zheng Ce Zheng Wei ZhouDenny Zhou Slav Petrov and Yonghui Wu.Palm 2 technical report. ArXiv, abs/2305.104032023.3 +[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Yitzhak Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo: An open-source framework for training large autoregressive vision-language models. ArXiv, abs/2308.01390, 2023. 3 +[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 22861-22872, 2024. 1, 3, 5, 7, 8 +[4] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. 3, 7 +[5] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 6 +[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 6, 7 +[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. ArXiv, abs/2005.14165, 2020. 1, 3 +[8] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Frédo Durand. Learning photographic global tonal adjustment with a database of input / output image pairs. In The Twenty-Fourth IEEE Conference on Computer Vision and Pattern Recognition, 2011. 7 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6 +[10] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diazorhan Firdat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res., 24:240:1-240:113, 2022. 3 + +[11] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness, 2022. 1 +[12] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1 +[13] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Ji-aqi Wang. Internl m-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model, 2024. 3 +[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 7 +[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3 +[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 3 +[17] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3855-3863, 2017. 5 +[18] Yulu Gan, Sungwoo Park, Alexander Schubert, Anthony Philippakis, and Ahmed M Alaa. Instructcv: Instruction-tuned text-to-image diffusion models as vision generalists. arXiv preprint arXiv:2310.00390, 2023. 1 +[19] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 3 +[20] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 3 +[21] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 5, 4 +[22] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text + +to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 5, 6 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[24] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and L. Sifre. Training compute-optimal large language models. ArXiv, abs/2203.15556, 2022. 3 +[25] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Huanzhang Dou, Yupeng Shi, Yutong Feng, Chen Liang, Yu Liu, and Jingren Zhou. Group diffusion transformers are unsupervised multitask learners. arXiv preprint arXiv:2410.15027, 2024.3 +[26] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775, 2024. 3 +[27] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 3 +[28] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1 +[29] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roktaschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020. 4 +[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 3 +[31] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiumiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024. 3 +[32] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 3 +[33] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 3, 5, 6 +[34] Ziyu Liu, Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Haodong Duan, Conghui He, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Mia-dpo: Multi-image augmented di + +rect preference optimization for large vision-language models. arXiv preprint arXiv:2410.17637, 2024. 3 +[35] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations, 2022. 5 +[36] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 3 +[37] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 3 +[38] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. 7 +[39] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 5 +[40] OpenAI. Gpt-4v(ision) system card. OpenAI, 2023. 3, 4 +[41] R OpenAI. Gpt-4 technical report. arXiv, pages 2303–08774, 2023. 1, 3 +[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 6, 2 +[43] Zhangyang Qi, Ye Fang, Zeyi Sun, Xiaoyang Wu, Tong Wu, Jiaqi Wang, Dahua Lin, and Hengshuang Zhao. Gpt4point: A unified framework for point-language understanding and generation. arXiv preprint arXiv:2312.02980, 2023. 3 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 6, 7, 3 +[45] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3 +[46] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 6 +[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 6 +[48] Litu Rout, Yujia Chen, Nataniel Ruiz, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng + +Chu. Rb-modulation: Training-free personalization of diffusion models using stochastic optimal control. arXiv preprint arXiv:2405.17401, 2024. 1, 8, 2 +[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5, 1, 3 +[50] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 1 +[51] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 5 +[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3 +[53] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 3, 6 +[54] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In The Twelfth International Conference on Learning Representations, 2023. 1, 3 +[55] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 1, 3 +[56] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want, 2023. 3 +[57] Zeyi Sun, Tong Wu, Pan Zhang, Yuhang Zang, Xiaoyi Dong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Bootstrap3d: Improving multi-view diffusion model with synthetic data, 2024. 3 +[58] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024.3 +[59] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2, 3, 4, 5, 6, 8 +[60] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable im + +age generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024.3 +[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. *ArXiv*, abs/2302.13971, 2023. 3 +[62] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[63] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1 +[64] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 3 +[65] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 1 +[66] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 4, 5, 6, 8, 2, 3 +[67] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 1, 3, 5, 7 +[68] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 1, 3 +[69] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 3, 6 +[70] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 5, 7 +[71] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 1, 3, 6 +[72] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 1, 3 +[73] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu + +Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1, 3 +[74] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2, 5, 7, 8 +[75] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 6 +[76] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 3 +[77] Long Xing, Qidong Huang, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Jinsong Li, Shuangrui Ding, Weiming Zhang, Nenghai Yu, Jiaqi Wang, Feng Wu, and Dahua Lin. Scalecap: Inference-time scalable image captioning via dual-modality debiasing, 2025. 3 +[78] Wenhan Yang, Robby T Tan, Jiashi Feng, Jiaying Liu, Zongming Guo, and Shuicheng Yan. Deep joint rain detection and removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1357-1366, 2017. 7 +[79] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 6 +[80] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 3 +[81] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 6, 7, 2, 3 +[82] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 8 +[83] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Haodong Duan, Songyang Zhang, Shuangrui Ding, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer: A vision-language large model for advanced text-image comprehension and composition, 2023. 3 +[84] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024.3 +[85] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and + +Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. arXiv preprint arXiv:2407.05282, 2024. 1, 6, 7, 2, 3 +[86] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5 \ No newline at end of file diff --git a/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/images.zip b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3eaab508c618de3f282c971a12c44cbcc7b5de33 --- /dev/null +++ b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a7ab79e781405d379e1226301f465448894bda0f3166c201f7163a4761643a9 +size 1076077 diff --git a/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/layout.json b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c6b620c1c0f44e945dcefe10adafcd311597c9e0 --- /dev/null +++ b/xpromptgeneralizableautoregressivevisuallearningwithincontextprompting/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8d03df4349050d954168b037cbf04cf951b6b732007c21670afd4d082c98a58 +size 496216 diff --git a/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/c568565b-91f6-457d-aad4-2c25aae796ae_content_list.json b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/c568565b-91f6-457d-aad4-2c25aae796ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56e1ab82eaef5ceded3a86f6b58c526d6ed76423 --- /dev/null +++ b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/c568565b-91f6-457d-aad4-2c25aae796ae_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03e5d81c2a0c761588a5471281dbced09cc373a300600b3af6f56e494536e90d +size 89652 diff --git a/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/c568565b-91f6-457d-aad4-2c25aae796ae_model.json b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/c568565b-91f6-457d-aad4-2c25aae796ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..048fe3ecfc3c1a95a18c3a544c74eb694e3b199e --- /dev/null +++ b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/c568565b-91f6-457d-aad4-2c25aae796ae_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc3f7591b7fa9b4f5b914b73c6f0fa63fedfe44d5806a2d3996eb2a6a091b946 +size 113271 diff --git a/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/c568565b-91f6-457d-aad4-2c25aae796ae_origin.pdf b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/c568565b-91f6-457d-aad4-2c25aae796ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a69227c5259ea7e2fced90ed1813bff3f8d08c3 --- /dev/null +++ b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/c568565b-91f6-457d-aad4-2c25aae796ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:892dd1aa4a61a6c9c95f1e7c065db0e8b7260f730cb26cbdc103284a274988b8 +size 1334353 diff --git a/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/full.md b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..42356a753a4ff29b04e0be785b7625aad8fda26a --- /dev/null +++ b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/full.md @@ -0,0 +1,398 @@ +# XTrack: Multimodal Training Boosts RGB-X Video Object Trackers + +Yuedong Tan $^{1,2}$ Zongwei Wu $^{1*}$ Yuqian Fu $^{2}$ Zhuyun Zhou $^{1}$ Guolei Sun $^{4}$ Eduard Zamfir $^{1}$ Chao Ma $^{3}$ Danda Paudel $^{2}$ Luc Van Gool $^{2,4}$ Radu Timofte $^{1}$ + +1 Computer Vision Lab, CAIDAS & IFI, University of Wurzburg 2 INSAIT, Sofia University + +$^{3}$ AI Institute, Shanghai Jiao Tong University $^{4}$ CVL, ETH Zurich + +# Abstract + +Multimodal sensing has proven valuable for visual tracking, as different sensor types offer unique strengths in handling one specific challenging scene where object appearance varies. While a generalist model capable of leveraging all modalities would be ideal, development is hindered by data sparsity, typically in practice, only one modality is available at a time. Therefore, it is crucial to ensure and achieve that knowledge gained from multimodal sensing – such as identifying relevant features and regions – is effectively shared, even when certain modalities are unavailable at inference. We venture with a simple assumption: similar samples across different modalities have more knowledge to share than otherwise. To implement this, we employ a classifier with weak loss tasked with distinguishing between modalities. More specifically, if the classifier "fails" to accurately identify the modality of the given sample, this signals an opportunity for cross-modal knowledge sharing. Intuitively, knowledge transfer is facilitated whenever a sample from one modality is sufficiently close and aligned with another. Technically, we achieve this by routing samples from one modality to the expert of the others, within a mixture-of-experts framework designed for multimodal video object tracking. During the inference, the expert of the respective modality is chosen, which we show to benefit from the multimodal knowledge available during training, thanks to the proposed method. Through the exhaustive experiments that use only paired RGB-E, RGB-D, and RGB-T during training, we showcase the benefit of the proposed method for RGB-X tracker during inference, with an average +3% precision improvement over the current SOTA. The source code is publicly available at https://github.com/supertyd/XTrack. + +# 1. Introduction + +Visual object tracking has made significant progress, driven by recent advancements in deep learning [4-6]. While RGB-based trackers have demonstrated strong, generaliz + +![](images/cf1df479e8457da9683694bf2091ead0f286e8f4e23b4f85ae65ba836e2b08c1.jpg) +(a) Conventional Single Modal Training for VOT + +![](images/f5342c852666bcdf4ed8faf31fc1033c2c88b3f538fd573ef382a5f42b44ec70.jpg) +(b) Ours with Cross Modal Knowledge Sharing +Figure 1. Motivation: (a) Existing tracking methods typically address each modality in isolation, tackling one appearance-related challenge at a time. This is mainly due to cross-modal domain gaps and the lack of a comprehensive multimodal dataset. Consequently, only modality-specific branches are activated during inference based on a priori knowledge of the input type, limiting the potential for cross-modal integration. (b) We leverage Mixture of Experts to achieve such knowledge sharing. This is made possible by integrating a weak auxiliary classification loss, which is achieved for the first time for an RGB-X tracker. + +able performance [5, 12, 33, 47], they still struggle with large appearance changes under challenging scenes [11, 20, 38]. Additional sensory input can enhance robustness in these scenarios, with each sensor addressing a specific challenge — such as Event cameras[46, 58] for motion awareness, depth sensors[68] for occlusion handling, and thermal imaging[28] for illumination variation — leading to increasing interest in multimodal object tracking [30, 53, 66]. + +However, current advances in multimodal tracking are + +limited by the lack of a comprehensive dataset that includes all modalities simultaneously. Most existing datasets treat each modality separately [26, 29, 68]. Consequently, there is significant research interest in developing multimodal tracking systems that can benefit from multimodal training using only paired data, while generalizing to any modality available during inference, achieving the same level of generalization as conventional RGB trackers [18, 19, 66]. + +Unfortunately, these practical concerns have been largely overlooked in previous works. Early multimodal tracking approaches [29, 30, 53, 68] were designed to handle only one additional modality, such as Event [45, 58], Depth [53], or Thermal data [28], making it inherently impossible to benefit from broader multimodal training. While recent works [18, 19] have attempted to transcend modality-specific architectures by creating general processing blocks, they typically rely on predetermined branches activated based on the input modality type, assuming prior knowledge of the modality during inference [35, 49]. This rigid separation of modality-specific parameters, while achieving unification, prevents any meaningful interaction between modalities during training. Such practice of strict modality isolation fails to capitalize on a crucial opportunity: the potential for cross-modal knowledge transfer [39]. Consider a complex RGB-Depth sequence involving fast-moving objects under low-lighting conditions - a model should naturally acquire knowledge that transcends modality-specific boundaries, developing insights transferable across different sensing domains. + +In this work, we present the first systematic approach to enable cross-modal knowledge sharing in RGB-X video object tracking. Our key insight is that similar samples across different modalities naturally share more transferable knowledge. More importantly, when feature similarity creates confusion between modalities, we posit that this "confusion" signals an optimal opportunity for knowledge sharing, as it indicates minimal domain gap and maximal communication between modality-specific attributes. + +To implement this insight, we introduce a novel approach using a classifier together with a weak loss to identify such "confusion", or more properly speaking, knowledge-sharing opportunities. We realize this through a carefully designed Mixture of Modal Experts (MeME) framework. Different from previous mixture-of-experts (MoE) work aiming for inference speedup [7, 67], we combine the MoE sparse architecture with modality-specific processing. More importantly, we design a soft router where experts originally trained for one modality can be leveraged by another when their features align sufficiently. Specifically, we first learn a dynamic routing function that acts as a classifier with explicit modal awareness. Based on the router's soft predictions, the input modality is directed to relevant experts for specialized analysis, creating a blended representation that + +![](images/9b5b8da9daa2460d4b9aa4d64ffdeb3c5e267d391ac414e39e6e8c0816c34a22.jpg) +Figure 2. Joint Benefits: Despite domain gaps, some samples across modalities share similar attributes, creating overlap in representation, as shown in the left. This overlap complicates strict modality classification, introducing ambiguity. For example, on the Event benchmark, SOTA methods like ViPT [66] perform worse when trained on multiple modalities than on Events alone. In contrast, we view this ambiguity as a chance for cross-modal knowledge sharing. Our approach leverages this potential, enabling effective multimodal training and consistent improvement. + +incorporates both shared and modality-specific knowledge. + +At the expert level, our MeME framework employs a shared unit for general knowledge alongside several modality-specific experts focusing on distinct attributes (e.g., geometry, motion, low-light adaptation). This attribute-based, flexible understanding enables comprehensive cross-modal intelligence, advancing our tracker's capability to handle complex scenarios without requiring prior modality information, which is the first of its kind for VOT. + +Through extensive experiments with paired RGB-E [45], RGB-D [53], and RGB-T [30] data during training, we show that our approach successfully preserves and transfers multimodal knowledge to enhance single-modality tracking during inference. Our method achieves an average precision improvement of 3 points over the current state-of-the-art across various RGB-X tracking scenarios, establishing a new paradigm for robust multimodal tracking in real-world conditions where sensor availability cannot be guaranteed. + +# 2. Related Work + +Multi-Modal Tracking: In the evolution of object tracking [6, 9, 23, 64], the model architecture has undergone a continuous transformation, transitioning from correlation filters [17, 23] to Siamese networks [2, 15, 31], and then to the current transformer architecture [4, 36] that integrates unified feature extraction and interaction. However, the performance and stability of object tracking are still limited when dealing with complex scenes. + +Multi-modal information, including depth [16, 54], event [59, 70], and thermal data [62, 65], can compensate for this deficiency and enhance the robustness of object tracking networks when dealing with objects with large appearance variation. For example, BAT [3] and TBSI [21] have designed a dual-transformer architecture specifically for RGBT object tracking, while DepthTrack [53] is tailored for + +RGBD object tracking. As can be seen, most of these works are modality-tailored, without being able to transfer the knowledge from one modal domain to another. Recently, there has been a growing interest in developing multi-modal object tracking, especially following the design paradigm of prompt learning, aiming to establish a unified yet efficient object tracking paradigm [18, 19, 49, 49, 55, 66]. However, these works adopt a rigid design, only activating the modality-specific parameter during inference, making it impossible to leverage joint benefits across all the modalities. + +Emergent Intelligence: Previous works have demonstrated the benefits of training on diverse modalities, including images, text, and 3D data, while others have highlighted the advantages of learning across different tasks. The ultimate goal is to create a single network capable of handling any type of input, an ability inherent to humans. Some researchers refer to this as a generalist model. For example, the Meta-Transformer [63] aligns various modalities into a unified feature space, while ImageBind [14] uses images to align data from different modalities. PolyViT [32] has pioneered unified training of a Vision Transformer (ViT) backbone with multiple modalities. + +These approaches highlight the importance of a joint learning space, where knowledge can be transferred between domains. In this paper, we aim to achieve this goal within the context of multimodal object tracking, addressing this challenge for the first time. + +Mixture-of-Experts: Mixture-of-Experts (MoE) has emerged as a powerful technique that enhances model capability and efficiency by delegating tasks to specialized expert networks. This architecture is particularly effective for handling complex and diverse data, as it dynamically selects the most appropriate expert based on the input's characteristics. Models such as VMoE [41] and DAMeX [22] have significantly improved image classification and object detection tasks by extending feed-forward network (FFN) layers within the vision transformer architecture. These models adjust the combination of their internal expert networks based on input features, thereby capturing complex patterns more effectively. DeepSeekMoE [7] exemplifies the use of MoE in large language models, improving language understanding and generation through the expert mixture mechanism. Other models like IMP [1] and LiMoE [39] apply MoE to unify tasks across different modalities, such as visual, auditory, and textual data processing. + +However, existing works primarily associate MoE's sparsity with task-specific needs, overlooking the joint benefits of cross-modal integration due to rigid task or modality-specific isolation. In this paper, we aim to introduce a more flexible and intelligent separation strategy that facilitates cross-modal knowledge transfer for VOT. + +# 3. Method + +# 3.1. Overall Pipeline + +In line with the current tracking paradigm [18, 66], we input both the template and the search region, along with their respective multi-modal patches, into the same patch embedding to obtain the RGB tokens $T_{rgb}$ and the auxiliary modal tokens $T_x$ . To ensure effective interaction between the two modalities, as illustrated in Fig. 3(left), we incorporate the Mixture of Modal Experts (MeME), which enables intertwining between RGB and X input. Our MeME takes place after each attention block (Attn) and the Feed-Forward Network (FFN) of our baseline foundation RGB trackers, built using transformers with frozen weights [11, 20, 34, 38]. + +Our goal is to leverage the joint benefits of the RGB and X modalities to enhance feature modeling and better handle large appearance variations across video frames. To achieve this, MeME is designed in a bidirectional manner, allowing the X modality to improve the RGB features, and vice versa. For simplicity and clarity, we show in the following section how the modal token $T_{x}$ improves the RGB token $T_{rgb}$ . + +As shown in Fig. 3(left), at the $l^{th}$ transformer block, our MeME takes the RGB token $T_{rgb}^{l}$ and the X token $T_{x}^{l}$ from the previous transformer layer outputs. We first merge these two feature sets and enhance the conventional feature modeling with transformer attention: + +$$ +T _ {r g b} ^ {\text {a t t n}} = T _ {r g b} ^ {l} + \operatorname {A t t n} \left(T _ {r g b} ^ {l}\right) + \operatorname {M e M E} \left(T _ {r g b} ^ {l}, T _ {x} ^ {l}\right). \tag {1} +$$ + +Similarly, we enable the cross-modal interaction again to complement the FFN processing. We have: + +$$ +T _ {r g b} ^ {l + 1} = T _ {r g b} ^ {a t t n} + F F N \left(T _ {r g b} ^ {a t t n}\right) + M e M E \left(T _ {r g b} ^ {a t t n}, T _ {x} ^ {a t t n}\right), \tag {2} +$$ + +where $T_{x}^{attn}$ is computed in the similar way as in Eq. 1. + +# 3.2. Mixture of Modal Experts (MeME) + +MeME operates on two key principles: (a) assigning the most appropriate and specialized experts based on the input modality, and (b) isolating the shared experts that are common across all modal inputs to facilitate cross-modal alignment and reasoning. By jointly incorporating both specialized and shared experts, MeME allows for the exploration of modal-specific clues with finer granularity thanks to the reduced redundancy within each specialized expert, and ensures emergent alignment across modalities, even when trained solely on paired data. + +Router Supervision: One key component of an MoE architecture is the routing function: + +$$ +y = \sum_ {i \in t o p _ {-} k} p _ {i} \left(T _ {x}\right) e _ {i} \left(T _ {x}\right), \tag {3} +$$ + +where we adopt top- $k$ experts for routing the input, $p_i$ is the conventional probability [7], $y$ is the output of MoE, and + +![](images/82779da8a80042b5feeff8361a0455cdd3320d197bc256158242dd7a53a8d8c0.jpg) +Figure 3. Mixture of Modal Experts (MeME): Recent work [7] has improved MoE by leveraging shared experts. However, it remains unclear which expert learns what, due to the implicit learning setting. In contrast, we make the whole protocol explicit by first gating the shared expert with edge initialization and second assigning specific experts (D/T/E) to particular modal inputs with the help of an additional classifier. More importantly, we employ a weak classification loss for knowledge sharing. Detailed advantages can be found in Tab. 1. + +![](images/1394735b79d520116f4729301be40702d29c490279550f9ca1b2f0e232102345.jpg) +(a) Conventional MoE + +![](images/206c9cab351de9d4f3b5bd405c0136d87fe055a88e1ca978f6b5cf02c30c823e.jpg) +(b) + Shared Expert + +![](images/32bb024a86a265e18fe0a1cd21990fddbff3b19cb46346299c777e811a317f43.jpg) +(c) Our MeME with Modal Expert & Edge-Gated Shared Expert + +$e_i$ is the expert. We supervise such a routing function by two losses: the widely adopted expert balance loss and our newly introduced classification loss: + +Expert Balance Loss: Balance loss is composed of two parts, the importance loss $\mathcal{L}_{Imp}$ and the loading loss $\mathcal{L}_{Load}$ . Let $I$ denote the vector of importance scores for all experts, where each element $I_{i}$ represents the importance score of expert $\epsilon_{i}$ . The importance score $I_{i}$ is calculated as the sum of the probabilities $p_i$ over all samples. We consider the normal distribution denoted as $\mathcal{N}(0,\sigma^2 I)$ , where the standard deviation $\sigma$ is defined as the ratio of gate noise to the number of experts $|\epsilon|$ . $\Phi(\cdot)$ is the cumulative distribution function (CDF) of this normal distribution. + +For the $\text{Load}_i$ of expert $e_i$ , we evaluate the CDF at the probabilities $p_i(T_x)$ , represented as $\Phi(p_i(T_x))$ , and sum these values over all inputs $T_x$ in a single mini-batch. The load loss $\mathcal{L}_{\text{Load}}$ is then defined as: + +$$ +\mathcal {L} _ {\text {L o a d}} = \frac {\operatorname {V a r} (\text {L o a d})}{\operatorname {M e a n} (\text {L o a d}) ^ {2}}; \operatorname {L o a d} _ {i} = \sum_ {T _ {x} \in B} \Phi \left(p _ {i} \left(T _ {x}\right)\right). \tag {4} +$$ + +Here, $\text{Load}_i$ represents the number of assignments allocated to expert $e_i$ , and $B$ stands for the batch. Similarly, we obtain the importance loss $\mathcal{L}_{\text{Imp}}$ , but without considering the CDF, by: + +$$ +\mathcal {L} _ {I m p} = \frac {\operatorname {V a r} (I m p)}{\operatorname {M e a n} (I m p) ^ {2}}; I m p _ {i} = \sum_ {T _ {x} \in B} p _ {i} \left(T _ {x}\right). \tag {5} +$$ + +Hence, our balance loss $\mathcal{L}_{\text{balance}}$ is formulated by: + +$$ +\mathcal {L} _ {\text {b a l a n c e}} = \mathcal {L} _ {\text {I m p}} + \mathcal {L} _ {\text {L o a d}}. \tag {6} +$$ + +Weak Classification Loss: To enable each expert in our MeME to benefit from multimodal information while maintaining a certain degree of modality specialization, we improve the conventional routing function with an additional classification loss. Compared to conventional MoE variations, our contributions are twofold: + +- Explicit Routing: Unlike conventional implicit routing [7, 67], we use explicit classification-based routing, directly + +Table 1. Benefits of our soft routing with weak classification loss. + +
Unified ModelRouting FunctionSensor SpecificationCross-Expert Learning
SOTA Trackers--Yes-
+ MoE [7, 67]YesImplicit-Yes
+ Classifier (Strong Loss)YesExplicitYes-
+ Classifier (Weak Loss)YesExplicitYesYes
+ +linking experts to modalities for disentangled sensor representations and transparent selection. + +- Flexible Cross-Expert Learning: Explicit routing is conventionally achieved by incorporating a strong classification loss. This ensures sensor-specific expertise but may hinder cross-modal learning under data scarcity. We mitigate this by relaxing and adopting a weak classification loss, allowing experts to integrate cross-modal features while retaining their identity. This enables effective knowledge sharing, which is the first of its kind. + +Our benefits are summarized in Tab. 1. Technically, to achieve this soft routing, we additionally employ a multiclass loss to constrain the data that each expert can process. Given a collection $M$ of modalities, indexed by $n$ , denoted as $\{m_{n = 1}^{|M|}\}$ . We introduce a mapping function $h:\{M\} \to \{\epsilon\}$ such that each modality $m_{n}$ is associated with several particular experts $e_i \in \epsilon$ , which ideally are not shared from one modality to another. The classification loss $\mathcal{L}_{cls}$ is computed using a binary cross-entropy loss, comparing the logits $p_i$ (representing the probability of selecting experts) with the labels $h(m_n)$ (indicating the target expert for tokens from modality $m_{n}$ ). We have $\mathcal{L}_{cls}$ : + +$$ +\mathcal {L} _ {c l s} = - \sum_ {n = 1} ^ {| M |} \sum_ {i = 1} ^ {| \epsilon |} 1 _ {\{h \left(m _ {n}\right) = i \}} \log p _ {i} \left(T _ {x}\right), \tag {7} +$$ + +where $p_i(T_x)$ is the probability, $1_{\{h(m_n) = i\}}$ is the indicator function that equals to 1 if $h(m_n) = i$ and 0 otherwise, and $|\epsilon|$ is the cardinality of the set of experts $\epsilon$ . + +Finally, during training and on top of conventional tracking, our model additionally leverages MoE loss $\mathcal{L}_{moe}$ to achieve our targeted soft routing setting: + +$$ +\mathcal {L} _ {m o e} = \mathcal {L} _ {c l s} + \lambda \cdot \mathcal {L} _ {\text {b a l a n c e}}, \tag {8} +$$ + +where $\lambda$ is the proportion hyperparameter. + +# 3.3. Efficient Expertized Reasoning and Gathering + +Experts with Low-Dimensional Reasoning: The goal of each specialized expert is to project the input feature into a dedicated space for further processing. Inspired by recent successes in tuning large language models [44, 67], we aim to enhance efficiency by mitigating low-dimensional feature decomposition and reconstruction. Specifically, we project each input token with channel $c$ to the significantly lower dimensional latent space $k$ ( $k << c$ ). This approach preserves the principal features while reducing computational cost, making the learning feasible on a single 24 GB GPU. + +Assume that our router decomposes the mixed modal input $M$ into subset features $m_{n}$ . Their respective $k_{th}$ low-dimensional matrices $m_{n,k}$ , are approximated by: + +$$ +m _ {n, k} = e _ {n} (m _ {n}), \tag {9} +$$ + +where $e_n$ mimics the lower dimension projection for key feature transformation. Note that all the $e_n$ structures are identical for each modality-specific expert but without weight sharing, allowing for distinct reasoning. + +Edge-Gated Shared Expert: Simultaneously, we compute the shared low-dimensional matrix $m_{s}$ from the shared expert $e_{s}$ . However, we believe that for the shared expert, a single projection through implicit learning might not be sufficient, especially when the training data size is relatively small compared to that used for training large foundation models. Therefore, we think it is necessary to incorporate additional human prior to guide the shared learning process. Since the objective is to find commonalities, we believe that high frequency is a naturally shared feature among these modalities, typically manifesting in the form of edges. + +Technically, we introduce a gating module at low latent space. A graphical illustration of this process can be found in Fig. 4(a). Let $m_{sk}$ be the input shared low-dimension feature. We obtain the output $Out_{1}$ by: + +$$ +X = \operatorname {N o r m} \left(m _ {s k}\right), +$$ + +$$ +\lambda = N \sigma_ {k} / (m _ {g k}), \tag {10} +$$ + +$$ +\operatorname {O u t} _ {1} = \left(\sigma \left(\operatorname {E d g e M i x} \left(X W _ {1}\right)\right) \cdot X W _ {2}\right) W _ {3} + m _ {s k}, +$$ + +where $\text{Norm}()$ denotes the conventional normalization. $\text{EdgeMix}()$ refers to the module, with Laplacian initialization, responsible for conducting token mixing, as shown in Fig. 4(a). Technically, we begin by transferring the token to a 2D feature form and then apply convolution to enable interaction between neighboring tokens. Unlike previous approaches, in our case, we additionally initialize the convolution with a Laplacian Filter, incorporating edge priors while remaining open to discovering more meaningful and learnable shared features. $W_{1}, W_{2}, W_{3}$ represent learnable parameters within the low-dimensional latent space. + +![](images/1417f436e9f4bf22ef17e852078b31b0fb760196aafbd76aa010bd5400210454.jpg) +(a) Edge-Gated Shared Expert +Figure 4. Details on experts and prompts. + +![](images/e85b9f6e63a4ab6443bc9df763c5f94df3a6ccbd605a47b0b038fe9ecac28bed.jpg) +(b) Prompting Module + +Fusion of Expertized Tokens: The next step aims to fuse previously separately reasoned tokens together into a global low-dimensional matrix $M_{k}$ . It should contain both modality-specific and modality-shared clues. Technically, we develop a shrinkage fusion design, where we first concatenate $m_{nk}$ together batch-wise, obtaining $y$ from Eq. (3). Then, we learn a joint approximation from the whole batch, which is further incorporated with the shared matrix $m_{s}$ . This pipeline can be expressed as follows: + +$$ +M _ {k} = \left(y W _ {4} + O u t _ {1}\right) W _ {5}, \tag {11} +$$ + +where $W_{4}, W_{5}$ are low latent space learnable parameters. + +# 3.4. Modal Prompting + +Finally, we aim to prompt the RGB token obtained from the frozen RGB foundation tracker and transform it to be modality-aware, making it more suitable for downstream and challenging cases. Following the same idea, we first project the RGB token into the low latent space $I_{k}$ . Then, we use the approximated modal low-dimensional matrix $M_{k}$ from Eq. 11 to prompt $I_{k}$ . Our motivation is to use the modality clues as a gating module to improve the RGB feature modeling. Specifically, we have: + +$$ +\begin{array}{l} X _ {i} = \operatorname {N o r m} \left(I _ {k}\right), X _ {m} = \operatorname {N o r m} \left(M _ {k}\right), \\ O u t _ {2} = \left(\left(X _ {i} W _ {5} \cdot \sigma \left(X _ {m} W _ {6}\right)\right) W _ {7} + I _ {k}\right) W _ {8}, \tag {12} \\ \end{array} +$$ + +where $W_{5}, W_{6}, W_{7}$ are the learning parameters in the low latent space, and $W_{8}$ projects the prompted token back to the initial embedding space. And, $Out_{2}$ is the output of MeME. + +# 4. Experiments + +We adopt [56] and [48] as RGB foundation trackers by freezing the parameters, forming XTrack-B and XTrack-L, respectively. To train our XTrack, only the parameters in MeME are learnable, as shown in Fig. 3. We set the batch size to 32, with a learning rate of 4e-4, training for 60 epochs. The learning rate is decreased by a factor of 10 after 78 epochs. In addition to the object tracking loss shared with the RGB Tracker, we incorporate a loss that guides + +MoE expert balance and specialization of modality-specific experts, as detailed in Sec. 3.2. + +Regarding dataset selection, we utilize DepthTrack [53] for RGB-Depth sequences, LasHeR [30] for RGB-Thermal sequences, and VisEvent [45] for RGB-Event sequences. Previous well-established modality-tailored works [18, 19, 55, 66], we train our model from all available modalities. We want to highlight that only one RGB-X pair (i.e., RGB-Depth, or RGB-Thermal, or RGB-Event) is available at a time, which makes our problem setting challenging. + +# 4.1. Depth Benchmark + +We compare our XTrack with two groups of competitors: 1) specific models only for tackling depth dataset [13, 24, 53, 57, 69], 2) SOTA generalizable models but with tailored depth parameters [18, 19, 27, 49, 66]. + +DepthTrack: DepthTrack [53] is a comprehensive, long-distance dataset consisting of 50 video sequences for test. We evaluate the accuracy of various trackers on this dataset using F-score, Recall (Re), and Precision (Pr) metrics. Our results show that: 1) XTrack outperforms all models specifically designed for the DepthTrack dataset, as well as unified models using depth-specialized parameters. 2) With an increase in model scale from XTrack-B to XTrack-L, we achieve notable improvements, setting new SOTA records. + +VOT-RGBD2022: The VOT-RGBD2022 dataset [26], which includes 127 video sequences, is the latest benchmark for RGB-D object tracking. We evaluate our model's performance using the metrics of expected average overlap (EAO), accuracy, and robustness. + +Note that VOT-RGBD2022 contains out-of-distribution samples, creating a domain gap between the DepthTrack training set and the VOT-RGBD2022 test set. However, as shown in Tab. 2, our method surpasses current state-of-the-art performance on VOT-RGBD2022 by an even larger margin than on DepthTrack. This result highlights our model's enhanced capability to manage domain gaps, a strength we attribute to our multimodal training. + +# 4.2. Thermal Benchmark + +We also compare our XTrack with two groups of competitors: 1) task specific ones and 2) the same group of unified models but specialized for RGB-Thermal. + +LasHeR: The LasHeR [30] dataset is a large-scale, high-quality dataset comprising 245 test sequences. We evaluate the performance of trackers using precision (Pr) and success plots (Sr). Notably, XTrack achieves SOTA on this dataset clearly outperforming all the competitors, without requiring any parameter shifting from one depth to thermal. + +RGBT234: The RGBT234 [29] dataset is a comprehensive video dataset, including 234 video sequences for testing, encompassing a variety of environmental challenges, including rainy conditions, nighttime scenes, and extreme + +Table 2. Performance on RGB-Depth datasets. + +
MethodDepthTrack [53]VOT-RGBD22 [26]
F-scoreRePrEAOAcc.Rob.
ATCAIS [24]47.645.550.055.976.173.9
DRefine [25]---59.277.576.0
KeepTrack [37]---60.675.379.7
DDiMP [24]48.556.950.3---
DMTrack [26]---65.875.885.1
DeT [53]53.250.656.065.776.084.5
OSTrack [56]52.952.253.667.680.383.3
SBT-RGBD [51]---70.880.986.4
SPT [68]53.854.952.765.179.885.1
ProTrack [55]57.857.358.365.180.180.2
ViPT [66]59.459.659.272.181.587.1
UnTrack [49]61.060.861.172.182.086.9
OneTracker [18]60.960.460.772.781.987.2
SDSTrack [19]61.460.961.972.881.288.3
XTrack-B61.562.061.874.082.188.8
XTrack-L64.864.365.474.082.888.9
+ +Table 3. Performance on RGB-Thermal datasets. + +
MethodLasHeR [30]RGBT234 [29]
PrSrMPRMSR
mfDiMP [60]44.734.364.642.8
JMMC [61]--79.057.3
CMPP [43]--82.357.5
APFNet [50]50.036.282.757.9
OSTrack [56]51.541.272.954.9
ProTrack [55]53.842.079.559.9
ViPT [66]65.152.583.561.7
UnTrack [49]64.651.384.262.5
SDSTrack [19]66.553.184.862.5
OneTracker [18]67.253.885.764.2
XTrack-B69.155.787.464.9
XTrack-L73.158.787.865.4
+ +weather scenarios. Results in Tab. 3 show our consistent improvement over SOTA methods across all the sequences. + +# 4.3.Event Benchmark + +VisEvent: As for the RGB-Event benchmark, we also compare our model with two kinds of trackers: event-specific ones and the same group of unified models but specialized for RGB-Event tracking. + +VisEvent [45] is a large-scale dataset containing RGB images and event data for object tracking, comprising 320 test sequences. We use precision (Pr) and success plots (Sr) to evaluate the performance of different trackers. The quantitative results in Tab. 4 validate our effectiveness, achieving new SOTA records on the Event benchmark. + +# 5. Ablation Studies and Discussion + +Without explicitly mentioning, all the experiments in this section are reported based on our XTrack-B. Please refer to the Supplementary for more comparisons and visuals. + +Table 4. Performance on RGB-Event datasets. + +
MethodVisEvent [45]
PrSr
MDNet_E [40]66.142.6
STARK_E [52]61.244.6
PrDiMP_E [10]64.445.3
LTMU_E [8]65.545.9
TransT_E [4]65.047.4
SiamRCNN_E [42]65.949.9
OSTrack [56]69.553.4
ProTrack [55]63.247.1
ViPT [66]75.859.2
UnTrack [49]75.558.9
OneTracker [18]76.760.8
SDSTrack [19]76.759.7
XTrack-B77.560.9
XTrack-L80.563.3
+ +Table 5. Key Component Analysis + +
Shared Spec.DepthTrackLasHeRVisEvent
F-scoreRePrPrSrPrSr
✓57.657.557.668.054.876.259.9
✓59.159.259.467.854.276.559.6
✓✓61.562.061.869.155.777.560.9
+ +**Shared Experts:** We conduct an ablation study on the shared experts. As shown in Tab. 5, we found that, for DepthTrack [53], removing the shared experts resulted in a $-3.9\%$ decrease in F-score. Removing the shared experts might lead to redundancy in the modality-specialized experts, reducing their effectiveness and hence leading to such a performance drop. This observation joins the conclusion from the previous work [7]. + +Modality-Specific Experts: In addition to the shared experts, our MeME also includes modality-specific experts to capture unique information from each modality. As shown in Tab. 5, we remove this design and process all modalities using a single expert. In such a case, all the experts are treated equally, losing the identification such as geometry from depth, motion from event, or temperature from thermal, which is not desirable. This can be also proven by the declined performance across various modalities, particularly in depth. Such a performance drop indicates the effectiveness and necessity of our decomposition and the collaboration of both modality-shared and modality-specific features. This design enables the effective modeling of unique characteristics and makes the different modalities benefit from each other. + +Number of Experts: We further conduct ablations on the number of experts, by setting it to 1, 2, and 3. Experimental results in Tab. 6 show that using two modality-specific experts achieves the best performance. With only one expert per modality, feature representation is limited, while using three experts introduces internal conflicts, which also degrade model performance. + +Table 6. Study on the Expert Number. + +
NumberDepthTrackLasHeRVisEvent
F-scoreRePrPrSrPrSr
160.061.260.668.854.976.059.8
261.562.061.869.155.777.560.9
360.960.160.568.255.075.759.7
+ +Table 7. Multimodal Training Benefits. + +
EDTVisEventDepthTrackLasHeR
PrSrF-scoreRePrPrSr
69.553.452.952.253.651.541.2
✓76.359.945.545.845.758.145.1
✓✓76.660.260.859.560.158.145.2
✓✓✓77.560.961.562.061.869.155.7
+ +Joint Training Benefits: Our model is designed for cross-modal knowledge transfer, enabling it to benefit from training datasets with diverse domain samples. To verify this, we conducted experiments by progressively expanding the training set with additional modalities. + +We first trained the model solely on event data and tested it on the event benchmark, mirroring previous modality-specific or unified models with event-only parameters activated. In this setting, as shown in Tab. 7, our model achieved a $+8\%$ improvement in precision over the baseline [56]. Interestingly, even when trained exclusively on event data, our model demonstrated strong zero-shot generalization to thermal data, achieving results comparable to ProTrack [55], which was trained and tested on thermal data. However, zero-shot transfer to the depth dataset proved more challenging, likely due to a greater domain gap between event and depth data than between event and thermal. This finding aligns with our model's underlying motivation to leverage transferable knowledge across modalities: event cameras, sensitive to light changes, share similarities with thermal cameras, which also handle illumination variation effectively. + +Moreover, we show that joint training with both depth and event samples further boosts our model's performance. However, we also observe that this setup does not necessarily enhance zero-shot performance on the thermal domain. This may be because the domain gap between event and depth data is too substantial. As a result, the model primarily focuses on bridging the domain gap rather than fostering cross-domain alignment. Finally, joint training with all available modalities further enhances performance across each individual modality. + +Soft Router: In our router design, we use classification loss to activate the most suitable modality-specific experts. A higher classification loss pushes the classifier toward strict separation, resulting in a rigid design with parallel, non-communicating branches - similar to current multimodal tracking models. Conversely, a lower classification loss introduces more ambiguity, approaching random expert as + +![](images/3e000c509cd9009e0ab4607957fec61f24cbfc4cdb5ff753e9d65e79bc2519a0.jpg) +Figure 5. Routing Choice. We present the top-k ( $k = 2$ ) decisions for activating the Depth, Event, and Thermal experts, using RGB-Event as input during inference. Our model dynamically selects the most suitable experts for different challenging scenarios, ensuring optimal object tracking performance despite appearance changes across diverse scenes. + +![](images/e512f8cd5177c1916139b4982723a6e787c35e9219ee33ff1a544e6fc0f75fc7.jpg) + +![](images/d2cb7764447fc81177be88b895ad62665db2b6bede144940c51a5e7c042c7ad3.jpg) + +Table 8. Rigid or Soft Classifier. + +
Prob.DepthTrackLasHeRVisEvent
F-scoreRePrPrSrPrSr
100%58.857.758.268.455.176.460.2
80%61.562.061.869.155.777.560.9
33%61.060.660.869.055.477.060.5
+ +Table 9. Ablation on MeME. + +
MethodDepthTrackLasHeRVisEvent
Ours61.569.177.6
w/ DeepSeek-MoE [7]60.561.476.4
w/ LLaMA-MoE [67]60.060.875.9
w/o Modal Prompt61.067.475.2
w/o Gating Function60.568.977.0
w/o Projection60.268.776.5
+ +signment, akin to extending current unified models by treating all modalities equally without differentiation. + +Tab. 8 presents the performance results for the two extreme cases: rigid separation and random assignment. Our final model outperforms both, achieving the best performance with a classification probability of, amazingly, $80.00\%$ under the 0.001 loss proportion. We observe that rigid classification significantly hinders performance, underscoring the need for trackers that effectively leverage multimodal training - a promising yet underexplored area. + +MeME variants: In further support of our claim in Tab. 1, we conduct ablation studies on our MeME module by replacing it with counterparts [7, 67] or removing key components. As shown in Tab. 9, our proposed design outperforms the variants, confirming the importance of each component in achieving superior performance. + +Visualization and Analysis: To better understand the routing mechanism, we provide in Fig. 5 the top-k choice visualization for different challenging scenes, all with RGB-Event only as input for inference. In our case, we set $k = 2$ : - In Scene 1, a low-light scenario, thermal sensing contributes to lighting awareness. Visualizations show that alongside the primary event expert, several thermal experts are activated, enhancing feature modeling and adapting to low-light-induced appearance changes. + +- Scene 2 focuses on car tracking, where activated foreground events are mixed with background events. To better model geometry and facilitate foreground-background separation, our network employs depth experts. Thermal experts are also triggered here, likely due to overexposure. + +- Scene 3 involves tracking with a heavily blurred image; our model primarily relies on event experts, aligning well with human intuition, as event cameras excel in such a task. + +Additionally, we observe that the routing choice can be influenced by network depth, as deeper layers may handle different features compared to shallower layers. + +# 6. Conclusion + +In this paper, we introduce a novel approach to multimodal knowledge sharing for visual tracking, addressing the core challenge of data sparsity in multimodal sensing and the significant factors leading to appearance discrepancies. Our key innovation is a weak classifier-based mechanism that facilitates cross-modal knowledge transfer when samples from different modalities share similar characteristics. By integrating this mechanism within a mixture-of-experts framework, we enable effective cross-modal knowledge sharing. Our method successfully bridges the gap between multimodal training and single-modality deployment, offering a practical solution to the challenge of leveraging multimodal data effectively. The experimental results demonstrated the effectiveness of our approach, achieving a significant overall improvement over SOTA methods across all RGB-X tracking scenarios. + +Acknowledgements. The authors sincerely thank the reviewers and all members of the program committee for their tremendous efforts and incisive feedback. This research was supported in part by the Alexander von Humboldt Foundation and in part by the Ministry of Education and Science of Bulgaria (support for INSAIT, under the Bulgarian National Roadmap for Research Infrastructure). C. Ma was supported in part by NSFC (62376156, 62322113). + +# References + +[1] Hassan Akbari, Dan Kondratyuk, Yin Cui, Rachel Hornung, Huisheng Wang, and Hartwig Adam. Alternating gradient descent and mixture-of-experts for integrated multimodal perception. NeurIPS, 36, 2024. 3 +[2] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In ECCVW. Springer, 2016. 2 +[3] Bing Cao, Junliang Guo, Pengfei Zhu, and Qinghua Hu. Bidirectional adapter for multimodal tracking. In AAAI, 2024. 2 +[4] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 1, 2, 7 +[5] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 1 +[6] Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, Rongrong Ji, Zhenjun Tang, and Xianxian Li. Siamban: Target-aware tracking with siamese box adaptive network. TPAMI, 45(4):5158-5173, 2022. 1, 2 +[7] Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y Wu, et al. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models. arXiv preprint arXiv:2401.06066, 2024. 2, 3, 4, 7, 8 +[8] Kenan Dai, Yunhua Zhang, Dong Wang, Jianhua Li, Huchuan Lu, and Xiaoyun Yang. High-performance long-term tracking with meta-updater. In CVPR, 2020. 7 +[9] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In CVPR, 2019. 2 +[10] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 7 +[11] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 1, 3 +[12] Zhihong Fu, Zehua Fu, Qingjie Liu, Wenrui Cai, and Yun-hong Wang. Sparsett: Visual tracking with sparse transformers. In IJCAI, 2022. 1 +[13] Yuan Gao, Chenglong Li, Yabin Zhu, Jin Tang, Tao He, and Futian Wang. Deep adaptive fusion network for high performance RGBT tracking. In ICCVW, 2019. 6 +[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 3 +[15] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. Siamcar: Siamese fully convolutional classification and regression for visual tracking. In CVPR, 2020. 2 +[16] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. Fast-dynamic-vision: Detection and tracking dynamic objects with event and depth sensing. In IROS, 2021. 2 + +[17] João F Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. TPAMI, 37(3):583-596, 2014. 2 +[18] Lingyi Hong, Shilin Yan, Renrui Zhang, Wanyun Li, Xinyu Zhou, Pinxue Guo, Kaixun Jiang, Yiting Chen, Jinglun Li, Zhaoyu Chen, et al. Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In CVPR, 2024. 2, 3, 6, 7 +[19] Xiaojun Hou, Jiazheng Xing, Yijie Qian, Yaowei Guo, Shuo Xin, Junhao Chen, Kai Tang, Mengmeng Wang, Zhengkai Jiang, Liang Liu, et al. Sdstrack: Self-distillation symmetric adapter learning for multi-modal visual object tracking. In CVPR, 2024. 2, 3, 6, 7 +[20] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 43(5):1562-1577, 2019. 1, 3 +[21] Tianrui Hui, Zizheng Xun, Fengguang Peng, Junshi Huang, Xiaoming Wei, Xiaolin Wei, Jiao Dai, Jizhong Han, and Si Liu. Bridging search region interaction with template for rgb-t tracking. In CVPR, 2023. 2 +[22] Yash Jain, Harkirat Behl, Zsolt Kira, and Vibhav Vineet. Damex: Dataset-aware mixture-of-experts for visual understanding of mixture-of-datasets. NeurIPS, 36, 2024. 3 +[23] Hamed Kiani Galoogahi, Ashton Fagg, and Simon Lucey. Learning background-aware correlation filters for visual tracking. In ICCV, 2017. 2 +[24] Matej Kristan, Aleš Leonardis, Jiří Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, Ondrej Drbohlav, et al. The eighth visual object tracking VOT2020 challenge results. In ECCVW, 2020. 6 +[25] Matej Kristan, Jiří Matas, Aleš Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Cehovin, Alan Lukežić, et al. The ninth visual object tracking vot2021 challenge results. In ICCVW, 2021. 6 +[26] Matej Kristan, Aleš Leonardis, Jiří Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, et al. The tenth visual object tracking vot2022 challenge results. In ECCVW, 2023. 2, 6 +[27] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 6 +[28] Chenglong Li, Hui Cheng, Shiyi Hu, Xiaobai Liu, Jin Tang, and Liang Lin. Learning collaborative sparse representation for grayscale-thermal tracking. TIP, 25(12):5743-5756, 2016. 1, 2 +[29] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. Rgb-t object tracking: Benchmark and baseline. PR, 96:106977, 2019. 2, 6 +[30] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for rgbt tracking. TIP, 31:392-404, 2021. 1, 2, 6 +[31] Wei Li, Zengfu Hou, Jun Zhou, and Ran Tao. Siambag: Band attention grouping-based siamese object tracking network for hyperspectral videos. TGRS, 61:1-12, 2023. 2 + +[32] Valerii Likhosherstov, Anurag Arnab, Krzysztof Marcin Choromanski, Mario Lucic, Yi Tay, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. TMLR, 2022. 3 +[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. NeurIPS, 35:16743-16754, 2022. 1 +[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 3 +[35] Hongjiao Liu, Jiayue He, Jinpeng Wang, Nan Su, Chunhui Zhao, Yiming Yan, Shou Feng, Ze Liu, Jianfei Liu, and Zi-long Zhao. Multi-band hyperspectral object tracking: Leveraging spectral information prompts and spectral scale-aware representation. In WHISPERS, 2023. 2 +[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2 +[37] Christoph Mayer, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Learning target candidate association to keep track of what not to track. In ICCV, 2021. 6 +[38] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 1, 3 +[39] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. NeurIPS, 35:9564-9576, 2022. 2, 3 +[40] Hyeonseob Nam and Bohyung Han. Learning multi-domain convolutional neural networks for visual tracking. In CVPR, 2016. 7 +[41] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. NeurlPS, 34:8583-8595, 2021. 3 +[42] Paul Voigtaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: Visual tracking by redetection. In CVPR, 2020. 7 +[43] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for rgb-t tracking. In CVPR, 2020. 6 +[44] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. NeurIPS, 36, 2024. 5 +[45] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. TCYB, pages 1-14, 2023. 2, 6, 7 +[46] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In CVPR, 2024. 1 + +[47] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1 +[48] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In CVPR, 2023. 5 +[49] Zongwei Wu, Jilai Zheng, Xiangxuan Ren, Florin-Alexandru Vasluianu, Chao Ma, Danda Pani Paudel, Luc Van Gool, and Radu Timofte. Single-model and any-modality for video object tracking. In CVPR, 2024. 2, 3, 6, 7 +[50] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for rgbt tracking. In AAAI, 2022. 6 +[51] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In CVPR, 2022. 6 +[52] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 7 +[53] Song Yan, Jinyu Yang, Jani Käpylä, Feng Zheng, Ales Leonardis, and Joni-Kristian Kämäräinen. Depthtrack: Unveiling the power of rgbd tracking. In ICCV, 2021. 1, 2, 6, 7 +[54] Jinyu Yang, Zhe Li, Song Yan, Feng Zheng, Ales Leonardis, Joni-Kristian Kämäräinen, and Ling Shao. Rgbd object tracking: An in-depth review. arXiv preprint arXiv:2203.14134, 2022. 2 +[55] Jinyu Yang, Zhe Li, Feng Zheng, Ales Leonardis, and Jingkuan Song. Prompting for multi-modal tracking. In ACMM, 2022. 3, 6, 7 +[56] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In ECCV, 2022. 5, 6, 7 +[57] Hui Zhang, Lei Zhang, Li Zhuo, and Jing Zhang. Object tracking in RGB-T videos using modal-aware attention network and competitive learning. Sensors, page 393, 2020. 6 +[58] Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Baocai Yin, and Bo Dong. Object tracking by jointly exploiting frame and event domain. In ICCV, 2021. 1, 2 +[59] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, 2022. 2 +[60] Lichao Zhang, Martin Danelljan, Abel Gonzalez-Garcia, Joost Van De Weijer, and Fahad Shahbaz Khan. Multi-modal fusion for end-to-end rgb-t tracking. In ICCVW, 2019. 6 +[61] Pengyu Zhang, Jie Zhao, Chunjuan Bo, Dong Wang, Huchuan Lu, and Xiaoyun Yang. Jointly modeling motion and appearance cues for robust RGB-T tracking. TIP, pages 3335-3347, 2021. 6 +[62] Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, and Xiang Ruan. Visible-thermal uav tracking: A large-scale benchmark and new baseline. In CVPR, 2022. 2 +[63] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 3 + +[64] Haojie Zhao, Dong Wang, and Hutchuan Lu. Representation learning for visual object tracking by masked appearance transfer. In CVPR, 2023. 2 +[65] Jinjian Zhao, Xiaohan Zhang, and Pengyu Zhang. A unified approach for tracking uavs in infrared. In ICCV, 2021. 2 +[66] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, 2023. 1, 2, 3, 6, 7 +[67] Tong Zhu, Xiaoye Qu, Daize Dong, Jiacheng Ruan, Jingqi Tong, Conghui He, and Yu Cheng. Llama-moe: Building mixture-of-experts from llama with continual pre-training. arXiv preprint arXiv:2406.16554, 2024. 2, 4, 5, 8 +[68] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. Rgbd1k: A large-scale dataset and benchmark for rgb-d object tracking. In AAAI, 2023. 1, 2, 6 +[69] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. Rgbd1k: A large-scale dataset and benchmark for rgb-d object tracking. In AAAI, 2023. 6 +[70] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for rgb-event transformer-trackers. In ICCV, 2023. 2 \ No newline at end of file diff --git a/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/images.zip b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..627111e5076698c9c24772b971e19452a03b4e08 --- /dev/null +++ b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:405fd4ff67dd5ce88ad4e56db4fb008136ae28addb8c428f0b0849ad9c8e6c56 +size 537461 diff --git a/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/layout.json b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3554de49004f8583789f6d146f8d69155f8496e2 --- /dev/null +++ b/xtrackmultimodaltrainingboostsrgbxvideoobjecttrackers/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96c40b49122194afd89f978a74fe8e61749185121f74a975ae19a2b2a87e05fc +size 454074 diff --git a/yolocountdifferentiableobjectcountingfortexttoimagegeneration/cfefcd61-b5c9-4ed3-a2d8-a13711519bfe_content_list.json b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/cfefcd61-b5c9-4ed3-a2d8-a13711519bfe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..30f6d728d80bc7b0a9f78d67191b3677d3cdebcc --- /dev/null +++ b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/cfefcd61-b5c9-4ed3-a2d8-a13711519bfe_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf5147af61130b0086ddf9f0b3a3f382a5cf3a56d32271220005bf7bdcbc339a +size 86400 diff --git a/yolocountdifferentiableobjectcountingfortexttoimagegeneration/cfefcd61-b5c9-4ed3-a2d8-a13711519bfe_model.json b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/cfefcd61-b5c9-4ed3-a2d8-a13711519bfe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8693704b87880825c5aef8c230c1640007b48ed0 --- /dev/null +++ b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/cfefcd61-b5c9-4ed3-a2d8-a13711519bfe_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e642ebbe678713e4bd60357872b759b5a9c1613970ff5909f47b24f5bcb33d2 +size 107347 diff --git a/yolocountdifferentiableobjectcountingfortexttoimagegeneration/cfefcd61-b5c9-4ed3-a2d8-a13711519bfe_origin.pdf b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/cfefcd61-b5c9-4ed3-a2d8-a13711519bfe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..74a02b1ce782c4e30ae5bf1cf381118600aef37c --- /dev/null +++ b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/cfefcd61-b5c9-4ed3-a2d8-a13711519bfe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d445f141ef7087e8d8491c11c046317cee888c2cc559abd9c56e34bdd39731cd +size 6135546 diff --git a/yolocountdifferentiableobjectcountingfortexttoimagegeneration/full.md b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d6af7cacd214687ee46c28964ed3c5547a4b1272 --- /dev/null +++ b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/full.md @@ -0,0 +1,405 @@ +# YOLO-Count: Differentiable Object Counting for Text-to-Image Generation + +Guanning Zeng $^{1*}$ Xiang Zhang $^{2}$ Zirui Wang $^{3}$ Haiyang Xu $^{2}$ +Zeyuan Chen $^{2}$ Bingnan Li $^{2}$ Zhuowen Tu $^{2}$ $^{1}$ Tsinghua University $^{2}$ UC San Diego $^{3}$ UC Berkeley + +SDXL + +SDXL with YOLO-Count Guidance + +![](images/fdc2e6758b3e56237bcfa9e52bf494b70b18cc097cc94efd9bf91cbddc931e10.jpg) + +![](images/7d6a68d766aba25e9c3dc786b692638d499f344610bbe8bcbb376f3e1f44c46a.jpg) +Amidst the morning fog, three pine cones rest upon an ancient stump carpeted in moss, encircled by blueberry bushes. + +![](images/630e7ef68381e8a1f452f18b98fa2c06b4d47a48a730e531fc88b72f57b82e12.jpg) + +![](images/3b6f73b9cc9c8f01b6834b695c73d7856deac6f6de9665d7166b89787f6c4b9a.jpg) +Five happy dogs on sandy shores, beneath a glowing sunset sky. + +![](images/7ded53ba01579347be51fd2bd8d0587248cc696bfbc0fe2d91a009a52deb6018.jpg) + +![](images/a71c16744e23a5b75749be4f1c4b00eae56c515020f5ebce3d24520383ff7e42.jpg) +A collection of twenty colorful marbles scattered across a pristine white surface. +Figure 1. Demonstrations of YOLO-Count's object quantity controllability for text-to-image generation. Incorporating YOLO-Count as a differentiable guidance module over a strong baseline (SDXL [41]) substantially improves alignment between text-specified object quantities and the generated images. + +![](images/0ff106b3428eb5930539902902caec1153b093afa84e1a44340b2d90519218fa.jpg) + +![](images/0c07d8d3f7d83f39626b209710860034256e04a193a06438c2117a1e57d9d7ca.jpg) +On a pristine white platter lie five sumptuous donuts, each enrobed in a sleek coat of white icing, their forms boasting a tempting golden-brown crumb. + +![](images/b528f11ad4ab17eabc3064e42a25ca3a31424c217eb841cc3903595516affba6.jpg) + +![](images/3c42748394e8212b2b55ffa2cd4a4c96d378bd04e6b3af836a90736dab09d3d5.jpg) + +# Abstract + +We propose YOLO-Count, a differentiable open-vocabulary object counting model that tackles both general counting challenges and enables precise quantity control for text-to-image (T2I) generation. A core contribution is the 'cardinality' map, a novel regression target that accounts for variations in object size and spatial distribution. Leveraging representation alignment and a hybrid strong-weak supervision scheme, YOLO-Count bridges the gap between open-vocabulary counting and T2I generation control. Its fully differentiable architecture facilitates gradient-based optimization, enabling accurate object count estimation and fine-grained guidance for generative models. Extensive experiments demonstrate that YOLO-Count achieves state-of-the-art counting accuracy while providing robust and effective quantity control for T2I systems. + +# 1. Introduction + +Text-to-image (T2I) generative models have achieved remarkable success in producing high-fidelity images from natural language descriptions. However, ensuring precise alignment with textual specifications, particularly regarding object quantity, remains a significant challenge. While prior research has improved adherence to object layout, attributes, and style through conditional training and guidance mechanisms, accurately controlling the number of objects synthesized within an image remains difficult. Unlike localized attributes, object quantity constitutes a global constraint, requiring models to establish numerical correspondence between language tokens and compositional objects. Consequently, conventional conditional training approaches such as ControlNet [52] are ill-suited for explicit quantity control. Moreover, the stochastic nature of the denoising process in T2I models introduces ambiguity in object differentiation, further complicating count consistency. Re + +cent conditional guidance methods, such as BoxDiff [46] and Ranni [12], address aspects of spatial layout, object attributes, and semantic panel conditioning. However, these methods lack a direct and principled mechanism for precise quantity control, leaving a critical gap in bridging linguistic numeracy and visual synthesis. + +In this work, we propose YOLO-Count, an open-vocabulary object counting model built on the YOLO architecture. YOLO-Count is a fully differentiable, regression-based model that demonstrates high accuracy, computational efficiency, and open-vocabulary capabilities. A key contribution is the introduction of the cardinality map, a novel representation that encodes object quantity while preserving awareness of object size and spatial location. Unlike traditional density maps, which apply Gaussian kernels at object centers, the cardinality map distributes quantity scores across object instances, improving accuracy and robustness to scale variation. Furthermore, YOLO-Count leverages representation alignment and a hybrid strong-weak supervision strategy, enabling the use of large-scale instance segmentation datasets without reliance on computationally expensive pre-trained visual encoders. + +Beyond generic object counting, we are motivated to apply YOLO-Count for precise control of object quantities in text-to-image (T2I) generation. This is achieved by employing YOLO-Count as a differentiable guidance module [5], where gradient signals from the counting model steer the generative process toward numerical consistency. While prior research has predominantly focused on guidance algorithms for attributes and layout, explicit quantity control remains underexplored. We argue that an ideal object counting model for T2I applications should possess four key properties: (1) full differentiability w.r.t. the input image; (2) open-vocabulary capability for diverse object categories; (3) cross-scale generalization to varying object sizes; (4) computational efficiency for practical deployment. + +Constructing such a model introduces several challenges. First, state-of-the-art counting approaches [2, 35] are often detection-based, producing outputs that preclude gradient propagation. Second, existing counting datasets such as FSC147 [37] or CARPK [18] are limited in scale and category diversity, hindering open-vocabulary generalization. Third, while large-scale vision encoders (e.g., CLIP [36] or GroundingDINO [32, 38]) can alleviate data limitations, they impose significant computational overhead. + +To address these issues, we integrate YOLO-Count with textual inversion [13, 50] to achieve precise quantity control in T2I generation. Extensive experiments demonstrate that YOLO-Count achieves state-of-the-art accuracy on counting benchmarks, outperforms density-based and detection-based counting models, and substantially improves object quantity controllability in T2I generation. + +Our contributions are summarized as follows: + +- We introduce the cardinality map, a novel regression target that improves object counting accuracy compared to density maps. +- We develop YOLO-Count, an efficient, open-vocabulary, and fully differentiable counting model that achieves state-of-the-art performance and enhances quantity control for T2I generation. +- We propose hybrid strong-weak supervision with representation alignment, enabling effective training using large-scale segmentation datasets without reliance on heavy visual encoders. + +# 2. Related Works + +# 2.1. Object Counting Models and Datasets + +Object counting models can be broadly classified according to their category scope into fixed-category counting models [14, 40, 44] and open-vocabulary counting models [2, 11, 35]. For controlling object quantities in generative tasks, open-vocabulary counting is essential, as it supports arbitrary object categories without retraining. Based on the type of supervision or guidance, counting models can be further divided into text-guided models [47, 58], visual-exemplar-guided models [20, 37], multimodal-guided models [2], and reference-less models [17, 31, 45]. For T2I integration, a purely text-guided counting model is preferable to ensure compatibility with prompt-driven generation. + +From a methodological perspective, counting models are typically divided into detection-based and regression-based approaches. Detection-based models [2, 18, 34] rely on explicit object detection, filtering instances via thresholds and enumerating discrete counts, which inherently produce nondifferentiable integer outputs. In contrast, regression-based models [3, 4, 27] predict continuous-valued maps such as density maps [10, 33] that represent pixel-wise contributions to the final count. This direct differentiability makes regression-based models particularly suitable for gradient-based control in generative pipelines. + +Finally, training datasets for object counting are categorized into fixed-category datasets [18, 21, 43] and open-vocabulary datasets [1, 37]. Open-vocabulary datasets provide images containing diverse object categories and instance counts, but are expensive to collect and annotate [37]. For example, the widely used FSC147 dataset includes only 3,659 training images, which limits scale and diversity. To address this, recent works [2, 22] incorporate large-scale pre-trained visual backbones (e.g., CLIP [36] and GroundingDINO [32]) and fine-tune them on smaller counting datasets to enhance open-vocabulary generalization. + +# 2.2. Controllable Text-to-Image Generation + +Controllable text-to-image (T2I) generation methods can be broadly categorized into two paradigms: training-based + +![](images/4f1e308275c9582a6557b88d0d1892a2c7957cf4091622488e73c44e95446879.jpg) +Figure 2. YOLO-Count Model Overview. YOLO-Count comprises a YOLO backbone, CLIP text encoder, vision-language path aggregation network (VLPAN), cardinality regression head and classification head. Built upon the YOLO-World [9] architecture, the cardinality head predicts a cardinality map. The final object quantity is obtained by summing over the cardinality map. + +methods [19, 52, 54] and guidance-based methods [5, 49, 53]. Training-based approaches, such as ControlNet [52], IP-Adapter [48], and GLIGEN [28], inject conditional inputs directly into the generative model through additional network branches or adapters. While effective, these methods rely on large-scale training datasets annotated with the corresponding conditions. In contrast, guidance-based approaches, including BoxDiff [46], Attend-and-Excite [8], and Separate-and-Enhance [6], control generation by manipulating the diffusion process at inference time, eliminating the need for retraining. Many of these methods exploit the interpretability of cross-attention mechanisms [30] to steer image synthesis. However, cross-attention is primarily effective for distinguishing object categories rather than differentiating multiple instances of the same category. As a result, existing controllable T2I techniques excel at localized attribute binding [15, 55] and layout control [56, 57], but struggle with enforcing global constraints such as precise object quantity. + +# 2.3. Object Quantity Control for T2I Models + +Research on explicit object quantity control in text-to-image (T2I) models remains limited. [25] pioneers the use of universal diffusion guidance for quantity control, representing the first attempt to directly address this challenge. [7] introduces an attention-based representation for counting objects, but their approach is constrained to controlling small quantities (ranging from 1 to 10). More recently, prompting approaches [42, 50] have been proposed to incorporate numerical cues into the text embedding space, enabling limited quantity control without modifying the underlying diffusion model. However, these methods still struggle with accurate control over larger counts. + +# 3. Methods + +# 3.1. Model Overview + +Our proposed YOLO-Count builds upon the YOLO-World architecture [9] and consists of three primary components: + +(1) a vision backbone, (2) a vision-language path aggregation network (VLPAN), and (3) prediction heads. Fig. 2 illustrates the overall pipeline and highlights our key architectural modifications. + +Vision Backbone. The vision backbone in YOLO-Count follows the design of YOLOv81 [23] and YOLO-World-L [9]. It comprises five stages of convolutional modules (ConvModules) and cross-stage partial layers (CSPLayers). Given an input image $I \in \mathbb{R}^{640 \times 640 \times 3}$ , the backbone extracts multiscale visual features at three resolutions: + +$$ +f ^ {0} = \left[ f _ {8 0 \times 8 0}, f _ {4 0 \times 4 0}, f _ {2 0 \times 2 0} \right] = \operatorname {V i s u a l B a c k b o n e} (I) +$$ + +Vision-Language Path Aggregation Network (VLPAN). The VLPAN is designed to fuse visual features with textual semantics and aggregate information across scales. Inheriting from YOLO-World, it employs both top-down and bottom-up pathways, but with key enhancements: (1) T-CSPLayers: standard CSPLayers are replaced by TCSPLayers, which integrate sigmoid attention blocks to modulate visual features based on precomputed CLIP text embeddings [36]. (2) Extended Top-Down Fusion: to better preserve fine-grained spatial details, an additional top-down pathway is introduced following the initial bidirectional aggregation, maximizing high-resolution feature utilization, which is critical for accurate counting regression. The enhanced VLPAN is formulated as: + +$$ +[ f ^ {1}, f ^ {2} ] = \mathrm {V L P A N} (f ^ {0}, f _ {\mathrm {T}}) +$$ + +where $f_{\mathrm{T}}$ denotes the CLIP text embedding of the category, $f^{1}$ and $f^{2}$ represent multimodal features for classification and counting regression, respectively. + +Prediction Heads. Following the VLPAN, several ConvModules are applied to text-aware visual features to aggregate multi-scale signals into a unified $80 \times 80$ resolution. The prediction stage then produces two parallel outputs: (1) a cardinality regression head, which predicts a + +dense cardinality map for differentiable counting, and (2) a classification head, trained with contrastive supervision to ensure robust open-vocabulary capability. These two outputs jointly enable YOLO-Count to provide accurate, differentiable count estimates while maintaining strong category generalization, as shown on the right side of Fig. 2. + +$$ +\left\{ \begin{array}{l} o _ {\mathrm {c l s}} = \mathrm {C l a s s i f i c a t i o n H e a d} (f ^ {1}) \\ \hat {y} _ {\mathrm {c n t}} = \mathrm {C o u n t i n g H e a d} (f ^ {2}) \end{array} \right. +$$ + +# 3.2. Cardinality Map Regression + +We introduce the concept of the cardinality map, a novel regression target designed to address the inherent ambiguities of density-based counting. + +Density-based counting models formulate the counting loss as: + +$$ +\mathcal {L} _ {\mathrm {c n t}} = \left| \hat {y} _ {\mathrm {c n t}} - y _ {\mathrm {d e n}} \right| \tag {1} +$$ + +where $y_{\mathrm{den}}$ is the density map. For an image containing $Q$ objects, $y_{\mathrm{den}}$ is constructed by placing $Q$ Gaussian kernels centered at object locations, with their sum equal to $Q$ . While regressing to $y_{\mathrm{den}}$ enables quantity prediction, this representation suffers from two key ambiguities. First, the Gaussian kernel's center can be placed anywhere within an object's extent. Second, the kernel radius is arbitrarily chosen and lacks physical meaning. These issues degrade model accuracy for objects with diverse sizes and shapes, as density maps fail to provide a consistent, unambiguous representation. + +To overcome these limitations, we replace $y_{\mathrm{den}}$ with a cardinality map $y_{\mathrm{car}}$ , defined using object masks. Given a binary mask $M_{i}$ for the $i$ -th object instance, with area $N_{i} = |M_{i}|$ , we uniformly distribute a value of 1 across all pixels within the object and sum contributions across all $K$ objects: + +$$ +y _ {\text {p i x e l c a r d i n a l i t y}} = \sum_ {i = 1} ^ {K} \frac {1}{N _ {i}} M _ {i} \tag {2} +$$ + +We then downsample this pixel-level cardinality map to a grid-based representation by summing within each grid cell: + +$$ +y _ {\operatorname {c a r}} (u, v) = \sum_ {(i, j) \in \Omega_ {u, v}} y _ {\text {p i x e l c a r d i n a l i t y}} (i, j) \tag {3} +$$ + +where $\Omega_{u,v}$ is the set of pixel coordinates within grid cell $(u,v)$ . By construction, the total sum of the cardinality map equals the true object count: + +$$ +\sum_ {u, v} y _ {\mathrm {c a r}} (u, v) = Q \tag {4} +$$ + +Unlike density maps, which concentrate mass at object centers and often ignore large portions of extended objects, + +the cardinality map uniformly covers the entire spatial extent of each object. This yields a unique, unambiguous representation that is robust to variations in object size and shape, making it better suited for differentiable regression-based counting. + +# 3.3. Representation Alignment + +Here we describe our method using the single-category counting scenario for simplicity, where the model is designed to count instances of a specific category specified by the user. To achieve this, we adapt the contrastive learning framework into a binary classification task, where each pixel is classified as either belonging to the target category or not. This additional branch aligns the visual and textual representations during training, ensuring the model effectively localizes instances of the specified category. Specifically, the classification loss is formulated as: + +$$ +\mathcal {L} _ {\mathrm {c l s}} = \operatorname {B C E L o s s} \left(\hat {y} _ {\mathrm {c l s}}, y _ {\mathrm {c l s}}\right) \tag {5} +$$ + +where $y_{\mathrm{cls}}(i,j) \in \{0,1\}$ is the binary ground truth label indicating whether pixel $(i,j)$ belongs to the target category, and $\hat{y}_{\mathrm{cls}}(i,j) \in [0,1]$ is the predicted probability obtained by projecting the visual feature $o_{\mathrm{cls}}$ and the text embedding $f_{\mathrm{T}}$ into a shared multimodal space and applying a sigmoid activation to their inner product, similar to SigLIP [51]. + +# 3.4. Hybrid Strong-Weak Training + +Training an object-counting model typically necessitates a specialized counting dataset, where each image contains multiple instances of a single category. Regression-based counting models rely on corresponding density maps, which are costly to produce and task-specific. To overcome this data limitation, we propose a hybrid strong-weak training technique that enables cardinality regression using both instance segmentation datasets and counting datasets. This approach comprises two stages: strong-supervision pretraining and weak-supervision finetuning. + +![](images/96a6d6e4fbda1c020b913e1dac7484c59237afe70f2b24d145ab9775735bb262.jpg) +Figure 3. Illustration of hybrid strong-weak supervision. Instance segmentation masks provide dense, per-grid category labels, while counting annotations offer sparse category labels limited to annotated points. By combining these two forms of supervision, we enable effective training of the object counting model using both precise (strong) and sparse (weak) annotations. + +# 3.4.1. Strong-Supervision Pretraining + +We first pretrain the model on instance segmentation datasets where precise per-instance masks are available. We + +construct cardinality maps $y_{\mathrm{car}}$ (as defined in Eq. (3)) and binary classification masks $y_{\mathrm{cls}} \in \{0,1\}^{H \times W}$ . The pretraining objectives are as follows: + +$$ +\mathcal {L} _ {\mathrm {c n t}} ^ {\mathrm {s t r o n g}} = | \hat {y} _ {\mathrm {c n t}} - y _ {\mathrm {c a r}} | \tag {6} +$$ + +$$ +\mathcal {L} _ {\text {t o t a l}} ^ {\text {s t r o n g}} = \alpha_ {1} \mathcal {L} _ {\text {c n t}} ^ {\text {s t r o n g}} + \beta_ {1} \mathcal {L} _ {\text {c l s}} ^ {\text {s t r o n g}} \tag {7} +$$ + +where $\alpha_{1}$ and $\beta_{1}$ are weighting coefficients. $\mathcal{L}_{\mathrm{cls}}^{\mathrm{strong}}$ is the same $\mathcal{L}_{\mathrm{cls}}$ defined in Eq. (5). This stage provides precise pixel-level supervision for both cardinality regression and category-specific classification, establishing a strong model initialization. + +# 3.4.2. Weak-Supervision Finetuning + +To better adapt YOLO-Count to dense counting scenarios, we perform weak-supervision finetuning on counting datasets that provide sparse point-level annotations. Each image is annotated with instance points $\mathcal{P} = \{(x_i,y_i)\}_{i = 1}^K$ , where $K$ is both the total count of objects in the image and the number of annotated points. Weak supervision comprises the following two components: + +(1) Sparse Classification Labels: Positive labels are derived from annotated points, forming $M_{\mathrm{pos}} \in \{0,1\}^{H \times W}$ where $M_{\mathrm{pos}}(i,j) = 1$ if $(i,j) \in \mathcal{P}$ . Negative labels $M_{\mathrm{neg}}$ are sampled from background regions. +(2) Total Count Consistency: The predicted total count must match the ground truth count $K$ . + +The weak supervision losses are then formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {c l s}} ^ {\text {w e a k}} = - \frac {1}{| \Omega |} \sum_ {p \in \Omega} [ M _ {\text {p o s}} (p) \log \hat {y} _ {\mathrm {c l s}} (p) + \tag {8} +$$ + +$$ +\left. M _ {\mathrm {n e g}} (p) \log \left(1 - \hat {y} _ {\mathrm {c l s}} (p)\right) \right] +$$ + +$$ +\mathcal {L} _ {\mathrm {c n t}} ^ {\text {w e a k}} = \left| \left(\sum_ {p} \hat {y} _ {\mathrm {c n t}} (p)\right) - K \right| \tag {9} +$$ + +$$ +\mathcal {L} _ {\text {t o t a l}} ^ {\text {w e a k}} = \alpha_ {2} \mathcal {L} _ {\text {c n t}} ^ {\text {w e a k}} + \beta_ {2} \mathcal {L} _ {\text {c l s}} ^ {\text {w e a k}}, \tag {10} +$$ + +where $\Omega = \{(i,j) \mid M_{\mathrm{pos}}(i,j) = 1 \vee M_{\mathrm{neg}}(i,j) = 1\}$ denotes the annotated pixel locations (see Fig. 3). + +This hybrid training scheme effectively leverages large-scale instance segmentation datasets for pretraining and adapts to limited counting datasets during finetuning, enabling robust and data-efficient model training. + +# 3.5. Counting-Controlled Generation + +Following [50], we adopt a textual inversion-based approach for counting-controlled generation. The text-to-image (T2I) model first synthesizes an initial image from a text prompt, which may not accurately reflect the desired object quantity. We then feed the generated image and its target category into the YOLO-Count model to estimate the + +predicted quantity and compute a guidance loss based on the deviation from the required count $Q_{\mathrm{req}}$ : + +$$ +\mathcal {L} _ {\text {g u i d e}} = \left| \left(\sum_ {p} \hat {y} _ {\text {c n t}} (p)\right) - Q _ {\text {r e q}} \right| \tag {11} +$$ + +We iteratively update a learnable counting token inserted into the text sequence via the gradient of $\mathcal{L}_{\mathrm{guide}}$ . This process continues until convergence, effectively steering the T2I model toward images with the desired object quantity. + +# 4. Experiments + +# 4.1. Setups + +# 4.1.1. Training Datasets + +FSC147 [37] is an object counting dataset comprising 6,135 images across 89 training classes, 29 validation classes, and 29 test classes, with no overlap between the splits. Following [2], we applied corrections to several images containing erroneous annotations to ensure label consistency. + +LVIS v1.0 [16] is a large-scale, long-tailed dataset containing 1,203 object categories, sharing images with MSCOCO [29]. Originally designed for instance segmentation with a large vocabulary, we employ its validation set to evaluate counting accuracy across diverse and open-vocabulary categories. + +# 4.1.2. Evaluation Metrics + +Following prior work [18, 21, 37], we employ mean absolute error (MAE) and root mean square error (RMSE) to evaluate counting performance. MAE measures the average absolute deviation between predicted and ground-truth counts, while RMSE penalizes larger errors more heavily, capturing both accuracy and precision. + +# 4.1.3. Benchmarks of Object Counting + +In addition to FSC147 and LVIS, we introduce two new benchmarks to evaluate open-vocabulary object counting accuracy. + +OpenImages V7 [26] and Objects365 [39] are large-scale open-vocabulary object detection datasets, containing 9 million images across 600 categories and 2 million images across 365 categories, respectively. Based on these datasets, we construct two novel counting benchmarks: OpenImg7-New and Obj365-New, specifically designed to evaluate generalization to unseen categories. + +To ensure category novelty, we first compute CLIP [36] embeddings for all text labels in both datasets. We then filter out categories whose maximum cosine similarity with any LVIS label embedding exceeds 0.7. This procedure yields 47 novel categories from OpenImages V7 and 51 from Objects365. Finally, we retain only images containing these filtered categories, resulting in benchmarks with 14,699 images (OpenImg7-New) and 22,724 images (Obj365-New). + +
ModelFSC-Test [37] +(Seen Categories)FSC-Val [37] +(Seen Categories)LVIS [16] +(Seen for Ours)OImg7-New [26] +(Unseen Categories)Obj365-New [39] +(Unseen Categories)
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
Non-Differentiable
CountGD [2]12.9898.3512.1447.514.8412.456.0929.923.5310.61
DAVE [35]14.90103.4215.4852.575.2911.485.3114.244.8913.22
Differentiable
CLIP-Count [22]17.78106.6218.7961.1810.8122.6114.0130.1615.4830.28
VLCounter [24]17.05106.1618.0665.138.9423.0415.3234.4118.0833.02
CounTX [1]15.88106.2917.1065.6113.7040.9417.9448.7318.7644.80
YOLO-Count (Ours)14.8096.1415.4358.361.656.083.7211.963.289.15
+ +Table 1. Comparison of counting accuracy with existing text-guided object counting models. + +These benchmarks provide a challenging evaluation protocol for assessing the open-vocabulary generalization ability of object counting models beyond the categories seen during FSC147 and LVIS training. + +# 4.1.4. Benchmarks of Controllable Generation + +To evaluate the precision of quantity-controlled text-to-image (T2I) generation, we construct two benchmarks: LargeGen and LargeGen-New. LargeGen is derived from the FSC147 dataset by selecting the 10 text categories with the highest average object counts per image, providing a benchmark for evaluating generation performance on seen categories. LargeGen-New is built from Obj365-New and OpenImg7-New, thereby assessing generation accuracy on novel categories. These benchmarks enable a systematic evaluation of quantity-guided generation, measuring both seen-category and novel-category performance when using object counting models to steer T2I generation. + +# 4.1.5. Training YOLO-Count Model + +For YOLO-Count, we first perform a 250-epoch strong-supervision pretraining on the LVIS dataset. The vision backbone is initialized with YOLOv8l weights, while all other modules are randomly initialized. We optimize a composite loss with coefficients $\alpha_{1} = 1.0$ for cardinality regression and $\beta_{1} = 0.1$ for category classification. The CLIP text encoder remains frozen during training, while all other parameters are updated with different learning rates: $5 \times 10^{-9}$ for the backbone (to preserve pre-trained visual representations) and $1 \times 10^{-5}$ for newly initialized modules. This learning rate strategy stabilizes training and facilitates adaptation of the YOLO backbone to the counting task. + +Subsequently, we perform weak-supervised finetuning on FSC147, leveraging the dataset-provided positive point annotations and manually annotating negative labels. For negative labels, we dot approximately 10 background points per image in the FSC147 training set. This annotation process is highly efficient, facilitated through a labeling interface, and requires only about 5 seconds per image on average. During finetuning, we retain a proportion $\gamma$ of LVIS + +data within each training batch to provide strong supervision and preserve the open-vocabulary capability ofYOLO-Count. The model is trained for up to 500 epochs with early stopping based on the mean absolute error (MAE) on the FSC147 validation set. + +# 4.1.6. Counting Control with Token Optimization + +We integrate YOLO-Count with SDXL-Turbo [41] using a differentiable token optimization strategy for quantity-controlled generation. The pipeline employs single-step inference to balance generation quality and computational efficiency. For each generation task, we iteratively optimize a count token embedding for up to 150 steps using a learning rate of $5 \times 10^{-3}$ , with early stopping applied if the guidance loss (Eq. (11)) plateaus for 20 consecutive steps. This process allows direct gradient-based feedback from YOLO-Count to refine the token embedding, effectively steering the T2I model toward producing images with the desired object quantity. + +# 4.2. Results of Object Counting + +We evaluate counting accuracy on the FSC147 validation and test sets, a widely used benchmark for object-counting models. For FSC147, we set $\alpha_{1} = 0$ , $\beta_{1} = 1$ , $\alpha_{2} = 1$ , $\beta_{2} = 0.1$ , and $\gamma = 0.05$ during finetuning. Following prior work such as CountGD [2] and CLIP-Count [22], we adopt an automatic clipping and aggregation strategy to handle + +![](images/9cc252bc5990062963171c41cf7343f6106b7af171c7b8140a48be98543697ad.jpg) +Figure 4. Counting results of YOLO-Count. + +images containing large object counts. As shown in Tab. 1, YOLO-Count achieves state-of-the-art performance among regression-based counting models. Our MAE and RMSE scores on FSC147 are competitive and closely approach those of CountGD, the current leading non-differentiable counting model. Visual examples in Fig. 4 further demonstrate that YOLO-Count produces accurate object counts. + +We further evaluate the open-vocabulary capability of YOLO-Count by testing its counting accuracy on unseen categories. For LVIS, OpenImg7-New, and Objects365-New, we set $\alpha_{1} = \beta_{1} = \alpha_{2} = \beta_{2} = 1$ and $\gamma = 0.5$ for finetuning. The results, shown in Tab. 1, demonstrate that exposure to more diverse training data significantly improves YOLO-Count's ability to generalize to novel categories. Notably, despite having the fewest parameters among the compared counting models, YOLO-Count achieves superior accuracy in open-vocabulary settings. + +In summary, Figs. 1 and 4 to 6 illustrate results obtained using the FSC147 checkpoint settings of Tab. 1. Figs. 7 and 8 present results based on the checkpoint settings for LVIS, OpenImg7, and Objects365 of Tab. 1. + +# 4.3. Results of T2I Quantity Control + +We evaluate the effectiveness ofYOLO-Count in guiding T2I models for accurate quantity control. We compare against two series of baselines: (1) extension of previous guidance-based methods [7, 25] to large quantities, and (2) the token optimization framework of [50] guided by alternative counting models, including CLIP-Count and CountGD. + +For each category in LargeGen and LargeGen-New, we generate 10 images per target quantity, with counts set to 25, 50, 75, and 100. Notably, since CountGD is non-differentiable, we cannot apply the gradient-based loss in Eq. (11). Instead, we employ a surrogate cross-entropy objective that encourages a high probability for the target count $Q_{\mathrm{req}}$ while suppressing probabilities for other counts. + +We manually evaluate the generated images by counting the objects and comparing them against the target $Q_{\mathrm{req}}$ . As shown in Fig. 5, YOLO-Count substantially improves generation accuracy compared to all baselines. + +![](images/4e2744d2c1a3b540b8232f518b21ae3f49de03d16b25ead96619386348a10f91.jpg) +Figure 5. Quantitative results for T2I quantity control. Our method significantly reduces the discrepancy between the requested and generated object quantities compared to previous approaches, both on seen and unseen categories. + +![](images/810d12290089f75486c4ec375465ec33114731c214da146618175f5afd03445f.jpg) + +Furthermore, Fig. 6 illustrates qualitative differences + +when generating scenes with large object counts. Under CountGD's surrogate loss $\mathcal{L}_{\mathrm{guide}}^{\mathrm{det}}$ , the T2I model fails to adjust object counts and suffers from degraded visual quality. Similarly, density-based counting guidance results in large deviations from $Q_{\mathrm{req}}$ , likely due to the domain gap between real counting datasets and synthetic T2I images. In contrast, YOLO-Count provides precise, differentiable guidance signals, enabling accurate quantity control while generalizing effectively to novel categories. Additional qualitative examples are provided in the Supplementary Material. + +![](images/5ab64ee04fbee635a916eebd5a01a0f412ee7594830657fa5d0177672657909a.jpg) +Figure 6. Comparison of counting-controlled generation by different models. + +# 4.4. Ablation Study + +We conduct an ablation study to evaluate the contribution of each key component inYOLO-Count. Specifically, we examine the impact of our cardinality map regression, representation alignment, hybrid strong-weak training strategy, and architectural modifications on the performance of regression-based object counting models. + +
ModelFSC-Test [37]FSC-Val [37]
MAE↓RMSE↓MAE↓RMSE↓
Baseline
YOLO-Count14.8096.1415.4358.36
...without
Pretraining18.42111.4519.5088.64
Weak-Supervision43.91150.4043.86124.33
Cardinality Map16.71107.2417.8776.42
Alignment17.01110.4117.5785.54
Additional VLPAN16.54106.3216.8984.40
+ +Table 2. Ablation Studies. w/o Pretraining: no LVIS data; w/o Weak-Supervision: no FSC147 data; w/o Cardinality Map: training on density maps directly; w/o Alignment: no classification branch; w/o Additional VLPAN: both YOLO-Count heads receive same VLPAN feature output. + +We begin by evaluating the contribution of each stage in + +the training pipeline. Specifically, we analyze counting accuracy under two conditions: (1) without strong-supervised pretraining and (2) without weak-supervised finetuning. In the first scenario (w/o pretraining), YOLO-Count is trained directly on FSC147 using density map supervision. In the second scenario (w/o finetuning), YOLO-Count is trained solely on LVIS with strong labels, without subsequent finetuning. As shown in Tab. 2, the absence of finetuning leads to significantly higher MAE and RMSE, highlighting the domain gap between LVIS and FSC147 in terms of object count distributions. Importantly, combining strong-supervised pretraining on LVIS with weak-supervised finetuning on FSC147 achieves substantial improvements over either stage alone, validating our hybrid training strategy. + +Next, we examine the impact of removing the cardinality regression component. In this variant (w/o cardinality), we pretrain on LVIS as in the default setting, but replace cardinality map regression with density map regression during FSC147 training. As observed in Tab. 2, the inclusion of cardinality regression yields consistently lower MAE and RMSE compared to density-based regression. Furthermore, visual results in Fig. 7 reveal that density-based regression often suffers from overlapping kernels within a single instance and exhibits a bias toward over-counting in images containing larger-sized objects. This demonstrates that cardinality regression mitigates ambiguities inherent in density maps and improves robustness across varying object scales. + +![](images/5a870a71f79c3d88ce534b5c70fcb79974368bc02cc94daeac34305b9fbd3c83.jpg) +Figure 7. Comparison of density vs. cardinality regression. + +# 4.5. Analysis on Size Bias of Counting Models + +We design an experiment to investigate the size bias exhibited by density-based regression methods when handling objects of varying scales. Specifically, we select images from the FSC147 validation and test sets containing no more than 30 objects. Each image is progressively down + +![](images/0b5dc5cac837edc2a754f2ec25a0855d202ec06227508fa706661b67f4b4b94c.jpg) + +![](images/b81e9090bc441434f3b162789ae68caf2069aadb6822bca8cfd3840e239e9f19.jpg) +Figure 8. Comparison of counting bias for detection, density-based regression and cardinality-based regression models in different image downscaling scales and number of objects. + +scaled with a scaling ratio ranging from 1.0 (original size) to 4.0, and then padded to restore the original image dimensions. While this process reduces the object sizes, the ground-truth counts remain unchanged. + +The processed images are then fed into several object-counting models: three density-based models (VL-Counter [24], CLIP-Count [22], and CountTX [1]), one detection-based model (CountGD [31]), and our proposed YOLO-Count model. For each model, we record the predicted counts and analyze their differences relative to predictions on the original (unscaled) images, providing a measure of counting stability under size variation. + +As illustrated in Fig. 8, density-based regression models consistently over-count as object sizes increase, with this bias becoming more pronounced for images containing larger objects. In contrast,YOLO-Count exhibits stability similar to detection-based models, maintaining accurate counts across different object scales. This result underscores the advantage of training regression-based models on cardinality maps, which eliminates kernel-based ambiguities and improves robustness to variations in object size. + +# 5. Conclusion + +In this paper, we introduced YOLO-Count, a novel open-vocabulary, regression-based object counting model that substantially improves object quantity control in text-to-image generation. By integrating cardinality regression, hybrid strong-weak supervision, and representation alignment, YOLO-Count achieves state-of-the-art counting accuracy, computational efficiency, and robust open-vocabulary generalization. Extensive experiments and ablation studies validate its effectiveness in overcoming the limitations of prior approaches, particularly in handling large object quantities and novel categories. Beyond advancing object counting, YOLO-Count provides a practical and differentiable mechanism for enhancing the controllability of text-to-image models, thereby enabling more precise and reliable multimodal generation and perception. + +Acknowledgment This work is supported by NSF Award IIS-2127544 and NSF Award IIS-2433768. We thank Yuheon Joh for insightful discussions and valuable feedback. + +# References + +[1] Niki Amini-Naieni, Kiana Amini-Naieni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. In BMVC, page 510, 2023. 2, 6, 8 +[2] Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting. In Advances in Neural Information Processing Systems, pages 48810-48837. Curran Associates, Inc., 2024. 2, 5, 6 +[3] Carlos Arteta, Victor Lempitsky, J Alison Noble, and Andrew Zisserman. Interactive object counting. In European Conference on Computer Vision, pages 504-518. Springer, 2014. 2 +[4] Carlos Arteta, Victor Lempitsky, and Andrew Zisserman. Counting in the wild. In European Conference on Computer Vision, pages 483-498. Springer, 2016. 2 +[5] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 843-852, 2023. 2, 3 +[6] Zhipeng Bao, Yijun Li, Krishna Kumar Singh, Yu-Xiong Wang, and Martial Hebert. Separate-and-enhance: Compositional finetuning for text-to-image diffusion models. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3 +[7] Lital Binyamin, Yoad Tewel, Hilit Segev, Eran Hirsch, Royi Rassin, and Gal Chechik. Make it count: Text-to-image generation with an accurate number of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13242-13251, 2025. 3, 7 +[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans. Graph., 42(4), 2023. 3 +[9] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16901-16911, 2024. 3 +[10] Siu-Yeung Cho, Tommy WS Chow, and Chi-Tat Leung. A neural-based crowd estimation by hybrid global learning algorithm. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 29(4):535-541, 1999. 2 +[11] Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16985-16995, 2024. 2 +[12] Yutong Feng, Biao Gong, Di Chen, Yujun Shen, Yu Liu, and Jingren Zhou. Ranni: Taming text-to-image diffusion for accurate instruction following. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4744-4753, 2024. 2 +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2023. 2 +[14] Mingyue Guo, Li Yuan, Zhaoyi Yan, Binghui Chen, Yaowei Wang, and Qixiang Ye. Regressor-segmenter mutual prompt learning for crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28380-28389, 2024. 2 +[15] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9380-9389, 2024. 3 +[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6 +[17] Michael Hobley and Victor Prisacariu. Learning to count anything: Reference-less class-agnostic counting with weak supervision. arXiv preprint arXiv:2205.10203, 2022. 2 +[18] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H. Hsu. Drone-based object counting by spatially regularized regional proposal network. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2, 5 +[19] Minghui Hu, Jianbin Zheng, Daqing Liu, Chuanxia Zheng, Chaoyue Wang, Dacheng Tao, and Tat-Jen Cham. Cocktail: Mixing multi-modality control for text-conditional image generation. In Advances in Neural Information Processing Systems, pages 32424–32444. Curran Associates, Inc., 2023. 3 +[20] Zhizhong Huang, Mingliang Dai, Yi Zhang, Junping Zhang, and Hongming Shan. Point segment and count: A generalized framework for object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17067-17076, 2024. 2 +[21] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 5 +[22] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 2, 6, 8 +[23] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. *Ultralytics yolov8*, 2023. 3 +[24] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2714-2722, 2024. 6, 8 +[25] Wonjun Kang, Kevin Galim, Hyung Il Koo, and Nam Ik Cho. Counting guidance for high fidelity text-to-image synthesis. + +In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 899-908, 2025. 3, 7 +[26] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 5, 6 +[27] Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2010. 2 +[28] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 3 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740-755. Springer, 2014. 5 +[30] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7817-7826, 2024. 3 +[31] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. arXiv preprint arXiv:2208.13721, 2022. 2, 8 +[32] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 2 +[33] Aparecido Nilceu Marana, Sergio A Velastín, LF Costa, and RA Lotufo. Estimation of crowd density using image processing. In IEE Colloquium on Image Processing for Security Applications (Digest No: 1997/074), pages 11-1. IET, 1997. 2 +[34] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In European Conference on Computer Vision, pages 348-365. Springer, 2022. 2 +[35] Jer Pelhan, Alan Lukeži?, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23293–23302, 2024. 2, 6 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 5 + +[37] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3394-3403, 2021. 2, 5, 6, 7 +[38] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the" edge" of open-set object detection. arXiv preprint arXiv:2405.10300, 2024. 2 +[39] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5, 6 +[40] Zenglin Shi, Pascal Mettes, and Cees GM Snoek. Focus for free in density-based counting. International Journal of Computer Vision, 132(7):2600-2617, 2024. 2 +[41] Stability AI. Sdsl-turbo: Text-to-image generative model, 2024. 1, 6 +[42] Wenfang Sun, Yingjun Du, Gaowen Liu, and Cees GM Snoek. Quota: Quantifying objects with text-to-image models for any domain. arXiv preprint arXiv:2411.19534, 2024. 3 +[43] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpcrowd: A large-scale benchmark for crowd counting and localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2 +[44] Yi Wang, Qiongyang Hu, and Lap-Pui Chau. Weakly-supervised crowd counting with token attention and fusion: A simple and effective baseline. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13456-13460. IEEE, 2024. 2 +[45] Yuzhe Wu, Yipeng Xu, Tianyu Xu, Jialu Zhang, Jianfeng Ren, and Xudong Jiang. Gca-sun: A gated context-aware swim-unet for exemplar-free counting. arXiv preprint arXiv:2409.12249, 2024. 2 +[46] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7452-7461, 2023. 2, 3 +[47] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15548-15557, 2023. 2 +[48] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3 +[49] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23174-23184, 2023. 3 +[50] Oz Zafar, Lior Wolf, and Idan Schwartz. Iterative object count optimization for text-to-image diffusion models. arXiv preprint arXiv:2408.11721, 2024. 2, 3, 5, 7 + +[51] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 4 +[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 1, 3 +[53] Peiang Zhao, Han Li, Ruiyang Jin, and S Kevin Zhou. Loco: Locally constrained training-free layout-to-image synthesis. arXiv preprint arXiv:2311.12342, 2023. 3 +[54] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. In Advances in Neural Information Processing Systems, pages 11127-11150. Curran Associates, Inc., 2023. 3 +[55] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems, pages 73969-73982. Curran Associates, Inc., 2023. 3 +[56] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22490-22499, 2023. 3 +[57] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6818-6828, 2024. 3 +[58] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Yu Guo, Zheng Wang, Xian Zhong, and Shengfeng He. Zero-shot object counting with good exemplars. In European Conference on Computer Vision, pages 368-385. Springer, 2024. 2 \ No newline at end of file diff --git a/yolocountdifferentiableobjectcountingfortexttoimagegeneration/images.zip b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..982803305e7b562bbd683e928d27c8e37da01806 --- /dev/null +++ b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad1006aa5759fa1cc96457b36aaf6c8c4d5c7e5cc0374154aecc37c6c466eaf6 +size 590415 diff --git a/yolocountdifferentiableobjectcountingfortexttoimagegeneration/layout.json b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a72e9282b31d19a0abc2b4cf22a91dc8cd8b1c7 --- /dev/null +++ b/yolocountdifferentiableobjectcountingfortexttoimagegeneration/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfc9c3f441f6b23629b53af36e7e72ef2acb4860c18fc3908e65a312651dca07 +size 417127 diff --git a/yoloerealtimeseeinganything/3dbff6c8-6a64-4ee0-abf9-c4b47ba4d2a6_content_list.json b/yoloerealtimeseeinganything/3dbff6c8-6a64-4ee0-abf9-c4b47ba4d2a6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3484df4abdd06e6792e2dee02a15377df23e9d2e --- /dev/null +++ b/yoloerealtimeseeinganything/3dbff6c8-6a64-4ee0-abf9-c4b47ba4d2a6_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fe94ccb5605dc9a12a9f699b256a9fbb062c98bcda84e2b01f8643d23f26fa0 +size 93381 diff --git a/yoloerealtimeseeinganything/3dbff6c8-6a64-4ee0-abf9-c4b47ba4d2a6_model.json b/yoloerealtimeseeinganything/3dbff6c8-6a64-4ee0-abf9-c4b47ba4d2a6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..87719034e66b37315df3fa18b220fef45d3a7178 --- /dev/null +++ b/yoloerealtimeseeinganything/3dbff6c8-6a64-4ee0-abf9-c4b47ba4d2a6_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa092fc419480206ea56eaac9436843a3760994665861baa6321204982aafa6d +size 117916 diff --git a/yoloerealtimeseeinganything/3dbff6c8-6a64-4ee0-abf9-c4b47ba4d2a6_origin.pdf b/yoloerealtimeseeinganything/3dbff6c8-6a64-4ee0-abf9-c4b47ba4d2a6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7a746a0c5857bb7eebb374f2a020531fdeba8fac --- /dev/null +++ b/yoloerealtimeseeinganything/3dbff6c8-6a64-4ee0-abf9-c4b47ba4d2a6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1185331eb00270e09eea3bd2cb40202903717b6493d1d8ee2454dc4a8a83116b +size 1573187 diff --git a/yoloerealtimeseeinganything/full.md b/yoloerealtimeseeinganything/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e2b6b2540bfabcc613c7d5f6606cd3232f4d9bec --- /dev/null +++ b/yoloerealtimeseeinganything/full.md @@ -0,0 +1,319 @@ +![](images/64b511208d704ee47fe116d4b0e9d32573663d80d55aadd6fde9d0ad69a3aa6a.jpg) + +# YOLOE: Real-Time Seeing Anything + +Ao Wang $^{1*}$ Lihao Liu $^{1*}$ Hui Chen $^{2}$ Zijia Lin $^{1}$ Jungong Han $^{3}$ Guiguang Ding $^{1,\dagger}$ $^{1}$ School of Software, Tsinghua University $^{2}$ BNRist, Tsinghua University $^{3}$ Department of Automation, Tsinghua University + +# Abstract + +Object detection and segmentation are widely employed in computer vision applications, yet conventional models likeYOLO series, while efficient and accurate, are limited by predefined categories, hindering adaptability in open scenarios. Recent open-set methods leverage text prompts, visual cues, or prompt-free paradigm to overcome this, but often compromise between performance and efficiency due to high computational demands or deployment complexity. In this work, we introduce YOLOE, which integrates detection and segmentation across diverse open prompt mechanisms within a single highly efficient model, achieving real-time seeing anything. For text prompts, we propose Re-parameterizable Region-Text Alignment (RepRTA) strategy. It refines pretrained textual embeddings via a re-parameterizable lightweight auxiliary network and enhances visual-textual alignment with zero inference and transferring overhead. For visual prompts, we present Semantic-Automated Visual Prompt Encoder (SAVPE). It employs decoupled semantic and activation branches to bring improved visual embedding and accuracy with minimal complexity. For prompt-free scenario, we introduce Lazy Region-Prompt Contrast (LRPC) strategy. It utilizes a built-in large vocabulary and specialized embedding to identify all objects, avoiding costly language model dependency. Extensive experiments show YOLOE's exceptional zero-shot performance and transferability with high inference efficiency and low training cost. Notably, on LVIS, with $3 \times$ less training cost and $1.4 \times$ inference speedup, YOLOE-v8-S surpasses YOLO-Worldv2-S by 3.5 AP. When transferring to COCO, YOLOE-v8-L achieves $0.6\mathrm{AP}^b$ and $0.4\mathrm{AP}^m$ gains over closed-set YOLOv8-L with nearly $4 \times$ less training time. Code and models are available at here. + +# 1. Introduction + +Object detection and segmentation are foundational tasks in computer vision [15, 48], with widespread applications + +![](images/d103abd0f140af40c7af5e0af41f2a647d3b587b1a70838a91e6320794a469e4.jpg) +Figure 1. Comparison of performance, training cost, and inference efficiency between YOLOE (Ours) and advanced YOLO-Worldv2 in terms of open text prompts. LVIS AP is evaluated on minival set and FPS w/ TensorRT and w/ CoreML is measured on T4 GPU and iPhone 12, respectively. The results highlight our superiority. + +![](images/75954bcdb1c887b79f913d7224040ce285671d13b615247bbc93581e30d05816.jpg) + +![](images/e5b7c39c61cebdbd9056738d77ddfb906c346d923272f9821075cb7ee5f98fcc.jpg) + +spanning autonomous driving [2], medical analyses [55], and robotics [8], etc. Traditional approaches likeYOLO series [1, 3, 21, 47], have leveraged convolutional neural networks to achieve real-time remarkable performance. However, their dependence on predefined object categories constrains flexibility in practical open scenarios. Such scenarios increasingly demand models capable of detecting and segmenting arbitrary objects guided by diverse prompt mechanisms, such as texts, visual cues, or without prompt. + +Given this, recent efforts have shifted towards enabling models to generalize for open prompts [5, 20, 49, 80]. They target single prompt type, e.g., GLIP [32], or multiple prompt types in a unified way, e.g., DINO-X [49]. Specifically, with region-level vision-language pretraining [32, 37, 65], text prompts are usually processed by text encoder to serve as contrastive objectives for region features [20, 49], achieving recognition for arbitrary categories, e.g., YOLO-World [5]. For visual prompts, they are often encoded as class embeddings tied to specified regions for identifying similar objects, by the interaction with image features or language-aligned visual encoder [5, 19, 30, 49], e.g., T-Rex2 [20]. In prompt-free scenario, existing methods typically integrate language models, finding all objects and generating the corresponding category names conditioned on region features sequentially [49, 62], e.g., GenerateU [33]. + +Despite notable advancements, a single model that sup- + +ports diverse open prompts for arbitrary objects with high efficiency and accuracy is still lacking. For example, DINO-X [49] features a unified architecture, which, however, incurs resource-intensive training and inference overhead. Additionally, individual designs for different prompts in separate works exhibit suboptimal trade-offs between performance and efficiency, making it difficult to directly combine them into one model. For example, text-prompted approaches often incur substantial computational overhead when incorporating large vocabularies, due to complexity of cross-modality fusion [5, 32, 37, 49]. Visual-prompted methods usually compromise deployability on edge devices owing to the transformer-heavy design or reliance on additional visual encoder [20, 30, 67]. Prompt-free ways, meanwhile, depend on large language models, introducing considerable memory and latency costs [33, 49]. + +In light of these, in this paper, we introduce YOLOE(ye), a highly efficient, unified, and open object detection and segmentation model, like human eye, under different prompt mechanisms, like texts, visual inputs, and prompt-free paradigm. We begin with YOLO models with widely proven efficacy. For text prompts, we propose a Reparameterizable Region-Text Alignment (RepRTA) strategy, which employs a lightweight auxiliary network to improve pretrained textual embeddings for better visualsemantic alignment. During training, pre-cached textual embeddings require only the auxiliary network to process text prompts, incurring low additional cost compared with closed-set training. At inference and transferring, auxiliary network is seamlessly re-parameterized into the classification head, yielding an architecture identical to YOLOs with zero overhead. For visual prompts, we design a Semantic-Activated Visual Prompt Encoder (SAVPE). By formalizing regions of interest as masks, SAVPE fuses them with multi-scale features from PAN to produce grouped prompt-aware weights in low dimension in an activation branch and extract prompt-agnostic semantic features in a semantic branch. Prompt embeddings are derived through aggregation of them, resulting in favorable performance with minimal complexity. For prompt-free scenario, we introduce Lazy Region-Prompt Contrast (LRPC) strategy. Without relying on costly language models, LRPC leverages a specialized prompt embedding to find all objects and a built-in large vocabulary for category retrieval. By matching only anchor points with identified objects against the vocabulary, LRPC ensures high performance with low overhead. + +Thanks to them, YOLOE excels in detection and segmentation across diverse open prompt mechanisms within one model, enjoying high inference efficiency and low training cost. Notably, as shown in Fig. 1, under $3 \times$ less training cost, YOLOE-v8-S significantly outperforms YOLOwWorldv2-S [5] by 3.5 AP on LVIS [14], with $1.4 \times$ and $1.3 \times$ inference speedups on T4 and iPhone 12, respectively. + +In visual-prompted and prompt-free settings, YOLOE-v8-L outperforms T-Rex2 by $3.3\mathrm{AP}_r$ and GenerateU by 0.4 AP with $2\times$ less training data and $6.3\times$ fewer parameters, respectively. For transferring to COCO [34], YOLOE-v8-M / L outperforms YOLOv8-M / L by $0.4 / 0.6\mathrm{AP}^b$ and $0.4 / 0.4\mathrm{AP}^m$ with nearly $4\times$ less training time. We hope that YOLOE can establish a strong baseline and inspire further advancements in real-time open prompt-driven vision tasks. + +# 2. Related Work + +Traditional detection and segmentation. Traditional approaches for object detection and segmentation primarily operate under closed-set paradigms. Early two-stage frameworks [4, 12, 15, 48], exemplified by Faster R-CNN [48], introduce region proposal networks (RPNs) followed by region-of-interest (ROI) classification and regression. Meanwhile, single-stage detectors [10, 35, 38, 56, 72] prioritizes speed through grid-based predictions within a single network. The YOLO series [1, 21, 27, 47, 59, 60] plays a significant role in this paradigm and are widely used in real world. Moreover, DETR [28] and its variants [28, 69, 77] mark a major shift by removing heuristic-driven components with transformer-based architectures. To achieve finer-grained results, existing instance segmentation methods predict pixel-level masks rather than bounding box coordinates [15]. For this, YOLACT [3] facilitates real-time instance segmentation through integration of prototype masks and mask coefficients. Based on DINO [69], MaskDINO [29] utilizes query embeddings and a high-resolution pixel embedding map to produce binary masks. + +Text-prompted detection and segmentation. Recent advancements in open-vocabulary object detection [13, 25, 61, 65, 68, 74-76] have focused on detecting novel categories by aligning visual features with textual embeddings. Specifically, GLIP [32] unifies object detection and phrase grounding through grounded pre-training on large-scale image-text pairs, demonstrating robust zero-shot performance. Grounding DINO [37] enhances this by integrating cross-modality fusion into DINO, improving alignment between text prompts and visual representations. YOLO-World [5] further shows the potential of pretraining small detectors with open recognition capabilities based on the YOLO architecture. YOLO-UniOW [36] builds upon YOLO-World by leveraging the adaptive decision-learning strategy. Similarly, several open-vocabulary instance segmentation models [11, 18, 26, 45, 63] learn rich visual-semantic knowledge from advanced foundation models to perform segmentation on novel object categories. For example, X-Decoder [79] and OpenSeeD [71] explore both the open-vocabulary detection and segmentation tasks. APE [54] introduces universal visual perception model that aligns and prompts all objects using various text prompts. + +Visual-prompted detection and segmentation. While + +![](images/04e9c385a0ac614f19ad65395e2aa63f5a8f9120577041ba42fec9ed1f0f31fe.jpg) +Figure 2. The overview of YOLOE, which supports detection and segmentation for diverse open prompt mechanisms. For text prompts, We design a re-parameterizable region-text alignment strategy to improve performance with zero inference and transferring overhead. For visual prompts, SAVPE is employed to encode visual cues with enhanced prompt embedding under minimal cost. For prompt-free setting, we introduce lazy region-prompt contrast strategy to provide category names for all identified objects efficiently by retrieval. + +text prompts offer a generic description, certain objects can be challenging to describe with language alone, such as those requiring specialized domain knowledge. In such cases, visual prompts can guide detection and segmentation more flexibly and specifically, complementing text prompts [19, 20, 23]. OV-DETR [67] and OWL-ViT [41] leverage CLIP encoders to process text and image prompts. DINOv [30] explores visual prompts as in-context examples for generic and referring vision tasks. T-Rex2 [20] integrates visual and text prompts by region-level contrastive alignment. For segmentation, SEEM [80] explores segmenting objects with various prompt types. SemanticSAM [31] excels in semantic comprehension and granularity detection, handling both panoptic and part segmentation. + +Prompt-free detection and segmentation. Existing approaches still depend on explicit prompts during inference for open-set detection and segmentation. To address this limitation, several works [33, 40, 49, 62, 66] explore integrating with generative language models to produce object descriptions for all found objects. For instance, GRiT [62] employs a text decoder for both dense captioning and object detection tasks. DetCLIPv3 [66] trains an object captioner on large-scale data, enabling model to generate rich label information. GenerateU [33] leverages the language model to generate object names in a free-form way. + +Closing remarks. To the best of our knowledge, aside from DINO-X [49], few efforts have achieved object detection and segmentation across various open prompt mechanisms within a single architecture. However, DINO-X entails extensive training cost and notable inference overhead, severely constraining the practicality for real-world edge deployments. In contrast, our YOLOE aims to deliver an efficient and unified model that enjoys real-time performance and efficiency with easy deployability. + +# 3. Methodology + +In this section, we detail designs of YOLOE. Building upon YOLOs (Sec. 3.1), YOLOE supports text prompts through RepRTA (Sec. 3.2), visual prompts via SAVPE (Sec. 3.3), and prompt-free scenario with LRPC (Sec. 3.4). + +# 3.1. Model architecture + +As shown in Fig. 2, YOLOE adopts the typical YOLOs' architecture [1, 21, 47], consisting of backbone, PAN, regression head, segmentation head, and object embedding head. The backbone and PAN extracts multi-scale features for the image. For each anchor point, the regression head predicts the bounding box for detection, and the segmentation head produces the prototype and mask coefficients for segmentation [3]. The object embedding head follows the structure of classification head in YOLOs, except that the output channel number of last $1 \times$ convolution layer is changed from the class number in closed-set scenario to the embedding dimension. Meanwhile, given text and visual prompts, we employ RepRTA and SAVPE to encode them as normalized prompt embeddings $\mathcal{P}$ , respectively. They serve as the classification weights and contrast with the anchor points' object embeddings $\mathcal{O}$ to obtain category labels. The process can be formalized as + +$$ +\operatorname {L a b e l} = \mathcal {O} \cdot \mathcal {P} ^ {T}: \mathbb {R} ^ {N \times D} \times \mathbb {R} ^ {D \times C} \rightarrow \mathbb {R} ^ {N \times C}, \tag {1} +$$ + +where $N$ denotes the number of anchor points, $C$ indicates the number of prompts, and $D$ means the feature dimension of embeddings, respectively. + +# 3.2. Re-parameterizable region-text alignment + +In open-set scenarios, the alignment between textual and object embeddings determines the accuracy of identified categories. Prior works usually introduce complex cross + +modality fusion to improve the visual-textual representation for better alignment [5, 37]. However, these ways incur notable computational overhead, especially with large number of texts. Given this, we present Re-parameterizable Region-Text Alignment (RepRTA) strategy, which improves pretrained textual embeddings during training through the reparameterizable lightweight auxiliary network. The alignment between textual and anchor points' object embeddings can be enhanced with zero inference and transferring cost. + +Specifically, with the text prompts of $T$ with length of $C$ , we first employ the CLIP text encoder [44, 57] to obtain pretrained textual embedding $P = \operatorname{TextEncoder}(T)$ . Before training, we cache all embeddings of texts in datasets in advance and the text encoder can be removed with no extra training cost. Meanwhile, as shown in Fig. 3.(a), we introduce a lightweight auxiliary network $f_{\theta}$ with only one feed forward block [53, 58], where $\theta$ indicates the trainable parameters and introduces low overhead compared with closed-set training. It derives the enhanced textual embedding $\mathcal{P} = f_{\theta}(P) \in \mathbb{R}^{C \times D}$ for contrasting with the anchor points' object embedding during training, leading to improved visual-semantic alignment. Let $K \in \mathbb{R}^{D \times D' \times 1 \times 1}$ be the kernel parameters of last convolution layer with input features $I \in \mathbb{R}^{D' \times H \times W}$ in the object embedding head, $\otimes$ be the convolution operator, and $R$ be the reshape function, we have + +$$ +\operatorname {L a b e l} = R _ {D \times H \times W \rightarrow H W \times D} (I \oplus K) \cdot \left(f _ {\theta} (P)\right) ^ {T}. \tag {2} +$$ + +Moreover, after training, the auxiliary network can be reparameterized with the object embedding head into the identical classification head of YOLOs. The new kernel parameters $K^{\prime}\in \mathbb{R}^{C\times D^{\prime}\times 1\times 1}$ for last convolution layer after re-parameterization can be derived by + +$$ +K ^ {\prime} = R _ {C \times D \rightarrow C \times D \times 1 \times 1} \left(f _ {\theta} (P)\right) \circledast K ^ {T}. \tag {3} +$$ + +The final predication can be obtained by Label $= I\circledast K^{\prime}$ which is identical to the original YOLO architecture, leading to zero overhead for deployment and transferring to downstream closed-set tasks. + +# 3.3. Semantic-activated visual prompt encoder + +Visual prompts are designed to indicate the object category of interest through visual cues, e.g., box and mask. To produce the visual prompt embedding, prior works often employ transformer-heavy design [20, 30], e.g., deformable attention [78], or additional CLIP vision encoder [44, 67]. These ways, however, introduce challenges in deployment and efficiency due to complex operators or high computational demands. Considering this, we introduce Semantic-Activated Visual Prompt Encoder (SAVPE) for efficiently processing visual cues. It features two decoupled lightweight branches: (1) Semantic branch outputs prompt-agnostic semantic features in $D$ channels without overhead of fusing visual cues, and (2) Activation branch + +![](images/5b2aade31cbcf617c7c4c2f7e7b8d3f8a8dff54e7d6e612d6cb8dc883ccf6247.jpg) +(a) + +![](images/538a29f76dc045c797d843e3c57efe8955653a4213c2fe3328a7b432aabbea6f.jpg) +Upsample +(b) + +![](images/0138b8e96ebfb06c2369e36b6388c2d09fe31aa87d647d5d80f8696942d982f8.jpg) +Figure 3. (a) The structure of lightweight auxiliary network in RepRTA, which consists of one SwiGLU FFN block [53]. (b) The structure of SAVPE, which consists of semantic branch to generate prompt-agnostic semantic features and activation branch to provide grouped prompt-aware weights. Visual prompt embedding can thus be efficiently derived by their aggregation. + +produces grouped prompt-aware weights by interacting visual cues with image features in much fewer channels under low costs. Their aggregation then leads to informative prompt embedding under minimal complexity. + +As shown in Fig. 3.(b), in the semantic branch, we adopt the similar structure as object embedding head. With multiscale features $\{P_3,P_4,P_5\}$ from PAN, we employ two $3\times 3$ convs for each scale, respectively. After upsampling, features are concatenated and projected to derive semantic features $S\in \mathbb{R}^{D\times H\times W}$ . In the activation branch, we formalize visual prompt as mask with 1 for indicated region and 0 for others. We downsample it and leverage $3\times 3$ conv to derive prompt feature $F_{V}\in \mathbb{R}^{A\times H\times W}$ . Besides, we obtain image features $F_{I}\in \mathbb{R}^{A\times H\times W}$ for fusion with it from $\{P_3,P_4,P_5\}$ by convs. $F_{V}$ and $F_{I}$ are then concatenated and utilized to output prompt-aware weights $\mathcal{W}\in \mathbb{R}^{A\times H\times W}$ , which is normalized using softmax within prompt-indicated region. Moreover, we divide the channels of $S$ into $A$ groups with $\frac{D}{A}$ channels in each. The channels in the $i$ -th group share the weight $\mathcal{W}_{i:i + 1}$ from the $i$ -th channel of $\mathcal{W}$ . With $A\ll D$ , we can process visual cues with image features in low dimension, bringing minimal cost. Furthermore, prompt embedding can be derived with aggregation of two branches by + +$$ +\mathcal {P} = \operatorname {C o n c a t} \left(G _ {1}, \dots , G _ {A}\right); G _ {i} = \mathcal {W} _ {i: i + 1} \cdot S _ {\frac {D}{A} * i: \frac {D}{A} * (i + 1)} ^ {T}. \tag {4} +$$ + +It can thus contrast with anchor points' object embeddings to identify objects with category of interest. + +# 3.4. Lazy region-prompt contrast + +In prompt-free scenario without explicit guidance, models are expected to identify all objects with names in the image. Prior works usually formulate such setting as a generative problem, where language model is employed to generate categories for dense found objects [33, 49, 62]. However, this introduces notable overhead, where language models, e.g., FlanT5-base [6] with 250M parameters in GenerateU [33] and OPT-125M [73] in DINO-X [49], are far from + +meeting high efficiency requirement. Given this, we reformulate such setting as a retrieval problem and present Lazy Region-Prompt Contrast (LRPC) strategy. It lazily retrieves category names from a built-in large vocabulary for anchor points with objects in the cost-effective way. Such paradigm enjoys zero dependency on language models, meanwhile with favorable efficiency and performance. + +Specifically, with pretrained YOLOE, we introduce a specialized prompt embedding and train it exclusively to find all objects, where objects are treated as one category. Meanwhile, we follow [16] to collect a large vocabulary which covers various categories and serve as the built-in data source for retrieval. One may directly leverage the large vocabulary as text prompts for YOLOE to identify all objects, which, however, incurs notable computational cost by contrasting abundant anchor points' object embeddings with numerous textual embeddings. Instead, we employ the specialized prompt embedding $\mathcal{P}_s$ to find the set $\mathcal{O}'$ of anchor points corresponding to objects by + +$$ +\mathcal {O} ^ {\prime} = \left\{o \in \mathcal {O} \mid o \cdot \mathcal {P} _ {s} ^ {T} > \delta \right\}, \tag {5} +$$ + +where $\mathcal{O}$ denotes all anchor points and $\delta$ is the threshold hyperparameter for filtering. Then, only anchor points in $\mathcal{O}'$ are lazily matched against the built-in vocabulary to retrieve category names, bypassing the cost for irrelevant anchor points. This further improves efficiency without performance drop, facilitating the real world application. + +# 3.5. Training objective + +During training, we follow [5] to obtain an online vocabulary for each mosaic sample with the texts involved in the images as positive labels. Following [21], we leverage task-aligned label assignment to match predictions with ground truths. The binary cross entropy loss is employed for classification, with IoU loss and distributed focal loss adopted for regression. For segmentation, we follow [3] to utilize binary cross-entropy loss for optimizing masks. + +# 4. Experiments + +# 4.1. Implementation details + +Model. For fair comparison with [5], we employ the same YOLOv8 architecture [21] for YOLOE. Besides, to verify its good generalizability on other YOLOs, we also experiment with YOLO11 architecture [21]. For both of them, we provide three model scales, i.e., small (S), medium (M), and large (L), to suit various application needs. Text prompts are encoded using the pretrained MobileCLIP-B(LT) [57] text encoder. We empirically use $A = 16$ in SAVPE, by default. + +Data. We follow [5] to utilize detection and grounding datasets, including Objects365 (V1) [52], GoldG [22] (includes GQA [17] and Flickr30k [43]), where images from COCO [34] are excluded. Beside, we leverage advanced + +SAM-2.1 [46] model to generate pseudo instance masks using ground truth bounding boxes from the detection and grounding datasets for segmentation data. These masks undergo filtering and simplification to eliminate noise [9]. For visual prompt data, we follow [20] to leverage ground truth bounding boxes for visual cues. In prompt-free tasks, we reuse the same datasets, but annotate all objects as a single category to learn a specialized prompt embedding. + +Training. Due to limited computational resource, unlike YOLO-World's training for 100 epochs, we first train YOLOE with text prompts for 30 epochs. Then, we only train the SAVPE for merely 2 epochs with visual prompts, which avoids additional significant training cost that comes with supporting visual prompts. At last, we train the specialized prompt embedding for only 1 epoch for prompt-free scenarios. During the text prompt training stage, we adopt the same settings as [5]. Notably, YOLOE-v8-S / M / L can be trained on 8 Nvidia RTX4090 GPUs in 12.0 / 17.0 / 22.5 hours, with $3 \times$ less cost compared with YOLOWorld. For visual prompt training, we freeze all other parts and adopt the same setting as in text prompt training. To enable prompt-free capability, we leverage the same data to train a specialized embedding. We can see that YOLOE not only enjoys low training costs but also show exceptional zero-shot performance. Besides, to verify YOLOE's good transferability on downstream tasks, we fine-tune our YOLOE on COCO [34] for closed-set detection and segmentation. We experiment with two distinct practical fin-tuning strategies: (1) Linear probing: Only the classification head is learnable and (2) Full tuning: All parameters are trainable. For Linear probing, we train all models for only 10 epochs. For Full tuning, we train small scale models including YOLOE-v8-S / 11-S for 160 epochs, and medium and large scale models including YOLOE-v8-M / L and YOLOE-11-M / L for 80 epochs, respectively. + +Metric. For text prompt evaluation, we utilize all category names from the benchmark as inputs, adhering to the standard protocol for open-vocabulary object detection tasks. For visual prompt evaluation, following [20], for each category, we randomly sample $N$ training images $(N = 16$ by default), extract visual embeddings using their ground truth bounding boxes, and compute the average prompt embedding. For prompt-free evaluation, we employ the same protocol as [33]. A pretrained text encoder [57] is employed to map open-ended predictions to semantically similar category names within the benchmark. In contrast to [33], we streamline the mapping process by selecting the most confident prediction, and eliminating the need for top-k selection and beam search. We use the tag list from [16] as the built-in large vocabulary with total 4585 category names, and empirically use $\delta = 0.001$ for LRPC, by default. For all three prompt types, following [5, 20, 33], evaluations are conducted on LVIS [14] in a zero-shot manner, which con + +Table 1. Zero-shot detection evaluation on LVIS. For fair comparisons, Fixed AP is reported on LVIS minival set in a zero-shot manner. The training time is for text prompts, based on 8 Nvidia V100 GPUs for [32, 65] and 8 RTX4090 GPUs for YOLO-World and YOLOE. The FPS is measured on Nvidia T4 GPU using TensorRT and on iPhone 12 using CoreML, respectively. Results are provided with text prompt (T) and visual prompt (V) type. For training data, OI, HT, and CH indicates OpenImages [24], HierText [39], and CrowdHuman [51], respectively. OG indicates Objects365 [52] and GoldG [22], and G-20M represents Grounding-20M [50]. + +
ModelPrompt TypeParamsTraining DataTraining TimeFPS T4 / iPhoneAPAPrAPcAPf
GLIP-T [32]T232MOG, Cap4M1337.6h- / -26.020.821.431.0
GLIPv2-T [70]T232MOG, Cap4M-- / -29.0---
GDINO-T [37]T172MOG, Cap4M-- / -27.418.123.332.7
DetCLIP-T [65]T155MOG250.0h- / -34.426.933.936.3
G-1.5 Edge [50]T-G-20M-- / -33.528.034.333.9
T-Rex2 [20]V-O365,OI,HT CH,SA-1B-- / -37.429.933.941.8
YWorldv2-S [5]T13MOG41.7h216.4 / 48.924.417.122.527.3
YWorldv2-M [5]T29MOG60.0h117.9 / 34.232.428.429.635.5
YWorldv2-L [5]T48MOG80.0h80.0 / 22.135.525.634.638.1
YOLOE-v8-ST/V12M / 13MOG12.0h305.8 / 64.327.9 / 26.222.3 / 21.327.8 / 27.729.0 / 25.7
YOLOE-v8-MT/V27M / 30MOG17.0h156.7 / 41.732.6 / 31.026.9 / 27.031.9 / 31.734.4 / 31.1
YOLOE-v8-LT/V45M / 50MOG22.5h102.5 / 27.235.9 / 34.233.2 / 33.234.8 / 34.637.3 / 34.1
YOLOE-11-ST/V10M / 12MOG13.0h301.2 / 73.327.5 / 26.321.4 / 22.526.8 / 27.129.3 / 26.4
YOLOE-11-MT/V21M / 27MOG18.5h168.3 / 39.233.0 / 31.426.9 / 27.132.5 / 31.934.5 / 31.7
YOLOE-11-LT/V26M / 32MOG23.5h130.5 / 35.135.2 / 33.729.1 / 28.135.0 / 34.636.5 / 33.8
+ +tains 1,203 categories. By default, Fixed AP [7] on LVIS minival subset is reported. For transferring to COCO, standard AP is evaluated, following [1, 21]. Besides, we measure the FPS for all models on Nvidia T4 GPU with TensorRT and mobile device iPhone 12 with CoreML. + +# 4.2. Text and visual prompt evaluation + +As shown in Tab. 1, for detection on LVIS, YOLOE exhibits favorable trade-offs between efficiency and zero-shot performance across different model scales. We also note that such results are achieved under much less training time, e.g., $3 \times$ faster than YOLO-Worldv2. Specifically, YOLOE-v8-S / M / L outperforms YOLOv8-Worldv2-S / M / L by $3.5 / 0.2 / 0.4$ AP, along with $1.4 \times / 1.3 \times / 1.3 \times$ and $1.3 \times / 1.2 \times / 1.2 \times$ inference speedups on T4 and iPhone 12, respectively. Besides, for rare category which is challenging, our YOLOE-v8-S and YOLOE-v8-L obtains significant improvements of $5.2\%$ and $7.6\%$ APr. Besides, compared with YOLO-Worldv2, while YOLOE-v8-M / L achieves lower APf, this performance gap primarily stems from YOLOE's integration of both detection and segmentation in one model. Such multi-task learning introduces a trade-off that adversely impact detection performance on frequent categories, as shown in Tab. 5. Besides, YOLOE with YOLO11 architecture also exhibits favorable performance and efficiency. For example, YOLOE-11-L achieves comparable AP with YOLO-Worldv2-L, but with notably $1.6 \times$ inference speedups on T4 and iPhone 12, highlighting the strong generalizability of our YOLOE. + +Moreover, the inclusion of visual prompts further amplifies YOLOE's versatility. Compared with T-Rex2, YOLOEV8-L yield the improvements of $3.3\mathrm{AP}_r$ and $0.9\mathrm{AP}_c$ , with $2\times$ less training data (3.1 M vs. Our: 1.4 M) and much lower training resource (16 Nvidia A100 GPUs vs. Our: 8 Nvidia RTX4090 GPUs). Besides, for visual prompts, while we only train SAVPE with other parts frozen for 2 epochs, we note that it can achieve comparable $\mathrm{AP}_r$ and $\mathrm{AP}_c$ with the text prompts for various model scales. This shows the efficacy of visual prompts in less frequent objects that text prompts often struggle to accurately describe, which is similar to the observation in [20]. + +Furthermore, for segmentation, we present the evaluation results on the LVIS val set with the standard $\mathrm{AP}^m$ reported in Tab. 2. It shows that YOLOE exhibits strong performance by leveraging both text prompts and visual prompts. Specifically, YOLOE-v8-M / L achieves 20.8 and $23.5\mathrm{AP}^m$ in the zero-shot manner, significantly outperforming YOLO-Worldv2-M / L that is fine-tuned on LVIS-Base dataset, by 3.0 and $3.7\mathrm{AP}^m$ , respectively. These results well show the superiority of YOLOE. + +# 4.3. Prompt-free evaluation + +As shown in Tab. 3, for prompt-free scenario, YOLOE also exhibits superior performance and efficiency. Specifically, YOLO-v8-L achieves 27.2 AP and $23.5\mathrm{AP}_r$ , outperforming GenerateU with Swin-T backbone by 0.4 AP and 3.5 $\mathrm{AP}_r$ , along with $6.3\times$ fewer parameters and $53\times$ inference speedups. It shows the effectiveness of YOLOE by + +Table 2. Segmentation evaluation on LVIS. We evaluate all models on LVIS val set with the standard $\mathrm{AP}^m$ reported. YOLOE supports both text (T) and visual cues (V) as inputs. $\dagger$ indicates that the pretrained models are fine-tuned on LVIS-Base data for segmentation head. In contrast, we evaluate YOLOE in a zero-shot manner without utilizing any images from LVIS during training. + +
ModelPromptAPmAPrmrAPcmcAPf
YWorld-M†T16.712.614.620.8
YWorld-L†T19.114.217.223.5
YWorldv2-M†T17.813.915.522.0
YWorldv2-L†T19.817.217.523.6
YOLOE-v8-ST/V17.7/16.815.5/13.516.3/16.720.3/18.2
YOLOE-v8-MT/V20.8/20.317.2/17.019.2/20.124.2/22.0
YOLOE-v8-LT/V23.5/22.021.9/16.521.6/22.126.4/24.3
YOLOE-11-ST/V17.6/17.116.1/14.415.6/16.820.5/18.6
YOLOE-11-MT/V21.1/21.017.2/18.319.6/20.624.4/22.6
YOLOE-11-LT/V22.6/22.519.3/20.520.9/21.726.0/24.1
+ +Table 3. Prompt-free evaluation on LVIS. Fixed AP is reported on the LVIS minival set, following the protocol in [33]. The FPS is measured on Nvidia T4 GPU with Pytorch [42]. + +
ModelBackboneParamsAPAPrAPcAPfFPS
GenerateU [33]Swin-T297M26.820.024.929.80.48
GenerateU [33]Swin-L467M27.922.325.231.40.40
YOLOE-v8-SYOLOv8-S13M21.019.121.321.095.8
YOLOE-v8-MYOLOv8-M29M24.722.224.525.345.9
YOLOE-v8-LYOLOv8-L47M27.223.527.028.025.3
YOLOE-11-SYOLO11-S11M20.618.420.221.393.0
YOLOE-11-MYOLO11-M24M25.521.625.526.142.5
YOLOE-11-LYOLO11-L29M26.322.725.827.534.9
+ +reformulating the open-ended problem as the retrieval task for a built-in large vocabulary and underscores its potential in generalizing across a wide range of categories without replying on explicit prompts. Such functionality also enhances YOLOE's practicality, enabling its application in a broader range of real-world scenarios. + +# 4.4. Downstream transferring + +As shown in Tab. 4, when transferring to COCO for downstream closed-set detection and segmentation, YOLOE exhibits favorable performance under limited training epochs in both two fine-tuning strategies. Specifically, for Linear probing, with less than $2\%$ of the training time, YOLOE-11-M / L can achieve over $80\%$ of the performance of YOLO11-M / L, respectively. This highlights the strong transferability of YOLOE. For Full tuning, YOLOE can further enhance the performance under limited training cost. For example, with nearly $4\times$ less training epochs, YOLOe-v8-M / L outperforms YOLOv8-M / L by $0.4\mathrm{AP}^m$ and 0.6 $\mathrm{AP}^b$ , respectively. Under $3\times$ less training time, YOLO-v8-S also obtains better performance compared with YOLOv8-S for both detection and segmentation. These results well + +Table 4. Downstream transfer on COCO. We fine-tune YOLOE on COCO and report the standard AP for both detection and segmentation. We experiment with two practical fine-tuning strategies, i.e., Linear probing and Full tuning. + +
ModelEpochs\( AP^b \)\( AP_{50}^b \)\( AP_{75}^b \)\( AP^m \)\( AP_{50}^m \)\( AP_{75}^m \)
Training from scratch
YOLOv8-S50044.761.448.736.658.038.6
YOLOv8-M30050.066.854.840.563.443.3
YOLOv8-L30052.469.357.242.366.044.9
YOLO11-S50046.663.350.637.859.740.0
YOLO11-M60051.568.555.741.565.043.9
YOLO11-L60053.370.158.242.866.845.5
Linear probing
YOLOE-v8-S1035.651.538.930.348.232.0
YOLOE-v8-M1042.259.246.335.555.637.7
YOLOE-v8-L1045.463.350.038.359.640.8
YOLOE-11-S1037.052.940.431.549.733.5
YOLOE-11-M1043.160.647.436.556.939.0
YOLOE-11-L1045.162.849.538.059.240.6
Full tuning
YOLOE-v8-S16045.061.649.136.758.339.1
YOLOE-v8-M8050.467.055.240.963.743.5
YOLOE-v8-L8053.069.857.942.766.545.6
YOLOE-11-S16046.262.950.037.659.340.1
YOLOE-11-M8051.368.356.041.564.844.3
YOLOE-11-L8052.669.757.542.466.245.2
+ +Table 5. Roadmap to YOLOE in terms of text prompts. The standard AP is reported on LVIS minival set in the zero-shot manner. The FPS is measured on Nvidia T4 GPU and iPhone 12 with TensorRT (T) and CoreML (C), respectively. + +
ModelEpochsAPAPrAPcAPfFPS (T / C)
YOLO-Worldv2-L10033.022.632.035.880.0 / 22.1
+ Fewer train. epochs3031.022.628.834.280.0 / 22.1
+ Global negative dict.3031.922.831.034.480.0 / 22.1
- Cross-modal. fusion3030.019.128.033.9102.5 / 27.2
+ MobileCLIP encoder3031.520.230.534.4102.5 / 27.2
+ RepRTA3033.529.532.035.5102.5 / 27.2
+ Segment. (YOLOE)3033.330.832.234.6102.5 / 27.2
+ +demonstrate that YOLOE can serve as a strong starting point for transferring to downstream task. + +# 4.5. Ablation study + +We further provide extensive analyses for the effectiveness of designs in our YOLOE. Experiments are conducted on YOLOE-v8-L and standard AP is reported on LVIS minival set for zero-shot evaluation, by default. + +Roadmap to YOLOE. We outline the stepwise progression from the baseline model YOLOv8-Worldv2-L to our YOLOE-v8-L in terms of text prompts in Tab. 5. With the initial baseline metric of $33.0\%$ AP, due to limited computational resource, we first reduce the training epochs to 30, leading to $31.0\%$ AP. Besides, instead of using empty + +![](images/85b529afccc2ce1b97a4819e655ddfdc08b8f8653a55c1fc21fd12ed99fef1e4.jpg) +(a) + +![](images/fa789d6b41e4a4d27973106371b317cfa539136d81e74a3347c82962fb149d9d.jpg) +(b) +Figure 4. (a) Zero-shot inference on LVIS. (b) Results with customized text prompt, where "white hat, red hat, white car, sunglasses, mustache, tie" are provided as text prompts. (c) Results with visual prompt, where the red dashed bounding box serves as the visual cues. (d) Results in prompt-free scenario, where no explicit prompt is provided. Please refer to the supplementary for more examples. + +![](images/39a5e45c14e8b49bbba828e8160acbe561c652c28644f1a655458fddde2984d9.jpg) +(c) + +![](images/eb6ac8305f56cdb804cf86469cf96b28c4a693b3ecbece206a5288db00b633e6.jpg) +(d) + +Table 6. Effective. of SAVPE. Table 7. Effective. of LRPC. + +
ModelAPAPrAPcAPf
Mask pool30.427.631.330.2
SAVPE31.929.432.531.7
A = 130.928.231.930.4
A = 1631.929.432.531.7
A = 3231.928.233.031.7
+ +
ModelLRPCAPAPrAPcAPfFPS
v8-SX21.019.121.421.056.5
δ = 1e-321.019.121.321.095.8
δ = 1e-421.019.121.321.066.1
δ = 1e-220.819.121.220.8106
v8-LX27.223.527.028.019.9
δ = 1e-327.223.527.028.025.3
+ +string as negative texts for grounding data, we follow [65] by maintaining a global dictionary to sample more diverse negative prompts. The global dictionary is constructed by selecting category names that appear more than 100 times in the training data. This leads to $0.9\%$ AP improvement. Next, we remove the cross-modality fusion to avoid costly visual-textual feature interaction, which results in $1.9\%$ AP degradation but with $1.28 \times$ and $1.23 \times$ inference speedups on T4 and iPhone 12, respectively. To address this drop, we utilize stronger MobileCLIP-B(LT) text encoder [57] to obtain better pretrained textual embeddings, which recovers AP to $31.5\%$ . Furthermore, we employ RepRTA to enhance the alignment between anchor points' object and textual embeddings, which leads to notable $2.3\%$ AP enhancement with zero inference overhead, showing its effectiveness. At last, we introduce the segmentation head and train YOLOE for detection and segmentation simultaneously. Although this leads to $0.2\%$ AP and $0.9\mathrm{AP}_f$ drop due to multi-task learning, YOLOE gains ability to segment arbitrary objects. + +Effectiveness of SAVPE. To verify the effectiveness of SAVPE for visual inputs, we remove the activation branch and simply leverage mask pooling to aggregate semantic features with the formulated visual prompt mask. As shown in Tab. 6, SAVPE significantly outperforms "Mask pool" by 1.5 AP. This is because "Mask pool" neglects the varying semantic importance at different positions within promptindicated region, while our activation branch effectively models such difference, leading to improved aggregation of semantic features and better prompt embedding for contrast. We also examine the impact of different group numbers, i.e., $A$ , in the activation branch. As shown in Tab. 6, performance can also be enhanced with only a group, i.e., $A = 1$ . Besides, we can achieve the strong performance of 31.9 AP under $A = 16$ , obtaining the favorable balance, where more groups lead to marginal performance difference. + +Effectiveness of LRPC. To verify the effectiveness of LRPC for prompt-free setting, we introduce the baseline that directly leverage the built-in large vocabulary as text prompts for YOLOE to identify all objects. Tab. 7 presents the comparison results. We observe that with the same performance, our LRPC obtains notably $1.7 \times / 1.3 \times$ inference speedups for YOLOE-v8-S / L, respectively, by lazily retrieving the categories for anchor points with found objects and skipping the numerous irrelevant ones. These results well highlight its efficacy and practicality. Besides, with different threshold $\delta$ for filtering, LRPC can achieve different performance and efficiency trade-offs, e.g., enabling $1.9 \times$ speedup for YOLOE-v8-S with only 0.2 AP drop. + +# 4.6. Visualization analyses + +We conduct visualization analyses for YOLOE in four scenarios: (1) Zero-shot inference on LVIS in Fig. 4.(a), where its category names are text prompts, (2) Text prompts in Fig. 4.(b), where arbitrary texts can be input as prompts, (3) Visual prompts in Fig. 4.(c), where visual cues can be drawn as prompts, and (4) No explicit prompt in Fig. 4.(d), where model identifies all objects. We can see that YOLOE performs well and can accurately detect and segment various objects in these diverse scenarios, further showing its efficacy and practicality in various applications. + +# 5. Conclusion + +In this paper, we present YOLOE, a single highly efficient model that seamlessly integrates object detection and segmentation across diverse open prompt mechanisms. Specifically, we introduce RepRTA, SAVPE, and LRPC to enable YOLOs to process textual prompt, visual cues, and prompt-free paradigm with favorable performance and low cost. Thanks to them, YOLOE enjoys strong capabilities and high efficiency for various prompt ways, enabling real-time seeing anything. We hope that it can serve as a strong baseline to inspire further advancements. + +# 6. Acknowledgments + +This work was supported by National Natural Science Foundation of China (Nos. 62525103, 624B2082, 62271281, 62441235). + +# References + +[1] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 1, 2, 3, 6 +[2] Daniel Bogdoll, Maximilian Nitsche, and J Marius Zöllner. Anomaly detection in autonomous driving: A survey. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4488-4499, 2022. 1 +[3] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9157-9166, 2019. 1, 2, 3, 5 +[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 2 +[5] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 1, 2, 4, 5, 6 +[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 4 +[7] Achal Dave, Piotr Dólar, Deva Ramanan, Alexander Kirillov, and Ross Girshick. Evaluating large-vocabulary object detectors: The devil is in the details. arXiv preprint arXiv:2102.01066, 2021. 6 +[8] Douglas Henke Dos Reis, Daniel Welfer, Marco Antonio De Souza Leite Cuadros, and Daniel Fernando Tello Gamarra. Mobile robot navigation using an object recognition software with rgbd images and the yolo algorithm. Applied Artificial Intelligence, 33(14):1290-1305, 2019. 1 +[9] David H Douglas and Thomas K Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica: the international journal for geographic information and geovisualization, 10 (2):112-122, 1973. 5 +[10] Chengjian Feng, Yujie Zhong, Yu Gao, Matthew R Scott, and Weilin Huang. Toed: Task-aligned one-stage object detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3490-3499. IEEE Computer Society, 2021. 2 +[11] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European conference on computer vision, pages 540-557. Springer, 2022. 2 +[12] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 2 +[13] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. International Conference on Learning Representation, 2022. 2 + +[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2, 5 +[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 2 +[16] Xinyu Huang, Yi-Jie Huang, Youcai Zhang, Weiwei Tian, Rui Feng, Yuejie Zhang, Yanchun Xie, Yaqian Li, and Lei Zhang. Open-set image tagging with multi-grained text supervision. arXiv preprint arXiv:2310.15200, 2023. 5 +[17] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 5 +[18] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2 +[19] Qing Jiang, Feng Li, Tianhe Ren, Shilong Liu, Zhaoyang Zeng, Kent Yu, and Lei Zhang. T-rex: Counting by visual prompting. arXiv preprint arXiv:2311.13596, 2023. 1, 3 +[20] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38–57. Springer, 2024. 1, 2, 3, 4, 5, 6 +[21] Glenn Jocher, Jing Qiu, and Ayush Chaurasia. *Ultralytics* YOLO, 2023. 1, 2, 3, 5, 6 +[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 5, 6 +[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 3 +[24] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, et al. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2(3):18, 2017. 6 +[25] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. International Conference on Learning Representation, 2022. 2 +[26] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. International Conference on Learning Representation, 2022. 2 +[27] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, + +Weiqiang Nie, et al. Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976, 2022. 2 +[28] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 2 +[29] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3041-3050, 2023. 2 +[30] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Jianwei Yang, Chunyuan Li, et al. Visual in-context prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861-12871, 2024. 1, 2, 3, 4 +[31] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, Lei Zhang, and Jianfeng Gao. Segment and recognize anything at any granularity. In European Conference on Computer Vision, pages 467-484. Springer, 2024. 3 +[32] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 1, 2, 6 +[33] Chuang Lin, Yi Jiang, Lizhen Qu, Zehuan Yuan, and Jianfei Cai. Generative region-language pretraining for open-ended object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13958-13968, 2024. 1, 2, 3, 4, 5, 7 +[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 2, 5 +[35] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2 +[36] Lihao Liu, Juexiao Feng, Hui Chen, Ao Wang, Lin Song, Jungong Han, and Guiguang Ding. Yolo-uniow: Efficient universal open-world object detection. arXiv preprint arXiv:2412.20645, 2024. 2 +[37] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 1, 2, 4, 6 +[38] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C + +Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 2 +[39] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049-1059, 2022. 6 +[40] Yanxin Long, Youpeng Wen, Jianhua Han, Hang Xu, Pengzhen Ren, Wei Zhang, Shen Zhao, and Xiaodan Liang. Capdet: Unifying dense captioning and open-world detection pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15233-15243, 2023. 3 +[41] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European conference on computer vision, pages 728-755. Springer, 2022. 3 +[42] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 7 +[43] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 5 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4 +[45] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 2 +[46] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5 +[47] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1, 2, 3 +[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 2 + +[49] Tianhe Ren, Yihao Chen, Qing Jiang, Zhaoyang Zeng, Yuda Xiong, Wenlong Liu, Zhengyu Ma, Junyi Shen, Yuan Gao, Xiaoke Jiang, et al. Dino-x: A unified vision model for open-world object detection and understanding. arXiv preprint arXiv:2411.14347, 2024. 1, 2, 3, 4 +[50] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, Yuda Xiong, Hao Zhang, Feng Li, Peijun Tang, Kent Yu, and Lei Zhang. Grounding dino 1.5: Advance the "edge" of open-set object detection, 2024. 6 +[51] Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123, 2018. 6 +[52] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 5, 6 +[53] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 4 +[54] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13193-13203, 2024. 2 +[55] Joseph Sobek, Jose R Medina Inojosa, Betsy J Medina Inojosa, SM Rassoulinejad-Mousavi, Gian Marco Conte, Francisco Lopez-Jimenez, and Bradley J Erickson. Medyolo: a medical image object detection framework. Journal of Imaging Informatics in Medicine, pages 1-9, 2024. 1 +[56] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 2 +[57] Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, and Oncel Tuzel. Mobile-clip: Fast image-text models through multi-modal reinforced training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15963–15974, 2024. 4, 5, 8 +[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[59] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2025. 2 +[60] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7464-7475, 2023. 2 +[61] Yu Wang, Xiangbo Su, Qiang Chen, Xinyu Zhang, Teng Xi, Kun Yao, Errui Ding, Gang Zhang, and Jingdong Wang. + +Ovlw-detr: Open-vocabulary light-weighted detection transformer. arXiv preprint arXiv:2407.10655, 2024. 2 +[62] Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. Grit: A generative region-to-text transformer for object understanding. In European Conference on Computer Vision, pages 207-224. Springer, 2024. 1, 3, 4 +[63] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18134-18144, 2022. 2 +[64] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multi-modal queried object detection in the wild. Advances in Neural Information Processing Systems, 36:4452-4469, 2023. +[65] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 1, 2, 6, 8 +[66] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27391-27401, 2024. 3 +[67] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European conference on computer vision, pages 106-122. Springer, 2022. 2, 3, 4 +[68] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14393-14402, 2021. 2 +[69] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2 +[70] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding, 2022. 6 +[71] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2 +[72] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. 2 +[73] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, + +Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.4 +[74] Tiancheng Zhao, Peng Liu, Xuan He, Lu Zhang, and Kyusong Lee. Real-time transformer-based open-vocabulary detection with efficient fusion head. arXiv preprint arXiv:2403.06892, 2024. 2 +[75] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16793-16803, 2022. +[76] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European conference on computer vision, pages 350-368. Springer, 2022. 2 +[77] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. International Conference on Learning Representation, 2021. 2 +[78] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection, 2021. 4 +[79] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15116-15127, 2023. 2 +[80] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in neural information processing systems, 36:19769-19782, 2023. 1, 3 \ No newline at end of file diff --git a/yoloerealtimeseeinganything/images.zip b/yoloerealtimeseeinganything/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8dcee7c7fc21596aa0bccefb902758d6df4a2ebd --- /dev/null +++ b/yoloerealtimeseeinganything/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43bfb07a6ed171363b682f543a9aa6ca9ee9a899fe8838aefb5df43ccf0a146e +size 667812 diff --git a/yoloerealtimeseeinganything/layout.json b/yoloerealtimeseeinganything/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..01fe847476b0a5a07f15255fe42fad1e24228b0e --- /dev/null +++ b/yoloerealtimeseeinganything/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da6eb7f5e503b4be6667f873a99b4192db977a45d9abf313bb06b25b7f459937 +size 453548 diff --git a/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/24258e3b-6f86-4fb2-aca3-b713b31c24a9_content_list.json b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/24258e3b-6f86-4fb2-aca3-b713b31c24a9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..37bf4e9151c048d66bd80d26d06d0d34e927737f --- /dev/null +++ b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/24258e3b-6f86-4fb2-aca3-b713b31c24a9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f15217f95050f5de6caea608a6b41ca4779a46c4fd789b03598467716ed4e548 +size 82089 diff --git a/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/24258e3b-6f86-4fb2-aca3-b713b31c24a9_model.json b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/24258e3b-6f86-4fb2-aca3-b713b31c24a9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e2d6b5e49969b1b009e6fcb12e786b3336608a4a --- /dev/null +++ b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/24258e3b-6f86-4fb2-aca3-b713b31c24a9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdbcb046f638c7d945e6d8a09e4dc55f0db0c15a421fc37a46373bafddab293f +size 103669 diff --git a/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/24258e3b-6f86-4fb2-aca3-b713b31c24a9_origin.pdf b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/24258e3b-6f86-4fb2-aca3-b713b31c24a9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a0d38926af5a95c051ae878591fd53253190cf43 --- /dev/null +++ b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/24258e3b-6f86-4fb2-aca3-b713b31c24a9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2998f4f77c8464067f6d494fac12175d51a98218bdfd81176afbf8a33fa6a704 +size 2621534 diff --git a/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/full.md b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e18bc661d5909cdf653847e41994be360e93cbf4 --- /dev/null +++ b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/full.md @@ -0,0 +1,337 @@ +# You Are Your Own Best Teacher: Achieving Centralized-level Performance in Federated Learning under Heterogeneous and Long-tailed Data + +Shanshan Yan $^{1}$ Zexi Li $^{3,4}$ Chao Wu $^{4}$ Meng Pang $^{5}$ Yang Lu $^{1,2*}$ Yan Yan $^{1,2}$ Hanzi Wang $^{1,2}$ + +1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, Xiamen, China + +$^{2}$ Fujian Key Laboratory of Sensing and Computing for Smart City, School of Informatics, Xiamen + +University, Xiamen, China + +3University of Cambridge 4Zhejiang University + +$^{5}$ School of Mathematics and Computer Sciences, Nanchang University, Nanchang + +yanshanshan@stu.xmu.edu.cn, zexi.li@zju.edu.cn, chao.wu@zju.edu.cn, + +mengpang@ncu.edu.cn, luyang@xmu.edu.cn, yanyan@xmu.edu.cn, hanzi_wang@163.com + +# Abstract + +Data heterogeneity, stemming from local non-IID data and global long-tailed distributions, is a major challenge in federated learning (FL), leading to significant performance gaps compared to centralized learning. Previous research found that poor representations and biased classifiers are the main problems and proposed neural-collapse-inspired synthetic simplex ETF to help representations be closer to neural collapse optima. However, we find that the neural-collapse-inspired methods are not strong enough to reach neural collapse and still have huge gaps to centralized training. In this paper, we rethink this issue from a self-bootstrap perspective and propose FedYoYo (You Are Your Own Best Teacher), introducing Augmented Self-bootstrap Distillation (ASD) to improve representation learning by distilling knowledge between weakly and strongly augmented local samples, without needing extra datasets or models. We further introduce Distribution-aware Logit Adjustment (DLA) to balance the self-bootstrap process and correct biased feature representations. FedYoYo nearly eliminates the performance gap, achieving centralized-level performance even under mixed heterogeneity. It enhances local representation learning, reducing model drift and improving convergence, with feature prototypes closer to neural collapse optimality. Extensive experiments show FedYoYo achieves state-of-the-art results, even surpassing centralized logit adjustment methods by $5.4\%$ under global long-tailed settings. The code is available at https://github.com/shanss132/FedYoYo. + +![](images/1d9cd291874bc9eff92a111cd244efbb4734eee730ec7be13be26355f4738e66.jpg) +Figure 1. Our method substantially closes the gap between centralized training and federated learning under heterogeneous data. Left: non-IID data with global long-tailed (LT) distribution, where non-IID $\alpha = 0.5$ and imbalance factor (IF) $= 0.01$ . Right: non-IID data with global balanced distribution (vanilla non-IID setting in the literature: the smaller $\alpha$ , the more non-IID). Baselines: vanilla centralized training, LA loss [21], FedAvg [20], FedLC [35]. + +![](images/72c225d6bbb14a801793accb96a7d8ae542b7d139b2b3748ce2e133c3b714954.jpg) + +# 1. Introduction + +Federated learning (FL) [11, 14, 15, 20] is a collaborative learning paradigm that builds machine learning models from distributed data sources without sharing the raw data, which is communication-efficient [20] and privacy-preserving [34]. FL is promising in a wide range of scenarios, like medical imaging [1], multi-media analysis [30], the Internet of Things [39], etc. One inherent and key challenge in FL is data heterogeneity, a critical bottleneck that significantly impacts the performance of FL methods [14, 16]. + +In practice, the overall data distribution in an FL system is often long-tailed, meaning that data heterogeneity arises from both the local non-IID data and the global long- + +tailed [23, 24, 40] data. This combination leads to a more severe form of heterogeneity, causing in sub-optimal local models and poor generalization of the aggregated global model. As a result, there is always a huge gap between federated methods and centralized training due to data heterogeneity, which becomes even more evident when clients' data are highly non-IID [11, 14, 16, 20]. + +Previous works tried to tackle non-IID data and improve the performances of FL by a large margin from FedAvg, but the centralized-federated gap is also dominant (Fig. 1). Recent studies [16, 17, 19, 28, 31] have shown that data heterogeneity can lead to poor feature representations and biased classifiers, which might be the main cause of bad generalization performances. The poor representations in local models further exacerbates the feature misalignment between the global model and other client models. Therefore, the key challenge is to learn better and more unified feature representations across clients with non-IID data distributions. Existing approaches, such as FedETF [16] and FedLoGe [31], attempt to mitigate the effects of heterogeneity by leveraging the theory of Neural Collapse [9, 12] and constructing a synthetic simplex equiangular tight frame (ETF) with maximal pairwise angles [22]. Neural collapse is a deep learning phenomenon, which depicts an ideal representation and classifier structure under balanced and sufficient training. Though these methods are inspired by neural collapse, we find they are not good enough to reach neural collapse optima under severe heterogeneity and global long-tailed distribution. An intuitive visualization is in Fig. 2: the pair-wise prototype angles of FedETF and FedLoGe are far away from the theoretical neural collapse optima. + +Therefore, rethinking the representation learning strategy in non-IID federated learning is needed. We notice that self-supervised learning (SSL) has been widely validated in large-scale representation learning tasks, demonstrating the ability to capture more robust feature representations. This suggests that we can think of supervised FL in an unsupervised way to gain better representations, but more tailored designs are needed. The question is how to effectively use class/label distributions under data heterogeneity. Existing SSL uses sample-wise contrastive or bootstrap methods and learns feature extractors instead of the whole model. But in supervised heterogeneous FL, class-biased local data and long-tailed global data will make features prioritize majority classes while neglecting minority ones, also the local model representations are not aligned, causing severe model drifts. + +In this paper, we find that self-bootstrap representation learning can release its full potential under the supervision of distribution-aware logit adjustment, and propose FedY-oYo: You Are Your Own Best Teacher for local clients. Unlike previous SSL, we learn logits as representations instead of features, and logit adjustment can serve as distribution + +guidance to improve the representation of minor classes and align the representations across clients. FedYoYo consists of two core components: Augmented Self-bootstrap Distillation (ASD) and Distribution-aware Logit Adjustment (DLA). ASD is inspired by BYOL [7] in SSL, but tailored designs are made in FL. We learn logits instead of features and use one local model as the teacher itself instead of two online and target models in BYOL. We use the logits of a weakly augmented sample as the teacher to guide the learning of a strongly augmented sample. DLA is a tailored version of logit adjustment for heterogeneous FL; considering the potential long-tailed global distribution, we realize a tradeoff between both local and global distributions. + +FedYoYo realizes a matched co-design of self-bootstrap learning and logit adjustment, reaching near-centralized performances under non-IID FL (Fig. 1). In Fig. 2, it can be seen that our FedYoYo has better representations closer to neural collapse optimality and reaches higher generalization, compared with ETF-based methods. + +Our main contributions are summarized as follows: + +- We propose FedYoYo, which addresses two challenges: data heterogeneity and the combined issue of global long-tailed and local non-IID data. It has two key components: Augmented Self-bootstrap Distillation and Distribution-Aware Logit Adjustment. +- FedYoYo achieves a new level of performance in FL, it is comparable to centralized training. FedYoYo closes the centralized-federated gap from $20\%$ to $1.3\%$ in vanilla $\alpha = 0.1$ data heterogeneity, and it even surpasses the centralized method by $5.4\%$ under global long-tailed distribution. +- We provide a new perspective on solving non-IID data and the caused poor representation issues in FL. We make FL more applicable and promising by reaching performances comparable to those of centralized. + +# 2. Proposed Method + +In this section, we introduce our proposed method FedY-oYo, which focuses on enhancing feature representations and mitigating feature misalignment under data heterogeneity. The method consists of two key components: Augmented Self-bootstrap Distillation (ASD) and Distribution-aware Logit Adjustment (DLA). ASD employs a self-bootstrap mechanism with logit adjustment as distribution guidance to enhance feature extraction, while DLA calibrates the classifier outputs using a fused global-local distribution to address client bias. Together, these components enable consistent and robust feature representation across clients, as illustrated in the overall framework Fig. 3. + +# 2.1. Preliminaries + +In the context of FL, consider a scenario with $K$ clients, where each client holds a non-IID local dataset $\mathcal{D}_k =$ + +![](images/2735419f0a28bd43282bdb6ca9e4f2d3328b1274986a9c1bfd1fe669138d7f1e.jpg) +0 0 0 +Max: $118.02^{\circ}$ Min: $18.25^{\circ}$ + +![](images/9e6f42507b84adf0ed03219458844ab142648207d4105229702294b18c81f5ee.jpg) +Max: $133.32^{\circ}$ Min: $41.20^{\circ}$ + +![](images/9a71aaa15568259b668a3d0774d8b727cb6a3a17045920d3b1695e888878d418.jpg) +Neural Collapse Optima: $96.38^{\circ}$ +Max: $142.91^{\circ}$ Min: $41.12^{\circ}$ + +![](images/4517dba8316f2dba49da0ae32ada50a6a6f96ea5eabc65770bc29352dc403ed5.jpg) +Max: $136.64^{\circ}$ Min: $50.25^{\circ}$ + +![](images/b1bdc48710840e6a56c9857480c2ed2e1b9ba09a2d95cd68bb37044c8310be97.jpg) +Mean: $78.32^{\circ}$ Acc: 60.76 +Imbalance rate $= 0.01$ & Heterogeneity $= 0.5$ +Max: $157.79^{\circ}$ Min: $11.82^{\circ}$ +Mean: $70.26^{\circ}$ Acc: 58.22 +(a) FedAvg +Figure 2. Visualization of neural collapse degrees and accuracy for global models on CIFAR-10. The optimal collapse angle $(96.38^{\circ})$ follows Equiangular Tight Frame (ETF) theory, which defines ideal class prototype angles for maximal separation. Arrows indicate prototype representations, and colors denote different categories. "Max", "Min", and "Mean" represent the largest, smallest, and average angles between class prototypes, while "Acc" denotes global model accuracy. The top row shows vanilla data heterogeneity, and the bottom row includes long-tailed federated non-IID settings. (a) FedAvg, (b) FedETF, (c) FedLoGe, and (d) FedYoYo. Our FedYoYo method reaches better neural collapse conditions and achieves the best performance. + +![](images/6251ee0fac66c41d0257c4b586ed09784bddf5f6161886c0d64c48dbb4f6e7fa.jpg) +Mean: $91.40^{\circ}$ Acc: 72.31 +Max: $152.16^{\circ}$ Min: $24.16^{\circ}$ +Mean: $83.09^{\circ}$ Acc: 68.25 +(b) FedETF + +![](images/a45186c53253090b0d0028b79640c141f6a17ae30eb5a11fa06a80535241ec90.jpg) +Mean: $78.32^{\circ}$ Acc: 76.54 +Max: $159.04^{\circ}$ Min: $39.09^{\circ}$ +Mean: $90.18^{\circ}$ Acc: 70.54 +(c) FedLoGe + +![](images/f0675dfb8272af9c39ee6e8049db0a199ad5dd454e8383cb63d5c003ecb5021b.jpg) +Mean: $93.97^{\circ}$ Acc: 81.70 +Max: $145.02^{\circ}$ Min: $46.45^{\circ}$ +Mean: $93.39^{\circ}$ Acc: 81.45 +(d) FedYoYo + +$\{(x_{i},y_{i})\mid 1\leq i\leq n_{k}\}$ , with $n_k$ representing the number of samples on client $k$ , and $x_{i}$ $y_{i}$ denoting the input data and its corresponding label for each sample $i$ . The model of client $k$ denotes $f_{k}(x)$ . The entire training dataset $\mathcal{D} =$ $\{\mathcal{D}_k\}_{k = 1}^K$ combines the data from all clients. In vanilla FL, the overall distribution of $\mathcal{D}$ is class-balanced [11, 14, 16], while in our paper, we also consider a more realistic scenario where the global dataset exhibits a long-tailed distribution across the $C$ classes [24]. In this global long-tailed distribution, we assume that classes are ordered by their sample frequencies such that if $i < j$ , then $N^{i}\geq N^{j}$ where $N^i$ is the total number of samples for class $i$ . For each class $c$ , let $n_k^c$ represent the number of samples of class $c$ on client $k$ leading to the global sample count of class $c$ as $N^{c} = \sum_{k = 1}^{K}n_{k}^{c}$ + +# 2.2. Augmented Self-bootstrap Distillation + +We adopt a self-supervised learning approach similar to BYOL [7], applying different augmentations to local data to capture more robust feature representations. Then we introduce Augmented Self-bootstrap Distillation, using the weakly augmented view as the teacher to provide guidance for the strongly augmented view, with the goal of enhanc + +ing feature representation quality. Unlike traditional self-distillation [37], which typically refers to knowledge transfer within a model such as from deeper to shallower layers, our approach distills knowledge across different views for the same model. Also, unlike BYOL, we learn logits instead of features and efficiently use one local model as the teacher itself instead of two online and target models in BYOL. Using logits as the representation can align the representation across clients since the logit space is the unified prediction space; also, we can calibrate the logits by considering both local and global distributions. Specifically, we employ Kullback-Leibler (KL) divergence to align the posterior distributions of the strong and weak augmentations. The distillation loss is then defined by minimizing the KL divergence as follows: + +$$ +\mathcal {L} _ {A S D} = \frac {1}{n _ {k}} \sum_ {i = 1} ^ {n _ {k}} K L \left(p \left(\bar {x} _ {i}\right) \| p \left(\tilde {x} _ {i}\right)\right), \tag {1} +$$ + +where $\overline{x}_i$ and $\widetilde{x}_i$ represent the weak and strong augmentations of the same sample. Specifically, we leverage the local model to learn from two augmented instances generated through weak and strong augmentations. For weak augmentations, we apply techniques including RandomCrop, Ran + +![](images/e0f2bae3f3822a8a5d2a26bc11dff296d4a30f17d32e0a608bf6a7de9a17bc33.jpg) +Figure 3. Overview of our proposed FedYoYo framework. On the server side, the estimated client distributions are aggregated to obtain an approximate global distribution. On the client side, client $k$ combines its local distribution $\pi_{k}$ with the global distribution $\pi_{g}$ using EMA updates for balance softmax. During local training, the model $f_{k}(x)$ processes weak and strong augmentations, using augmented self-bootstrap distillation loss $(\mathcal{L}_{ASD})$ and distribution-aware logit adjustment loss $(\mathcal{L}_{DLA})$ to enhance representation learning. + +domHorizontalFlip, and RandomRotation. For strong augmentations, we follow [4]. Notably, to avoid introducing noise, only correctly classified samples from the weak augmentation are used as the teacher to guide the learning of the corresponding student, rather than including all samples. Importantly, we adopt Eq. (7) as the adjusted softmax $p(x)$ in Eq. (1) to improve the performance of minority classes. + +The ASD technique not only captures richer features by leveraging a bootstrap style but also alleviates client drift by focusing on correctly classified samples. Thus, the local models and local data can be the best teachers for themselves in heterogeneous FL. This targeted guidance helps the local models converge more closely to optimal representations, bridging the performance gap between global and local models. Furthermore, by integrating an adjusted softmax, we effectively mitigate the biases caused by the long-tailed data, resulting in more stable and robust model performance across diverse clients. + +# 2.3. Distribution-aware Logit Adjustment + +As mentioned, data heterogeneity often leads to biases in both feature representation and the classifier itself. Such biased classifiers may undermine the effectiveness of representation learning, thereby negatively affecting the local self-bootstrap distillation process. Therefore, it is essential to calibrate the classifier on each client during training to guarantee consistent outputs across all clients. To achieve this, we adopt a balanced softmax for the model output, and the resulting probability distribution is: + +$$ +p (x) = \frac {n ^ {y} \exp \left(f (x , y) / T\right)}{\sum_ {y ^ {\prime} = 1} ^ {C} n ^ {y ^ {\prime}} \exp \left(f (x , y ^ {\prime}) / T\right)}, \tag {2} +$$ + +where $n^y$ is the sample count of class $y$ and $T$ is the temperature coefficient (set to 1.5 in experiments) that controls the smoothness of the output. + +In federated long-tailed classification, global data imbalance exacerbates client-side heterogeneity, making sample count an unreliable measure of class representation. Minority classes can occupy large feature space regions despite few samples, while majority classes may show limited + +spread due to overlapping features. This reliance on sample count leads to biased class representation and degrades global model performance. + +To address this, we propose a global-distribution-aware weight allocation strategy using the Pearson Correlation Coefficient to analyze local sample relationships for more adaptive class weighting. Inspired by previous work [3], this method derives a local weight distribution that better reflects data variability. For class $c$ , we first compute the correlation matrix $R_{cb}$ within each batch $b$ , which measures the pairwise relationships among all samples in the batch. The matrix $R_{cb}$ is defined as: + +$$ +R _ {c b} = \left[ R _ {c b} ^ {i, j} \right] _ {i, j \in b _ {k} ^ {c}}, \tag {3} +$$ + +$$ +R _ {c b} ^ {i, j} = \frac {\left(h _ {i} - \mu_ {c}\right) \left(h _ {j} - \mu_ {c}\right) ^ {\mathrm {T}}}{\left\| h _ {i} - \mu_ {c} \right\| _ {2} \cdot \left\| h _ {j} - \mu_ {c} \right\| _ {2}}, \tag {4} +$$ + +where $b_{k}^{c}$ denotes the set of samples from class $c$ in a single batch, $h_{i}$ and $h_{j}$ represent the feature vectors of samples $i$ and $j$ , $\mu_{c}$ is the mean feature vector of class $c$ in the local data, i.e., the class prototype. Next, we compute the effective prior distribution iteratively across batches during training. The effective prior distribution $\pi_{k}^{c}$ is computed as: + +$$ +\pi_ {k} ^ {c} = \left\{\sum_ {b = 1} ^ {B} \frac {1}{a _ {c b} R _ {c b} a _ {c b} ^ {\mathrm {T}}} \right\}, \tag {5} +$$ + +where $B$ denotes the total number of training batches, and $a_{cb} = \left(\frac{1}{|b_k^c|},\frac{1}{|b_k^c|},\dots ,\frac{1}{|b_k^c|}\right)\in \mathbb{R}^{1\times |b_k^c |}$ Finally, we obtain the local prior distribution as $\pi_k = \{\pi_k^1,\pi_k^2,\dots ,\pi_k^c\}$ + +Additionally, for data heterogeneity, we directly perform logit adjustment using the locally estimated distribution $\pi_{k}$ . However, in the case of federated long-tailed learning, the influence of the global distribution must also be considered. To address this, we approximate the global distribution $\pi_{g}$ by aggregating client-side $\pi_{k}$ at the server by FedAvg, enabling more balanced model training. The final fused distribution $\pi_{mix}$ is defined as: + +$$ +\pi_ {m i x} \leftarrow (1 - \gamma) \cdot \pi_ {g} + \gamma \cdot \pi_ {k}, \tag {6} +$$ + +where $\gamma$ controls the degree of integration, which is discussed in Sec. 3.4. + +Then the distribution-fused balanced softmax is expressed as: + +$$ +p (x) = \frac {\pi_ {m i x} ^ {y} \exp (f (x , y) / T)}{\sum_ {y ^ {\prime} = 1} ^ {C} \pi_ {m i x} ^ {y ^ {\prime}} \exp (f (x , y ^ {\prime}) / T)}. \tag {7} +$$ + +Thus, we propose a distribution-aware logit adjustment loss function, denoted as DLA. The loss function is formulated as follows: + +$$ +\mathcal {L} _ {D L A} = - \frac {1}{2 n _ {k}} \sum_ {i = 1} ^ {2 n _ {k}} \log (p (\hat {x} _ {i})) \tag {8} +$$ + +where $\hat{x}_i$ represents the $i$ -th augmented sample (both weakly and strongly augmented), with $2n_k$ accounting for the two types of augmented samples. Since self-bootstrap distillation involves two augmented views, we apply $\mathcal{L}_{DLA}$ to both views to ensure balanced loss contributions. + +# 2.4. Training + +Our method's local loss function combines two components: distillation loss $\mathcal{L}_{ASD}$ and classifier balancing loss $\mathcal{L}_{DLA}$ . The $\mathcal{L}_{ASD}$ loss, as defined earlier, improves local representation learning. To mitigate classifier bias and its negative impact on feature representation learning, we use the $\mathcal{L}_{DLA}$ loss to adjust the local classifier. During client-side training, $\mathcal{L}_{DLA}$ is applied to learn from hard labels (i.e., the one-hot label), with both augmented batches processed simultaneously by the model. Thus, the total loss function is defined as: + +$$ +\mathcal {L} _ {a l l} = \mathcal {L} _ {D L A} + \lambda \mathcal {L} _ {A S D}, \tag {9} +$$ + +where $\lambda$ is a hyperparameter that controls the weight of the distillation loss. The impact of this hyperparameter is discussed in Sec. 3.4. For the server aggregation, we conduct FedAvg to local models. + +# 2.5. Privacy Discussion + +Since local distribution estimation occurs entirely client-side, no privacy risk arises in the non-IID scenario. However, privacy concerns may emerge when incorporating global distribution information in federated long-tailed learning—a common challenge in federated learning, not specific to our method, as previous approaches such as FedGrab [32] and FedLoGe [31] have also utilized local data distributions. If privacy protection is required, differential privacy (DP) [6] can be applied by adding noise to uploaded distributions. A comprehensive discussion on federated learning privacy is beyond this work's scope; thus, we briefly address it here. + +# 3. Experiments + +# 3.1. Experimental Setup + +Datasets and models. We first evaluate our model on CIFAR-10/100, where the heterogeneity of the client data is controlled using the concentration parameter $\alpha$ of the Dirichlet distribution (the smaller $\alpha$ , the more heterogeneous data). To further verify the robustness of our method under the more heterogeneous scenarios induced by real-world long-tailed distributions, we conduct experiments on several standard long-tailed datasets: CIFAR-10/100-LT, SVHN-LT, and ImageNet-LT. The imbalance factor (IF) is used to control the degree of imbalance. ImageNet-LT is a long-tailed version of ImageNet, with the largest and smallest categories containing 1,280 and 5 images. For CIFAR-10/100-LT and SVHN-LT, we utilize the ResNet-8 model, while for ImageNet-LT, we employ the ResNet-50 model. The detailed data distribution of CIFAR-10/100-LT is provided in Appendix.A. Implementation of baseline methods. We select three categories of state-of-the-art (SOTA) baseline methods for comparison: (1) Heterogeneity-oriented methods (FedProx [14], FedETF [16], FedLC [35], and CCVR [19]) and federated long-tailed methods (CReFF [24], Fed-Grab [32], BalanceFL [27], FedIC [23], RUCR [8], and FedLoGe [31]); (2) Federated distillation methods, including FedDF [18], FedFTG [38], FedGen [42], and DaFKD [29]; (3) Long-tailed methods like $\tau$ -norm [10], AREA [3], and LWS [10]. We also present the performance of centralized learning (CL) methods under both heterogeneous and long-tailed settings as an oracle upper bound. + +Federated environment and local training. We follow the experimental setup in previous federated long-tailed literature [24]. We train for 300 rounds to reach sufficient convergence. All models are implemented in PyTorch and trained on NVIDIA GeForce 3090 GPUs. + +# 3.2. Comparison with State-of-the-art Methods + +Results on vanilla non-IID settings. All results are reported in Tab. 1. Our FedYoYo achieves the best performance under different $\alpha$ . It is notable that our method surpasses FedETF and FedLC in extreme non-IID settings. FedETF is the state-of-the-art method using neural-collapse-inspired classifiers. The results show our method has better generalization than FedETF and Fig. 2 shows that our method can realize better neural collapse optimality than FedETF. FedLC incorporates logit adjustment in FL, but our co-design of self-bootstrap and logit adjustment can reach better performances than logit adjustment solely (i.e., FedLC). Furthermore, our approach successfully reduces the performance gap with the centralized baseline, demonstrating its effectiveness in mitigating the challenges posed by non-IID data distribution. + +Table 1. Top-1 test accuracy (\%) of FedYoYo and FL methods with different $\alpha$ . We compare our method on the vanilla non-IID setting without long-tailed distribution on CIFAR-10/100. + +
MethodCIFAR-10CIFAR-100
α=0.01α=0.1α=0.01α=0.1
Centralized90.0368.23
FedAvg [20]60.7672.4652.4658.18
FedProx [14]55.2270.7152.9160.37
CCVR [19]63.2574.1856.9561.82
FedDF [18]60.8379.4844.4951.69
FedGen [42]52.0071.8237.7243.44
FedFTG [38]51.5479.0747.8753.93
DaFKD [29]59.2881.6747.1552.17
FedLC [35]68.4982.0358.1160.44
FedETF [16]72.3183.1058.1861.87
FedYoYo81.7088.8264.4967.59
+ +Results on global long-tailed and non-HID settings. CIFAR-10/100-LT and SVHN-LT: As summarized in Tab. 2, our method achieves the highest test accuracy across all datasets with varying imbalance factors (IFs). Heterogeneity-oriented methods generally perform similarly to FedAvg, as they focus on data heterogeneity but overlook global class imbalance. Imbalance-oriented methods like AREA perform better than FedAvg in certain cases but still lag behind our approach, likely because they primarily address imbalance without accounting for inter-client data heterogeneity. Notably, our method consistently shows significant performance gains over centralized learning with LA loss. ImageNet-LT: We further validate our method on the more challenging ImageNet-LT dataset. In Tab. 3, we report accuracies for three class groups: many-shot (over 100 samples per class), medium-shot (20-100 samples), and few-shot (fewer than 20 samples). Our method consistently outperforms others in all groups, particularly in few-shot classes, where it achieves $15.44\%$ accuracy—an $8.04\%$ improvement over the baseline. This showcases the effectiveness of our approach in enhancing few-shot class performance while maintaining strong accuracy in many-shot classes. + +# 3.3. Analysis of FedYoYo + +Global-to-local model gap reduction. We visualize the average gains from local training and global aggregation in Appendix.B. Our method significantly reduces the gap between global and local models compared to FedAvg, enhancing the gains from global aggregation. This demonstrates that our approach benefits the global model, enabling faster adaptation to local data, even under the challenging non-IID conditions caused by long-tailed distributions. As local models converge toward optimal consistency, the aggregated model also achieves superior performance. Furthermore, as shown in Fig. 4, our method surpasses FedAvg in creating more balanced across-client + +![](images/11cfdadf1c7be67cf02042f0e99ec3a131dffe926e0320f61a7d7ab93be56ec1.jpg) +Figure 4. Comparison of per-class local accuracy after receiving the global model and performing local updates. Bars represent the local data distribution, while lines indicate accuracy. "Softmax" means FedAvg with vanilla softmax. + +![](images/931eefa43f93005efb66209d545bf5905cefb76c866263c071c80f0e1dfce735.jpg) +Figure 5. The t-SNE visualization of feature spaces of FedAvg [20], FedLC [35], and our FedYoYo. + +performance. Overall, our approach exhibits strong adaptability in heterogeneous and long-tailed scenarios, resulting in notable improvements for both global and local models. + +Effectiveness of estimated global distribution. Our estimated class distribution closely aligns with the oracle distribution. To further verify this, we tracked the $\ell_2$ distance between the estimated and original data distributions across training epochs in Appendix.B. Feature level analysis. In Fig. 5, we present a visualization of the feature representations extracted by the global model. Compared to other methods, our approach achieves more compact intra-class features and better inter-class separability. Additionally, we compare the similarity between features of the global model and those of the local models in Appendix.B. Results show that our method effectively reduces the feature discrepancy between local and global models, demonstrating its ability to alleviate client drift and improve local representation learning. Computation Cost. We evaluated three methods using ResNet-18 on CIFAR-100-LT (IF=100) in terms of GFLOPs per round. Compared to FedAvg's 1003.72G and FedGrab's 2371.36G, our method, FedYoYo, achieves better performance with a more efficient computational cost of 1721.79G. + +# 3.4. Ablation Study + +The necessity of distribution fusion. We compare the performance of the real distribution and the estimated distribution and further evaluate the effect of applying a fusion strategy to each. As illustrated in Fig. 6, after applying the fusion strategy, significant improvements are observed in both cases. Notably, under high heterogeneity settings (e.g., + +Table 2. Top-1 test accuracy (%) of FedYoYo and SOTA methods on CIFAR-10/100-LT, SVHN-LT with different IFs and $\alpha = {0.5}$ . The best performance is in bold, and the second is underlined. Red text indicates improved performance compared to the centralized learning with LA loss. ↑ indicates improved accuracy compared with the underlined (best baseline in each setting). + +
MethodCIFAR-10-LTCIFAR-100-LTSVHN-LT
IF=100IF=50IF=10IF=100IF=50IF=10IF=100IF=50IF=10
Centralized Methods
Centralized68.1776.1580.3238.6140.3050.3484.8686.5590.14
Centralized w/ LA Loss [21]77.2780.2386.1541.5744.0256.0688.1790.2393.45
Heterogeneity-oriented FL methods
FedAvg [20]58.2263.4777.9531.6436.1446.8481.3184.4087.31
FedProx [14]55.6860.7276.6432.7335.8245.7782.0486.3390.13
CCVR [19]68.1372.9881.5634.7337.6848.9282.3486.4891.26
FedLC [35]69.0271.9379.0931.7538.3647.5481.1785.2387.43
FedETF [16]68.2572.3581.6033.1937.8648.7183.0287.0790.86
Imbalance-oriented FL methods
τ-norm [10]40.7041.3151.6619.2920.8232.5170.6574.6378.58
LWS [10]37.4639.8249.3018.1820.1033.8171.1873.2578.15
AREA [3]64.3365.1678.9736.3437.8348.5982.8785.2490.35
Federated long-tailed methods
BalanceFL [27]49.3454.1772.0326.0329.2840.1675.4179.1385.16
CReFF [24]70.5173.6079.8532.9034.6643.4285.4787.6491.16
FedIC [23]66.4967.5571.2133.6736.7441.9384.9486.8290.53
Fed-Grab [32]70.6375.4485.2133.5344.0155.8790.3991.1194.56
RUCR [8]55.3260.2475.5627.6133.8141.2973.4582.2087.23
FedLoGe [31]70.5474.8584.5442.6347.6658.3685.0587.0689.68
FedYoYo81.45 (+4.18)83.85 (+3.62)87.94 (+1.79)46.13 (+4.56)50.83 (+6.81)60.16 (+4.10)91.73 (+3.56)92.38 (+2.15)95.10 (+1.65)
↑ 10.82↑ 8.41↑ 2.73↑ 3.50↑ 3.17↑ 1.80↑ 1.34↑ 1.27↑ 0.54
+ +Table 3. Top-1 test accuracy (\%) of FedYoYo and SOTA methods on ImageNet-LT with $\alpha = 0.1$ . + +
MethodImageNet-LT
ManyMediumFewAll
FedAvg [20]34.9219.187.4123.85
FedProx [14]34.2517.066.7322.57
CCVR [19]36.7220.249.2625.49
FedLC [35]36.0321.146.5723.23
FedETF [16]35.9419.917.0723.97
τ-norm [10]30.8114.575.2219.58
LWS [10]37.2323.47.5025.37
AREA [3]39.8323.518.5326.03
BalanceFL [27]30.6319.267.2021.87
CReFF [24]37.6121.4810.0226.91
FedIC [23]36.2220.59.7625.71
Fed-Grab [32]41.1624.4214.2930.56
RUCR [8]30.4715.775.2220.06
FedLoGe [31]40.7724.1613.2729.73
FedYoYo42.0525.7815.4431.41
+ +$\alpha = 0.1)$ , the estimated distribution with the fusion strategy even outperforms the real distribution, demonstrating that the fusion approach not only mitigates the impact of data heterogeneity but also enhances model generalization. This confirms the value of fusing global and local distributions in strengthening model robustness. + +Ablation studies on all components of FedYoYo. In Tab. 4, we present a comprehensive ablation study on the CIFAR-100-LT dataset, evaluating key components: RandAug, ASD, and DLA. RandAug refers to the use of strong + +augmentation. A vanilla model without any components reaches an accuracy of $33.34\%$ , similar to FedAvg. When RandAug is combined with ASD, accuracy improves by $8.65\%$ , highlighting ASD's crucial role in enhancing feature learning and knowledge transfer. Notably, without DLA, many-shot classes see a significant performance boost, but gains for few-shot classes are minimal. When all components are combined, the overall accuracy improves by $12.79\%$ , with gains of $8.42\%$ for medium-shot and $21.13\%$ for few-shot classes. These findings show that DLA mitigates classifier bias and balances the distillation process. Additionally, ASD and DLA reinforce each other, working synergistically to deliver consistent performance improvements across head, medium, and tail classes. + +Impact of data augmentation. In this section, we apply various augmentations to the training samples to assess the effectiveness of ASD. As shown in Tab. 5, we compare four augmentation variants in FedYoYo, and the results indicate that the weak-strong self-bootstrap distillation outperforms other augmentations. Moreover, other augmentation methods were examined in Appendix.C. + +Influence of fusion coefficient $\gamma$ . Fig. 7a shows the effect of the fusion coefficient $\gamma$ on model accuracy across different imbalance factors (IF). A larger $\gamma$ emphasizes the global distribution. Initially, increasing $\gamma$ improves accuracy, but excessive values (e.g., 1.0) cause a decline, indicating that relying solely on global or local distributions is suboptimal. Our approach balances global and local information, achieving better overall performance. + +Influence of loss weight $\lambda$ . As shown in Fig. 7b, despite in + +Table 4. Ablation study of our method's key components on the CIFAR-100-LT dataset with IF = 100 and $\alpha = 0.5$ . + +
RandAugASDDLAManyMediumFewAll
56.8931.947.5033.34
✓60.1830.4912.5035.71
✓51.4634.4321.1336.40
✓✓65.3142.7113.9341.99
✓✓56.0342.5127.3341.69
✓✓✓60.2647.0028.6346.13
+ +Table 5. Top-1 test accuracy $(\%)$ accuracy comparison with different augmentation policies on CIFAR100-LT with IF = 100 and $\alpha = 0.5$ + +
View1(Teacher)View2(Student)Accuracy
Weak augmentationWeak augmentation33.74
Strong augmentationStrong augmentation42.92
Weak augmentationStrong augmentation46.13
Strong augmentationWeak augmentation41.06
+ +creasing $\lambda$ from 3.0 to 5.5, the model performance remains relatively stable with minor variations in accuracy across different values. This suggests that the model is not sensitive to the choice of $\lambda$ , implying that consistent performance can be maintained even with different weights. Therefore, we recommend choosing $\lambda$ within this range, as it provides stable and reliable performance. + +# 4. Related Works + +# 4.1. Data Heterogeneity in Federated Learning + +Our paper focuses on data heterogeneity in federated learning that includes local non-IID heterogeneity and global long-tailed data heterogeneity. Local data heterogeneity causes client drift, which is often mitigated by regularization methods like FedProx [14], SCAFFOLD [11], and MOON [13]. To tackle classifier bias in non-IID settings, prior works explore strategies such as classifier retraining [19], prototype-based classifier rebalancing [5, 36] and local calibration [35]. However, global long-tailed distributions further exacerbate cross-client heterogeneity, making adaptation more difficult. Recent methods focus on server-side calibration [23, 24] or local model adjustment using global distribution [32], but their effectiveness remains limited, highlighting the need for more robust solutions. + +# 4.2. Feature Representation Learning in FL + +Data heterogeneity in federated learning leads to poor feature representations and biased classifiers, causing misalignment between global and client models. Under global long-tailed distributions, classifier bias further worsens this misalignment. Anchor-based methods [33, 41] attempt feature alignment, while neural-collapse-inspired approaches + +![](images/8940472e55c2c4175e26ffe1c38f7b35db36c1a8438d555ea1c7abbf91fe2f32.jpg) +(a) CIFAR-100-LT, $\alpha = 0.1$ + +![](images/36b8b13079465e3ad27d252916bff78eab644587d824c2d62d09df641c922980.jpg) +(b) CIFAR-100-LT, $\alpha = 0.5$ + +![](images/2a1b03862bc700fc4e291497d62fbcb5c220282ddf0b9c280a983ca6fa423622.jpg) +Figure 6. Top-1 test accuracy $(\%)$ of our method using different distribution estimation strategies on CIFAR-100-LT with varying $\alpha$ . Light blue bars show results using the ground-truth distribution, while dark blue bars represent the Pearson coefficient-based estimation. Red texts highlight the performance gains from the fused distribution strategy. +(a) Influence of $\gamma$ +Figure 7. (a) Comparison of different fusion ratios $\gamma$ . We report the performance across various fusion ratios. (b) Comparison of different values of $\lambda$ . + +![](images/1329e1106c5cc46b93bb6786f485f81e5aac82915ef0e9cd84c5333129171b00.jpg) +(b) Influence of $\lambda$ + +employ ETF-based methods [16, 31] or feature regularization [26] to enhance generalization. Others use weighting in client aggregation [25] to mitigate bias. However, these methods yield limited improvements in feature representation. Existing self-supervised federated methods typically utilize contrastive learning [2, 43] or bootstrap methods [7], often overlooking minority classes and disproportionately emphasizing majority classes in supervised heterogeneous federated scenarios. In contrast, our method explicitly learns logits as representations and employs logit adjustment guided by distribution information. This strategy effectively enhances minority-class representation, aligns client features, and improves the global model's generalization and performance. + +# 5. Conclusions + +We propose FedYoYo, a novel federated learning method addressing challenges posed by heterogeneous and long-tailed data distributions. FedYoYo integrates Augmented Self-bootstrap Distillation (ASD) and Distribution-aware Logit Adjustment (DLA). ASD employs weakly augmented samples as self-teachers to guide strongly augmented samples, enhancing local feature extraction under client data diversity. DLA leverages both local and global distributions to calibrate logits, providing effective guidance signals for representation learning. Extensive experiments on CIFAR-10-LT, CIFAR-100-LT, and ImageNet-LT demonstrate FedYoYo's state-of-the-art performance, surpassing even centralized baselines in global long-tailed scenarios. + +# 6. Acknowledgements + +This study was supported in part by the National Natural Science Foundation of China under Grants 62376233, 62431004, U21A20514, 62372388, and 62466036; in part by the Natural Science Foundation of Fujian Province under Grant 2024J09001; in part by the High-level and Urgently Needed Overseas Talent Programs of Jiangxi Province under Grant 20232BCJ25024; in part by the Zhejiang Provincial Key Research and Development Project under Grant 2023C01043 and Engineering Research Center of Integration and Application of Digital Learning Technology, Ministry of Education; and in part by Xiaomi Young Talents Program. + +# References + +[1] Mohammed Adnan, Shivam Kalra, Jesse C Cresswell, Graham W Taylor, and Hamid R Tizhoosh. Federated learning and differential privacy for medical image analysis. Scientific Reports, 12(1):1953, 2022. 1 +[2] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. 8 +[3] Xiaohua Chen, Yucan Zhou, Dayan Wu, Chule Yang, Bo Li, Qinghua Hu, and Weiping Wang. Area: Adaptive reweighting via effective area for long-tailed classification. In The IEEE/CVF International Conference on Computer Vision, pages 19220-19230, 2023. 4, 5, 7 +[4] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 4 +[5] Yutong Dai, Zeyuan Chen, Junnan Li, Shelby Heinecke, Lichao Sun, and Ran Xu. Tackling data heterogeneity in federated learning with class prototypes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7314-7322, 2023. 8 +[6] Cynthia Dwork, Aaron Roth, et al. The algorithmic foundations of differential privacy. Foundations and Trends® in Theoretical Computer Science, 9(3-4):211-407, 2014. 5 +[7] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 2, 3, 8 +[8] Wenke Huang, Yuxia Liu, Mang Ye, Jun Chen, and Bo Du. Federated learning with long-tailed data via representation unification and classifier rectification. IEEE Transactions on Information Forensics and Security, 2024. 5, 7 +[9] Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. Understanding dimensional collapse in contrastive self-supervised learning. In The International Conference on Learning Representations, 2022. 2 + +[10] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In The International Conference on Learning Representations, 2020. 5, 7 +[11] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 1, 2, 3, 8 +[12] Alexander C Li, Alexei A Efros, and Deepak Pathak. Understanding collapse in non-contrastive siamese representation learning. In European Conference on Computer Vision, pages 490–505. Springer, 2022. 2 +[13] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10713–10722, 2021. 8 +[14] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 1, 2, 3, 5, 6, 7, 8 +[15] Zexi Li, Tao Lin, Xinyi Shang, and Chao Wu. Revisiting weighted aggregation in federated learning with neural networks. In International Conference on Machine Learning, pages 19767-19788. PMLR, 2023. 1 +[16] Zexi Li, Xinyi Shang, Rui He, Tao Lin, and Chao Wu. No fear of classifier biases: Neural collapse inspired federated learning with synthetic and fixed classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5319-5329, 2023. 1, 2, 3, 5, 6, 7, 8 +[17] Xinting Liao, Weiming Liu, Chaochao Chen, Pengyang Zhou, Fengyuan Yu, Huabin Zhu, Binhui Yao, Tao Wang, Xiaolin Zheng, and Yanchao Tan. Rethinking the representation in federated unsupervised learning with non-iid data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22841-22850, 2024. 2 +[18] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in Neural Information Processing Systems, 33:2351-2363, 2020. 5, 6 +[19] Mi Luo, Fei Chen, Dapeng Hu, Yifan Zhang, Jian Liang, and Jiashi Feng. No fear of heterogeneity: Classifier calibration for federated learning with non-iid data. Advances in Neural Information Processing Systems, 34:5972-5984, 2021. 2, 5, 6, 7, 8 +[20] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1, 2, 6, 7 +[21] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In The International Conference on Learning Representations, 2021. 1, 7 + +[22] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 2 +[23] Xinyi Shang, Yang Lu, Yiu-ming Cheung, and Hanzi Wang. Fedic: Federated learning on non-iid and long-tailed data via calibrated distillation. In IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2022. 2, 5, 7, 8 +[24] Xinyi Shang, Yang Lu, Gang Huang, and Hanzi Wang. Federated learning on heterogeneous and long-tailed data via classifier re-training with federated features. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, pages 2218–2224, 2022. 2, 3, 5, 7, 8 +[25] Zebang Shen, Juan Cervino, Hamed Hassani, and Alejandro Ribeiro. An agnostic approach to federated learning with class imbalance. In The International Conference on Learning Representations, 2021. 8 +[26] Yujun Shi, Jian Liang, Wenqing Zhang, Chuhui Xue, Vincent YF Tan, and Song Bai. Understanding and mitigating dimensional collapse in federated learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5): 2936-2949, 2023. 8 +[27] Xian Shuai, Yulin Shen, Siyang Jiang, Zhihe Zhao, Zhenyu Yan, and Guoliang Xing. Balancefl: Addressing class imbalance in long-tail federated learning. In The ACM/IEEE International Conference on Information Processing in Sensor Networks, pages 271-284. IEEE, 2022. 5, 7 +[28] Ha Min Son, Moon-Hyun Kim, Tai-Myoung Chung, Chao Huang, and Xin Liu. Feduv: uniformity and variance for heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5863–5872, 2024. 2 +[29] Haozhao Wang, Yichen Li, Wenchao Xu, Ruixuan Li, Yufeng Zhan, and Zhigang Zeng. Dafkd: Domain-aware federated knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20412-20421, 2023. 5, 6 +[30] Chenrui Wu, Haishuai Wang, Xiang Zhang, Zhen Fang, and Jiajun Bu. Spatio-temporal heterogeneous federated learning for time series classification with multi-view orthogonal training. In Proceedings of the ACM International Conference on Multimedia, pages 2613-2622, 2024. 1 +[31] Zikai Xiao, Zihan Chen, Liyinglan Liu, Yang Feng, Jian Wu, Wanlu Liu, Joey Tianyi Zhou, Howard Hao Yang, and Zuozhu Liu. Fedloge: Joint local and generic federated learning under long-tailed data. In The International Conference on Learning Representations, 2024. 2, 5, 7, 8 +[32] Zikai Xiao, Zihan Chen, Songshang Liu, Hualiang Wang, Yang Feng, Jin Hao, Joey Tianyi Zhou, Jian Wu, Howard Yang, and Zuozhu Liu. Fed-grab: Federated long-tailed learning with self-adjusting gradient balancer. Advances in Neural Information Processing Systems, 36, 2024. 5, 7, 8 +[33] Rui Ye, Zhenyang Ni, Chenxin Xu, Jianyu Wang, Siheng Chen, and Yonina C Eldar. Fedfm: Anchor-based feature matching for data heterogeneity in federated learning. IEEE Transactions on Signal Processing, 71:4224-4239, 2023. 8 +[34] Xuefei Yin, Yanming Zhu, and Jiankun Hu. A comprehensive survey of privacy-preserving federated learning: A tax- + +onomy, review, and future directions. ACM Computing Surveys, 54(6):1-36, 2021. 1 +[35] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In International Conference on Machine Learning, pages 26311-26329. PMLR, 2022. 1, 5, 6, 7, 8 +[36] Jianqing Zhang, Yang Liu, Yang Hua, and Jian Cao. Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning. In Proceedings of the AAAI conference on Artificial Intelligence, pages 16768-16776, 2024. 8 +[37] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3713-3722, 2019. 3 +[38] Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10174-10183, 2022. 5, 6 +[39] Tuo Zhang, Lei Gao, Chaoyang He, Mi Zhang, Bhaskar Krishnamachari, and A Salman Avestimehr. Federated learning for the internet of things: Applications, challenges, and opportunities. IEEE Internet of Things Magazine, 5(1):24-29, 2022. 1 +[40] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10795-10816, 2023. 2 +[41] Tailin Zhou, Jun Zhang, and Danny HK Tsang. Fedfa: Federated learning with feature anchors to align features and classifiers for heterogeneous data. IEEE Transactions on Mobile Computing, 2023. 8 +[42] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In International Conference on Machine Learning, pages 12878-12889. PMLR, 2021. 5, 6 +[43] Weiming Zhuang, Yonggang Wen, and Shuai Zhang. Divergence-aware federated self-supervised learning. In The International Conference on Learning Representations, 2022. 8 \ No newline at end of file diff --git a/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/images.zip b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e4a55452b89f871e9fe4f128a63134f20b593fa6 --- /dev/null +++ b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1db1f22ed1c78357a3aa044dd47f8ba0eda321619b07edbdf5ecf675985b612 +size 618336 diff --git a/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/layout.json b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..528430a600fa40015a5c369c4a6de75a0dc6cc53 --- /dev/null +++ b/youareyourownbestteacherachievingcentralizedlevelperformanceinfederatedlearningunderheterogeneousandlongtaileddata/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efdb3731bf166d7e718b780d467584187a71a758de15b722fb079dffaf25a010 +size 460908 diff --git a/yourtextencodercanbeanobjectlevelwatermarkingcontroller/a9813057-9d7e-495c-bd8d-0a15137ba2fa_content_list.json b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/a9813057-9d7e-495c-bd8d-0a15137ba2fa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9cdce13baaebcdc26575c17f29283692e40e9c1b --- /dev/null +++ b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/a9813057-9d7e-495c-bd8d-0a15137ba2fa_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da74390336a6edb2105b9cc182715dbf550879a1691a73ce63449e36266448bd +size 81961 diff --git a/yourtextencodercanbeanobjectlevelwatermarkingcontroller/a9813057-9d7e-495c-bd8d-0a15137ba2fa_model.json b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/a9813057-9d7e-495c-bd8d-0a15137ba2fa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..124a205ccf3837d5b96e7307482bd38815685a81 --- /dev/null +++ b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/a9813057-9d7e-495c-bd8d-0a15137ba2fa_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b2607fee6cf1f51b2c261c9de471edad7e9f456300a81ecd6e2f45881e53d0f +size 98326 diff --git a/yourtextencodercanbeanobjectlevelwatermarkingcontroller/a9813057-9d7e-495c-bd8d-0a15137ba2fa_origin.pdf b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/a9813057-9d7e-495c-bd8d-0a15137ba2fa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ea138e3c83699067a2163e0b8e83f31a0b59be20 --- /dev/null +++ b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/a9813057-9d7e-495c-bd8d-0a15137ba2fa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b947f07d431293ba46d58043fe87acdbd922d5fcfbb21eb2e32e8074a39168c7 +size 10952394 diff --git a/yourtextencodercanbeanobjectlevelwatermarkingcontroller/full.md b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3eef91a652f76eef78af8d1d3ad5f652a59d0237 --- /dev/null +++ b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/full.md @@ -0,0 +1,325 @@ +# Your Text Encoder Can Be An Object-Level Watermarking Controller + +Naresh Kumar Devulapally $^{1\dagger}$ Mingzhen Huang $^{1}$ Vishal Asnani $^{2}$ Shruti Agarwal $^{2}$ Siwei Lyu $^{1}$ Vishnu Suresh Lokhande $^{1\dagger}$ ${}^{1}$ University at Buffalo, SUNY ${}^{2}$ Adobe Research + +{devulapa, mhuang33, siweilyu, vishnulo}@buffalo.edu {vasnani, shragarw}@adobe.com + +![](images/776885a642ae90076ce3c33855a4e5fd26d53496de0ac29601553755110a344e.jpg) + +![](images/7684af477d774eb5551ab60fc60967a0002d7a7e4b607f5766d5d5d7441e0265.jpg) +A $[\mathrm{dog}\mathcal{W}_{*}^{1}]$ and a cat $[\mathrm{cat}\mathcal{W}_{*}^{2}]$ in a park. + +![](images/8b92f0087e1d3365675bc7999307afec99172b4498a7fd6aeef2194fe5b06187.jpg) + +![](images/701e03e1529c270e54a4965cacb015eaf61e77105d8686a2279e952d6eb685c7.jpg) +Figure 1. Text prompt-controlled object-level watermarking: (Left Image): Our method embeds a watermarking token $\mathcal{W}_*$ into text-to-image generation, allowing users to watermark a full image or selected objects in an image. By leveraging cross-attention maps, any subset of prompt tokens $\{\mathcal{P}_i,\mathcal{P}_j,\ldots \}$ can be targeted while preserving non-watermarked regions. (Right Image): We also demonstrate personalized object watermarking using Styled UNet and Textual Inversion. + +# Abstract + +Invisible watermarking of AI-generated images can help with copyright protection, enabling detection and identification of AI-generated media. In this work, we present a novel approach to watermark images of T2I Latent Diffusion Models (LDMs). By only fine-tuning text token embeddings $\mathcal{W}_{*}$ , we enable watermarking in selected objects or parts of the image, offering greater flexibility compared to traditional full-image watermarking. Our method leverages the text encoder's compatibility across various LDMs, allowing plug-and-play integration for different LDMs. Moreover, introducing the watermark early in the encoding stage improves robustness to adversarial perturbations in later stages of the pipeline. Our approach achieves $99\%$ bit accuracy (48 bits) with a $10^{5} \times$ reduction in model parameters, enabling efficient watermarking. Code can be found at github.com/naresh-ab/object_watermark. + +# 1. Introduction + +As debates around copyright and intellectual property intensify, the need for effective watermarking solutions increases + +[33]. With generative AI models producing indistinguishable images from human-made art, maintaining authorship provenance in digital media is crucial. Efforts are underway, from legislative measures like the Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence [1], to industry initiatives that aim to watermark all AI-generated content [24] [30]. These developments highlight the growing importance of watermarking as a key area of study. + +As a result, multiple Generative AI (GenAI) watermarking methods have been proposed [7, 9, 33] to invisibly watermark the entire image while generating the image. However, lots of current watermark methods [3, 6] can only encode a watermark to an open-box generative model as they need to access the latent space. Moreover, similar to other invisible watermarking techniques [2, 6, 7], these methods also suffer with the inherent trade-off between quality and robustness of the embedded signal. Owing to the high imperceptibility requirements for the high-quality image generation, these watermarks more often lacks robustness [9, 35]. To improve the watermark imperceptibility, partial watermarking has been proposed to limit the watermark related changes only to selected "non-salient" objects or regions in the image [22]. In our paper, we bring such + +partial watermarking to text-to-image generation pipeline. Our partial watermarking is not only for improving the quality of the watermarked image, but also to give the user the flexibility to watermark only selected object of interest in the generated image without accessing the latent space. This is crucial for scenarios where the user would like to protect the unique object in the image, see Fig. 1. + +Specifically, we propose a token-based watermarking approach that can embed watermarks into selected objects or partial regions of an image. Unlike previous works that have explored architectural modifications, such as the addition of secret encoders [3], distortion layers [6], and adapters [5], our watermarking method focuses solely on learning a watermark embedding. This allows for a prét-à-porter style of training [28], leveraging pre-trained models with minimal adjustments. Additionally, compared with prior works [3, 5, 6], our model is blind, meaning that the entire generative model is neither modified nor requires additional training. A simple watermark token is learned by our model and can be plugged into any diffusion model for watermark generation. Previous works [7, 33] have investigated image watermarking without altering the architecture. Although effective, they lack the convenience of prét-à-porter training, which offers the key advantage of personalizing large models with minimal retraining while using the model's core components as-is. In contrast, our approach enables users to apply watermarking directly at the prompt stage without modifying core model components such as the UNet or image decoder. We introduce new pseudotokens $\mathcal{W}_*$ into the model's vocabulary, where the $\mathcal{W}_*$ can be used as an input text prompt to be applied on any T2I diffusion models for generation watermarked images. This also enables users to selectively watermark specific image regions by leveraging the cross-attention between the special token and visual region in the image where the watermark needs to be embedded. By integrating watermarking functionality early in the text-to-image pipeline, our approach performs in-generation watermarking, offering greater robustness against image manipulation attacks compared to post-processing techniques while improving the overall quality of the watermarked images. + +# We make the following key contributions: + +1. We introduce a novel pseudo-token with watermarking capabilities, learning a new embedding vector for it. We investigate the impact of applying this pseudo-token conditioning at different timesteps of the diffusion process, finding that timesteps closer to the VAE encoder in LDMs during forward process offer enhanced image generation quality. +2. We introduce object-level watermarking, allowing for selective watermarking with greater precision than traditional whole-image approaches. +3. We offer plug-and-play integration with various Stable + +Diffusion (SD) variants by embedding the watermark directly via the text encoder. Our approach achieves $99\%$ bit accuracy with a $10^{5} \times$ reduction in model parameters, enabling efficient watermarking with a throughput of 48 bits. + +# 2. Related Works + +Text-Guided Image Generation Diffusion Models: Diffusion models have revolutionized text-to-image (T2I) generation, surpassing GANs in fidelity and diversity. By iteratively denoising random noise into images conditioned on text prompts, these models enable fine-grained control over generation. Frameworks like DALL-E 2, Imagen [13], and Stable Diffusion [31] democratize high-quality image synthesis, while diffusion-based editing methods [12] enable localized modifications. However, their widespread adoption raises challenges, including copyright infringement and harmful content generation, necessitating controllable synthesis mechanisms such as watermarking. + +Textual Inversion for Personalized Image Generation: Textual Inversion [8] embeds visual or stylistic concepts into T2I models through learnable tokens, enabling personalization without altering model parameters. This lightweight approach has been extended to style transfer [34]. In this work, we find that training the special token not only allows for flexible concept learning but also could acts as a watermark controller. + +Constraint-Based Control in Text-to-Image Generation: Text-to-image generation has gained significant popularity, especially to incorporate constraints that enhance control over generated content. Common constraints include spatial or layout constraints, which dictate object placement and dimensions within the generated image to meet predefined spatial requirements [17]. Another type, multi-view consistency constraints, ensures scene consistency across perspectives, preserving layout, lighting, and depth even in complex views like outdoor or stylized scenes [36]. Attribute preservation constraints maintain prompt-specified attributes (e.g., color, texture, size) in the generated image, ensuring semantic alignment. Finally, watermarking constraints embed an invisible watermark to verify ownership while preserving image quality. Watermarks can be embedded through polytope constraints defined by orthogonal Gaussian vectors or high-frequency components, allowing watermarking without image degradation. Constraint-based generation methods, which offer explicit control over outputs, draw inspiration from classifier training techniques that optimize using soft penalties [4]. Recent approaches such as [11] have achieved notable success with convergence guarantees and adaptability across data regimes. + +In-generation Image watermarking in T2I generation: With the increasing popularity of diffusion models, recent research has explored embedding watermarks within the + +diffusion process, either by manipulating the noise schedule to encode watermark information or by conditioning the model to output watermarked images. Tree-ring [33] proposed to encode a watermark into the noise space, it can be simply detected by applying a DDIM [31] inversion to obtain the noise. However, current in-generation watermark methods [6, 7, 26] cannot inject a watermark in a local region, e.g. an object. As many diffusion-based image editing method have been emerging for a long time, the object-level watermarking is long overdue. Recently, WAM [29] discusses localized watermarking and proposes a training mechanism for watermarked region detection from watermarked image. However, WAM still requires segmentation masks to localize watermark during watermark embedding in both training and inference. + +Blind v/s Non-Blind Watermarking: Mentioned in [16], blind detection methods verify ownership using only the watermarked image, eliminating reliance on auxiliary metadata. Our work adopts this practical approach, ensuring compatibility with standard detection workflows. + +Motivated by these insights, we explore similar methods in text-to-image generation, where constraints are often enforced as soft penalties via gradient-based optimization. Current diffusion-based watermarking methods lack spatial control, limiting object-centric applications. Recent advances in localized diffusion editing [12] remain unexplored for watermark integration. By unifying textual inversion with constraint-based control, we propose the first framework for object-level in-generation watermarking in T2I generation, addressing traceability and granularity. + +# 3. Problem Statement + +Our problem setting uses Latent Diffusion Models (LDMs) [27] for image generation. We consider an in-generation watermarking scenario, as defined in Sec. 2, where the aim is to utilize existing modules within the LDM pipeline to watermark images while generation. + +Watermark Embedder: Given an input text prompt $(\mathcal{P} = \{p_0, p_1, \ldots, p_n\})$ and/or an image $\mathcal{I}$ , the aim of an in-generation watermark embedder $(W_e)$ is to generate a watermarked image $\mathcal{I}_w = LDM(\mathcal{I} \mid W_e, \mathcal{P})$ . Each embedder $W_e$ is associated with a watermark key $(m)$ (a bit string containing 0, 1 similar with prior works [6, 7]) of length $k$ , i.e., $m \in \{0, 1\}^k$ . From here we shall use $W_{e,m}$ to denote a watermark embedder and $\mathcal{I}_{w,m}$ to denote a watermarked image. + +Watermark Detector: Given a watermarked image $\mathcal{I}_{w,m}$ , watermark detector $D_w(\cdot)$ is a neural network that predicts the key $m$ from $\mathcal{I}_{w,m}$ . Our method utilizes Blind Watermarking scenario mentioned in Sec. 2 where $D_w(\cdot)$ only requires the watermarked image $\mathcal{I}_{w,m}$ as input without the need for any additional metadata. + +Attack Module: Attack module $\mathcal{A}(\cdot)$ represents unin + +tentional (or) intentional transforms applied to the watermarked image $\mathcal{I}_{w,m}$ that could result in the loss of watermark key during watermark detection. A successful attack module results in imperfect watermark detection, i.e., $D_w(\mathcal{A}(\mathcal{I}_{w,m}))\neq m$ . This module can usually be seen in the form of image compression by public platforms or an intentional malicious attacker that aims to remove watermark from $\mathcal{I}_{w,m}$ . Resistance to an attack module is a fundamental requirement of a robust watermark embedder $W_{e}$ + +Full-Image and Object-Level Watermarking: Given a text prompt for generation, full-image watermarking refers watermarking entire image and object-level watermarking refers to the scenario that watermarks specific objects $O_{i}, \ldots O_{j}$ in the image selected by a user. Information about $O_{i}$ can be provided in various ways. For example [7] uses post-processing techniques to localize watermark on various regions using masks. However, using such post-processing methods introduces additional computational overhead and may be susceptible to removal or modification. Instead, integrating object-level watermarking into the generative process enables seamless and robust embedding within the chosen object $O_{i}, \ldots O_{j}$ regions. + +Our problem setting considers both image and object-level watermarking scenarios with a constraint that the information for watermarking a specific object $O_{i}$ shall not be provided in the form of any additional information such as segmentation masks. This information can only be obtained right from the text prompt. Specifically, for our setting, a text prompt "[A photo of a cat $\mathcal{W}_{*}]$ " denotes full-image watermarking and a text prompt "A photo of a [cat $\mathcal{W}_{*}]$ " denotes that the object cat is to be watermarked. + +# 4. Method + +Given an image, $\mathcal{I}$ , our training pipeline aims to generate $\mathcal{I}_{w,m}$ where $m$ is the watermark key, $m\in \{0,1\} ^k$ . We utilize a differentiable watermark detector from [6] $D_w(I_{w,m})$ that permits gradient flow. + +It is known that training larger modules of an LDM pipeline, such as VAE Encoder/Decoder or the UNet, is computationally expensive and do not provide a lightweight, seamless way to integrate watermarking into various other LDM pipelines. There is a need for a unified watermark embedder that can be integrated with various LDM variants. + +# 4.1. Token Embeddings for Watermarking + +Our method utilizes the relatively under-explored Text Encoder of the LDM pipeline [19]. We introduce a new token in the Text Encoder, denoted by $\mathcal{W}_{*}$ , and fine-tune the text embeddings of $\mathcal{W}_{*}$ to watermark $\mathcal{I}$ . In addition to significant lower parameter requirement compared with prior works [6, 7], this token $\mathcal{W}_{*}$ act as the watermark trigger + +![](images/13766bb9eac10dab3a01bc13bb18987c8db18d6829b9babe5a1ac5f582a64bc1.jpg) +Figure 2. $\mathcal{W}_{*}$ training pipeline. (Left) To find $\mathcal{W}_{*}$ token embeddings, we use an Img2Img generation pipeline. $\mathcal{D}$ and $\mathcal{D}_w$ represent VAE decoder in the LDM, and Watermark Detector respectively. While training, we send the input image through LDM encoder to retrieve the latent $z_0$ , we then add a forward diffusion noise of $\tau^{*}$ timesteps, followed by iteratively denoising $z_{\tau^{*}}$ using Classifier-Free Guidance [11] from $[\tau^{*} \to 0]$ to retrieve $z_{0,w}^{\prime}$ . During the denoising process, we train for $\mathcal{W}_{*}$ token embeddings. (Middle) We use latent matching loss to control the trajectory of watermarked latents and bit loss to find $\mathcal{W}_{*}$ token embeddings. (Right) We then use trained $\mathcal{W}_{*}$ embeddings to generate watermarked images. + +that can be seamlessly integrated into the text encoder of any LDM pipeline. + +Our method initially aims to train $\mathcal{W}_{*}$ token embeddings, similar with Textual Inversion [8], to generate watermarked latents $z_{t,w} = \epsilon_{\theta}(z_t,t,\{\mathcal{P},\mathcal{W}_*\})$ . However, we identify the need to carefully select the optimal noise timesteps during $\mathcal{W}^*$ training as different choices impact image quality, watermark robustness and image generation time. + +# 4.2. Optimal Timestep and Latent Loss + +We perform empirical studies to observe the effect of different noise timesteps during forward diffusion process while $\mathcal{W}_{*}$ training. As depicted in Fig. 2 (middle) (and in the section H. of the supplement), we observe a trade-off between image quality and watermarking performance when choosing different noise timesteps. A relatively large timestep $(t\sim T)$ would degrade the image quality while a relatively small timestep $(t\sim 0)$ would lead a lower bit accuracy but enhanced image generation quality. This observation is consistent with prior findings, as mentioned in [15, 28], which highlight the impact of noise strength on conditioning fidelity. We identify an optimal timestep $\tau^{*} = 8$ that balances this trade-off between image quality and watermark bit accuracy. During $\mathcal{W}_{*}$ training, we add a forward noise of $\tau^{*}$ to the image followed by iterative de-noising to generate watermarked latent $z_{0,w}^{\prime}$ . $z_{0,w}^{\prime}$ is then passed into the VAE decoder $Dec(\cdot)$ to generate a watermarked image $I_{w,m}$ . $D_w$ takes $I_{w,m}$ as input, $D_w(I_{w,m}) = m'$ , to train $\mathcal{W}_{*}$ token embeddings on watermark loss $\mathcal{L}_w$ . We utilize $BCE$ loss to embed a specific bit key $m$ using $\mathcal{W}_{*}$ . + +$$ +\mathcal {L} _ {w} = B C E \left(D _ {w} \left(D e c \left(z _ {0, w} ^ {\prime}\right), m\right). \right. \tag {1} +$$ + +While the optimal timestep $\tau^{*}$ preserves overall information in $\mathcal{I}$ and significantly reduces the time taken for $I_{w,m}$ generation, we still observe visible corruption in $I_{w,m}$ that hurts imperceptibility watermarking constraint. As a remedy, we employ latent matching loss $\mathcal{L}_z$ that aids our method to perform invisible watermarking. + +$$ +\mathcal {L} _ {z} = \min _ {\mathcal {W} _ {*}} \mathbb {E} _ {t} \left[ \left\| z _ {t} ^ {*} - z _ {t} ^ {\prime} (\mathcal {W} _ {*}) \right\| _ {2} ^ {2} \right]. \tag {2} +$$ + +Our method uses $\mathcal{L} = \alpha \mathcal{L}_w + \beta \mathcal{L}_z$ to train $\mathcal{W}_{*}$ token embeddings for watermarking where $\alpha$ and $\beta$ are tunable hyperparameters. + +# 4.3. Object-level watermarking + +In the original LDM [27], the query feature $Q$ and key feature $K$ from cross attention operation defines a token-wise attention mask $M$ where $M = \text{softmax}(Q \cdot K^T) / \sqrt{d}$ . As the architecture of the LDM remains unchanged in our model, we inherit this capability to generate corresponding attention mask from text tokens. Such feature is the key in our model unlocking new possibilities for object-level watermarking utilizing these cross-attention maps. + +Once the watermarking token embeddings are fine-tuned, we proceed to generate images using a standard text-to-image LDM model, but with an ability to perform object-level watermarking. We utilize cross-attention maps at each timestep $t$ to localize watermark on any chosen object $O_{i}$ right from the text prompt $\mathcal{P}$ as a token $\mathcal{P}_i(\in \mathcal{P})$ . Specifically, we utilize [10] to obtain cross-attention map $M_{t}$ of $O_{i}$ given by: + +$$ +z _ {t - 1} ^ {\prime}, M _ {t, O _ {i}} \leftarrow \operatorname {L D M} \left(z _ {t} ^ {\prime}, \mathcal {P} _ {i}, t\right) \tag {3} +$$ + +Algorithm 1 Object-level Watermarking during Text-to-Image Generation in Latent Diffusion Models + +Input: LDM Decoder $\mathcal{D}$ ; Watermark Token $\mathcal{W}_*^i$ for object $i$ ; Text Prompt $\mathcal{P}$ ; overlay strength $\pi(t)$ . + +1: Select objects to watermark, defining the subset $\{\mathcal{P}_0^*,\mathcal{P}_1^*,\ldots ,\mathcal{P}_k^*\}$ with corresponding watermarking tokens $\{\mathcal{W}_0^*,\mathcal{W}_1^*,\dots ,\mathcal{W}_k^*\}$ from the text prompt $\mathcal{P} = \{\mathcal{P}_0,\mathcal{P}_1,\dots ,\mathcal{P}_n\}$ . + +2: $z_{T}\in \mathcal{N}(0,I)$ + +3: for $t = T, \dots, 0$ do + +4: for each token $\mathcal{P}_i^* \in \{\mathcal{P}_0^*, \mathcal{P}_1^*, \dots, \mathcal{P}_k^*\}$ do + +5: $\mathcal{M}_{\mathcal{P}_i^*}^{(t)} / \mathcal{M}_{\mathcal{W}_i^*}^{(t)}\gets$ Attention map of $\mathcal{P}_i^* /\mathcal{W}_i^*$ + +6: $\mathcal{M}_{\mathcal{W}_*}^{(t)} \gets \pi(t) \cdot \mathcal{M}_{\mathcal{W}_*}^{(t)}$ (Adjust watermark-strength per timestep) + +7: $\mathcal{M}_{\mathcal{P}_i}^{(t)}\gets (1 - \alpha)\cdot \mathcal{M}_{\mathcal{P}_i}^{(t)} + \alpha \cdot \mathcal{M}_{\mathcal{W}_*}^{(t)}$ (Overlay attention map of $\mathcal{W}_*^i$ on that of $\mathcal{P}_*^i$ ) + +8: end for + +9: $\hat{\epsilon} = \epsilon_{\theta}(z_t,t,\psi (\mathcal{P}),\mathcal{M}_{\mathcal{P}^{0:k}}^{(t)},\mathcal{M}_{\mathcal{W}^{0:k}}^{(t)})$ + +10: $z_{t - 1} = \sqrt{\frac{\alpha_{t - 1}}{\alpha_t}} z_t + \left(\sqrt{\frac{1}{\alpha_{t - 1}} - 1} -\sqrt{\frac{1}{\alpha_t} - 1}\right)\hat{\epsilon}$ + +11: end for +12: Output: Decoder output of $z_0$ , that is $\mathcal{D}(z_0)$ + +Given a text prompt with multiple tokens $\mathcal{P} = \{\mathcal{P}_0,\mathcal{P}_1,\dots ,\mathcal{P}_n\}$ , the user has the flexibility to choose specific object tokens $\{\mathcal{P}_i,\mathcal{P}_j,\ldots \}$ to watermark. During the image generation process, we extract the cross-attention maps for the objects intended for watermarking from the UNet. These are then combined with the attention maps of the corresponding watermarking tokens $\mathcal{W}_{*}^{i},\mathcal{W}_{*}^{j},\ldots$ to precisely localize the watermark on each object allowing us to seamlessly perform multi-object watermarking. + +To bring in the effect of optimal timestep $\tau^{*}$ (seen during training Sec. 4.2), we introduce a watermark overlay strength controller $\pi(t)$ . $\pi(t)$ could be set to a step function with values $0$ (if $t > \tau^{*}$ ) and $1$ (if $t \leq \tau^{*}$ ) which exactly mimics the generation scenario seen during training. In addition to the step function, we also test with a smoothing function that brings the effect of $\tau^{*}$ by giving more weight to timesteps closer to the VAE decoder. At these timesteps we observe increased reliability of the attention maps enhancing the precision of the watermark placement. The complete procedure is outlined in Algorithm 1. + +# 5. Experiments + +Datasets: We evaluate our method on two datasets, namely, MS-COCO [18] and WikiArt [32]. We use a subset of 2,000 images from MS-COCO dataset, similar to [7], for training $\mathcal{W}_{*}$ token embeddings. Our method is evaluated on 1000 validation captions from MS-COCO and WikiArt each via image-to-image generation and on 100 prompts from [6] for text-to-image generation. + +Evaluation Metrics: In line with prior work [6, 7], we assess our watermarking method using the following metrics: + +(a) Robustness, measured by Bit Accuracy under various attacks (mentioned in Tab. 1). Bit accuracy represents the percentage of correctly detected bits in the watermark key $m$ output from the Detector $D_{w}(\cdot)$ . In addition to Bit Accuracy, we also compare our method with techniques [33] that use True Positive Rate as the metrics in supplement (section D). These methods do not embed a bit string $m$ . +(b) Imperceptibility, measured by Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [19, 23], and Fréchet inception distance (FID) [14], comparing watermarked and non-watermarked images. +(c) Parameter Efficiency, where our method trains only the textual embeddings, with parameter usage significantly reduced $(10^{5} \times)$ compared to baselines [6, 7], as the T2I pipeline and watermark decoder remain frozen. + +Watermark Detector: For our experiments we utilize two watermark detectors from [6] and [7]. + +# 5.1. Full-Image Watermarking results + +We evaluate full-image watermarking by integrating $\mathcal{W}_{*}$ into the Stable-Diffusion v1.5 [31], using the AquaLora watermark detector [6]. Our method's performance is compared to existing watermarking baselines in LDMs, focusing on robustness and imperceptibility. Tab. 1 presents the results, showing that our approach requires significantly fewer parameters while maintaining high bit accuracy under attacks. We categorize attacks encountered after watermarking into two categories, namely, basic image processing attacks such as Rotation, Resize, Crop, JPEG compression etc as listed in Tab. 1, and adversarial attacks such as DiffPure [21], WMAntacker [35] etc. We provide implementation details of attack module in supplement (section I). + +We consider that robustness to adversarial attacks including Diff-Pure [21] and SDEdit [20], while beneficial, is not essential for watermarking techniques. These attack methods are complex and go beyond common attacks like crop, resize, and rotations. However, our watermarking method integrated into the denoising process demonstrates enhanced robustness to both basic and adversarial attacks outperforming several in-generation watermarking techniques thereby pushing the benchmark for robust watermarking. When compared to AquaLORA [6], while using the same watermark detector, we see an increase in bit accuracy on both common attacks by over $20\%$ , and over 7dB in PSNR. Additionally, our method retains high robustness and personalization when used alongside a personalized textual inversion token and/or a personalized UNet. + +Summary. Our Full-image watermarking achieves improved robustness, lower parameter requirements, and enhanced imperceptibility compared to baselines. + +
MethodI.W.O.W.L.P.# Params ↓ImperceptibilityRobustness to basic attacks (BA):Adversarial attacks (BA):
PSNR ↑SSIM ↑FID ↓None ↑Brightness ↑Contrast ↑Blur ↑Crop ↑Rot. ↑JPEG ↑SDEdit ↑WMAttacker ↑DiffPure ↑
WikiArt (48 bits) - In-Generation Watermarking
Stable Sig. [7]✓XX105+31.570.8824.710.990.930.870.780.790.700.550.580.530.52
LaWa [26]✓XX105+32.520.9318.230.990.990.980.940.930.830.890.680.760.78
TrustMark [2]XXX105+39.900.9715.830.990.980.990.930.890.870.860.680.770.73
RoSteALS [3]✓XX105+32.680.8816.630.980.960.940.880.880.750.800.750.720.72
AquaLoRA [6]✓XX105+31.460.9217.270.940.910.910.810.900.580.760.680.670.66
WAM [29]✓✓X105+36.460.9716.270.970.930.920.840.920.760.840.720.710.72
Ours + SDstyle✓✓✓76835.880.9316.720.990.970.990.940.970.940.930.810.830.84
Ours + TI✓✓✓76836.890.9415.980.990.960.990.950.980.950.920.820.840.86
Ours + SD✓✓✓76839.920.9714.890.990.980.990.970.980.950.950.850.870.88
MS-COCO (48 bits) - In-Generation Watermarking
Stable Sig. [7]✓XX105+31.930.8824.740.990.940.890.810.820.720.590.620.630.57
LaWa [26]✓XX105+33.530.9517.450.990.990.980.940.930.860.900.710.790.79
TrustMark [2]XXX105+40.980.9814.890.990.980.990.950.910.890.880.690.780.74
RoSteALS [3]✓XX105+32.680.8816.630.990.980.930.890.870.780.810.760.720.75
AquaLoRA [6]✓XX105+31.460.9217.270.950.910.900.800.920.680.780.670.700.68
WAM [29]✓✓X105+36.980.9716.850.980.940.920.860.940.770.860.730.730.72
Ours + SDstyle✓✓✓76836.990.9316.720.990.980.990.940.970.940.930.810.830.84
Ours + TI✓✓✓76836.890.9415.230.990.960.990.950.980.960.940.830.840.86
Ours + SD✓✓✓76840.920.9714.830.990.980.990.970.980.970.960.860.880.89
+ +# 5.2. Key results on Object Watermarking + +Object Watermarking and Identification: We present the effectiveness of object-level watermarking, as illustrated in Fig. 3, where we apply watermarks to up to three distinct objects within a single image, mentioned in Algorithm 1. Using the watermark detector $D_w$ , and without the need for any additional information such as segmentation masks, heatmaps are generated by evaluating the bit accuracy of small patches, each covering approximately $10\%$ of the image area. Each patch's bit accuracy score is determined by $D_w$ , and these scores are aggregated across all patches to produce a comprehensive heatmap. In this map, a score of 1 represents a bit accuracy of $100\%$ , indicating high fidelity in watermark retrieval, while a score of 0.5 indicates retrieved watermark key does not match $m$ indicating no watermark. When watermarking a single object, the resulting heatmap shows precise retrieval of the watermark. For images with two or three watermarked objects, the detector reliably identifies each watermark, and accurate retrieval is observed. In cases involving four or more objects (bottom-right of Fig. 3), the heatmap confirms that the watermark is detected across most of the image, likely due to the accuracy of attention maps during inference. For images with multiple objects, separate heatmaps can be generated for each object to enhance clarity in visualization. The overall heatmap presented in Fig. 3 is an aggregate of individual heatmaps detailed in Fig. 6. Further technical details on the heatmap generation process are available in the supp. (§B). + +Table 1. Comparison to watermark baselines. (I.W.: In-generation Watermarking, O.W.: Object-level Watermarking, L.P.: Less than $10^{5}$ parameters, BA: Bit Accuracy) We compare our method several baselines. In addition to watermark invisibility and robustness of watermarking, we present the number of parameters used for training. We see that our method uses $10^{5} \times$ lower parameters. We see that early integration for watermarking improves robustness to attacks, we see a consistent trend of this improvements in basic image processing attacks and adversarial attacks. We also present the performance of our method in the presence of personalization using fine-tuned UNet and Textual Inversion. From the above table we see that our method can be plugged into various LDM pipelines. More details on specific implementation of attacks can be found in the supplement. + +
Bit Accuracy
NoneBright.Cont.BlurRot.JPEG
Single object
Segment + white bg0.990.970.960.970.960.97
Segment + style bg0.950.920.900.960.940.93
Crop object (0.8 × size)0.960.940.920.950.920.93
Crop object (0.5 × size)0.920.910.900.910.920.90
Crop object (0.4 × size)0.900.890.880.890.890.90
Multiple objects
1 object0.990.970.960.970.950.98
2 objects (no overlap)0.940.930.950.950.940.96
3 objects (no overlap)0.900.890.900.900.900.99
2 objects (overlap ≥ 40%)0.790.760.700.800.740.74
+ +Table 2. Object-level watermarking robustness performance with attacks. We evaluate the performance of our method to perform object-level watermarking in the presence of attacks. We clearly see robust performance under rotation and various crops within watermarked objects, for single and up to three objects. + +Stress Tests: Object Size and Multiple Object Watermarking: We evaluate the sensitivity of the watermark detection to small object sizes and multiple objects. As shown in Tab. 2, even with a $40\%$ crop of the object, the detector achieves $89\%$ bit accuracy. This accuracy remains robust to transformations. We also test watermark detection across multiple objects, with and without overlap. For non-overlapping objects, detection remains above $90\%$ without attacks and above $89\%$ under attacks. However, performance decreases when the overlap exceeds $40\%$ , particularly in the case of multiple overlapping objects. Our qualitative results indicate that single-object watermarking is not + +![](images/df2208ecc6133eb0fa51e5fd1418ce1b51f0d941330ec926dfece6378aa181e4.jpg) +A [cat $\mathcal{W}_{*}]$ playing with a ball in a park. + +![](images/5965ecd817567d8de28808949ddfb399ca103c26461142a13b8a7ef468aadd09.jpg) +[Two parrots $\mathcal{W}_*$ ] sitting on a stick, city street background. + +![](images/f1917a4ca55cbd2707c9265aab9170f41fd9c23297c49028fe2bfed58d0b59c6.jpg) +A toy figure of [naruto $\mathcal{W}_{*}]$ standing on a table. + +![](images/e36e8d03c8c49fb2a73f567eecfc33c6207de9efcfb2c6b1c326e81d3cfac82d.jpg) +A portrait of a [cyborg $\mathcal{W}_*$ ] in a golden suit. + +![](images/2ffc78d3c2ef0845e27fb7b4fc2ac0c79e36b5fd2032243c17c720caccbc64c0.jpg) + +![](images/2564c226bce39e403148d50a29ad17a9a63a6c88e86b356feb536e0f3a0ad7b5.jpg) +A [motorcyclist $\mathcal{W}_{*}]$ riding down the street with a [dog $\mathcal{W}_{*}]$ in a basket. + +![](images/1b37db6622081f670089ebc9ae7b61325011beeedb3e8ac5f569825a94f4e4ae.jpg) +A [mother duck $\mathcal{W}_*$ ] and [her babies $\mathcal{W}_*$ ] are standing near the water. + +![](images/832ef7d004332d5820449dcd988bd7267e88cbbe7f180ec2e6a4281a5fe28554.jpg) +A [cat $\mathcal{W}_*$ ], a [dog $\mathcal{W}_*$ ] and a [horse $\mathcal{W}_*$ ] in a park. + +![](images/8eb427660eaa13d797ae35af96d1dd98ec666a5f3eaa1759921af030abf9b5ab.jpg) +[A trio of dogs $\mathcal{W}_{*}]$ sitting in a red convertible. +Figure 3. Qualitative results and watermark heatmaps. We show qualitative results of our watermarking approach for up to three objects in a single image, all embedded within the T2I generation pipeline. As described in Sec. 4.3, object-level watermarking is controlled directly via the text prompt $\mathcal{P}$ . (Top row) shows single-object watermarking with corresponding heatmaps, achieving high bit accuracy within the selected object and 0 outside. (Second row, third row) show results for two and more than three objects, respectively, with accurate retrieval and high bit accuracy. (Bottom row, last column) illustrates a case where imperfect attention maps cause watermarking to leak outside objects, yet high bit accuracy is still achieved. + +affected by object overlap, but bit accuracy decreases with significant overlap due to multiple layers of interference. + +# 5.3. Applications: Integration with LDM Pipelines and Textual Inversion Compatibility + +We demonstrate the versatility of our method by integrating the plug-in watermark token $\mathcal{W}_{*}$ across diverse text-to-image (T2I) generation pipelines, assessing both watermark robustness and imperceptibility. The schematic in Fig. 2 illustrates this integration process, where $\mathcal{W}_{*}$ is loaded into the text encoder of different LDMs, including pipelines that employ personalized or fine-tuned diffusion models. Additionally, $\mathcal{W}_{*}$ can be combined with Textual Inversion tokens, as shown in qualitative results in Fig. 3. Our method achieves a high PSNR of $35\mathrm{dB}$ and bit accuracy above $92\%$ , outperforming the Stable Signature [7] in robustness while maintaining watermark invisibility and resistance to attacks. + +# 5.4. Ablation Studies + +Training with SAM Segmentation Masks: Our watermarking pipeline relies on attention maps generated by prompts, though these maps can sometimes lack precision + +(e.g., spillover beyond the intended object, seen in bottom-right scenario of Fig. 3). This can pose challenges for applications like medical imaging, where precise watermark localization is crucial. In case of a need for such high precision localization of watermark target, our method is flexible to utilize segmentations masks within watermark generation and need not rely on post-generation localization. In this ablation study, we incorporate SAM (Segment-Anything) segmentation masks directly into our generation process as an alternative to attention maps, particularly in Step 4 of Algorithm 1. Our results (Fig. 5) show that SAM masks significantly improve watermark localization accuracy. + +Additional analysis on more complex scenes with multiple overlapping objects: We clarify that our setting is indeed very challenging: We retrieve multiple watermarks from overlapping regions using the same detector that is natively pretrained to retrieve single watermark. This setting is known to cause a major drop in robustness (Fig. 4 of [25]) and remains an open research challenge on how to extract more than one watermark from an image region. To further support this, we provide a quantitative analysis. + +No drop is expected if we are permitted to use more than + +![](images/5b293bc94a996a6ba8a61bcb0ecf11bf449abc263540864d7eb134c596a81a72.jpg) + +![](images/2eaa61c3ccc499881e2721af7852e38e2b5ea89e7d9dea4e44e5b64de48b2b0b.jpg) + +![](images/26d8975f4431a0b32219eee3510f29361314ea04ca15638de1b29c67f825470c.jpg) +Figure 4. Plug-and-Play ability. We present our method's ability to be plugged into any combination of personalized T2I model. Above image shows four such combinations where we use Textual Inversion style tokens and styled T2I pipelines can be seen. It can be observed that the object-level control and watermarking ability of our method is preserved across all these pipelines. + +![](images/ac5ffd0cbc2fa67974354a73bae2ab03aeed058bf19400341890b896c03807f7.jpg) +Figure 5. Watermark heatmap enhancement using SAM. Our watermarking is embedded into the generation pipeline using Attention maps. We test the performance of our watermark detection by plugging our method into P2P and SAM. + +one watermark detector. Even with the same detector 36 bits can be retrieved with 0.85 Acc. with $\geq 40\%$ overlap. + +S.K.: Same Key, D.K.: Different Keys, S.D.: Same Detector, D.D.: Different Detector, O: $\geq 40\%$ Overlap. + +
Retrieval Accuracy
S.K. + O + S.D.0.99
D.K. + O + D.D.0.99
D.K. + O + S.D.0.79
+ +![](images/2b1a657874391b4a12837b082293980acb06cd182d008b832d82454284b22edd.jpg) +Left: Retrieval accuracy under different configurations. Right: Performance on the $(D.K. + O + S.D.)$ setting under varying bit counts. + +Ablation on Number of bits for Watermarking. We perform an ablation on number of bits that can be embedded into image by our method and present our results as a plot in the supplement (section C.). We observe that our method can watermark with $>91\%$ bit accuracy for 128 bits under various attacks. + +# 6. Limitations + +As our method performs in-generation watermarking. For localizing the watermark on an object, the method relies on Cross Attention maps for each token in the input prompt $\mathcal{P}$ . Hence, the method relies strongly on the accuracy of these cross attention maps. Our ablation studies conducted on using the segmentation masks from SAM enhance the localization of watermark on top of the object. However, relying on an external segmentation model such as SAM + +![](images/4889a80aa66844ea6f96bb616d220203b4f44309f45304d9f1ce6d0156211071.jpg) +Watermarked Image + +![](images/2273996886b736816058eac045c3b1c42b0724f0f8937ca2501abf6285bd6a76.jpg) +Plug-In $\mathcal{W}_{*}$ with P2P + +![](images/1d70a3303b31ddc08ccf582fa794e2e350a8cde7840530a52e1781ddf644634a.jpg) +Plug-In $\mathcal{W}_{*}$ with SAM + +![](images/91f7aa6334cddeb1d3e5e12e213b498be4259a215eb205c05d95b15afedcddd2.jpg) + +![](images/b7d15087b171f16384f8be06fa566c4a6eeb125a9aae9f9446fa23d1a07d1faa.jpg) + +![](images/7ce4df87e042aba590cb5234132388283b4a7f6da277bc13a16f43320fdcf04d.jpg) + +![](images/ecba487f3c5bf7b2a1278e18328f3f229dccb8ae5931470c915d9fb84fbe7808.jpg) + +![](images/9899b70e267623ae468fc575e14ac62f9fe532dff0900639256e62eb3c358254.jpg) + +![](images/65c5b455932c4660b86fecc4e5bdfa303d12bbb95781828212f961c6baf70ed1.jpg) +Figure 6. Multiple Object watermarking heatmap breakdown. Our goal is to embed watermark in specific regions of the image while not corrupting the entire image. In the above example, a user can choose to watermark either the dog or the car or both. Our method provides the versatility to choose to watermark any object(s) in an image while preserving other regions. + +![](images/3c7321415d2d94142f3b8db11b951a36c3104e8fd1d96852635bd4cff6dba0ec.jpg) + +![](images/1089c9c0c057d7a3271b51552039623c98240ef918da019e669918205bbeed32.jpg) + +![](images/5b75ed3875434a5c782984c1fa61818e3e1d21b782a292cf0cf0955b20d4ae4f.jpg) + +could be undesirable and overly relying on Cross Attention maps could be a limitation when the attention maps are not properly defined. + +# 7. Conclusions + +In this paper, we propose a novel in-generation watermarking technique to integrate watermarking into the latent within the denoising process of T2I generation. Our watermarking technique provides watermarking control directly from text and fine-tunes token embeddings of a single token. Our method contributes to a novel application of object-level watermarking within T2I generation. We show that our early watermarking technique shows improvements in watermarking robustness across several post generation attacks. Our method aims to motivate future research towards training-free watermarking, controllable watermarking with any T2I generation pipeline. + +Acknowledgments Prof. Lokhande thanks support provided by University at Buffalo Startup funds, Adobe Research Gift and internal funding from the University at Buffalo's Research and Economic Development office. + +# References + +[1] Joseph R Biden. Executive order on the safe, secure, and trustworthy development and use of artificial intelligence. 2023. 1 +[2] Tu Bui, Shruti Agarwal, and John Collomosse. Trustmark: Universal watermarking for arbitrary resolution images, 2023. 1, 6 +[3] Tu Bui, Shruti Agarwal, Ning Yu, and John Collomosse. Rosteals: Robust steganography using autoencoder latent space, 2023. 1, 2, 6 +[4] Giannis Daras, Maurizio Delbracio, Hossein Talebi, Alexandros G. Dimakis, and Peyman Milanfar. Soft diffusion: Score matching for general corruptions, 2022. 2 +[5] Weitao Feng, Jiyan He, Jie Zhang, Tianwei Zhang, Wenbo Zhou, Weiming Zhang, and Nenghai Yu. Catch you everything everywhere: Guarding textual inversion via concept watermarking, 2023. 2 +[6] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. AquaLoRA: Toward white-box protection for customized stable diffusion models via watermark LoRA. In Proceedings of the 41st International Conference on Machine Learning, pages 13423-13444. PMLR, 2024. 1, 2, 3, 5, 6 +[7] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22409-22420, 2023. 1, 2, 3, 5, 6, 7 +[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 2, 4 +[9] David Haden. Now you see it... now you don't: Detecting ai use via image watermarking. Information Today, 40(9): 31-33, 2023. 1 +[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022. 4 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 2, 4 +[12] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Paralleledits: Efficient multi-aspect text-driven image editing with attention grouping, 2024. 2, 3 +[13]Imagen-Team-Google. Imagen 3,2024.2 +[14] Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation, 2024. 5 +[15] Maomao Li, Yu Li, Yunfei Liu, and Dong Xu. Exploring iterative manifold constraint for zero-shot image editing, 2025. 4 +[16] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques, 2021. 3 +[17] Hanwen Liang, Yuyang Yin, Dejia Xu, Hanxue Liang, Zhangyang Wang, Konstantinos N. Plataniotis, Yao Zhao, + +and Yunchao Wei. Diffusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models, 2024. 2 +[18] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 5 +[19] Yuan Ma, Kewen Liu, Hongxia Xiong, Panpan Fang, Xiaojun Li, Yalei Chen, and Chaoyang Liu. Perception-oriented single image super-resolution via dual relativistic average generative adversarial networks, 2020. 3, 5 +[20] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations, 2022. 5 +[21] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification, 2022. 5 +[22] Athanasios Nikolaidis and Ioannis Pitas. Region-based image watermarking. IEEE Transactions on image processing, 10(11):1726-1740, 2001. 1 +[23] Jim Nilsson and Tomas Akenine-Möller. Understanding ssm, 2020. 5 +[24] Jacob Noti-Victor. Regulating hidden ai authorship. 2024. 1 +[25] Aleksandar Petrov, Shruti Agarwal, Philip H. S. Torr, Adel Bibi, and John Collomosse. On the coexistence and ensembling of watermarks, 2025. 7 +[26] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking, 2024. 3, 6 +[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3, 4 +[28] Litu Rout, Yujia Chen, Nataniel Ruiz, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Rb-modulation: Training-free personalization of diffusion models using stochastic optimal control, 2024. 2, 4 +[29] Tom Sander, Pierre Fernandez, Alain Durmus, Teddy Furon, and Matthijs Douze. Watermark anything with localized messages, 2024. 3, 6 +[30] Nathalie A Smuha, Emma Ahmed-Rengers, Adam Harkens, Wenlong Li, James MacLaren, Riccardo Piselli, and Karen Yeung. How the eu can achieve legally trustworthy ai: a response to the european commission's proposal for an artificial intelligence act. Available at SSRN 3899991, 2021. 1 +[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 2, 3, 5 +[32] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 5 +[33] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 5 + +[34] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 2 +[35] Xuandong Zhao, Kexun Zhang, Zihao Su, Saastha Vasan, Ilya Grishchenko, Christopher Kruegel, Giovanni Vigna, Yu-Xiang Wang, and Lei Li. Invisible image watermarks are provably removable using generative ai, 2024. 1, 5 +[36] Jinghao Zhou, Tomas Jakab, Philip Torr, and Christian Rupprecht. Scene-conditional 3d object stylization and composition, 2023. 2 \ No newline at end of file diff --git a/yourtextencodercanbeanobjectlevelwatermarkingcontroller/images.zip b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..200b307dc06cf7587cc7ea3347a957361c408f79 --- /dev/null +++ b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcec54b7aa4718c8f084a836aa870db3b2cf7c5d4e00a51174fe0afe211382e7 +size 737533 diff --git a/yourtextencodercanbeanobjectlevelwatermarkingcontroller/layout.json b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf7749f700b5be4a9eed6bc2022acfb11946174b --- /dev/null +++ b/yourtextencodercanbeanobjectlevelwatermarkingcontroller/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd8d11e8c18ed3e912f88d00c95b8653e0ee2d04286a6f0dbf50bd9fd44b5cf2 +size 486979 diff --git a/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/5467ef8e-c8b4-4c7d-b7ef-a1694767ed15_content_list.json b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/5467ef8e-c8b4-4c7d-b7ef-a1694767ed15_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..293a5ff37734e49a1718f1427ff75c643be669d4 --- /dev/null +++ b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/5467ef8e-c8b4-4c7d-b7ef-a1694767ed15_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13b9fa801c773c14d83c32e93249030f22c1345c5ca31a3ea8b41272cc943865 +size 70308 diff --git a/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/5467ef8e-c8b4-4c7d-b7ef-a1694767ed15_model.json b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/5467ef8e-c8b4-4c7d-b7ef-a1694767ed15_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c299bdd8f030f66615387c7369d83c20093c93dc --- /dev/null +++ b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/5467ef8e-c8b4-4c7d-b7ef-a1694767ed15_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12b6ec4a1962dbd86244a199976c05d340a6019bc3d7701d2218c47326e54968 +size 87115 diff --git a/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/5467ef8e-c8b4-4c7d-b7ef-a1694767ed15_origin.pdf b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/5467ef8e-c8b4-4c7d-b7ef-a1694767ed15_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7b4d25e68a6816acb86256258e8b52cc242f0d52 --- /dev/null +++ b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/5467ef8e-c8b4-4c7d-b7ef-a1694767ed15_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ccd38324a242d9d72ab79b3eb16dd11e464feeaeea2b909f4b5e9954dc93d14 +size 31120419 diff --git a/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/full.md b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/full.md new file mode 100644 index 0000000000000000000000000000000000000000..57dc7c603b1f3bb9d6150990f831ab24d8ae0c80 --- /dev/null +++ b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/full.md @@ -0,0 +1,292 @@ +# You Share Beliefs, I Adapt: Progressive Heterogeneous Collaborative Perception + +Hao Si Ehsan Javanmardi Manabu Tsukada * The University of Tokyo + +{si-hao, ejavanmardi, mtsukada}@g.ecc.u-tokyo.ac.jp + +# Abstract + +Collaborative perception enables vehicles to overcome individual perception limitations by sharing information, allowing them to see further and through occlusions. In real-world scenarios, models on different vehicles are often heterogeneous due to manufacturer variations. Existing methods for heterogeneous collaborative perception address this challenge by fine-tuning adapters or the entire network to bridge the domain gap. However, these methods are impractical in real-world applications, as each new collaborator must undergo joint training with the ego vehicle on a dataset before inference, or the ego vehicle stores models for all potential collaborators in advance. Therefore, we pose a new question: Can we tackle this challenge directly during inference, eliminating the need for joint training? To answer this, we introduce Progressive Heterogeneous Collaborative Perception (PHCP), a novel framework that formulates the problem as few-shot unsupervised domain adaptation. Unlike previous work, PHCP dynamically aligns features by self-training an adapter during inference, eliminating the need for labeled data and joint training. Extensive experiments on the OPV2V dataset demonstrate that PHCP achieves strong performance across diverse heterogeneous scenarios. Notably, PHCP achieves performance comparable to SOTA methods trained on the entire dataset while using only a small amount of unlabeled data. + +# 1. Introduction + +Recent advances in Vehicle-to-Everything (V2X) have improved the capability of autonomous driving systems[12, 40, 42]. V2X enables vehicles to interact with the surrounding environment, creating a more intelligent and interconnected traffic ecosystem. With the support of V2X, Collaborative Perception allows vehicles to share perception information[23, 25, 32], thus expanding the range of perception and empowering the ability to see through occlusion[9]. This achieves a more precise and robust environmental per + +![](images/c95f7e429301a07da79d0d5dc2d5c99a05b22df07c13fae23398f14a2f039c81.jpg) +(a) +Figure 1. Illustration of different heterogeneous collaborative perception patterns. Fig.1a corresponds to existing methods requiring joint training before collaboration. Fig.1b shows our PHCP process. It utilizes a small amount of unlabeled data from other agents to complete self-training and collaborate with others. + +![](images/72ad674680e6c7563d9e4b8b4ef52f08b115bd2edf81bd475cf6befe6fba4afb.jpg) +(b) + +ception in complex urban traffic environments. + +Despite the potential of collaborative perception in improving the performance of autonomous driving systems, it faces a major challenge in the real world: heterogeneity[14, 19, 20, 22, 36, 39]. Autonomous vehicles produced by different manufacturers typically adopt different sensor configurations and perception models, resulting in significant differences in the encoded intermediate features in the semantic space (domain gap)[36]. This problem makes collaborative perception between intelligent agents difficult and may lead to ineffective information sharing and feature fusion. + +Many existing methods enhance heterogeneous collaborative perception by training feature encoders and detection heads[19, 36] or by learning adapters[20] for new collaborators. While these approaches improve performance, they + +remain limited in their generalization to newly added agents. As shown in Fig.1a, this limitation implies that whenever a new agent joins the collaboration, the model or adapter must be trained, and additional parameters must be stored. However, with the continuous emergence of new sensors and perception models, it becomes impractical for autonomous vehicles to maintain dedicated adapters or models for every potential collaborator. This highlights the need for a more flexible and scalable heterogeneous collaborative perception framework. To address this challenge, we propose a new question: Can the ego vehicle dynamically change its model parameters during the inference stage for different collaborators without relying on joint training or any prior knowledge of them? + +Solving this problem presents several key challenges: (1) Lack of supervision in the inference stage. Unlike the training stage, the inference process lacks additional annotations. Therefore, a core challenge is leveraging unsupervised learning methods to fine-tune the model and achieve feature alignment. (2) Constraints on data volume and computational time. Given the real-time requirements of collaborative perception, we aim to minimize both the training data amount and fine-tune iterations after the collaboration relationship is confirmed. + +To address these challenges, we propose a Progressive Heterogeneous Collaborative Perception(PHCP) framework based on self-training and few-shot unsupervised domain adaptation. This framework aligns features in semantic space by only a small amount of unlabeled data from collaborating agents in the inference stage. The overall workflow consists of two stages. In stage I, the agent generates pseudo labels from its predictions and transmits them with intermediate features to the ego vehicle. Upon receiving this information, the ego vehicle fine-tunes an adapter by a few-shot learning approach. In stage II, the agent transmits only the intermediate features, while the ego vehicle uses the well-trained adapter to transform the features and then perform fusion and prediction. We conduct extensive experiments on the OPV2V dataset, and the results demonstrate that our method outperforms the direct collaboration baseline by approximately $30\%$ in heterogeneous collaborative perception scenarios. Our main contributions are summarized as follows: + +- To the best of our knowledge, this is the first work to tackle the domain gap in heterogeneous collaborative perception during the inference process. +- We provide a new perspective to address this challenge through few-shot unsupervised domain adaptation. We propose a novel collaborative perception scheme and framework, generating effective supervision labels during inference and enabling the ego vehicle to fine-tune the adapter and resolve the domain gap issues within a few iterations, + +![](images/377c5f774861c329daca73fd0d32717a7d9ae8ba0a0fa0edccd7d33b5b37df26.jpg) +Figure 2. Comparison of different collaborative perception schemes. + +- We conducted extensive experiments across various scenarios, demonstrating the robustness of our method in heterogeneous collaborative perception. + +# 2. Related Work + +Collaborative perception. Collaborative perception improves perception performance by fusing shared information among connected vehicles. As shown in the Fig.2, collaborative perception can be categorized into early, late, and intermediate collaboration depending on the type of shared data[9]. In early collaboration, raw sensor data such as LiDAR point clouds and RGB images are directly fused at the data level after calibration and spatial alignment, resulting in a more comprehensive environmental perception[1, 4, 8]. Although this method performs well, the transmission of large amounts of raw data from the vehicle poses challenges to the V2X network. In late collaboration, object-level fusion is used, where agents only share perception results. After post-processing steps like spatial transformation and non-maximum suppression, the final detection results are generated[2, 7, 24, 41]. This method is simple to implement and has low bandwidth requirements but is more sensitive to noise and localization errors, leading to relatively lower performance. Intermediate collaboration is a feature-level fusion, gaining increasing attention in recent years. It transmits feature-level data while sharing and preserving semantic information, effectively reducing data transmission volume[10, 17, 27]. It's a trade-off solution between early and late collaboration. Recent studies[5, 11, 15, 16, 18, 26, 38] have focused on this scheme. + +F-Cooper[3] is the pioneering work in this field, introducing a low-level voxel fusion approach and a spatial feature fusion strategy. V2VNet[28] leverages graph neural networks to model agent communication, achieving a balance between accuracy and bandwidth requirements. AttFusion[35] introduces a single-head self-attention fusion module to capture spatial relationships within feature maps. CoBEVT[21] explores an alternative approach by replacing LiDAR with cameras, proposing a fused axial attention module that effectively integrates camera-based BEV features, enhancing perception performance in a LiDAR-free scenario. However, these methods assume that all agents share the same model and parameters[19, 20, 36], which is unrealistic in the real world. So, recent studies have shifted their focus toward heterogeneous collaborative perception, aiming to address this challenge. + +Heterogeneous Collaborative perception.V2X-ViT[34] was the first to address the heterogeneity problem. It treats vehicle-to-infrastructure and vehicle-to-vehicle as two fusion types and proposes the Heterogeneous Multi-Agent Attention module to perform fusion among heterogeneous agents. HM-ViT[31] extends heterogeneous collaborative perception to multi-sensor settings, introducing a heterogeneous 3D graph attention module that effectively fuses BEV features from different modalities. MPDA[36] focuses on model-level heterogeneity and employs a learnable feature resizer to align features, along with a sparse cross-domain transformer to bridge the domain gap. PnPDA[20] introduces a novel plug-and-play domain adapter, effectively bridging the domain gap without destructing the models. HEAL[19] proposes a new framework that establishes a unified feature space for all agents, ensuring that when a new agent joins the collaboration, it only needs to be aligned to this shared space rather than adapting to every collaborator. Although these methods address the domain gap caused by heterogeneity, they rely on prior training on the dataset and lack the flexibility to adapt to newly introduced agents. + +# 3. Method + +In this paper, we explore achieving progressive collaboration in heterogeneous scenarios without relying on joint training or prior knowledge of other agents. Our method first generates pseudo labels from agents to facilitate self-training for the ego. This process is completed within a very limited number of frames. Subsequently, the agent only needs to share intermediate features, while the ego performs feature fusion and final prediction, thereby enhancing collaboration efficiency and addressing the domain gap problem. Fig.3 illustrates the main workflow. + +# 3.1. Problem Formulation + +The collaborative perception at the feature level can be divided into two stages. In the first stage, all agents use their + +own encoders to process the raw data and generate feature maps, which are then shared. In the second stage, ego vehicles perform fusion on all agents' features, and then the detection head is used to predict the results. Consider N agents in the scene, where $\chi_{i}$ represents the ith agent. The process can be described as follows: + +$$ +\mathbf {F} _ {\mathrm {i}} = \operatorname {e n c} _ {i} \left(d _ {i}\right), \quad \mathbf {F} _ {\mathrm {i}} \in S \tag {1} +$$ + +$$ +\mathbf {P} _ {\mathrm {i}} = \operatorname {h e a d} _ {i} \left(\operatorname {f u s e} _ {i} \left(\mathbf {F} _ {\mathbf {1}}, \mathbf {F} _ {\mathbf {2}}, \dots , \mathbf {F} _ {\mathbf {n}}\right)\right) \tag {2} +$$ + +where $d_{i}$ denotes $\chi_{i}$ 's original data, $P_{i}$ is the final prediction result. $enc_{i}$ is feature encoder of $\chi_{i}$ , which outputs the feature map $F_{i}$ . The feature fusion module $fuse_{i}$ is responsible for spatial-temporal alignment and feature integration. $head_{i}$ is the classification head for $\chi_{i}$ , predicting the target position and confidence score based on the feature map. If heterogeneity is considered, the intermediate features $F_{i}$ generated by the encoding may belong to different semantic spaces. To address this, we can use feature adapters to map the features of different agents to the ego's feature space. + +$$ +\mathbf {F} _ {\mathrm {i}} ^ {\prime} = \Phi_ {i \rightarrow e g o} (\mathbf {F} _ {\mathrm {i}}), \quad \mathbf {F} _ {\mathrm {i}} ^ {\prime} \in S _ {e g o}, \mathbf {F} _ {\mathrm {i}} \in S _ {i} \tag {3} +$$ + +Our target is to address the heterogeneity problem by achieving adaptive adjustment of the adapter $\Phi_{i\rightarrow eg_o}$ during the collaborative perception inference stage with only a small amount of unlabeled data. + +# 3.2. Feature Adapter + +Our approach employs a feature adapter to bridge the domain gap between the source $S_{i}$ and target domains $S_{ego}$ . Since we perform training on a small dataset, the adapter structure must remain simple to prevent overfitting. Through visual analysis of feature maps in Fig.4, we observed that while the overall feature distributions of the source and target domains are similar, they show misalignment in channel dimensions and certain critical regions. We found CBAM[30] to be particularly well-suited for this task. CBAM effectively addresses this issue by incorporating both channel and spatial attention mechanisms. The channel attention module(CAM) directs the network to focus on important channels, while the spatial attention module(SAM) enhances attention to crucial spatial regions. CBAM's lightweight design also introduces minimal computational overhead, making it an efficient and practical choice for our method. + +# 3.3. PHCP Scheme + +We adopt a novel collaborative perception approach between intermediate and late collaboration, ensuring the generation of pseudo-labeled data without introducing excessive communication overhead. Specifically, after establishing the collaboration relationship, in the first $k$ frames, the agent simultaneously sends features and detection results to the ego, and the ego uses this information to update the feature + +![](images/f80aa2951ad287a0f0140ec52f3566f93abd7f7673b1fce3f3227b7e105e727d.jpg) +Stage I: ego fine-tune adapter by self training +Stage II: ego inference + +![](images/0c39aad90a6e46225d425508109b78e81b725cfa57468a922ad7cbeb0ec54ee0.jpg) +Figure 3. Overview of our proposed heterogeneous collaboration process. In Stage I, the ego performs self-training using pseudo labels to fine-tune the adapter. In Stage II, the ego applies the adapter to transform features and then perform fusion and prediction. +(a) PointPillar feature map + +![](images/5a509feba0d7159fae9f51ac1a08547f5bbb131286c05b313f1b334766c9b7c0.jpg) +(b) SECOND feature map + +![](images/85f93ebe897a5af22053bd630f1e36d83ac0d69c88afa89ed26eda64200b26c8.jpg) +(c) PointPillar feature map, $c = 29$ +Figure 4. Illustration of domain gap of different encoders. We visualized the encoder's output features by displaying each channel's absolute values. Figure 4a and Figure 4b illustrate the misalignment in the channel dimension. Figures 4c and 4d explain the differences in the area of interest + +![](images/1ede67e7e98ade81a6c79241c47af33f96bfbff463cd28bc4bdd9a4918159654.jpg) +(d) SECOND feature map, $c = 29$ + +adapter. We define this as the stage I. The updated method will be introduced later. After k frames, the collaboration process follows the normal intermediate collaboration process, with the agent only sending features to the ego. This corresponds to stage II. We determine the value of k by re + +ferring to typical values in few-shot learning, choosing 1, 5, or 10. + +# 3.4. PHCP Process + +This progressive collaboration process consists of two stages. In the first stage, the ego trains its own adapter using features and pseudo-labels provided by agents through a few-shot self-training way. In the second stage, the ego aligns the features of the agents by the adapter and then performs fusion and prediction. + +Stage I. We aim to optimize the ego's adapter without prior knowledge about agents in this stage. Due to the lack of labels, we use agents' predictions as pseudolabels. In particular, we select high-quality predictions as pseudo-labels based on confidence scores, and we also record the corresponding confidence scores as soft labels instead of one-hot labels. In this way, we can create dataset $\mathcal{D}_i = \{(d_1, p_1), \ldots, (d_k, p_k) \mid d \in S_i\}$ as a small training sample dataset(the support set). Our model optimization strategy builds upon a two-stage fine-tuning approach (TFA)[29]due to its simplicity and efficiency. We introduce further enhancements to improve its adaptability within our collaborative perception framework. Considering that the detectors of ego and agent have already been trained separately before the collaboration begins, we only need to focus on the second fine-tuning stage. In the second stage, we fix all the parameters of the fusion network and the detector head and only fine-tune the adapters. This is because the ego needs to use the same fusion network and detector head for different agents. When collaborating with multiple agents simultaneously, fine-tuning the model on multiple datasets may introduce mutual interference, reducing its adaptability to each agent and degrading overall performance. However, + +Algorithm 1 Stage I: adapter fine-tuning +1: for each scenario do +2: Input: $\{F_i\}, \mathbf{F}_i = \{d_1,..d_k\} \quad \triangleright$ shared feature from each agent +3: for each agent $\chi_i$ except ego do +4: $\mathcal{H}_i = \{\mathbf{F_j} \mid \chi_j$ is homogenous to $\chi_i\}$ aggregate all agents that are homogeneous to the current +5: $\mathbf{P_i} = \text{head}_i(\text{fuse}_i(\mathcal{H}))$ +6: $\mathcal{D}_i = \{\mathcal{H}_i, \mathbf{P_i}\}$ +7: end for +8: for each agent $\chi_i$ except ego do +9: $\mathbf{L_i}, \mathcal{H}_i \gets \mathcal{D}_i$ +10: $\mathbf{P_i} = \text{head}_e(\text{fuse}_e(\mathcal{H}_i))$ +11: $\Phi_{i\rightarrow ego} \gets \nabla Loss(\mathbf{L_i}, \mathbf{P_i})$ +12: end for +13: Output: $\Phi_{i\rightarrow ego} \triangleright$ feature adapters from agents to ego +14: end for + +the adapter is created independently by the ego for each agent, and by only adjusting the adapter, isolation can be achieved, and interference can be reduced. In the fine-tuning process, we adopt a scheduler with a warm-up mechanism. We select one scenario for hyperparameter tuning, while the remaining scenarios are for testing. We train a total of 20 rounds, and due to the small training data size, the overall computational overhead is very low. + +Stage II. This stage follows immediately after the first stage. We consider the features shared by agents as the query set $\mathcal{Q}_i = \{d_{k + 1},\ldots ,d_N\mid d\in \mathcal{S}_i\}$ . We believe that after the training in the first stage, the adapter can effectively address the domain gap problem. Subsequently, the transformed features $\mathcal{Q}_i^\prime$ and the features from the ego's encoder can be fused to obtain the final detection results. We use the aver + +Algorithm 2 Stage II: Inference +1: for each scenario do +2: Input: $\{F_i\}, F_i = d_{k+1}, \ldots, d_N$ ▷ shared feature from each agent +3: for each agent $\chi_i$ except ego do +4: $F_i' = \Phi_{i \to ego}(F_i)$ +5: $F_i = F_i \cup F_i'$ +6: end for +7: $\mathbf{P}_{\mathrm{ego}} = \text{head}_{\mathrm{ego}}(fuse_{\mathrm{ego}}(\mathcal{F}_i))$ +8: Output: $\mathbf{P}_{\mathrm{ego}}$ ▷ ego final prediction +9: end for + +age AP value of all scenarios as the final evaluation metric, referred to as mean scenario AP (mSAP). This is because we observe significant differences in the data distribution among different scenarios. In Fig.5a, most of the objects + +![](images/b4a72766310c874c2772446497d7e6d31c7eac89bbc6c15d33a8022daed11209.jpg) +(a) broad shared field of view + +![](images/ddb43ee265af13dd942093faf40a76704b9705af07ca5f8a8f2eaf610ec0a871.jpg) +(b) narrow shared field of view +Figure 5. Detection performance under different ego vehicle field of view conditions + +are concentrated on one side of the ego vehicle, and even without receiving information from other agents, the final detection results are not affected. However, in Fig.5b, the objects are mainly distributed on the side of agent vehicles. If the ego vehicle cannot effectively utilize the feature information from agents, it may result in many false negatives. Therefore, mSAP, as a comprehensive performance evaluation metric, can more fairly measure the detection capability of the model in different scenarios. + +# 4. Experiments + +# 4.1. Datasets and Evaluation + +We use OPV2V[35] and its supplementary dataset, OPV2V-H[19] as the dataset for model training and evaluation. OPV2V is collected by the simulation tools OpenCDA[33] and CARLA[6], covering various complex driving scenarios, including challenging situations such as severe occlusion. We use the default dataset splitting ratio. The training and validation sets are used to train the baseline model for homogeneous collaborative perception, and the test set is used to evaluate the performance of heterogeneous collaborative perception. We further divide the test dataset according to the scenario settings, resulting in 16 specific scenarios. For each scenario, we split them into support sets $S$ and query sets $Q$ , and the sizes are determined by the number of shots. We adopt mSAP at Intersection-over-Union (IoU) thresholds of 0.3, 0.5, and 0.7 to evaluate the model performance. + +# 4.2. Experimental Setups + +We adopt two types of agents with different LiDAR encoders: agent LP using the PointPillars[13] encoder and agent LS using the SECOND[37] module. Feature fusion is performed using the multi-scale pyramid fusion method proposed by HEAL[19]. The overall training and validation procedure of the model is as follows: on the OPV2V dataset, we conducted homogeneous collaborative perception training for LP and LS respectively using two RTX 6000ADA GPUs, so we get base models $\mathcal{M}_{LP} = \{\text{Enc}_{LP}, \text{Fuse}_{LP}, \text{Head}_{LP}\}$ and $\mathcal{M}_{LS} = \{\text{Enc}_{LS}, \text{Fuse}_{LS}, \text{Head}_{LS}\}$ . Subsequently, on the dataset $S$ , we extract intermediate features by EncLS + +
MethodAP@IoU 0.5 in ScenariosmSAP@IoU
123456789101112131415160.30.50.7↑
Direct Fusion83.231.968.269.083.954.754.386.859.935.040.240.534.055.482.072.559.759.553.0
PHCP(Ours)96.298.496.294.296.395.282.697.095.295.988.996.280.685.690.989.092.992.485.9
+ +Table 1. Comparison between our methods and the direct fusion baseline + +
MethodAP@IoU 0.7 in ScenariosmSAP@IoU
123456789101112131415160.30.50.7↑
F-Cooper[3]84.150.869.756.379.465.557.076.161.257.658.060.948.449.672.677.391.283.163.4
CoBEVT[21]78.767.279.571.882.071.262.287.470.371.266.460.454.166.883.979.794.891.272.0
AttFusion[35]84.068.587.071.492.678.663.784.072.873.574.476.365.175.783.784.993.089.977.3
V2X-ViT[34]81.774.990.879.288.784.865.792.190.181.378.487.673.183.490.682.694.392.382.8
PHCP(Ours)89.292.690.487.492.294.167.195.289.694.784.791.570.783.889.080.892.892.387.1
HEAL[19]89.193.497.691.393.696.676.297.293.097.589.794.586.987.793.190.095.595.191.7
+ +and generate prediction results using model $Head_{LS}$ as pseudo-labels. Then, we perform fine-tuning on the submodel $\{\Phi()_{LS\rightarrow LP},Fuse_{LP},Head_{LP}\}$ with this pseudolabeled data. Only the adapter is trainable, while the parameters of other modules are fixed. Finally, we combine the adapter with $M_{LP}$ to form a new model $\mathcal{M}_{LP}^{\prime} = \{Enc_{LP},\Phi()_{LS\rightarrow LP},Fuse_{LP},Head_{LP}\}$ , and then evaluated it on the dataset $\mathcal{Q}$ . + +Optimization Strategy. We employ a scheduler that combines the warmup and multi-step decay strategies. In the warmup stage, we use a linear strategy, where the warmup factor was set to 0.001, and the warmup lasted for 8 epochs. During the actual training process, the learning rate decayed at the 12th and 16th epochs with a decay factor of 0.1. Our base learning rate was set to 0.005, the batch size was set to 1, and the training was completed within 20 epochs. + +Comparison. Our direct fusion baseline method uses models LS and LP, which have not undergone joint training, for direct collaborative perception. This setting aligns with real-world heterogeneous collaboration, as we have no prior knowledge of each other's model structures or parameters before collaboration begins. Additionally, we compare our approach with other collaborative perception methods, which were fine-tuned on the training dataset before testing. For models that do not support heterogeneous collaborative perception, such as F-Cooper[3], V2X-ViT[34], AttFusion[35], and CoBEVT[21], we first train the models in homogeneous scenarios and then perform 40 fine-tuning epochs in heterogeneous scenarios. For models that support heterogeneous collaborative perception, like HEAL[19], we follow the training methods provided by the authors. + +# 4.3.Quantitative evaluation + +Main performance comparison. Tab.1 presents a comparison between our method and the direct fusion baseline across various scenarios. Our method consistently outperforms the baseline by $33.2\%$ , $32.9\%$ , and $32.9\%$ in AP@0.3, AP@0.5, and AP@0.7. Our method still achieves strong results even in some challenging scenarios where the baseline performs poorly. Tab.2 further compares our approach with other collaborative perception methods. Except for HEAL, which achieves the SOTA performance, our method surpasses all others. It is important to note that while other methods utilize the full training dataset with labels, our approach achieves comparable performance using only a minimal amount of unlabeled data. + +Computational cost. As detailed in Tab.3, the training phase takes between 1.5 and 2.4 seconds depending on the number of shots. The inference takes 0.07 seconds. Importantly, the few-shot training process is executed only once when a new collaborative relationship is established. So the overall computational cost meets the real-time requirements. + +Table 2. Comparison between PHCP and other methods. Notably, we achieve these results using only a small amount of unlabeled data. + +
StageSettingTime(s)Mem(MB)
Trainingshots=1, bs=11.491290
shots=2, bs=21.712314
shots=4, bs=42.174552
shots=5, bs=52.395604
Inference-0.07798
+ +Table 3. Computation cost of training and inference. All training configurations use 20 iterations; shots indicates the number of shots used for training, and bs is the batch size. + +![](images/63874b0257d611c7ed02faa3e551c484249a043a8364a3bb257e61d034df360d.jpg) + +![](images/dfad7fa8315e0535408eaf6fce3e540f8e743c4cdea13ccc745e4d65f7dafecb.jpg) + +![](images/c50fba0707771476c486dceaa6cbbc1bc9790b881f1b69ac8bf410b6881f6e93.jpg) + +![](images/ea9cfab27b89cb190248e8d340a6d4e1b67bb25771746928f1191918a4e7cda9.jpg) +Figure 6. Visual comparison of three scenarios between our method and the baseline. Each scenario involves two, three, and four collaborative agents participating in collaborative perception. The red and green boxes represent the prediction and ground truth. + +![](images/ed1acfd6463a4fc812336d665bd8bc8f254ba29465e3c7791ad974239097928c.jpg) + +![](images/b7244a44c6741e2cfa960457f0c3b333dc3bfb62bcb8c1c1e95bf2148bc301e1.jpg) + +![](images/9f3f777d4a3af1b419e3096275faedbbffc9890d18bcc8dd57421878fb814836.jpg) +Figure 7. Feature map visualization comparison between our method and baseline + +# 4.4. Qualitative results + +Visualization of detection results. Fig.6 presents the qualitative visualizations of our method and the baseline. We select three scenarios with varying numbers of collaborative agents. Due to the lack of effective feature fusion, the baseline method has many false negatives outside the ego vehicle's field of view. In contrast, our approach successfully detects these objects. + +Visualization of feature map. Fig.7 shows a collaborative scenario involving three agents. By comparing the feature maps, we observe that our method achieves more effective feature alignment with the ego vehicle than the baseline. + +# 4.5. Ablation studies + +Trade-off between performance and number of shots. A larger number of shots represents more training data, which usually leads to more accurate prediction results. However, it also increases the waiting time for initiating collaboration. Therefore, we analyze the relation between the number of shots and model performance. In our experiments, 0-shot means direct collaboration and serves as the baseline. Fig.8 illustrates the relation between the number of shots and the model's performance. Even when relying only on single-frame data, our model still achieves a performance improvement of approximately $50\%$ . As the number of shots increases, the AP@IoU 0.7 also improves. This suggests that more training data can better optimize the adapter module, enhance the quality of fused feature maps, and thus achieve more precise localization and higher prediction accuracy. + +Quality of the pseudo label. The quality of pseudo labels impacts the final performance. We evaluate the pseudo labels on the test set and obtained an AP@0.5 of 0.86 and an AP@0.7 of 0.78; $80\%$ of the predictions have confidence scores greater than 0.5. Then we conduct an analysis of different filtering strategies. We set a series of confidence thresholds to filter pseudo-labels or directly used the confidence scores as soft labels. We evaluate the performance using mSAP and AP in the worst-case scenarios(wSAP) for a comprehensive comparison. From Tab.4, retaining only high-quality pseudo labels results in insufficient positive samples, limiting the model's learning capacity. Conversely, low-quality pseudo labels introduce noise, leading to a de- + +![](images/17485fe65117a25270291a3b86317d738f0ed9d3cf3a8c8c400cba0342521ca0.jpg) +Figure 8. Performance of different number of shots + +cline in overall model accuracy. However, the impact of pseudo-label quality on the final results remains relatively minor. We attribute this to the spatial attention mechanism in the adapter, which effectively focuses on target regions. + +
0.20.50.7soft
mSAP@0.785.085.985.885.4
wSAP@0.766.168.067.767.0
+ +Comparison between heterogeneous and homogeneous settings. From Tab.5, while our baseline model performs well under homogeneous conditions, its performance drops when applied directly to heterogeneous scenarios. This contrast highlights that even strong homogeneous collaboration models cannot effectively handle heterogeneity. In contrast, our method effectively addresses this challenge and maintain robust performance in heterogeneous settings. + +Table 4. Performance under different pseudo label confidence levels + +
MethodmSAP@IoU
0.30.50.7
Direct Fusion(baseline)59.759.553.0
PHCP92.992.485.9
SECOND (homogeneous)94.894.290.5
PointPillar (homogeneous)96.295.893.1
+ +Table 5. Heterogeneous vs. homogeneous collaborative perception. + +Generalization across scenarios. In this study, we conduct separate training and testing on each scenario because each + +![](images/aa8433e078dcd1d420269495c9f84a49630b37f915ce00f59d70d14aafd54905.jpg) +Figure 9. Cross-scenario evaluation results. The figure presents test results across different scenarios. The values have been normalized using the diagonal elements. The overall generalization performance is strong except for a few particularly challenging scenarios where all models perform poorly. + +scenario corresponds to a new collaborative relationship, and our model validation aims to assess its adaptability in each new heterogeneous collaboration environment. We also performed cross-scenario testing to verify the model's generalization ability and ensure it does not overfit a single scenario. Specifically, we train a base model for each scenario and used its AP@IoU 0.5 inference results as a baseline. Then, without any additional fine-tuning, we directly test the same model on other scenarios to analyze its generalization ability in different environments. The results in Fig.9 indicate that although the model's performance slightly declines in some complex scenarios, it remains stable in most scenarios. This result further validates the adaptability and generalization of our model. + +# 5. Conclusion + +We introduce PHCP, a novel heterogeneous collaborative perception framework designed to address feature misalignment during the inference stage of collaborative perception, where training on a dataset with collaborators is unrealistic, and collaborators' prior knowledge is unavailable. The key idea is to utilize pseudo labels generated by other agents and perform few-shot self-training on the ego vehicle to fine-tune the adapter module, thereby bridging the domain gap between the ego and agents. Extensive experiments on the OPV2V dataset demonstrate the effectiveness of our approach, outperforming baseline methods across various heterogeneous collaborative perception scenarios. Compared to other methods trained on the entire dataset, our approach achieves performance comparable to SOTA methods while using only a small amount of unlabeled data. + +# Acknowledgments + +# References + +[1] Eduardo Arnold, Mehrdad Dianati, Robert de Temple, and Saber Fallah. Cooperative perception for 3d object detection in driving scenarios using infrastructure sensors. IEEE Transactions on Intelligent Transportation Systems, 23(3): 1852-1864, 2020. 2 +[2] Dian Chen and Philipp Kähenbuhl. Learning from all vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17222-17231, 2022. 2 +[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 3, 6 +[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 2 +[5] Jiaxun Cui, Hang Qiu, Dian Chen, Peter Stone, and Yuke Zhu. Coopernaut: End-to-end driving with cooperative perception for networked vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17252-17262, 2022. 2 +[6] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 5 +[7] Chen Fu, Chiyu Dong, Christoph Mertz, and John M Dolan. Depth completion via inductive fusion of planar lidar and monocular camera. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10843-10848. IEEE, 2020. 2 +[8] Hongbo Gao, Bo Cheng, Jianqiang Wang, Keqiang Li, Jianhui Zhao, and Deyi Li. Object classification using cnn-based fusion of vision and lidar in autonomous vehicle environment. IEEE Transactions on Industrial Informatics, 14(9):4224-4231, 2018. 2 +[9] Yushan Han, Hui Zhang, Huifang Li, Yi Jin, Congyan Lang, and Yidong Li. Collaborative perception in autonomous driving: Methods, datasets, and challenges. IEEE Intelligent Transportation Systems Magazine, 15(6):131-151, 2023. 1, 2 +[10] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 2 +[11] Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake lidar in 3d detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2023. 2 + +This work was supported by JST, CRONOS, Japan Grant Number JPMJCS24K8. +[12] Sou Kitajima, Hanna Chouchane, Jacobo Antona-Makoshi, Nobuyuki Uchida, and Jun Tajima. A nationwide impact assessment of automated driving systems on traffic safety using multiagent traffic simulations. IEEE Open Journal of Intelligent Transportation Systems, 3:302-312, 2022. 1 +[13] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 5 +[14] Jinlong Li, Baolu Li, Xinyu Liu, Runsheng Xu, Jiaqi Ma, and Hongkai Yu. Breaking data silos: Cross-domain learning for multi-agent perception from independent private sources. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 18414-18420. IEEE, 2024. 1 +[15] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 2 +[16] Yiming Li, Juexiao Zhang, Dekun Ma, Yue Wang, and Chen Feng. Multi-robot scene completion: Towards task-agnostic collaborative perception. In Conference on Robot Learning, pages 2062–2072. PMLR, 2023. 2 +[17] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 2 +[18] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 2 +[19] Yifan Lu, Yue Hu, Yiqi Zhong, Dequan Wang, Siheng Chen, and Yanfeng Wang. An extensible framework for open heterogeneous collaborative perception. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 5, 6 +[20] Tianyou Luo, Quan Yuan, Guiyang Luo, Yuchen Xia, Yu-jia Yang, and Jinglin Li. Plug and play: A representation enhanced domain adapter for collaborative perception. In European Conference on Computer Vision, pages 287-303. Springer, 2024. 1, 3 +[21] Hao Xiang Wei Shao Bolei Zhou Jiaqi Ma Runsheng Xu, Zhengzhong Tu. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. In Conference on Robot Learning (CoRL), 2022. 3, 6 +[22] Congzhang Shao, Guiyang Luo, Quan Yuan, Yifu Chen, Yilin Liu, Kexin Gong, and Jinglin Li. Hetecooper: Feature collaboration graph for heterogeneous collaborative perception. In European Conference on Computer Vision, pages 162-178. Springer, 2024. 1 +[23] Steven E Shladover. Opportunities and challenges in cooperative road vehicle automation. IEEE Open Journal of Intelligent Transportation Systems, 2:216-224, 2021. 1 +[24] Sanbao Su, Yiming Li, Sihong He, Songyang Han, Chen Feng, Caiwen Ding, and Fei Miao. Uncertainty quantification of + +collaborative detection for self-driving. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5588-5594. IEEE, 2023. 2 +[25] Johan Thunberg, Galina Sidorenko, Katrin Sjoberg, and Alexey Vinel. Efficiently bounding the probabilities of vehicle collision at intelligent intersections. IEEE Open Journal of Intelligent Transportation Systems, 2:47-59, 2021. 1 +[26] Binglu Wang, Lei Zhang, Zhaozhong Wang, Yongqiang Zhao, and Tianfei Zhou. Core: Cooperative reconstruction for multiagent perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8710-8720, 2023. 2 +[27] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8187-8196, 2023. 2 +[28] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer vision-ECCV 2020: 16th European conference, Glasgow, UK, August 23-28, 2020, proceedings, part II 16, pages 605-621. Springer, 2020. 3 +[29] Xin Wang, Thomas E. Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. 2020. 4 +[30] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3 +[31] Hao Xiang, Runsheng Xu, and Jiaqi Ma. Hm-vit: Hetero-modal vehicle-to-vehicle cooperative perception with vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 284–295, 2023. 3 +[32] Hua Xie, Yunjia Wang, Xieyang Su, Shengchun Wang, and Liang Wang. Safe driving model based on v2v vehicle communication. IEEE Open Journal of Intelligent Transportation Systems, 3:449-457, 2022. 1 +[33] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 5 +[34] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 3, 6 +[35] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583-2589. IEEE, 2022. 3, 5, 6 +[36] Runsheng Xu, Jinlong Li, Xiaoyu Dong, Hongkai Yu, and Jiaqi Ma. Bridging the domain gap for multi-agent perception. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 6035-6042. IEEE, 2023. 1, 3 + +[37] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 5 +[38] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception. In Proceedings of the IEEE/CVF international conference on computer vision, pages 23383-23392, 2023. 2 +[39] Melih Yazgan, Thomas Graf, Min Liu, Tobias Fleck, and J Marius Zöllner. A survey on intermediate fusion methods for collaborative perception categorized by real world challenges. In 2024 IEEE Intelligent Vehicles Symposium (IV), pages 2226-2233. IEEE, 2024. 1 +[40] Syed Adnan Yusuf, Arshad Khan, and Riad Souissi. Vehicle-to-everything (v2x) in the autonomous vehicles domain-a technical review of communication, sensor, and ai technologies for road user safety. Transportation Research Interdisciplinary Perspectives, 23:100980, 2024. 1 +[41] Wenyuan Zeng, Shenlong Wang, Renjie Liao, Yun Chen, Bin Yang, and Raquel Urtasun. Dsdnet: Deep structured self-driving network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 156-172. Springer, 2020. 2 +[42] Liuhui Zhao and Andreas A Malikopoulos. Enhanced mobility with connectivity and automation: A review of shared autonomous vehicle systems. IEEE Intelligent Transportation Systems Magazine, 14(1):87-102, 2020. 1 \ No newline at end of file diff --git a/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/images.zip b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e008c5f4a4bf2317475380670e6065569d5c0248 --- /dev/null +++ b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a14b3e19bc0da6d8d94c93c883315f7c96214aaea8a929cec3600819b5f02323 +size 648035 diff --git a/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/layout.json b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..94f2294f3a9d222253ddeb23888bf702d75c3baf --- /dev/null +++ b/yousharebeliefsiadaptprogressiveheterogeneouscollaborativeperception/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bbc2ddbb6bc1c1bd2b0b5664d95d9fab6fd1659ba7129b2ae3833172ef703ed +size 334285 diff --git a/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/4bf51bb3-0b1c-40e9-80ac-090885fb49f2_content_list.json b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/4bf51bb3-0b1c-40e9-80ac-090885fb49f2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb69dac3089400e2ca2432be62e540d7b96e9b7b --- /dev/null +++ b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/4bf51bb3-0b1c-40e9-80ac-090885fb49f2_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88c63022dbc10a9bce415d4509ada47f77f146c512b536df77d382b414fa0e30 +size 80873 diff --git a/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/4bf51bb3-0b1c-40e9-80ac-090885fb49f2_model.json b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/4bf51bb3-0b1c-40e9-80ac-090885fb49f2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..090ca08d64377d45779b2469fbd7ec5b6913e05e --- /dev/null +++ b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/4bf51bb3-0b1c-40e9-80ac-090885fb49f2_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f011e3bbb07340c6cfb8b301a270247712df1d4ffbf59536baf40cd23c956839 +size 100958 diff --git a/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/4bf51bb3-0b1c-40e9-80ac-090885fb49f2_origin.pdf b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/4bf51bb3-0b1c-40e9-80ac-090885fb49f2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..41d9e8e639b58b6661981c17be82963beb4cf228 --- /dev/null +++ b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/4bf51bb3-0b1c-40e9-80ac-090885fb49f2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df4faccd257b1942eb35765a76594539a8facbfaf44bfa4c906505a14c03bacf +size 8819887 diff --git a/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/full.md b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1294e8b79d6e5588b44a29e485c64ac4439c9937 --- /dev/null +++ b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/full.md @@ -0,0 +1,331 @@ +# You Think, You ACT: The New Task of Arbitrary Text to Motion Generation + +Runqi Wang $^{1,2*}$ Caoyuan Ma $^{1,2*}$ Guopeng Li $^{3*}$ Hanrui Xu $^{2}$ Yuke Li $^{4}$ Zheng Wang $^{1,2\dagger}$ $^{1}$ National Engineering Research Center for Multimedia Software, Wuhan University + $^{2}$ School of Computer Science, Wuhan University + $^{3}$ StepFun + $^{4}$ University of Maryland College Park +* equal contribution, † corresponding to wangzwhu@whu.edu.cn + +![](images/2d941910e09d4c54297be8f67132f2088b7a0ff33523ae8a06cea73e7a8eabce.jpg) +Figure 1. (a) "Action Label to Motion" generates motion with a certain action label. (b) "Action Text to Motion" generates motion with an explicit Action Text (texts contain action labels). (c) "Arbitrary Texts to Motion" encompasses a new Scene Text to Motion task, where scene text refers to events or situations and does not contain explicit action labels. Understanding these scene texts and generating corresponding reactive motions is a multi-solution task. + +# Abstract + +Text to Motion aims to generate human motions from texts. Existing settings rely on limited Action Texts that include action labels (e.g., "walk, bend"), which limits flexibility and practicability in scenarios difficult to describe directly. This paper extends limited Action Texts to arbitrary ones. Scene texts without explicit action labels can enhance the practicality of models in complex and diverse industries such as virtual human interaction, robot behavior generation, and film production, while also supporting the exploration of potential implicit behavior patterns. However, newly introduced Scene Texts may yield multiple reasonable output results, causing significant challenges in existing data, framework, and evaluation. + +To address this practical issue, we first create a new dataset HUMANML3D++ by extending texts of the well-annotated dataset HUMANML3D. Secondly, we propose a simple yet effective framework that extracts action instructions from arbitrary texts and subsequently generates motions. Furthermore, we also benchmark this new setting with multi-solution metrics to address the inadequacies of + +existing single-solution metrics. Extensive experiments indicate that Text to Motion in this realistic setting is challenging, fostering new research in this practical direction. More details are available in https://github.com/RunqiWang77/TAAT.github.io. + +# 1. Introduction + +Text to Motion (T2M) denotes generating motions from natural language, reducing labor costs in industries requiring motion capture and manual editing. While current T2M methods have demonstrated effectiveness in controllable settings, some real-world applications such as Open-world games and virtual assistants often demand greater flexibility, where characters must dynamically respond to unrestricted and diverse scene inputs from users, rather than following a direct action command. For example, in an openworld game, a player might warn an NPC, "Watch out! Left punch incoming!" The NPC must interpret the intent and dodge or block, rather than execute a "punch" command. Despite this need, existing T2M methods still relies on ex + +
NameSub.MotionTextAct. ClassScene
Action LabelAMASS [34]34411k--X
NTU-120RGB+D[27]106114k-120X
UESTC [15]11826k-40X
NTU RGB+D [41]-56k-60X
BABEL [39]34466k-250X
HumanAct12 [8]121k-12X
Text DatasetKIT-ML [38]1113k6k-X
HumanML3D [9]34415k45k-X
Ours34415k135k-✓
+ +Table 1. Dataset comparison. Sub. refers to the number of humans included and Act. Class denotes the variety of action classes (not applicable to datasets with Action Texts). Our dataset excels with the most extensive annotated text content, especially with a substantial amount of Scene Texts. + +plicit action information (such as Action Labels [4, 8, 29, 31, 35, 56] and Action Texts [1, 11, 36, 44, 45, 57, 58] in Figure 1(a-b)), which are difficult to apply in less directly describable and more flexible settings, as illustrated in Figure 1(c). Therefore, exploring the generation of potential motions from arbitrary texts is important. + +Generating motion from arbitrary texts (Arbitrary Text to Motion) includes scene texts that describe events or situations without explicit action labels, contrasting with the "Action Text (label) to Motion" approach [3, 17, 19, 45, 58], which is a fixed one-to-one alignment of a precise action text and a certain kind of motion pattern (e.g., a description "walking" directly corresponds to walking pattern). Instead, "Arbitrary Text to Motion" introduces a multisolution paradigm, allowing multiple plausible motion patterns to be generated from a single scene text (e.g., in Figure 1(c), a scene text can generate four ("walk", "run", "bend down", "take a photo") different motion patterns). Consequently, datasets, frameworks, and evaluation metrics designed for deterministic one-to-one alignment in "Action Text to Motion" are unsuitable for new multi-solution task. + +Dataset. As shown in Table 1, existing datasets [8, 15, 34, 38, 39, 41] primarily focus on explicit Action Text annotations to describe motions, while Scene Text remains scarce. Furthermore, the multi-solution characteristics of Scene Text pose challenges in semantic correspondence. To address these issues, we introduce the HUMANML3D++ by adding Scene Texts to evaluate the "Arbitrary Text to Motion". To ensure both the scale and semantic coherence with motions, we use a LLM with a carefully designed prompting strategy and causal contextual guidance to generate Scene Texts. For data quality, we employ a hybrid denoising process that integrates automated filtering with manual verification. Our dataset comprises about $45\mathrm{k}$ Action Texts, $135\mathrm{k}$ Scene Texts, and $15\mathrm{k}$ motion sequences, providing a foundation for the "Arbitrary Text to Motion". + +Framework. Previous coupled frameworks [3, 11, 36, 45, 57, 58, 61] are limited to a simple and deterministic mapping between Action Text and motion, struggling to handle the complex Scene Text semantic parsing and motion plausibility. We propose a Think and Act framework for Arbitrary Text (TAAT). During the Think Stage, we leverage the optimal solution for multi-solution problems, i.e., LLM for cognitive processing and scene interpretation, exploring multiple action instructions derived from a single textual input. In the Act Stage, we improve the transformer to enable deterministic motion generation, avoiding the omission of action. Consequently, our framework decouples both cognitive processing and action execution, breaking the limitations of traditional T2M paradigms. + +Evaluation. The existing evaluation metrics fail to accommodate multi-solution scenarios. Specifically, metrics(e.g. "R-Precision, MM-Dist" [9]) assume a unique ground truth for generated motions, neglecting the possibility of output motions that match the Scene Text but differ from the dataset's GT. We further validate the dataset's ground truth with existing metrics to demonstrate that "R-Precision, MM-Dist" cannot effectively evaluate multisolution issues (Sec 5). We introduce two novel metrics, Hit Accuracy and Mean Hit Distance, which are designed to assess the validity of generated results in a manner aligned with the multi-solution characteristic of our approach, along with a comprehensive model performance analysis of different models across various application settings. + +Our main contributions can be summarized as follows: + +- We first introduce the Arbitrary Text to Motion task, expanding existing Text to Motion to more practical and flexible settings. +- We construct a comprehensive dataset with over 135k Scene Text annotations and propose a novel think-and-act framework to infer potential motions from Scene Texts. +- We develop a multi-solution evaluation metric system. Extensive experiments demonstrate our method generates more coherent motions than existing approaches. + +# 2. Related Work + +Human motion generation is an important direction in the field of multimodal learning [14, 22, 46, 47, 53, 60]. Human motion generation also supports diverse multimodal inputs, including Action Texts [1, 11, 25, 28, 36, 44, 45, 57–59, 62], action labels [4, 8, 29, 31, 35, 56], incomplete posture [6, 12, 45], music [2, 20, 21, 24, 42], images [40], and more [16, 26, 32, 33, 43, 51, 52, 60]. In common human motion generation, deterministic, action-centric texts serve as the primary inputs. Our task shifts to scene-based texts, enabling more responsive and adaptable motions from general descriptions. This novel perspective broadens the scope of T2M, increasing flexibility and enhancing usability. + +![](images/72ae30f1e56848e13cdd73bf7f2a9d517cdf51f120dbbcd19a08ddbd41b69b35.jpg) +Figure 2. Dataset overview. (a) We construct a novel HUMANML3D++ dataset through causal contextual guiding prompts and postfiltering. This is the first dataset with dual-text support. (b) Detailed data examples. + +# 2.1. Text to Motion Generation + +Arbitrary Text to Motion broadens traditional Text to Motion task, accommodating more diverse and flexible scenarios. Previous Text to Motion methodologies typically establish a deterministic one-to-one mapping between text and motion. Existing methods can be grouped into two main categories. Diffusion-based methods [3, 17-19, 45, 50, 58] utilize long Markov chains of reverse diffusion steps to generate samples, which can be computationally expensive and slow. Additionally, they cannot produce sequences of variable lengths. Furthermore, fine-grained motion generation methods [18, 50] require lengthy and complex training data or additional inputs like obstacles, trajectories, and keyframes, compromising practicality. Other approaches [10, 11, 30, 57, 61] use VAEs and transformers to model sequence relationships. However, due to limitations in model scale and training data, transformers may demonstrate suboptimal performance on unseen samples, resulting in unclear outputs. Additionally, HUMANOMATO [30] requires pre-training text-motion alignment models on specific datasets, while Momask [11] presents motions that are constrained and lack diversity, necessitating a target length input for generation and a pre-trained text-to-length model for sampling. Our approach leverages transformers' temporal modeling capabilities and LLMs' robust language generation and zero-shot transfer abilities. This synergy enables effective handling of complex textual inputs without the need for additional training data or auxiliary models. + +# 2.2. Extended Text to Motion Generation + +Recently, some research has incorporated interactive factors into the tasks of human motion generation. For instance, [16, 23, 37, 52, 55] integrated environmental elements into human motion generation, while [5, 7, 26, 43, 49, 51, 54] utilized textual guidance to facilitate the generation of in- + +teractive motions for pairs or groups. However, 3D data is challenging to obtain and lacks user-friendliness. Moreover, certain emotional or event states cannot be directly represented through 3D data. Compared to other methods, our approach offers greater flexibility in inputs and can cover a wider range of scenarios and contexts. + +# 3. Dataset: HUMANML3D++ + +We introduce HUMANML3D++, a novel dataset designed to support more flexible Arbitrary Text to Motion by providing both Action and Scene Texts for motion data. As shown in Table 1, HUMANML3D++ includes over 135k scene texts paired with matching motions, making it the first annotated motion dataset with dual-text support. + +Dataset Construction. (1) Data Collection: we select HumanML3D [9] as our foundation dataset. The selection is based on the dataset's inclusion of high-quality action annotations and a large volume of motion data, aiming to reduce annotation costs and ensure diversity. (2) Prompt Strategy Design: we design different prompts and employ expert scoring to assess the generated outcomes of various prompts. A comparison under different prompts is presented in Table 7. Ultimately, we choose the causal context-guided prompt as the final strategy for generating data. (3) Data Generation: For each motion in the source dataset, which comprises 3-5 Action Texts, we use LLMs to generate more than six Scene Texts per motion to ensure comprehensive data coverage due to the inherent multi-solution nature of the problem. + +Data Quality Control. To ensure high quality and reliability of the data, we implement a series of denoising and validation measures. (1) Filtering of Invalid Content: We use length checks and fixed-format matching to filter out invalid generated content. (2) Text Formatting: All texts are processed uniformly according to the format standards + +![](images/8b34d4489ad8f57a9a35e51cb2ad12b5397e67d78774da547dc6f616d3051fd7.jpg) +Figure 3. Pipeline overview. (Dataset) We extend HUMANML3D [9] to the novel HUMANML3D++ with Scene Texts. (TAAT) We utilize a fine-tuned LLM to generate multiple reasonable response action instructions for a single Scene Text. And we generate each action in an action instruction individually and utilize the code generated in the previous stage to guide the generation in the subsequent stage. (Evaluation) We introduce two new metrics, Hit Accuracy and Mean Hit Distance, to measure this multi-solution task. + +of HUMANML3D. (3) Cross-validation with Large Language Models: We employ a LLM, independent from the data collection phase, for cross-validation. This LLM comprehensively evaluates the generated scene text annotations based on predefined criteria, determining their appropriateness and flagging any anomalous texts for review. (4) Manual Evaluation and Iterative Optimization: We invite 20 participants to assess a randomly selected subset of data (15% of the total). All participants undergo standardized training. Each scene text is evaluated by at least three participants. A scene text is deemed acceptable if all three evaluators agree it is contextually appropriate, free from ambiguity and mismatches. The iterative improvement process continues until the validation accuracy exceeds $95\%$ . + +# 4. Method + +Given scene or action textual inputs, our objective is to generate realistic human motion $\mathbf{X} = [x_{1}, x_{2}, \ldots, x_{t}]$ , where each $x_{t} \in \mathbb{R}^{d}$ represents the human body pose in a $d$ -dimensional space at frame $t$ . As illustrated in Figure 3, the framework adopts a dual-phase architectures. In the Think phase, we harness the emergent properties of large language models to delineate the underlying relationship between textual inputs and action instructions. We synthesize the temporal sequence modeling capabilities inherent in Transformers with discrete action representations, establishing a bijective relationship in the Act phase. This dual-phase architecture decouples text comprehension from motion generation, enabling thorough scene understanding before generating precise motion sequences, which is effective for handling the multi-solution nature of the task. + +# 4.1. Think Model + +Modeling and learning from arbitrary texts, especially Scene Texts, differ significantly from previous tasks due to the multi-solution nature. This can be mathematically formulated as $p(\mathbf{X} \mid \mathbf{T})$ , where $\mathbf{T}$ represents the arbitrary text and $\mathbf{X}$ represents the corresponding motion sequence. The + +mapping from $\mathbf{T}$ to $\mathbf{X}$ is conditional and non-deterministic, allowing multiple motion sequences to arise from the same text. For this complex situational comprehension, large language models represent an optimal approach due to their advanced linguistic understanding and heightened contextual sensitivity when engaging with complex textual data. Through the refinement of pre-trained architectures using structured input-output pairs that encompass Scene Texts and associated responses, we can more adeptly leverage the vast training data inherent to LLMs. Formally, the objective can be defined as minimizing the following loss function over the dataset $\mathcal{D}$ : + +$$ +\mathcal {L} _ {\mathrm {L L M}} = \sum_ {(Q, A) \in \mathcal {D}} \mathbb {E} _ {Q} \left[ \log p (A \mid \theta_ {\mathrm {L L M}} (Q) \right], \tag {1} +$$ + +$$ +\mathcal {D} = \left\{\left(Q _ {i}, A _ {i}\right) \right\} _ {i = 1} ^ {N}, \tag {2} +$$ + +where $p(A \mid Q; \theta_{\mathrm{LLM}})$ represents the probability of generating the correct response $A$ given the question $Q$ under the LLM parameters $\theta_{\mathrm{LLM}}$ . $\mathcal{D}$ denotes the data used for fine-tuning, with $Q$ is Scene Text input and $A$ is Action Texts GT output. To mitigate resource consumption, we employ LoRA [13] to fine-tune the LLM. In alignment with established practices for large language models, we utilize cross-entropy loss to enforce similarity between the predicted tokens $\mathcal{A}$ and the ground truth tokens $\mathcal{A}^{gt}$ , formally represented as follows: + +$$ +\mathcal {L} _ {\mathrm {t o k e n}} = \operatorname {C E} (\mathcal {A}, \mathcal {A} ^ {\mathrm {g t}}) = - \sum_ {i = 1} ^ {n} \mathcal {A} _ {i} ^ {\mathrm {g t}} \log (\mathcal {A} _ {i}), \tag {3} +$$ + +where $\mathcal{A}_i$ represents the predicted probability distribution over tokens at position $i$ , and $\mathcal{A}_i^{\mathrm{gt}}$ is the ground truth token distribution. The fine-tuned LLMs extract action instruction sets, formulated as $A = \theta_{\mathrm{LLM}}(c,P) = (a_1,a_2,\ldots ,a_x)$ , where $P$ is the prompt used in inference, incorporating additional requirements for extracting the action instruction. The extracted instructions are then fed into the Act module to generate the corresponding motion. + +# 4.2. ACT Model + +We employ VQ-VAE [48] to discretely encode the motion, enabling efficient motion representation learning. The encoder and decoder are denoted as $E$ and $D$ . For a human motion $\mathbf{X} = [x_{1}, x_{2}, \ldots, x_{t}]$ , the latent feature $\mathbf{Z} = E(\mathbf{X})$ is represented as $\mathbf{Z} = [z_{1}, z_{2}, \ldots, z_{t / l}]$ , where $l$ is the temporal downsampling rate of the encoder $E$ . Quantization of each latent feature $z_{i}$ entails its mapping to the nearest element $c_{k}$ within the codebook $C$ , defined as: + +$$ +\hat {z} _ {i} = \underset {c _ {k} \in C} {\arg \min } \| z _ {i} - c _ {k} \| _ {2}. \tag {4} +$$ + +We use the motion tokenizer similar to T2M-GPT [57]. With a learned motion VQ-VAE, a human motion sequence $\mathbf{X}$ can be mapped to discrete indices $\mathbf{I} = [i_1,i_2,\dots ,i_{t / l},\mathrm{End}]$ , where $i_{i}\in [1,2,\ldots ]$ represents an index from the learned codebook. Each index $i_j$ corresponds to a codebook entry $c_{i_j}$ , defining the quantized representation for that segment of the motion sequence. The generation is formulated as an autoregressive next-index prediction task: given previous $t - 1$ indices $(\mathbf{I} < t)$ and scene text condition $c$ , our objective is to predict the next indices $p(i_t|\theta_{\mathrm{LLM}}(c),\mathbf{I} < t)$ , where $\theta$ represents the Think Model trained in the first stage, with its parameters frozen. The training optimization goal is defined by denoting the likelihood of the full sequence as $p(\mathbf{I}|c) = \prod_{i = 1}^{T / l}p(I_i|\theta_{\mathrm{LLM}}(c))$ . We directly maximize the log-likelihood of the data distribution: + +$$ +\mathcal {L} _ {\text {t r a n s}} = \mathbb {E} _ {\mathbf {I} \sim p (\mathbf {I})} [ - \log p (\mathbf {I} | \theta_ {\mathrm {L L M}} (c)) ]. \tag {5} +$$ + +For the several sets of action instructions obtained, motions are generated by following an autoregressive process over each action instruction sets. For an action instruction sets $A = \theta_{\mathrm{LLM}}(c,P) = (a_1,a_2,\ldots ,a_x)$ , we define the generation of each action with the following: + +$$ +\mathbf {I} _ {x} = \left\{ \begin{array}{l l} f \left(\left\{a _ {x}, \text {n u l l} \right\}\right) & \text {i f} x = 1, \\ f \left(\left\{a _ {x}, \mathbf {I} _ {x - 1} [ - n: ] \right\}\right) & \text {i f} x > 1, \end{array} \right. \tag {6} +$$ + +where $a_{x}$ denotes the current action, and $\mathbf{I}_{x - 1}[-n:]$ indicates the last $n$ indices generated by the previous action. For the first action, an empty ID list is used. For all subsequent actions, the action instruction and the most recent $n$ indices from the prior action are used to predict the next index. The generation for each action instruction $a_{x}$ begins from the initial embedding and proceeds autoregressively until the model predicts the End token. If an End token of the current action instruction is encountered, we replace the action instruction to update it. Upon completion of each action segment, the indices are concatenated to form $\mathbf{I}_{\mathrm{total}} = [i_1,i_2,\dots ,i_k]$ , where $k$ is the number of action segments. Finally, the entire sequence $\mathbf{I}_{\mathrm{total}}$ is decoded by the VQ-VAE to generate a cohesive and smooth motion. + +# 5. Metrics + +Current evaluation metrics [9] have limitations when applied to multi-solution task. (1) R-precision and MM + +
TaskR-Precision↑MM-Dist↓Diversity↑
Action Texts0.797±.0022.974±.0089.503±.065
Scene Texts0.665±.0033.945±.0008.435±.069
+ +Table 2. Comparison of metrics for GT motion in two tasks reveals poor accuracy, highlighting the failure of current metric reliability. + +Dist are limited by their requirement for the model's output to match a single ground truth, neglecting the possibility of output motions that match the Scene Text but differ from the dataset's GT. (2) According to [9], "R-Precision" and "MM-Dist" rely on Motion & Text feature extractors for alignment. We retrain the feature extractors on HumanML3D++, with the results presented in Table 2. Experiments demonstrate that, when utilizing scene text as input, the existing evaluation system exhibits a misjudgment rate approaching $40\%$ even when the generated motions are the same as the ground truth, indicating a failure in the reliability of the metrics. Inspired by QA task, we evaluate whether the model covers reference answers (HA) and measure the relevance of all generated outputs (not a single output as previous metrics) to the reference answers (MHD): + +$M$ represents the total number of instances in the test. For each instance $i$ , the model generates $N$ responses $R_{i} = \{R_{i}^{(1)}, R_{i}^{(2)}, \ldots, R_{i}^{(N)}\}$ . $G$ represents the ground truth, and the distance between the $j$ -th response of instance $R_{i}$ and the ground truth $G$ is quantified by $d_{ij} = \mathrm{dist}(R_{i}^{(j)}, G_{i})$ . A threshold, $\theta$ , is employed to determine whether a response qualifies as a "hit" based on its distance to the ground truth. We define an indicator function $\delta: \mathcal{R} \times \{G\} \to \{0, 1\}$ to determine whether a response meets the "hit" criterion: + +$$ +\delta_ {i j} = \delta \left(R _ {i} ^ {(j)}, G _ {i}\right) = \left\{ \begin{array}{l l} 1, & \text {i f} \operatorname {d i s t} \left(R _ {i} ^ {(j)}, G _ {i}\right) \leq \theta , \\ 0, & \text {i f} \operatorname {d i s t} \left(R _ {i} ^ {(j)}, G _ {i}\right) > \theta . \end{array} \right. \tag {7} +$$ + +Hit Accuracy(HA): Hit Accuracy measures the ratio of hits among total data. For a given text, the model generates N responses motion, each compared to ground truth motion. A response qualifies as a hit if its distance to the ground truth exceeds a defined threshold $\theta$ . + +$$ +\text {H i t A c c u r a c y} = \frac {1}{M} \sum_ {i = 1} ^ {M} \max _ {j = 1, \dots , N} \delta \left(R _ {i} ^ {(j)}, G _ {i}\right). \tag {8} +$$ + +Mean Hit Distance (MHD): Mean Hit Distance measures the average hit distance across multiple instances. For each instance, when hit, the hit distance is defined as the average distance of the qualifying responses; otherwise, a representative distance of the instance's responses is used. + +$$ +\mathrm {M H D} = \frac {\sum_ {i = 1} ^ {M} \sum_ {j = 1} ^ {N} \delta_ {i j} \cdot d _ {i j} + \sum_ {i = 1} ^ {M} \left(\mathbb {I} \left(\sum_ {j = 1} ^ {N} \delta_ {i j} = 0\right) \cdot \min d _ {i j}\right)}{\sum_ {i = 1} ^ {M} \sum_ {j = 1} ^ {N} \delta_ {i j} + \sum_ {i = 1} ^ {M} \mathbb {I} \left(\sum_ {j = 1} ^ {N} \delta_ {i j} = 0\right)}. \tag {9} +$$ + +We ensure a fair comparison with existing methods by employing consistent metrics for Action Texts to Motion Task. + +
MethodTypeTrainTestHit Accuracy ↑FID ↓Mean Hit Distance ↓Diversity ↑MModality ↑
TM2T [10]Zero-shot○●65.52.201±.0201.4077.286±.0752.600±.094
Trained●●68.81.394±.0001.3578.181±.0002.701±.000
MDM [45]Zero-shot○●64.40.827±.0531.4258.249±.0582.804±.052
Trained●●73.40.435±.0291.2478.634±.0572.901±.055
MLD [3]Zero-shot○●61.80.897±.0261.5479.289±.0963.018±.028
Trained●●72.49.408±.0601.3136.962±.0633.086±.130
MotionDiffuse [58]Zero-shot○●70.21.514±.0001.2367.907±.0001.813±.000
Trained●●77.92.688±.0001.1367.703±.0003.191±.000
T2M-GPT [57]Zero-shot○●73.90.516±.0421.1789.396±.2322.499±.348
Trained●●79.60.316±.0151.0818.627±.0802.620±.067
TAAT(Ours)Zero-shot○●75.10.488±0.0061.2128.552±.0952.957±.070
w/o filter●●78.60.420±.0191.1958.870±.0433.167±.053
Trained●●79.90.379±.0141.0758.950±.0953.046±.070
+ +Table 3. Experiment Results on Model Zero-Shot Ability and Scene Text to Motion. Our model demonstrates optimal performance with new scene texts in Zero-Shot experiment. After being trained on HUMANML3D++ (Trained row), models show improved results on scene texts. Our TAAT excels in both Hit Accuracy and Mean Hit Distance, indicating a superior understanding of scene texts. Furthermore, the achieved FID and Diversity metrics suggest that we can generate high-quality motions that align well with real-world motions. + +![](images/8bab19c010d7262f58033b12f8524c1739e5bab06e5cde9c92bf5e4d37bf540f.jpg) +Figure 4. Visual comparisons on scene texts. Our model generates context-appropriate responses (pointing to the sunset, taking a photo), while other models display irrelevant actions (looking down) or remain inactive (standing still). + +![](images/ab434273bc69efaf71bfa97ac00df3824ebe90a885a03eb392801234eeb44475.jpg) +Scene Text: A person sees a beautiful sunset while camping in the wilderness. + +# 6. Experiment + +# 6.1. Dataset and Implementation Details + +Dataset. We use the Scene Text annotations of HUMANML3D++ to benchmark the model's performance on the Scene Text to Motion, while HUMANML3D [9] is used to evaluate performance on the Action Text to Motion. + +Implementation Details. We use the LLaMA model as the base language model and fine-tune it using LoRA [13]. The maximum number of training epochs is set to 10, with a maximum sample size of 100,000 and a maximum input sequence length of 2048. The initial learning rate is set to $10^{-4}$ , and a cosine learning rate scheduler is employed, with a warmup ratio of 0.1. Additionally, the validation set is allocated $10\%$ of the data, and the rank for LoRA [13] is set to 8. Given memory constraints, the batch size is set to 16. The optimization process uses the AdamW optimizer + +with a weight decay parameter of 0.01. For the generative model, we utilize 18 transformer layers with a hidden dimension of 1,024 and 16 attention heads. The transformer is optimized using AdamW optimizer, a batch size of 128, an initial learning rate of $1 \times 10^{-4}$ for 150K iterations, and a decay to $5 \times 10^{-6}$ for an additional 150K iterations. + +# 6.2. Scene Text to Motion + +Table 3, Figure 4 and Figure 5 show the model's ability to learn Scene Texts and generate responsive motions. All models are retrained and tested on HUMANML3D++. Our TAAT demonstrates the best Hit Accuracy and Mean Hit Distance, indicating effective extraction of potential motions from Scene Texts and generation of contextually aligned responses. This is achieved through scene interpretation, utilizing LLMs' deep language comprehension and diverse candidate generation capabilities to produce multi- + +Scene Text: A mountain climber reaches the summit of the mountain, and he is greatly amazed by the scenery all around. + +Scene Text: A person suddenly comes across a bear emerging from the bushes while on a mountain path. + +![](images/a13459d9130a0eb833d93387c05bac6da7918f60674d81016978d7ff135f822c.jpg) + +![](images/70da20a69ff227dacf26507906fea8ca7cad00835edfbbc586d38b5b06b8a5c7.jpg) + +![](images/45a3502d29511e0727cab8533d4bc70da6a5865539a5aa7ad3eed241d603989f.jpg) + +![](images/aa16355893913c203167085c72f73d5992b375d3051907b792561dc9867bcb58.jpg) + +Scene Text : A person suddenly hears a loud bang on the street. + +Scene Text : A person discovers a small squirrel leaping on + +Figure 5. For a Scene Text, TAAT can generate various reasonable motions. +![](images/7494e310b6a70188eeada995cd86801ec8c2a582dd1efc8fa77f0414057e9249.jpg) +(Turn around and look around) (Crouch down and press closely against the ground.) + +![](images/295afb1f35c100a2cbb3b4b45eb16381db777b8d63afcd91976a9dbf414bdbe0.jpg) + +![](images/998420c78c06741334a6ac4823c7acba76aae7d22f62f2c901bd00d51ac915da.jpg) +(jump up for a clearer view) + +![](images/956592309caa024b8d3de6e559e82cfa65622dc7d05a225cf2718983411b8cd8.jpg) +(run closer and point to friends) + +ple plausible action interpretations. Moreover, our model achieves the highest Diversity and is ranked second in FID and MModality, indicating its ability to produce diverse realistic human poses. T2M-GPT [57]exhibits poor performance in MModality, indicating that the model is still learning limited mapping relationships within the dataset. Consequently, it struggles to generate diverse and reasonable reactive motions for the same Scene Texts. + +# 6.3. Action Texts to Motion + +Table 4 and Figure 6 present the results on Action Texts to Motion. Both training and test are performed on HumanML3D [9]. In this task, we use the Think model to directly extract action instructions and generate motions. Although not specifically designed for the Action Texts to Motion, it performs well in Diversity and FID, generating diverse and realistic motions. This is achieved by independently sampling each action instruction rather than for the entire sentence, increasing the number of choices. Furthermore, as shown in Figure 6, the ACT model's generation logic ensures the complete generation of all actions, suitable for multi-action generation. + +# 6.4. Zero-shot Experiments + +Table 3 illustrates our testing on the model's zero-shot capabilities. All models are trained on the HUMANML3D [9] and tested on the HUMANML3D++ to test whether the current models can understand and generate motion from Scene Texts without pretraining. Table 3 shows that existing models exhibit a decrease in multiple metrics when directly using Scene Texts inputs without pretraining. This indicates that these models primarily learn action knowledge from the existing dataset and struggle to generalize to unseen scene + +
MethodsR-Precision (Top-3) ↑FID ↓Diversity ↑
Real motion0.797±.0020.002±.0009.503±.065
TM2T [10]0.740±.0031.067±.0029.188±.002
MDM [45]0.611±.0070.544±.0449.599±.086
MLD [3]0.772±.0020.473±.0139.724±.082
MotionDiffuse [58]0.782±.0010.630±.0019.410±.049
T2M-GPT [57]0.685±.0030.140±.0069.844±0.095
TAAT(Ours)0.696±.0030.461±.00610.038±.095
+ +Table 4. Experiment on Action Texts to Motion. Our TAAT works well in FID, Diversity, demonstrating that our model can generate realistic and diverse motions that are close to real human motions. Furthermore, TAAT is suitable for multi-action generation in Fig 6. + +
TM2T [10]MDM [45]MLD [3]MD [58]T2M-GPT [57]Ours
A1.1340.2830.4240.8840.3760.027
B0.327-0.1098.9352.0580.176-0.082
+ +Table 5. FID variation across three experiments. Our model exhibits minimal variation in FID across different tasks. (A) shows each model's FID change from Zero-shot to Action Text; (B) from Scene Text to Action Text. + +texts. Table 5 presents the variations in FID scores of different models across three experiments(data from table 3, 4). Despite not being trained explicitly on Scene Texts, our model achieves the best FID and Hit Accuracy and exhibits a comparatively minor decrease in FID when presented with Scene Texts inputs. Furthermore, it achieves the best FID and satisfactory Diversity, demonstrating the ability to generate high-quality and diverse human motions. + +# 6.5. Ablation Study + +Dataset Filtering. We perform an ablation study by removing the dataset filtering, as shown by comparing the "w/o fil + +![](images/9ce0e6762e7c1dd8fe10d2a0f0ba142476e1a7b359ba23a2f8b58f84b37f56c3.jpg) +Action Texts: A person walked forward, then jumped, and then bent down + +![](images/70e94eb81f7a80027efb2477d649b632fb2d5daa06620d11271b31bffe3f338e.jpg) +Figure 6. Visual results on Action Texts. Only our model performs all actions in the correct sequence, while other models exhibit issues such as missing actions (MDM [45], MLD [3]), sequence disorder (T2M-GPT [57]), and spatial relationship errors (MLD [3]). + +ter” and “Retrain” rows in the last two lines of Table 3. Applying the filter increases Hit Accuracy and reduces Mean Hit Distance, indicating that the filtering effectively eliminates noise and irrelevant samples, thereby enhancing the overall quality and relevance of the dataset. + +Comparison of different LLMs. We use LLMs to construct HUMANML3D++. To evaluate the generation capabilities of different LLMs, we employ multiple LLMs to generate the same set of 100 scene texts based on the corresponding action texts, which are then evaluated by 23 assessors. As performance differences are negligible, we do not factor in minor variations when selecting a LLM. We ultimately choose Gemini considering both quality and cost. + +
LLMExcell.↑Sut.↑Inap.↓
GPT-3.516.3%74.2%9.5%
GPT-4o25.9%73.7%0.4%
Llama-3-8B18.3%74.6%7.1%
Gemini (Ours)23.5%75.9%0.6%
+ +Prompt design. Table 7 illustrates the impact of different prompt designs on the Scene Text generation. Acc. denotes the accuracy of the generated texts, while Score represents the human evaluation score. Including all features (last row) achieves the highest performance, indicating that our prompt design enhances the robustness and effectiveness of the generated outputs in dataset construction. In contrast, omitting specific features, particularly Verbs or Examples, leads to a noticeable decline in output quality. + +Index Length. We evaluate the impact of index lengths on the generated motions in Figure 7 and found that an optimal index length of 7 yielded the highest generation quality. A shorter index often leads to discontinuities and unnatural postures, while a longer index tends to result in excessive + +Table 6. "Excell," "Sut," and "Inap." represent the ability of different models to generate scene texts from action texts, as perceived by users, corresponding to "Excellent," "Suitable," and "Inappropriate" levels, respectively. + +
w/ VerbQuantityFew-shotCausalityAcc.↑Score↑
○●●●58.21.80
●○●●76.43.75
●●○●37.61.40
●●●○81.83.80
●●●●98.45.00
+ +Table 7. Comparison of prompt strategies. $w/Verb$ indicates verb usage restrictions; Quantity denotes quantity requirements; Few-shot indicates inclusion of examples; and Causality reflects descriptions of causal relationships. + +pose repetition and increased computational costs. + +![](images/92b1381531c945e42f773a648810db580893b9c19ea44785d221a8ad1548d033.jpg) +Figure 7. Impact of index lengths on motion generation. + +# 7. Conclusion + +In summary, this study introduces a novel task: inferring potential motions from Arbitrary Texts (including those with no explicit action labels), which has not been previously explored. Additionally, we propose a new dataset HumanML3D++ and a more practical think-and-act framework TAAT. To improve evaluation, we introduce multisolution metrics specifically designed for this novel task. We conduct extensive experiments to fully investigate the performance and zero-shot capabilities of existing models across the two tasks: Action Texts to Motion and Scene Texts to Motion. Our research establishes an essential foundation for future investigations in this domain. + +# Acknowledgements + +This work is supported by Hubei Provincial Key Research and Development Program (2024BAB050, 2024BAB039). + +# References + +[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), 2019. 2 +[2] Andreas Aristidou, Anastasios Yiannakidis, Kfir Aberman, Daniel Cohen-Or, Ariel Shamir, and Yiorgos Chrysanthou. Rhythm is a dancer: Music-driven motion synthesis with global structure. IEEE Transactions on Visualization and Computer Graphics, 2022. 2 +[3] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 6, 7, 8 +[4] Bruno Degardin, Joao Neves, Vasco Lopes, Joao Brito, Ehsan Yaghoubi, and Hugo Proença. Generative adversarial graph convolutional networks for human action synthesis. 2022. 2 +[5] Christian Diller and Angela Dai. Cg-hoi: Contact-guided 3d human-object interaction generation. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2024. 3 +[6] Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. 2 +[7] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In European Conference on Computer Vision (ECCV), 2024. 3 +[8] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, 2020. 2 +[9] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 3, 4, 5, 6, 7 +[10] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision(ECCV), 2022. 3, 6, 7 +[11] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3 +[12] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-between. ACM Transactions on Graphics (TOG), 2020. 2 + +[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 4, 6 +[14] Zechao Hu, Zhengwei Yang, Hao Li, Yixiong Zou, Fengbin Zhu, and Zheng Wang. Unified category and style generalization for instance-level sketch retrieval. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2025. 2 +[15] Yanli Ji, Feixiang Xu, Yang Yang, Fumin Shen, Heng Tao Shen, and Wei-Shi Zheng. A large-scale rgb-d database for arbitrary-view human action recognition. In Proceedings of the 26th ACM international Conference on Multimedia, 2018. 2 +[16] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3 +[17] Peng Jin, Yang Wu, Yanbo Fan, Zhongqian Sun, Wei Yang, and Li Yuan. Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs. Advances in Neural Information Processing Systems, 2024. 2, 3 +[18] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[19] Hanyang Kong, Kehong Gong, Dongze Lian, Michael Bi Mi, and Xinchao Wang. Priority-centric human motion generation in discrete latent space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3 +[20] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 2019. 2 +[21] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 2 +[22] Fei Li, Wenxuan Liu, Jingjing Chen, Ruixu Zhang, Yuran Wang, Xian Zhong, and Zheng Wang. Anomize: Better open vocabulary video anomaly detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2 +[23] Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[24] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[25] Han Liang, Jiacheng Bao, Ruichi Zhang, Sihan Ren, Yuecheng Xu, Sibei Yang, Xin Chen, Jingyi Yu, and Lan Xu. + +Omg: Towards open-vocabulary motion generation via mixture of controllers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[26] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, 2024. 2, 3 +[27] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 2019. 2 +[28] Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-vocabulary text-to-motion generation. In European Conference on Computer Vision, 2024. 2 +[29] Qiujing Lu, Yipeng Zhang, Mingjian Lu, and Vwani Roychowdhury. Action-conditioned on-demand motion generation. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2 +[30] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 3 +[31] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. 2022. 2 +[32] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 2 +[33] Yue Ma, Yulong Liu, Qiyuan Zhu, Ayden Yang, Kunyu Feng, Xinhua Zhang, Zhifeng Li, Sirui Han, Chenyang Qi, and Qifeng Chen. Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning. arXiv preprint arXiv:2506.05207, 2025. 2 +[34] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, 2019. 2 +[35] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[36] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision(ECCV), 2022. 2 +[37] Huaijin Pi, Sida Peng, Minghui Yang, Xiaowei Zhou, and Hujun Bao. Hierarchical generation of human-object interactions with diffusion probabilistic models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3 +[38] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. Big Data, 2016. 2 + +[39] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[40] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jamei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 2 +[41] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 2 +[42] Jiangxin Sun, Chunyu Wang, Huang Hu, Hanjiang Lai, Zhi Jin, and Jian-Fang Hu. You never stop dancing: Non-freezing dance generation via bank-constrained manifold projection. Advances in Neural Information Processing Systems, 2022. 2 +[43] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, 2023. 2, 3 +[44] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision(ECCV), 2022. 2 +[45] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 3, 6, 7, 8 +[46] Likai Tian, Zhengwei Yang, Zechao Hu, Hao Li, Yifang Yin, and Zheng Wang. Expressiveness is effectiveness: self-supervised fashion-aware clip for video-to-shop retrieval. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024. 2 +[47] Likai Tian, Jian Zhao, Zechao Hu, Zhengwei Yang, Hao Li, Lei Jin, Zheng Wang, and Xuelong Li. Ccin: Compositional conflict identification and neutralization for composed image retrieval. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2 +[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 2017. 5 +[49] Runqi Wang, Caoyuan Ma, Jian Zhao, Hanrui Xu, Dongfang Sun, Haoyang Chen, Lin Xiong, Zheng Wang, and Xuelong Li. Leader and follower: Interactive motion generation under trajectory constraints. arXiv preprint arXiv:2502.11563, 2025.3 +[50] Yin Wang, Zhiying Leng, Frederick WB Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[51] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 2, 3 + +[52] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3 +[53] Haiyang Xie, Zhengwei Yang, Huilin Zhu, and Zheng Wang. Striking a balance: Unsupervised cross-domain crowd counting via knowledge diffusion. In Proceedings of the 31st ACM International Conference on Multimedia, 2023. 2 +[54] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter- $\mathbf{x}$ : Towards versatile human-human interaction analysis. In CVPR, 2024. 3 +[55] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. InterDiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023. 3 +[56] Ping Yu, Yang Zhao, Chunyuan Li, Junsong Yuan, and Changyou Chen. Structure-aware human-action generation. 2020. 2 +[57] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 5, 6, 7, 8 +[58] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 2, 3, 6, 7 +[59] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottongpt: Finetuned llms are general-purpose motion generators. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 2 +[60] Zhimin Zhang, Bi'an Du, Caoyuan Ma, Zheng Wang, and Wei Hu. Behave your motion: Habit-preserved cross-category animal motion transfer. arXiv preprint arXiv:2507.07394, 2025. 2 +[61] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3 +[62] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 \ No newline at end of file diff --git a/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/images.zip b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3d5adaf150a1cc2fe5dc612b6ed7aa0eac394b94 --- /dev/null +++ b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbe2ef9b72db922a14d5d1562ed24c6f7737a22425c7671f705b14b70ff52cc5 +size 727521 diff --git a/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/layout.json b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..809759686ab76bf73804039bb4d3e34821aac20d --- /dev/null +++ b/youthinkyouactthenewtaskofarbitrarytexttomotiongeneration/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c01ab62ac24322e301e901e658171e6800f9b3874273c2b71d9a0cdce5b79d1 +size 401666 diff --git a/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/c8281586-b056-41a9-b08b-fbc46de86111_content_list.json b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/c8281586-b056-41a9-b08b-fbc46de86111_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1d0b471df74e66e3c69196bfb5d2a4e358a654fe --- /dev/null +++ b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/c8281586-b056-41a9-b08b-fbc46de86111_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5d74a1b8e4bfc7d6c87ac7a3b0a22c38fd3273e1b28f254ac671891376acf9f +size 87181 diff --git a/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/c8281586-b056-41a9-b08b-fbc46de86111_model.json b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/c8281586-b056-41a9-b08b-fbc46de86111_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39c84aa0af564426c6b9aa202fba59853c595dbf --- /dev/null +++ b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/c8281586-b056-41a9-b08b-fbc46de86111_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccd751a2f5afe7ee35b76df9fd4dc8a4d0ca703480053330a338d16139608533 +size 107602 diff --git a/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/c8281586-b056-41a9-b08b-fbc46de86111_origin.pdf b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/c8281586-b056-41a9-b08b-fbc46de86111_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fb9b9286541bc0167e4d185c70e7f0822dd474a6 --- /dev/null +++ b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/c8281586-b056-41a9-b08b-fbc46de86111_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e36aa84deae649a98a0bc5c716bbb0e81517c70c15b97a1f41d0a45ab41544e +size 552601 diff --git a/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/full.md b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..66cabd969e1d03e5b2015cbdeb80f3ff9eac058c --- /dev/null +++ b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/full.md @@ -0,0 +1,283 @@ +# Zero-AVSR: Zero-Shot Audio-Visual Speech Recognition with LLMs by Learning Language-Agnostic Speech Representations + +Jeong Hun Yeo $^{1,*}$ Minsu Kim $^{1,*}$ Chae Won Kim $^{1}$ Stavros Petridis $^{2}$ Yong Man Ro $^{1,\dagger}$ $^{1}$ KAIST, $^{2}$ Imperial College London + +{sedne246, ms.k, ymro}@kaist.ac.kr + +# Abstract + +We explore a novel zero-shot Audio-Visual Speech Recognition (AVSR) framework, dubbed Zero-AVSR, which enables speech recognition in target languages without requiring any audio-visual speech data in those languages. Specifically, we introduce the Audio-Visual Speech Romanizer (AVRomanizer), which learns language-agnostic speech representations by predicting Roman text. Then, by leveraging the strong multilingual modeling capabilities of Large Language Models (LLMs), we propose converting the predicted Roman text into language-specific graphemes, forming the proposed Cascaded Zero-AVSR. Taking it a step further, we explore a unified Zero-AVSR approach by directly integrating the audio-visual speech representations encoded by the AV-Romanizer into the LLM. This is achieved through finetuning the adapter and the LLM using our proposed multitask learning scheme. To capture the wide spectrum of phonetic and linguistic diversity, we also introduce a Multilingual Audio-Visual Romanized Corpus (MARC) consisting of 2,916 hours of audio-visual speech data across 82 languages, along with transcriptions in both language-specific graphemes and Roman text. Extensive analysis and experiments confirm that the proposed Zero-AVSR framework has the potential to expand language support beyond the languages seen during the training of the AV-Romanizer. Code is available at https://bit.ly/zero-avsr. + +# 1. Introduction + +Humans rely on multi-modal information for effective communication, combining verbal cues (e.g., spoken words), nonverbal signals (e.g., facial expressions, gestures), and paralinguistic auditory cues (e.g., tone of voice) to convey meaning effectively. Notably, the correlation between auditory speech and visual speech (i.e., lip movements) can significantly enhance speech comprehension, particularly in noisy environments. Leveraging this advantage, numerous + +studies [1-14] have explored Audio-Visual Speech Recognition (AVSR) using deep learning techniques. AVSR can be viewed as a multi-modal fusion of two uni-modal systems; Auditory Speech Recognition (ASR) [15-17] and Visual Speech Recognition (VSR) [18-24]. Prior works have demonstrated that jointly modeling both audio and visual speech representations can significantly improve the speech recognition performance of each uni-modal system, particularly in challenging noisy environments. + +Despite significant advancements, AVSR has primarily been developed and evaluated on English language corpora. To extend its capabilities to a wider range of languages, recent efforts have proposed multilingual audio-visual speech datasets [25, 26], and have developed and evaluated multilingual AVSR methods [27-29] on such databases. However, despite these performance improvements, existing approaches still face limitations in language expansion. This is because current multilingual AVSR datasets [25, 26] contain only nine languages, and it remains challenging to obtain a sufficient amount of labeled audio-visual data with transcriptions for diverse languages. To address this limitation, our main explorations include: 1) learning language-agnostic audio-visual speech representations to facilitate language expansion and 2) proposing a multilingual audio-visual database encompassing 82 languages, significantly expanding the linguistic diversity of existing datasets. + +In this paper, we present a novel AVSR method that exhibits zero-shot language recognition ability $^1$ , i.e., a scenario in which no speech data in the target language is used at training. To achieve this, we aim to represent all languages in Roman text, converting language-specific graphemes into language-agnostic pronunciations. Here, we note that pre-trained Large Language Models (LLMs) [30-32] already possess knowledge of modeling + +this Roman-grapheme mapping, and propose to leverage this comprehensive multilingual processing ability of LLMs in our proposed zero-shot AVSR framework. Concretely, we propose an Audio-Visual Speech Romanizer (AVRomanizer), which predicts Roman text (i.e., language-agnostic) from input multilingual audio-visual speech data. We then demonstrate that the proposed AV-Romanizer can be directly employed to achieve zero-shot AVSR by converting the predicted Roman text into language-specific graphemes using a pre-trained LLM, even when the target language is not used during the training of the AVRomanizer. We refer to this usage case as Cascaded Zero-AVSR, where the proposed AV-Romanizer and a pre-trained LLM form a cascaded system. + +Going one step further, we explore a model that holds considerable promise, namely Zero-AVSR, where the LLM is also fine-tuned to match the specific use case for zero-shot recognition, thereby improving performance. Specifically, Zero-AVSR is trained in a multi-task fashion composed of two tasks. The first task involves aligning the AV-Romanizer and LLM by fine-tuning an adapter and the LLM on seen languages, enabling the audio-visual speech representation obtained from the AV-Romanizer to be seamlessly embedded into the learned text space of the LLM. The second task focuses on learning to deromanize, where the LLM is trained to convert Roman texts into language-specific graphemes using only text data for both seen and unseen languages. Therefore, Zero-AVSR constructs knowledge connecting the audio-visual speech representation with the LLM's embedding space of seen languages at the first task, while extending the learned knowledge to more languages, including unseen languages, at the second task. + +In order to train the proposed Zero-AVSR, a sufficient amount of audio-visual speech data is essential to capture the full spectrum of phonetic and linguistic diversity. In this context, we introduce a Multilingual Audio-Visual Romanized Corpus (MARC), which provides Roman transcriptions for approximately 2,916 hours of audio-visual data across 82 languages, along with their corresponding transcriptions in language-specific graphemes. + +The contributions of this paper can be summarized as: + +- We explore Zero-AVSR, the first zero-shot AVSR framework, designed to operate in scenarios where no speech training data is available for the target language. +- We propose the MARC dataset, which comprises Roman transcriptions for 2,916 hours of audio-visual speech data across 82 languages. +- We explore two types of zero-shot AVSR frameworks, one is Cascaded Zero-AVSR, which works with any LLM without fine-tuning, even in API form. The other is Zero-AVSR, an improved framework by finetuning the LLM with the proposed multi-task learning. + +# 2. Related Work + +# 2.1. Audio-Visual Speech Recognition + +Recently, AVSR has gained significant attention for its practical benefits, particularly in enhancing robust speech recognition in noisy environments. Along with the development of large-scale audio-visual datasets [33, 34], early work [2] introduced end-to-end AVSR frameworks based on Bidirectional Gated Recurrent Units (BGRUs). Subsequently, researchers improved these architectures by incorporating Transformer [1, 35] and Conformer [4, 36], leading to notable performance gains. Concurrently, other researchers have explored multimodal learning strategies, including self-supervised learning [6, 19, 28], leveraging knowledge from pre-trained ASR models in AVSR [9, 12], and harnessing the context-modeling capabilities of LLMs for speech recognition [11, 22]. + +Despite these remarkable advances, most existing AVSR research has primarily focused on English. To address this gap, some studies have begun exploring the effectiveness of AVSR in multilingual contexts, leveraging newly introduced multilingual audio-visual databases [25, 26] spanning nine languages. Building on the progress made in English-based AVSR, recent multilingual AVSR approaches [14, 28, 29] have expanded its progress into multilingual AVSR models. In particular, because obtaining labeled multilingual audio-visual data is challenging, self-supervised learning [28] has shown notable promise for improving performance by leveraging abundant unlabeled multilingual audio-visual data. + +While these efforts have successfully extended AVSR's effectiveness to nine languages, it remains challenging to expand its impact to additional languages. Unlike previous work, which mainly focused on improving performance on publicly available multilingual databases, we explore language expansion in multilingual AVSR by constructing a zero-shot AVSR framework. The proposed framework can recognize speech in the target language without requiring any speech training data for that specific language. + +# 2.2. Zero-Shot Speech Recognition + +Recent research in speech recognition has advanced toward supporting multiple languages by employing effective methods such as self-supervised learning, which has been validated on English ASR. This rapid progress has been made possible by the use of large-scale multilingual audio datasets and carefully designed training methods [38, 39]. However, obtaining a sufficient amount of labeled data with transcriptions for all languages remains a challenge. + +To overcome this limitation, researchers have begun to explore zero-shot speech recognition. Early works [40- 42] proposed unsupervised ASR methods leveraging both unlabeled audio and text data. A more efficient ap + +
RomanizerDe-romanizerTarget Language (CER↓)Avg (w/o Eng)
AraDeuEllSpaFraItaPorRus
Llama3.1-70B [30]Llama3.1-70B [30]34.98.416.912.314.010.011.110.913.8
Llama3.1-70B [30]GPT-4o-mini [31]24.63.110.64.02.81.86.66.17.2
Uroman [37]Llama3.1-70B [30]15.04.15.89.84.24.94.414.87.4
Uroman [37]GPT-4o-mini [31]20.01.78.61.02.21.32.08.64.8
GPT-4o-mini [31]Llama3.1-70B [30]11.83.65.512.45.52.93.17.46.0
GPT-4o-mini [31]GPT-4o-mini [31]13.11.97.10.72.51.21.93.94.0
+ +Table 1. Reconstruction test results on MuAViC to evaluate the effectiveness of different methods in romanization and de-romanization. + +proach that requires only unlabeled text was proposed by [43]. This method utilizes language-agnostic allophones, which are subsequently converted into language-specific phonemes [44]. Then, by using a grapheme-to-phoneme (G2P) system [45], they generate words from phoneme sequences. More recently, [46] demonstrates that using a romanized form instead of phonemes offers a simpler yet effective alternative. + +Building on these findings, we also aim to develop a zero-shot AVSR framework by employing romanized text. Unlike [46], which relies on language-specific language models to decode Roman text into language-specific graphemes, we demonstrate that a pre-trained Large Language Model (LLM) can effectively perform this task, thereby eliminating the need for training multiple language-specific language models. + +# 2.3. Speech Large Language Model + +LLMs have significantly impacted Natural Language Processing (NLP), as demonstrated by their ability to perform a wide range of tasks [47]. Building on this success, recent research efforts have begun exploring their effectiveness in other modalities [48-50]. In particular, integrating speech with LLMs has effectively leveraged these models' language understanding capabilities, leading to substantial improvements in the speech domain. To achieve this, simple and effective adaptation methods such as LoRA [51] and a window-level Q-former [52], have been proposed to align audio-based speech with LLMs. Concurrently, these approaches have expanded the range of applications, enabling multi-task capabilities [52, 53] and multi-modal processing [11]. Similar to these research trends, we also explore a unified Zero-AVSR framework that fine-tunes a pre-trained LLM using QLoRA [54], enabling it to directly accept and process encoded audio-visual speech features. + +# 3. Method + +We propose a novel multilingual AVSR framework, ZeroAVSR, that enables zero-shot audio-visual speech recognition even though speech training data for the target language is unavailable. Specifically, we propose the AV-Romanizer, which predicts language-agnostic speech representations + +(i.e., Roman) from input audio-visual speech. Subsequently, we propose to leverage LLMs to convert these representations back into language-specific graphemes. + +# 3.1. Multilingual Audio-Visual Romanized Corpus + +While it is known that different languages share common pronunciation features at the phoneme level [55-57], the current publicly available AVSR datasets [26] may be insufficient for representing the phonemic diversity of diverse languages and for developing language-agnostic audiovisual speech representations. To address this limitation, we propose the Multilingual Audio-Visual Romanized Corpus (MARC). MARC is driven by mixing existing audiovisual corpora, LRS3 [33] and MuAViC [26], and unlabeled audio-visual datasets, VoxCeleb2 [58] and AVSpeech [59]. For the unlabeled datasets, we employ pre-trained language identification and ASR models [39] to obtain the language ID and the language-specific graphemes for each data point, similar to [9, 23, 29]. During language identification, as the pre-trained model may misclassify the language, we filter out annotations with prediction probabilities below a predetermined threshold to reduce the errors. Finally, we convert the transcriptions of the above datasets into Roman text. + +Prior to this, we evaluate which method is most suitable for romanizing transcriptions between using a romanization tool [37] and LLMs. To this end, we perform a reconstruction test, where the ground-truth transcriptions are transformed into Roman text by using either romanization tool or one of the LLMs, and then they are de-romanized into the original transcriptions. However, as romanization is not fully reversible, the de-romanization process is non-trivial. One may construct language-specific lexicons to de-romanize Roman text [46], but this approach cannot fully capture the complex few-to-many mapping of de-romanization. Instead, we employ LLMs as a de-romanizer and find that they already possess the capability to transform between Roman and graphemes freely. Table 1 shows the reconstruction test results using different methods as a romanizer and de-romanizer on MuAViC dataset. Through the test, we found that GPT-4o-mini [31] achieves the best performance when it is utilized as both romanizer and de-romanizer. Therefore, we romanize the transcriptions of the proposed MARC with GPT-4o-mini. + +![](images/01fe5ae76a83d7e7f0323ace1c7cb35fa5bbfe6901afc676fa0edaa1dd298225.jpg) +(a) Audio-Visual Romanizer + +![](images/5200fd4e458da2812752a665e55f5884861b8dd5bf97fccc9a4b31f73189f8f4.jpg) +(b) Cascaded Zero-shot AVSR +Figure 1. Illustration of (a) Audio-Visual Speech Romanizer (AV-Romanizer): It is pre-trained to learn language-agnostic representations by predicting romanized text through CTC loss. (b) Cascaded Zero-shot AVSR: By providing instructions along with the predicted Roman text from the proposed AV-Romanizer, diverse LLMs can be employed to predict graphemes from the predicted Roman text. Please note that, as the AV-Romanizer has learned to generate pronunciation information (Roman text) from the input speech, it can convert even unseen languages into Roman text. Since LLMs already contain information about the target language (key assumption of this paper), they can then convert this Roman text into the target language. + +The resulting dataset, MARC, is composed of 82 languages, and includes 2,916 hours of audio-visual data, and text transcriptions in both language-specific graphemes and Roman text. Detailed information about the dataset can be found in the Supplementary. + +# 3.2. Audio-Visual Speech Romanizer + +If we train an acoustic model to predict the pronunciation of input speech instead of predicting language-specific graphemes, we can employ a fixed number of characters for diverse languages and potentially represent languages even not used during training. Previous approaches [43-45] have often relied on mapping allophones to phonemes. While these approaches can model language-agnostic representations, recent work [46] demonstrates that using a Romanized form offers a simpler yet effective alternative. We also adopt Roman text as our language-agnostic representation. + +The proposed Audio-Visual Speech Romanizer (AV-Romanizer) predicts Roman text for capturing pronunciation as it sounds from the input audio-visual speech, regardless of language, as shown in Fig. 1a. It comprises four main components: an audio encoder $\mathcal{F}_a$ , a visual encoder $\mathcal{F}_v$ , a transformer $\mathcal{B}$ , and a linear layer for predicting Roman text. Given the training sample $(x_a, x_v, y)$ , where $x_a$ represents the log-filterbank features extracted from the input audio, $x_v$ denotes the video capturing lip movements, and $y$ is the target Roman text, we first encode the audio-visual speech features. Specifically, the audio features $f_a = \mathcal{F}_a(x_a) \in \mathbb{R}^{T \times D}$ and the visual features $f_v = \mathcal{F}_v(x_v) \in \mathbb{R}^{T \times D}$ are extracted using the audio and visual encoders, respectively. Note that the lengths of audio and visual features are synchronized through pre-processing + +and stripped convolution. Next, we concatenate the audio and visual features along the channel dimension and then these combined features are fed into a transformer encoder to encode the audio-visual speech features $f_{av}$ . This process can be formulated as follows: $f_{av} = \mathcal{B}((f_a\oplus f_v)W)$ , where $\oplus$ denotes concatenation along the channel dimension, and $W\in \mathbb{R}^{2D\times D}$ is a weight matrix of a linear layer applied to preserve the original dimensionality. Finally, a linear layer is applied to predict the output roman tokens $\hat{y}$ . The proposed AV-Romanizer is trained with Connectionist Temporal Classification (CTC) [60] objective function. + +After training on sufficient data that covers a diverse range of phones, the AV-Romanizer can predict Roman text by capturing how input audio-visual speech is pronounced, even though the language was not used during training. + +# 3.3. Cascaded Zero-shot AVSR + +If we can transform the Roman text into language-specific graphemes, then the speech recognition process is complete. To construct the complete AVSR pipeline, we cascade the AV-Romanizer and an LLM. The AV-Romanizer transcribes input audio-visual speech into Roman text. Then, the predicted Roman text is de-romanized by a pre-trained LLM. Here, the zero-shot language recognition ability arises. Even though a language has not been used to train the AV-Romanizer, it can still predict how the input speech sounds like, in Roman text. By taking the predicted text, a pre-trained LLM can convert the text into the target language, which is within its knowledge. We named this cascaded form as Cascaded Zero-AVSR and depict in Fig. 1b. + +Specifically, an instruction to convert the Roman text into the target language is incorporated along with the pre + +![](images/542e649293be492310f05acef8deeb4278869d1e564faceb750c592ba7c2386b.jpg) +Figure 2. Illustration of the proposed multi-tasks to train Zero-AVSR. (a) Task 1: Alignment between the AV Romanizer and a LLM. By using the paired audio-visual speech inputs and ground truth roman text, we align the audio-visual speech representations into LLM space (i.e., text). (b) Task 2: Learning to de-romanize. By leveraging abundant text-only data, we fine-tune the LLM to de-romanize input roman text into language-specific graphemes. This approach enables us to cover a wider range of languages, as text data is more readily available than audio-visual speech data for diverse languages. + +![](images/8660fbe9a665a81821c5d36b43f9858aa7622be9680ea9171ffea00febfda5a4.jpg) + +dicted Roman text from the AV-Romanizer. This text is then used as input for the LLMs as shown in Fig. 1b. Since this approach does not require finetuning the LLMs and the input is purely text-based, diverse LLMs can be employed for Cascaded Zero-AVSR, even in API form. + +# 3.4. Zero-shot AVSR with Multi-task Training + +Although we have seen the potential of our proposed AV-Romanizer to interact with LLMs at the text level as a cascaded system, recent works [52, 53] have demonstrated that systems that directly integrate speech representations with LLMs can achieve optimal performance. Motivated by this, we explore an unified model, namely Zero-AVSR, which directly integrates audio-visual speech representations with LLMs instead of using the predicted Roman text itself. To this end, Zero-AVSR is trained on two tasks to enable zero-shot language recognition scenarios. The first task aims to align the audio-visual speech features encoded by the AV-Romanizer with the text embeddings of LLMs. The second task involves learning to de-romanize languages, encompassing both seen and unseen languages. + +# 3.4.1. Task 1: Aligning the AV-Romanizer with LLMs + +In this task, our goal is to align the audio-visual speech representations obtained from the AV-Romanizer with the text embeddings of LLMs. Since audio-visual speech has a higher temporal resolution compared to text and thereby contains redundant information, a length compressor is typically employed to reduce the computational burden, especially when incorporating with LLMs [11, 22, 61, 62]. + +As shown in Fig. 2a, given the audio-visual features $f_{av}$ extracted at the penultimate layer before the classification head of the AV-Romanizer, we first apply a length compressor to halve the length of the features. Then, an adapter + +maps the compressed audio-visual speech features into the LLM's embedding space. Finally, the embedded audio-visual speech features are concatenated with the text embedding of the instruction to form the input to the LLM. By setting the target as language-specific graphemes, the model is trained using a typical language modeling objective. Here, the pre-trained weights of the LLM including token embeddings and AV-Romanizer are kept frozen, while only the LoRA weights at the LLMs, length compressor, and the adapter are finetuned. As the task 1 requires audio-visual speech data, it can only be performed using seen languages. The knowledge for unseen languages will be extended in the task 2. + +# 3.4.2. Task 2: Learning to De-romanize + +In this task, our goal is to train LLMs to de-romanize diverse languages. As the task 1 is performed using only seen languages, task 2 is essential to prevent LLMs from forgetting their multilingual ability which is the key for zero-shot language recognition. Specifically, as shown in Fig. 2b, the input is set to purely text, consisting of the instruction and the Roman text of all languages, including both seen and unseen languages. The target output is then set to language-specific graphemes, constructing a transformation task from Roman text to language-specific graphemes. During task 2 training, only the LoRA weights of the LLM are finetuned. + +# 4. Experimental Setup + +# 4.1. Dataset + +Multilingual Audio-visual Romanized Corpus (MARC) is the proposed dataset designed for zero-shot audio-visual speech recognition by labeling and integrating data from + +
MethodModalityAV Training HoursSupport # LangsTarget Language (WER(%)↓)Avg (w/ Eng)
UnlabeledLabeledAraDeuEllSpaFraItaPorRusEng
Non-Zero Shot Multilingual AVSR Models
AV-HuBERT [6]AVSR1,759*1,200989.452.046.217.420.320.822.144.71.735.0
u-HuBERT [63]AVSR1,759*(452*)1,200989.352.146.417.320.521.221.944.41.935.0
XLAVS-R 300M [28]AVSR1,200(436K)1,200981.744.724.310.914.412.813.232.72.426.3
XLAVS-R 2B [28]AVSR1,200(436K)1,200979.344.419.09.112.310.611.225.01.723.6
Zero-Shot Multilingual ASR/AVSR Models
MMS Zero-shot [46]ASR(436K)(40K)1,078+84.931.547.917.733.619.035.542.835.738.9
Cascaded Zero-AVSRAVSR1,759*2,91682+82.129.347.216.328.921.620.242.92.930.2
Zero-AVSRAVSR1,759*2,91682+81.427.838.413.114.315.915.432.61.525.2
+ +Table 2. Comparisons with state-of-the-art methods on MuAViC. Note that an asterisk (*) denotes the use of English-only data, and values in parentheses indicate the amount of audio-only data employed. + +four existing datasets: LRS3 [33], MuAViC [26], VoxCeleb2 [58], and AVSpeech [59]. Details for each source dataset and MARC can be found in the Supplementary. + +# 4.2. Implementation details + +Pre-processing. We resample all video and audio to 25 fps and $16\mathrm{kHz}$ , respectively. We extract facial landmarks using the ReinaFace detector [9] and crop the mouth region to a size of $96\times 96$ . For text data, we apply fairseq's text normalization. Finally, for data augmentation, we perform random cropping into a size of $88\times 88$ and horizontal flipping following [4, 6] during all training processes. + +Architecture. For the AV-Romanizer, we adopt the AV-HuBERT [6] architecture, which comprises a visual encoder, an audio encoder, an audio-visual fusion module, and a transformer encoder. Additionally, we employ a linear projection layer for predicting Roman text. The visual encoder uses a ResNet-18 with a 3D convolution layer, while the audio encoder consists of a single linear layer. The transformer encoder, composed of 24 layers, features a model dimension of 1024, a feed-forward dimension of 4096, and 16 attention heads. For the length compressor, a 1D convolution with a kernel size of 2 and a stride of 2 is utilized. For Zero-AVSR, we employ Llama3.2-3B [30] as the decoder and finetune it using QLoRA [51, 54], while GPT-4o-mini [31] is employed for Cascaded Zero-AVSR. + +Training and Evaluation. For training the AV-Romanizer, we employ a three-stage scheduler with 10K warmup steps, 40K hold steps, and 50K decay steps, along with a peak learning rate of 1e-4. Training is performed on 8 RTX 3090 GPUs with gradient accumulation set to 8. Also, the audio is randomly perturbed with acoustic noise sampled from MUSAN [64] with 0 dB SNR. For training the Zero-AVSR with an LLM, a cosine scheduler is employed with 0.5K warmup steps and 29.5K decay steps, using 7 A6000 GPUs with gradient accumulation set to 9. After training Zero-AVSR, we evaluate its performance using beam search with a width of 2 and a temperature of 0.3. + +# 5. Experimental Results + +# 5.1. Comparison with the state-of-the-art methods + +Although our primary goal is language expansion, it is also crucial to ensure that the proposed methods perform on seen languages at a level comparable to state-of-the-art approaches for real-world applications. To validate this, we first evaluate the effectiveness of our proposed methods, Cascaded Zero-AVSR and Zero-AVSR, which are trained using audio-visual data from all languages in MARC. Table 2 summarizes the performance of our methods compared to several state-of-the-art approaches on the MuAViC [26] dataset. We organize the comparison into two groups: (i) multilingual audio-visual speech recognition models that support a fixed set of nine languages, and (ii) methods that operate using Roman characters and cover a wider range of languages. + +Among the multilingual models (Non-zero shot), both AV-HuBERT [6] and u-HuBERT [63] achieve an average WER of $35.0\%$ across nine target languages. In contrast, XLAVS-R [28] models highlight the benefits of increased model capacity and the use of unlabeled multilingual audiovisual data, along with large-scale audio-only data (436K hours) during pre-training. Their 300M and 2B variants achieve average WERs of $26.3\%$ and $23.6\%$ , respectively. We observe high WERs on Arabic (Ara) for all methods. + +The audio-only zero-shot model, MMS Zero-shot [46], leverages 436K hours of unlabeled audio data for pretraining and then fine-tunes on 40K hours of labeled data spanning 1,078 languages. As a result, it supports over 1,078 languages but exhibits a relatively high average WER of $38.9\%$ . By learning language-agnostic audio-visual speech representations and integrating the multilingual language understanding capabilities of LLMs, the proposed methods, Cascaded Zero-AVSR and Zero-AVSR, outperform the previous approach, achieving average WERs of $30.2\%$ and $25.2\%$ , respectively. Compared to existing multilingual AVSR approaches that support only 9 languages, the proposed methods significantly expand AVSR capabilities to a broader range of languages by training on the + +
MethodUnseen Lang.Target Language (CER(%)↓)Avg (w/o Eng)
AraDeuEllSpaFraItaPorRus
Cascaded Zero-AVSR (GPT-4o-mini)Ara67.617.123.28.314.310.210.020.620.8
Deu53.746.922.38.213.79.79.820.225.8
Ell55.617.045.38.114.510.09.621.223.5
Spa54.817.222.519.014.310.310.120.520.4
Fra55.217.222.08.444.69.910.020.621.4
Ita54.217.522.58.313.723.910.020.920.9
Por54.417.122.68.314.210.137.721.022.9
Rus55.217.122.68.414.010.19.940.021.9
Zero-AVSR (Llama3.2-3B)Ara76.516.221.66.87.47.07.718.519.7
Deu56.952.922.47.48.26.97.718.726.4
Ell53.916.162.16.88.16.87.818.424.3
Spa57.816.724.419.78.57.58.620.219.0
Fra55.316.021.26.954.67.27.918.120.7
Ita61.617.222.67.18.125.17.718.620.7
Por58.016.324.77.57.77.444.018.822.3
Rus57.716.022.77.17.76.97.645.421.5
+ +Table 3. The zero-shot language AVSR performances of Cascaded Zero-AVSR and Zero-AVSR on MuAViC. We train 8 AVRomanizers, setting each language as an unseen language, and evaluate their zero-shot performance, shown in blue-colored cells. + +proposed MARC, while achieving the comparable performances. In addition, Zero-AVSR achieves state-of-the-art performances in English (Eng) and German (Deu), with WERs of $1.5\%$ and $27.8\%$ , respectively. + +# 5.2. Effectiveness on Unseen Languages + +To validate the effectiveness of the proposed methods on unseen languages, we train eight different models for each of Cascaded Zero-AVSR and Zero-AVSR. In each experiment, we train the model by excluding the audio-visual data of the target unseen language. For example, if Spanish (Spa) is the target unseen language, the AV-Romanizer is trained without using any Spanish data and then evaluated on 8 languages, including both seen and unseen languages. Given the dominant status of English, we exclude its results from our analysis to focus on other relatively low-resource languages, allowing us to better understand the effectiveness of Zero-AVSR in zero-shot speech recognition. + +For the Cascaded Zero-AVSR, we first predict the Roman text by using the AV-Romanizer and then de-romanize it using GPT-4o-mini [31]. The zero-shot AVSR performances (CER) of Cascaded Zero-AVSR for Arabic (Ara), German (Deu), Greek (Ell), Spanish (Spa), French (Fra), Italian (Ita), Portuguese (Por), and Russian (Rus) are $67.6\%$ , $46.9\%$ , $45.3\%$ , $19.0\%$ , $44.6\%$ , $23.9\%$ , $37.7\%$ , and $40.0\%$ , respectively. This indicates that even without training on any speech data, the Cascaded Zero-AVSR model can still predict transcriptions in unseen languages. For instance, the Cascaded Zero-AVSR can achieve a $62.3\%$ character-level accuracy when predicting Portuguese (Por) speech. + +For the Zero-AVSR, we employ Llama3.2-3B [30], a state-of-the-art open-source LLM, as the decoder. The zero + +
Train Dataset# Train Language# Train HoursUnseen Language\( \mathbf{CER(\%)}\downarrow \)
Zero-shotAvg (w/o Eng)
MuAViC [26]8745Rus62.348.3
81,944Rus61.028.3
+ MARC402,418Rus49.525.1
812,793Rus40.021.9
MuAViC [26]8724Ita34.849.0
81,921Ita28.525.9
+ MARC402,395Ita26.023.7
812,770Ita23.920.9
MuAViC [26]8698Spa35.546.3
81,851Spa29.925.4
+ MARC402,326Spa20.522.6
812,700Spa19.020.4
+ +Table 4. Impact of the proposed MARC on Cascaded Zero-AVSR, evaluated by varying data amount and number of languages. + +shot AVSR performances (CER) for the eight languages are as follows: Ara $76.5\%$ , Deu $52.9\%$ , Ell $62.1\%$ , Spa $19.7\%$ , Fra $54.6\%$ , Ita $25.1\%$ , Por $44.0\%$ , and Rus $45.4\%$ . Although the Cascaded Zero-AVSR and Zero-AVSR models are not directly comparable due to their different LLMs, we can confirm that incorporating speech features into the LLM and finetuning the model (i.e., Zero-AVSR) significantly improves performance on seen languages. Furthermore, zero-shot AVSR performance is also substantially improved when compared to the Cascaded Zero-AVSR using the same LLM, Llama3.2-3B, whose results are provided in the Supplementary. This result highlights the potential of the Zero-AVSR; If a better LLM is employed, its performance can be significantly enhanced. + +# 5.3. Ablation study + +# 5.3.1. Effectiveness of the MARC Dataset + +To validate the effectiveness of the proposed MARC dataset, we conducted a gradual data addition experiment starting with the MuAViC portion of the MARC dataset. Specifically, we conduct four experiments for each of three languages, Russian (Rus), Italian (Ita), and Spanish (Spa), totaling 12 experiments. We measure the performance of the Cascaded Zero-AVSR and each language is treated as a target unseen language. The results are shown in Table 4. + +First, we train the AV-Romanizer using only MuAViC portion of the MARC dataset, and achieve average CERs (i.e., including both unseen and seen languages) of $48.3\%$ , $49.0\%$ , and $46.3\%$ for Rus, Ita, and Spa, respectively. Next, by incorporating the remaining audio-visual data of the MARC with the same 8 languages in the MuAViC portion, we achieve significantly improved average CERs of $28.3\%$ , $25.9\%$ , and $25.4\%$ for Rus, Ita, and Spa, respectively. The results demonstrate the effectiveness of using MARC dataset and show that incorporating more data is beneficial for improving AVSR performance. + +Then, to assess the impact of language diversity on + +
MethodTarget Language (CER(%)↓)Avg (w/o Eng)
AraDeuEllSpaFraItaPorRus
Llama3.2-3B [30]69.324.256.613.927.917.514.956.535.7
Mistral-7B [32]71.226.732.813.523.916.217.736.529.6
Llama3.1-8B [30]61.421.739.414.820.414.114.430.927.3
Llama3.1-70B [30]56.118.126.79.515.611.111.123.321.3
GPT-4o-mini [31]54.017.322.58.114.210.49.921.219.5
+ +modeling language-agnostic audio-visual speech representations, we train the AV-Romanizer using audio-visual speech data from 40 and 81 languages in the MARC dataset. That means the model is trained on more than just the 8 evaluation languages. The results show that incorporating data from 41 languages improves the performance for both unseen (i.e., zero-shot) and seen languages. Especially, the zero-shot performance for Russian is greatly improved from $61.0\%$ CER to $49.5\%$ CER. This suggests that incorporating a wider range of diverse phonetic information through the use of multiple languages can significantly enhance zero-shot speech recognition performance. Similarly, employing a total of 81 languages further improves both the zero-shot performance and the average performance. Again, this result confirms the importance of employing diverse languages and sufficient phonetic coverage in learning language-agnostic speech representations. + +# 5.3.2. Ablation Study using Different LLM Models + +As the Cascaded Zero-AVSR does not finetune the LLM, its performance is heavily dependent on the choice of LLM used, as expected. To investigate the performance of Cascaded Zero-AVSR according to different types of LLMs, we perform ablation study by employing 5 different LLMs. The AV-Romanizer is trained on all 82 languages. Table 5 shows the ablation results. We can confirm that employing a larger model, which has a better ability to transform between Roman and language-specific graphemes, yields better performance. Notably, GPT-4o-mini [31] achieves the best performance among the baselines. Therefore, we employ GPT-4o-mini as the de-romanizer for other Cascaded Zero-AVSR experiments. Please note that we employ Llama3.2-3B [30] for training Zero-AVSR as it has a feasible number of parameters. + +# 5.3.3. Ablation Study on the Impact of Language Family + +Languages within the same language family [65] often exhibit similar grammatical structures and pronunciation patterns. Consequently, it is reasonable to expect that incorporating more data from languages within the same group as the target unseen language will enhance zero-shot speech recognition performance. To evaluate this hypothesis, we conduct ablation studies starting with a baseline model trained without data from the Romance family, the Turkic + +Table 5. Performances of Cascaded Zero-AVSR using different types of LLMs. The AV-Romanizer is trained on all 82 languages. + +
Train Dataset# Train LanguageUnseen Languages\( \mathbf{{CER}}\left( \% \right) \downarrow \)
SpaItaFraPor
Baseline67Spa, Ita, Fra, Por40.134.255.848.0
• Adding Data from Different Language Family
+ Turkish (51 hrs)68Spa, Ita, Fra, Por38.133.661.048.2
+ Korean (129 hrs)69Spa, Ita, Fra, Por38.433.659.749.5
• Adding Data from Romance Subgroup
+ Spanish (51 hrs)68Ita, Fra, Por11.528.156.644.0
+ Italian (129 hrs)69Fra, Por11.011.151.341.9
+ +Table 6. Ablation study on the impact of using data from the same language family (Romance) on zero-shot speech recognition. + +family, and Korean. We then incrementally add data to the baseline model and build two variants: one incorporating Spanish and Italian (Romance languages) and another adding Turkish and Korean (non-Romance languages), ensuring that the same amount of data is added to each. By comparing these two variants, we can assess the impact of linguistic similarity on zero-shot speech recognition performance. We report and analyze the zero-shot AVSR performances on Romance languages, including Spanish (Spa), Italian (Ita), French (Fra), and Portuguese (Por). + +The ablation results are shown in Table 6. The first row shows the zero-shot AVSR performance of the baseline model. By gradually adding non-Romance languages, Turkish and Korean sequentially, the zero-shot performances are slightly improved for Spa and Ita, while there is no improvement for Fra and Por. In contrast, we can confirm that the zero-shot speech recognition performances for Romance languages are greatly improved by adding Spa and Ita, compared to adding the non-Romance languages. When Spa is added, the Ita and Por performances are improved to $28.1\%$ and $44.0\%$ CERs, respectively, while the performance of Fra remains unchanged. Furthermore, adding Italian data on top of the Spa added model results in an additional improvement, ultimately achieving $51.3\%$ and $41.9\%$ CERs for Fra and Por. These findings provide evidence that incorporating languages similar to the target language is more effective in zero-shot speech recognition. + +# 6. Conclusion + +We presented a zero-shot AVSR framework, Zero-AVSR, which extends language support beyond seen languages. The key idea behind this is 1) learning language-agnostic speech representations using Roman text, and 2) employing the LLMs to generate language-specific graphemes. To this end, we introduced the AV-Romanizer to convert audio-visual speech input into Roman text, which is then de-romanized into native scripts by LLMs. Depending on whether it involves fine-tuning an LLM, we explored two variants of frameworks, Cascaded Zero-AVSR and Zero-AVSR. Finally, we introduced the MARC dataset, comprising audio-visual speech data for 82 languages. Extensive experiments verified the effectiveness of Zero-AVSR. + +# Acknowledgments + +This work was partly supported by two funds: the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF2022R1A2C2005529), and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2022-000124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities). + +# References + +[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 44(12):8717-8727, 2018. 1, 2 +[2] Stavros Petridis, Themos Stafylakis, Pingehuan Ma, Feipeng Cai, Georgios Tzimiropoulos, and Maja Pantic. End-to-end audiovisual speech recognition. In 2018 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 6548-6552. IEEE, 2018. 2 +[3] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In 2019 IEEE automatic speech recognition and understanding workshop (ASRU), pages 905–912. IEEE, 2019. +[4] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with conformers. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7613-7617. IEEE, 2021. 2, 6 +[5] Sucheng Ren, Yong Du, Jianming Lv, Guoqiang Han, and Shengfeng He. Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13325-13333, 2021. +[6] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. arXiv preprint arXiv:2201.02184, 2022. 2, 6 +[7] Joanna Hong, Minsu Kim, Daehun Yoo, and Yong Man Ro. Visual context-driven audio feature enhancement for robust end-to-end audio-visual speech recognition. In Interspeech, 2022. +[8] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. arXiv preprint arXiv:2201.10439, 2022. +[9] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-avsr: Audio-visual speech recognition with automatic labels. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 2, 3, 6, 1 + +[10] Joanna Hong, Minsu Kim, Jeongsoo Choi, and Yong Man Ro. Watch or listen: Robust audio-visual speech recognition with visual corruption modeling and reliability scoring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18783–18794, 2023. +[11] Umberto Cappellazzo, Minsu Kim, Honglie Chen, Pingchuan Ma, Stavros Petridis, Daniele Falavigna, Alessio Brutti, and Maja Pantic. Large language models are strong audio-visual speech recognition learners. arXiv preprint arXiv:2409.12319, 2024. 2, 3, 5 +[12] Andrew Rouditchenko, Yuan Gong, Samuel Thomas, Leonid Karlinsky, Hilde Kuehne, Rogerio Feris, and James Glass. Whisper-flamingo: Integrating visual features into whisper for audio-visual speech recognition and translation. arXiv preprint arXiv:2406.10082, 2024. 2 +[13] Alexandros Haliassos, Rodrigo Mira, Honglie Chen, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Unified speech recognition: A single model for auditory, visual, and audio-visual inputs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[14] Sanath Narayan, Yasser Abdelaziz Dahou Djilali, Ankit Singh, Eustache Le Bihan, and Hakim Hacid. Visper: Multilingual audio-visual speech recognition. arXiv preprint arXiv:2406.00038, 2024. 1, 2 +[15] Dario Amodei, Sundaram Ananthanarayanan, Rishita Anubhai, Jingliang Bai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, Qiang Cheng, Guoliang Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin. In International conference on machine learning, pages 173-182. PMLR, 2016. 1 +[16] Suyoun Kim, Takaaki Hori, and Shinji Watanabe. Joint ctc-attention based end-to-end speech recognition using multitask learning. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 4835-4839. IEEE, 2017. +[17] Rohit Prabhavalkar, Takaaki Hori, Tara N Sainath, Ralf Schlüter, and Shinji Watanabe. End-to-end speech recognition: A survey. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 32:325-351, 2023. 1 +[18] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nature Machine Intelligence, 4(11):930-939, 2022. 1 +[19] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. arXiv preprint arXiv:2212.06246, 2022. 2 +[20] Minsu Kim, Jeong Hun Yeo, Jeongsoo Choi, and Yong Man Ro. Lip reading for low-resource languages by learning and combining general speech knowledge and language-specific knowledge. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15359-15371, 2023. +[21] Minsu Kim, Jeonghun Yeo, Se Jin Park, Hyeongseop Rha, and Yong Man Ro. Efficient training for multilingual visual speech recognition: Pre-training with discretized visual speech representation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1311-1320, 2024. + +[22] Jeong Hun Yeo, Seunghee Han, Minsu Kim, and Yong Man Ro. Where visual speech meets language: Vsp-llm framework for efficient and context-aware visual speech processing. arXiv preprint arXiv:2402.15151, 2024. 2, 5 +[23] Jeong Hun Yeo, Minsu Kim, Shinji Watanabe, and Yong Man Ro. Visual speech recognition for languages with limited labeled data using automatic labels from whisper. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10471-10475. IEEE, 2024. 3 +[24] Alexandros Haliassos, Andreas Zinonos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Braven: Improving self-supervised pre-training for visual and auditory speech recognition. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 11431-11435. IEEE, 2024. 1 +[25] Elizabeth Salesky, Matthew Wiesner, Jacob Bremerman, Roldano Cattoni, Matteo Negri, Marco Turchi, Douglas W Oard, and Matt Post. The multilingual tex corpus for speech recognition and translation. arXiv preprint arXiv:2102.01757, 2021. 1, 2 +[26] Mohamed Anwar, Bowen Shi, Vedanuj Goswami, Wei- Ning Hsu, Juan Pino, and Changhan Wang. Muavic: A multilingual audio-visual corpus for robust speech recognition and robust speech-to-text translation. arXiv preprint arXiv:2303.00628, 2023. 1, 2, 3, 6, 7 +[27] Andreas Zinonos, Alexandros Haliassos, Pingchuan Ma, Stavros Petridis, and Maja Pantic. Learning cross-lingual visual speech representations. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1 +[28] HyoJung Han, Mohamed Anwar, Juan Pino, Wei-Ning Hsu, Marine Carpuat, Bowen Shi, and Changhan Wang. Xlavs r: Cross-lingual audio-visual speech representation learning for noise-robust speech perception. arXiv preprint arXiv:2403.14402, 2024. 2, 6 +[29] Maxime Burchi, Krishna C Puvvada, Jagadeesh Balam, Boris Ginsburg, and Radu Timofte. Multilingual audiovisual speech recognition with hybrid ctc/rnn-t fast conformer. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10211-10215. IEEE, 2024. 1, 2, 3 +[30] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 6, 7, 8 +[31] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 6, 7, 8 +[32] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1, 8 + +[33] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 2, 3, 6, 1 +[34] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 87-103. Springer, 2017. 2 +[35] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2 +[36] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer for speech recognition. arXiv preprint arXiv:2005.08100, 2020. 2 +[37] Ulf Hermjakob, Jonathan May, and Kevin Knight. Out-of-the-box universal romanization tool uroman. In Proceedings of ACL 2018, system demonstrations, pages 13-18, 2018. 3 +[38] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR, 2023. 2 +[39] Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, et al. Scaling speech technology to 1,000+ languages. Journal of Machine Learning Research, 25(97):1-52, 2024. 2, 3 +[40] Da-Rong Liu, Kuan-Yu Chen, Hung-Yi Lee, and Lin-shan Lee. Completely unsupervised phoneme recognition by adversarially learning mapping relationships from audio embeddings. arXiv preprint arXiv:1804.00316, 2018. 2 +[41] Kuan-Yu Chen, Che-Ping Tsai, Da-Rong Liu, Hung-Yi Lee, and Lin-shan Lee. Completely unsupervised speech recognition by a generative adversarial network harmonized with iteratively refined hidden markov models. arXiv preprint arXiv:1904.04100, 2019. +[42] Alexei Baevski, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. Unsupervised speech recognition. Advances in Neural Information Processing Systems, 34, 2021. 2 +[43] Xinjian Li, Florian Metze, David R Mortensen, Alan W Black, and Shinji Watanabe. Asr2k: Speech recognition for around 2000 languages without audio. arXiv preprint arXiv:2209.02842, 2022. 3, 4 +[44] Xinjian Li, Siddharth Dalmia, Juncheng Li, Matthew Lee, Patrick Littell, Jiali Yao, Antonios Anastasopoulos, David R Mortensen, Graham Neubig, Alan W Black, et al. Universal phone recognition with a multilingual allophone system. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8249-8253. IEEE, 2020. 3 +[45] Xinjian Li, Florian Metze, David R Mortensen, Shinji Watanabe, and Alan W Black. Zero-shot learning for grapheme to phoneme conversion with language ensemble. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2106-2115, 2022. 3, 4 + +[46] Jinming Zhao, Vineel Pratap, and Michael Auli. Scaling a simple approach to zero-shot speech recognition. arXiv preprint arXiv:2407.17852, 2024. 3, 4, 6 +[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3 +[48] Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, et al. On generative spoken language modeling from raw audio. Transactions of the Association for Computational Linguistics, 9:1336-1354, 2021. 3 +[49] Siddique Latif, Moazzam Shoukat, Fahad Shamshad, Muhammad Usama, Yi Ren, Heriberto Cuayáhuitl, Wenwu Wang, Xulong Zhang, Roberto Togneri, Erik Cambria, et al. Sparks of large audio models: A survey and outlook. arXiv preprint arXiv:2308.12792, 2023. +[50] Se Jin Park, Chae Won Kim, Hyeongseop Rha, Minsu Kim, Joanna Hong, Jeong Hun Yeo, and Yong Man Ro. Let's go real talk: Spoken dialogue model for face-to-face conversation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024. 3 +[51] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 6 +[52] Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. *Salmon: Towards generic hearing abilities for large language models.* arXiv preprint arXiv:2310.13289, 2023. 3, 5 +[53] Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shiliang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou. Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models. arXiv preprint arXiv:2311.07919, 2023. 3, 5 +[54] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized lIms. Advances in neural information processing systems, 36:10088-10115, 2023. 3, 6 +[55] Tanja Schultz and Alex Waibel. Language-independent and language-adaptive acoustic modeling for speech recognition. Speech Communication, 35(1-2):31-51, 2001. 3 +[56] Ngoc Thang Vu, David Imseng, Daniel Povey, Petr Moticek, Tanja Schultz, and Hervé Boullard. Multilingual deep neural network based acoustic modeling for rapid language adaptation. In 2014 IEEE international Conference on acoustics, speech and signal processing (ICASSP), pages 7639-7643. IEEE, 2014. +[57] Minsu Kim, Jeongsoo Choi, Dahun Kim, and Yong Man Ro. Textless unit-to-unit training for many-to-many multilingual speech-to-speech translation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024. 3 +[58] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 3, 6, 1 + +[59] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 3, 6, 1 +[60] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369-376, 2006. 4 +[61] CHEN CHEN, Ruizhe Li, Yuchen Hu, Sabato Marco Siniscalchi, Pin-Yu Chen, EngSiong Chng, and Chao-Han Huck Yang. It's never too late: Fusing acoustic information into large language models for automatic speech recognition. In The Twelfth International Conference on Learning Representations, 2024. 5 +[62] Yuchen Hu, CHEN CHEN, Chao-Han Huck Yang, Ruizhe Li, Chao Zhang, Pin-Yu Chen, and EngSiong Chng. Large language models are efficient learners of noise-robust speech recognition. In The Twelfth International Conference on Learning Representations, 2024. 5 +[63] Wei-Ning Hsu and Bowen Shi. u-hubert: Unified mixed-modal speech pretraining and zero-shot transfer to unlabeled modality. Advances in Neural Information Processing Systems, 35:21157-21170, 2022. 6 +[64] David Snyder, Guoguo Chen, and Daniel Povey. Musan: A music, speech, and noise corpus. arXiv preprint arXiv:1510.08484, 2015. 6, 2 +[65] Kenneth Katzner and Kirk Miller. The languages of the world. Routledge, 2002. 8, 1 +[66] Rosana Ardila, Megan Branson, Kelly Davis, Michael Hen-cretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. arXiv preprint arXiv:1912.06670, 2019. 2 \ No newline at end of file diff --git a/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/images.zip b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..031ce1598f778264bc6eaf3d87e60e1a323503d2 --- /dev/null +++ b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c41bd2fe73e421c46c34c32c1bb24dd69036729b2e9de40cde7068115d6bedac +size 455989 diff --git a/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/layout.json b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c720e2de364b60d6048162caa4683ca9957d91ce --- /dev/null +++ b/zeroavsrzeroshotaudiovisualspeechrecognitionwithllmsbylearninglanguageagnosticspeechrepresentations/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3257eaa0df49cbd079bf9f0d79431f1c74d5e36f95f392928c9eb01280a9166 +size 364006 diff --git a/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/348d3cba-bca8-4d06-8d2b-4af61751a081_content_list.json b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/348d3cba-bca8-4d06-8d2b-4af61751a081_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..29c3e4534d90859c16e2a8f4e934f9a36e2d40ff --- /dev/null +++ b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/348d3cba-bca8-4d06-8d2b-4af61751a081_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:267b33ad047bb2fdeaa4387598454ddcb048a49f06eab4f0dbc661e7cdc2f7e5 +size 82652 diff --git a/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/348d3cba-bca8-4d06-8d2b-4af61751a081_model.json b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/348d3cba-bca8-4d06-8d2b-4af61751a081_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7bfca08060c4d4dcde89e6d984860e00e2659481 --- /dev/null +++ b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/348d3cba-bca8-4d06-8d2b-4af61751a081_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfead250e86369e593b005b1e46a2be55a558491aeaf51d5280552caa7b22f15 +size 103005 diff --git a/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/348d3cba-bca8-4d06-8d2b-4af61751a081_origin.pdf b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/348d3cba-bca8-4d06-8d2b-4af61751a081_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c0c880de4d4fbb5b81747327c7603ff9833d873c --- /dev/null +++ b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/348d3cba-bca8-4d06-8d2b-4af61751a081_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:448dddc9fb764340bf4ce5440adc3e0ceeb80e9fb1438cbf0e28a67dd1e4c46d +size 4546714 diff --git a/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/full.md b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..367ab8060dac78517b358dfd3342421ded78af35 --- /dev/null +++ b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/full.md @@ -0,0 +1,310 @@ +# ZeroKey: Point-Level Reasoning and Zero-Shot 3D Keypoint Detection from Large Language Models + +Bingchen Gong1 Diego Gomez1 Abdullah Hamdi2 Abdelrahman Eldesokey3 +Ahmed Abdelreheem3 Peter Wonka3 Maks Ovsjanikov1 + +$^{1}$ École Polytechnique $^{2}$ Visual Geometry Group, University of Oxford $^{3}$ King Abdullah University of Science and Technology (KAUST) + +gongbingchen@gmail.com + +![](images/39abd729b4fccea4c8fa416f5bb8678c1243fd2e0960dd7012a0e1fb1b7d0aae.jpg) +Figure 1. Zero-shot 3D Keypoint Detection. Without any ground truth labels or supervised training, our method leverages the point-level reasoning embedded within MLLMs to extract and name salient keypoints on 3D models. The figure illustrates how our approach achieves competitive performance compared to CLIP-DINOiser [50] baselines, highlighting the potential of integrating language models with vision tasks for enhanced 3D shape understanding. + +# Abstract + +We propose a novel zero-shot approach for keypoint detection on 3D shapes. Point-level reasoning on visual data is challenging as it requires precise localization capability, posing problems even for powerful models like DINO or CLIP. Traditional methods for 3D keypoint detection rely heavily on annotated 3D datasets and extensive supervised training, limiting their scalability and applicability to new categories or domains. In contrast, our method utilizes the rich knowledge embedded within Multi-Modal Large Language Models (MLLMs). Specifically, we demonstrate, for the first time, that pixel-level annotations used to train recent MLLMs can be exploited for both extracting and naming salient keypoints on 3D models without any ground truth labels or supervision. Experimental evaluations demonstrate that our approach achieves competitive performance + +on standard benchmarks compared to supervised methods, despite not requiring any 3D keypoint annotations during training. Our results highlight the potential of integrating language models for localized 3D shape understanding. This work opens new avenues for cross-modal learning and underscores the effectiveness of MLLMs in contributing to 3D computer vision challenges. + +# 1. Introduction + +Multimodal Large Language Models (MLLMs) have been shown to seamlessly integrate visual and text representations with great success [3, 27-29, 34, 59]. Models like GPT-4 with vision capabilities, OFA [48], and Flamingo [3] have demonstrated the ability to process visual and textual data, leading to advancements in vision-language understanding. Furthermore, related efforts have been made to + +endow MLLMs with 3D or spatial reasoning [18, 51, 58]. When trained at scale, these MLLMs enable tackling a wide array of complex tasks, ranging from comprehensive image or 3D shape description to answering detailed questions about visual content. + +Unfortunately, despite this tremendous progress, existing MLLMs still struggle with tasks that require precise points or pixel-level reasoning, such as localization, counting, or salient keypoint detection [8, 37, 56]. This type of reasoning focuses on understanding and interpreting visual input at a fine-grained level using text. In general, we can observe an increased level of difficulty when going from complete objects to object parts and, finally, to specific points or small regions. Even advanced models like GPT-4o [34] and Claude Sonnet 3.5 [4], which have demonstrated impressive capabilities in various computer vision tasks, still struggle with point-level understanding. + +This challenge arises because point-level reasoning requires models to accurately identify and analyze local details, such as landmarks on a face, joints in a human pose, or intricate components in a 3D shape. However, the architectures, training methods, and dataset annotations of most existing approaches are primarily designed to capture global visual properties and potentially link them to textual descriptions. Since text captions are most often crawled from the Internet, and thus rarely contain precise point or region-level annotations, point-level reasoning capabilities of most MLLMs remain very limited. + +Addressing this problem necessitates a shift in how models are designed and trained. One potential solution is to use more powerful visual encoders trained on real-world data encompassing a variety of global and local features. Interestingly, scaling alone to more crawled data or model complexity does not seem to alleviate this problem [35]. In contrast, very recently, a family of MLLMs has been introduced that was trained with specially-designed auxiliary training data and tasks aimed at pixel-level image annotations alongside global image captions [12]. When trained with this additional data source, the resulting model (dubbed Molmo) exhibits impressive localized reasoning abilities (counting, localization, etc.) surpassing other MLLMs. + +Inspired by these recent developments, we propose investigating MLLMs endowed with point-level reasoning in the context of 3D shape understanding and specifically for zero-shot keypoint detection. Given a 3D model from an arbitrary category, our main task is to localize and name salient keypoints on this model. Traditional methods for 3D keypoint detection heavily rely on annotated 3D datasets and extensive supervised training [5, 49, 55]. This severely limits their applicability to specific data-rich object categories. At the same time, recent general-purpose 3D LLMs are typically trained with explicit objectives that align 3D data representations with vision-language models [18], en + +abling global tasks like classification. However, since they are based on standard vision-language models with limited localization capabilities, none of these methods can accurately perform 3D keypoint localization based on a language prompt, see the supplementary material for more details. + +Even in the presence of a powerful MLLM, we note that localization or keypoint detection from images to 3D environments presents a significant challenge. Since both detections and point-level captions tend to be noisy, developing a robust approach requires reliable multi-view aggregation, backprojection, and filtering mechanisms. In this paper, we propose a comprehensive zero-shot 3D keypoint detection method that exploits the point-level reasoning capabilities of recent MLLMs [12] while integrating 3D consistency and being completely category agnostic. To the best of our knowledge, ours is the first robust method for zero-shot 3D keypoint detection. + +Furthermore, we analyze the 3D understanding encoded in Molmo through our method by leveraging Schelling Points and evaluating the describability of keypoints. Through this study, we characterize the strengths and limitations of the 3D awareness imparted to models through training with pixel-level annotations. Our evaluations demonstrate that, despite not having access to ground truth data, our method can predict points similar to those identified by human annotators by selecting appropriate text prompts. These two tasks could serve as valuable evaluations for the point-level reasoning capability of multi-layered language models (MLLMs). + +Additionally, we developed two text-image baselines for detecting 3D keypoints from text prompts and conducted comparisons with these baselines. + +To summarize, our main contributions are as follows: + +- Identify the challenge of zero-shot point-level reasoning in 3D computer vision. +- Propose the first zero-shot 3D keypoint detection method that operates without 3D annotations or training and applies to arbitrary shape categories, establishing a baseline for zero-shot 3D keypoint detection in this new area. +- Evaluate several MLLMs in our context, highlighting the importance of auxiliary point-level training to solve the task at hand. + +# 2. Related Work + +# 2.1. Multimodal Large Language Models + +The remarkable success of vision-language models like CLIP [36] and large language models like GPT-3 [7] has motivated the integration of multimodal capabilities into large language models for visual understanding. Multimodal Large Language Models (MLLMs) have been extended to process images alongside text, achieving remarkable success in vision-language tasks [3, 27-29, 34, 59]. + +![](images/220049c8b362b36c4b562dc0f35aa20c282fddb9c2631cc5e87c15376012e7b8.jpg) +Figure 2. ZeroKey Pipeline. Our proposed ZeroKey employs MLLM Molmo for zero-shot keypoint detection on 3D objects by 1) rendering multiple views for a given shape, 2) leveraging MLLM reasoning in each view using point-specific prompts, and 3) aggregating the results through clustering, eliminating the need for annotated training data for 3D keypoint detection. + +Notable models such as GPT-4 with vision capabilities, MiniGPT-4 [59], LLaVA [28, 29], Flamingo [3], and BLIP [27] have demonstrated advanced abilities in integrating visual and textual information, enabling tasks ranging from detailed image captioning to complex visual reasoning. These models leverage large-scale pre-training on extensive image-text datasets to capture both global semantics and fine-grained details in 2D images. Despite their success with 2D data, extending these capabilities to 3D data remains a significant challenge due to the elevated cost and complexity of collecting large amounts of high-quality 3D data. Consequently, there is no dominant MLLM capable of processing 3D data with natural language understanding. To address this gap, recent efforts have focused on aligning models trained on small amounts of 3D data with large vision-language models. For instance, ULIP [52] seeks to align 3D models with vision-language representations to enhance cross-modal understanding. Similarly, Hong et al. [18] propose training a 3D Large Language Model by leveraging a pre-trained vision-language model, aiming to bridge the modality gap and facilitate better 3D comprehension through multimodal learning. While these "3D LLMs" are promising, the granularity we are targeting in work at the point-level 3D reasoning is still beyond their capabilities [18, 52]. See the supplementary material for more details. + +# 2.2. Keypoint Detection + +Keypoint detection has been extensively studied in 3D due to its significance in shape representation, matching, and abstraction applications [9, 13, 21, 41, 44, 47, 53]. However, the definition of a keypoint has remained task-specific and subjective. Even human annotations of keypoints often differ, as highlighted in [10, 54]. Recent work [17] has shown that image diffusion models can develop an unsupervised understanding of keypoints as salient features. Additionally, + +it was demonstrated in [42] that the CLIP foundation model [36] can reasonably interpret keypoints with appropriate visual prompts. This paper investigates the understanding of keypoints as foundational elements of objects and shapes within recent multimodal large language models (MLLMs). + +# 2.3. Lifting from 2D to 3D + +A number of studies have explored multi-view approaches to enhance 3D data analysis by leveraging 2D representations [11, 14, 16, 22, 23, 25, 30, 32, 33, 46]. A key challenge in this area is combining features from different viewpoints to represent local 3D points or voxels while preserving essential geometric details. Some methods address this by averaging features across views [23, 25], propagating labels between views [46], or learning from reconstructed neighboring points [22]. Others organize points within a unified grid structure [31] or merge multi-view features with 3D voxel information [11, 19]. VointNet [15] operates within the Voint cloud framework, retaining the original point cloud's compactness and descriptive 3D properties. It employs learned aggregation of multi-view features applied independently to each point. Recently, there's a growing trend of utilizing foundational models with zero-shot understanding capabilities—such as the Segment Anything Model [24]—to achieve zero-shot 3D comprehension by lifting 2D predictions into 3D space [1, 2, 20, 40]. Our work differs from the Zero-shot 3D segmentation in that our final goal is a fine-grained single 3D point representing an unseen text query on an unseen new 3D object. As we show below, existing vision-language models struggle with this task, thus necessitating a novel adapted solution. + +# 3. Motivation + +Localization and naming of points in an image or a 3D shape is an extremely challenging problem. Solving such + +a problem requires a localized understanding of the visual data and tight integration between text. Vision-Language Models (VLMs) are thus a strong candidate to tackle this problem. However, as mentioned above, most such models are trained using global or object-level semantics which are insufficient for point localization tasks. Furthermore, the recent MLLMs that incorporate 3D data [18, 51, 58] are typically trained with explicit alignment against pre-trained traditional vision-language models, and thus inherit their limitations in point-level reasoning. As a result, such models are unable to solve the zero-shot keypoint detection problem, which is the main focus of our work. + +Existing methods for the 3D keypoint detection problem typically formulate the problem as either a supervised learning task, by exploiting the ground truth annotations, e.g., in the KeypointNet dataset [54]. Alternatively, several few-shot approaches have been developed, e.g., [49], which are based on transferring known keypoints to new instances. + +Rather than training an entirely new 3D model, our goal is to exploit the improved point-level reasoning capabilities of very recent VLMs and demonstrate their utility in 3D shape understanding tasks. We base our approach on the recent open-source MOLMO model [12], which incorporated a specially designed pixel-level annotation task into its training. Thanks to a well-chosen design and this additional curated dataset the model has been shown to perform significantly better than even the largest closed-source models in localized image understanding. + +Our key motivation is to evaluate the effectiveness of such a model in the context of 3D shape understanding and specifically for zero-shot keypoint detection. We first introduce the problem of zero-shot keypoint detection and naming. Then, we develop a pipeline based on prompting the model and multi-view aggregation. We show that the resulting approach leads to unprecedented localized 3D shape understanding, opening the door to a range of applications. In this paper, we present the very first baseline for solving the zero-shot 3D keypoint detection problem. We leverage the powerful 2D spatially-aware VLM, Molmo [12], to generate precise 2D keypoint predictions from several points of view. These points are back-projected into a 3D shape, and after aggregation, we obtain a 3D keypoint prediction. + +# 4. Method + +Problem Setting and Overview. We consider the Zero-Shot Keypoint Detection problem for 3D shapes. Specifically, given a 3D shape, we aim to automatically generate a set of salient points corresponding to the shape's semantic parts. Our solution comprises three main components: first, we prompt a MLLM with the shape, asking the model to generate a list of names for possible candidate keypoints. Then, for each candidate, we ask the model to detect the precise coordinates of the point in a given image. Finally, + +we back-project those detected points into 3D and aggregate them to get the location of 3D keypoints in a given shape. + +# 4.1. Generating Text Candidates for Salient Points + +To detect points using a MLLM, we first need to assign names to these points. The definition of keypoints can vary depending on the application. However, these points are typically common to a given shape, salient, and carry significant semantic meaning. This information can be retrieved from MLLMs by prompting the model with the shape class in text form or by using an image of the shape when the MLLM supports multimodal input. Given the shape class label $c$ , we prompt the MLLM to generate a list of candidate keypoints: $\mathcal{K} = \{k_1, k_2, \dots, k_N\}$ , where $N$ is the number of keypoints generated. For example, for the class "airplane," the MLLM might generate keypoints names such as "nose," "wing tip," and "tail." In our experiments, we utilize GPT-4o [34] as our MLLM backbone for this step. We prompt it with a front-facing rendering of the 3D model along with the request text: "List possible salient keypoints (in text)." The GPT-4o model is instructed to provide a candidate list of possible names for keypoints in JSON format. + +# 4.2. Prompting Molmo to Detect 2D Keypoints + +To detect the precise 2D coordinates of each candidate keypoint, we utilize Molmo [12], a state-of-the-art MLLM capable of localizing points in images based on natural language prompts. Given a set of candidate keypoint names $\mathcal{K} = \{k_1,k_2,\dots ,k_N\}$ and a set of images (views) of the 3D shape $\mathcal{V} = \{\mathbf{V}_1,\mathbf{V}_2,\dots ,\mathbf{V}_M\}$ , we prompt Molmo to detect the location of each keypoint in each image. We define the detection function as: + +$$ +\mathbf {p} _ {i, j} = \operatorname {M o l m o} \left(\mathbf {V} _ {j}, k _ {i}\right), i = 1, \dots , N; j = 1, \dots , M, \tag {1} +$$ + +where $\mathbf{p}_{i,j} \in \mathbb{R}^2$ are the 2D coordinates of keypoint $k_{i}$ in image $\mathbf{V}_j$ as predicted by Molmo. The prompt to Molmo consists of the image $\mathbf{V}_j$ and the instruction to localize the keypoint $k_{i}$ . For example: + +"Point to the left wing tip in this image." + +This leverages Molmo's capability to understand natural language instructions and perform point-level localization. + +# 4.3. Zero-Shot 3D Keypoint Detection + +After obtaining the 2D keypoint detections from Molmo, we reconstruct the 3D coordinates of each keypoint in shape by back-projecting the 2D points into 3D space and aggregating the results from multiple views. Fig. 3 shows aggregated points from different numbers of views. + +Back-Projection of 2D Keypoints. Given the camera projection matrices $\{\mathbf{C}_j\}_{j=1}^M$ corresponding to each image $\mathbf{V}_j$ , + +![](images/f3a9302ec4706076b3014025c7e4043f6f1da73e55faddb39ad66fcc2ccd8abc.jpg) +(a) 6 Views + +![](images/3941f8fc1312f7644bb5de8c7359884496c08bbabc5d82c9e14bc57fa6bf60d5.jpg) +(b) 26 Views + +![](images/bb91b4fe8a9117aa15c33fe77b4d78666570a59c99e5fa72e184de5880f16196.jpg) +(c) 46 Views +Figure 3. The number of rendered views versus the detected keypoints after aggregation. This figure shows how varying the number of rendered views affects the total number of keypoints detected by ZeroKey. The prompt here is "corner of the table". + +we can back-project the 2D points $\mathbf{p}_{i,j}$ into 3D space. Assuming a pinhole camera model, each camera projection matrix $\mathbf{C}_j$ maps 3D points $\mathbf{P} \in \mathbb{R}^3$ to 2D points $\mathbf{p} \in \mathbb{R}^2$ : + +$$ +\mathbf {p} _ {i, j} \sim \mathbf {C} _ {j} \mathbf {P} _ {i, j}, \quad i = 1, \dots , N; \quad j = 1, \dots , M, \tag {2} +$$ + +where $\sim$ denotes equality up to scale. To back-project $\mathbf{p}_{i,j}$ into 3D space, we compute the ray $\mathbf{r}_{i,j}(t)$ corresponding to $\mathbf{p}_{i,j}$ as $\mathbf{r}_{i,j}(t) = \mathbf{C}_j^{-1}(\mathbf{p}_{i,j},t)$ , where $t$ represents the depth along the ray. We intersect the ray with the 3D shape $S$ to find the 3D point $\mathbf{P}_{i,j} = \operatorname{Intersect}(\mathbf{r}_{i,j},S)$ . During the rasterization process, we cache the depth $t$ for each pixel and utilize that information for back-projection, eliminating the need for expensive intersection calculations. + +Soft Voting Weights. To handle sharp angular intersections and stabilize the mapping process, we back-project a $h \times h$ patch $\mathbb{S}_{i,j}$ in the region centered at $\mathbf{p}_{i,j}$ . We exclude all pixels that do not intersect with the mesh, then compute the mean of the back-projected points within the patch to obtain a stable back-projection of $\mathbf{P}_{i,j}$ . We also assign a soft-voting weight $\mathbf{W}_{i,j}$ to each $\mathbf{P}_{i,j}$ derived from the Gaussian kernel centered at $p_{i,j}$ : + +$$ +\mathbf {W} _ {i, j} = \sum_ {\mathbf {p} _ {m, n} \in \mathbf {N} _ {i, j}} \frac {1}{\sigma \sqrt {2 \pi}} \exp \left(- \frac {\left\| \mathbf {p} _ {i , j} - \mathbf {p} _ {m , n} \right\| ^ {2}}{2 \sigma^ {2}}\right), \tag {3} +$$ + +where $\mathbf{N}_{i,j} = \{\mathbf{p}_{m,n}|\mathbf{p}_{i,j}\in \mathbb{S}_{m,n}\}$ . The point $\mathbf{P}_{i,j}$ gains higher weights when its corresponding pixel $\mathbf{p}_{i,j}$ in the patch $\mathbb{S}_{m,n}$ is closer to the center pixel $\mathbf{p}_{m,n}$ . The $\sigma$ is set to $\frac{h}{3}$ in our implementation. + +Aggregation of 3D Keypoints. For each of the $N$ keypoints $k_{i}$ , we collect the set of back-projected 3D points from all views: $\mathcal{P}_i = \{\mathbf{P}_{i,j} \mid j = 1, \dots, M\}$ . To obtain the final 3D keypoint location $\hat{\mathbf{P}}_i$ , we aggregate the set $\mathcal{P}_i$ . A straightforward method is to compute the weighted mean of the points $\hat{\mathbf{P}}_i = \frac{\sum_{j=1}^{M} \mathbf{W}_{i,j} \mathbf{P}_{i,j}}{\sum_{j=1}^{M} \mathbf{W}_{ij}}$ . However, naive summarization often produces unsatisfactory results. This occurs because back-projection may become unstable when the angle between the ray and the mesh intersection is sharp. In such cases, the ray may intersect different parts of the object, leading to inconsistencies. Molmo [12] can also gen + +erate multiple keypoints for one prompt, resulting in ambiguity and inconsistencies in the summarized coordinates. Patchify back projection can help to address the issue of sharp angular intersections by providing a weight for each back-projected point. One can also extract the confidence weighting from Molmo's feature map and apply it to the predictions. Given that the predictions from Molmo can produce multiple points for each prompt and may seem noisy from different perspectives, we treat Molmo as a black box in our approach. We aggregate all the points into a single point cloud, which contains an unknown number of key points. Subsequently, we apply hierarchical density-based spatial clustering to this aggregated point cloud. + +Filtering and Refinement. Specifically, we apply HDBSCAN to the set of predicted points $\mathcal{P}_i$ obtained from multiple views to identify dense clusters that represent consistent predictions across views. HDBSCAN operates by defining the mutual reachability distance $d_{\mathrm{m}}$ between two points $\mathbf{P}_{i,j}$ and $\mathbf{P}_{i,l}$ as: + +$$ +d _ {\mathrm {m}} = \max \left(\operatorname {c o r e} _ {k} \left(\mathbf {P} _ {i, j}\right), \operatorname {c o r e} _ {k} \left(\mathbf {P} _ {i, l}\right), d \left(\mathbf {P} _ {i, j}, \mathbf {P} _ {i, l}\right)\right), \tag {4} +$$ + +where $\mathrm{core}_k(\mathbf{P}_{i,j})$ is the core distance of point $\mathbf{P}_{i,j}$ , defined as the distance to its $k$ -th nearest neighbor, and $d(\mathbf{P}_{i,j}, \mathbf{P}_{i,l})$ is the Euclidean distance between points $\mathbf{P}_{i,j}$ and $\mathbf{P}_{i,l}$ . We apply each point $\mathbf{P}_{i,j}$ corresponding weights $\mathbf{W}_{i,j}$ obtained from Eq.3. Throughout our experiments, we consistently set $\min \mathrm{Pts} k = 10$ . By employing the mutual reachability distance with this fixed $\min \mathrm{Pts}$ parameter, we effectively filter out noise and robustly identify the most stable clusters among the predicted points. + +# 5. Practical 3D Understanding with MLLMs + +Gauging to what extent MLLMs are able to reason about 3D structures remains an underexplored question with significant implications and applications. Molmo [12] has demonstrated strong 2D spatial reasoning abilities. We find that this reasoning generalizes to some degree to 3D. In this section, we propose to probe Molmo's 3D reasoning capabilities by leveraging Schelling points and describability. Through this, we aim to characterize the strengths and limitations of its 3D reasoning, specifically the strengths and limitations of the 3D reasoning induced by training with pixel-level annotations. + +# 5.1. Schelling Points on 3D Surface + +Schelling points, as introduced by Schelling [39], are specific focal points that people are likely to choose independently due to their prominence. In visual perception, these keypoints stand out due to distinct features. We hypothesize that the salient points people select often have clear, widely understood names in language. Therefore, an effective MLLM should be able to better predict such salient points by using its knowledge of both language and vision. + +![](images/d5cbbb979c8ca4be8121588c926d3aa866408c8ac6304e22f615de605a0028a4.jpg) +We ask Molmo to Describe Molmo predicts point +Figure 4. We ask Molmo to describe the green point, and using this as a prompt ZeroKey predicts the blue point. We show that salient points, given by the Schelling Points paper [10], are more easily describable and consistently retrievable than arbitrary points. Arbitrary points can cause ZeroKey to not find any points at all. + +To illustrate this idea, we analyze ZeroKey's ability to generate salient points that align with human-selected Schelling points. Using a dataset [10] that captures human-selected Schelling points, we prompt ZeroKey with natural language descriptions provided by ChatGPT, simulating natural language descriptions that humans might use. This aims to replicate the human process of identifying and naming keypoints based on visual and linguistic salience. An illustration and the detailed settings can be found in the supplementary material. Our findings show that ZeroKey's generated points align with high-density regions reported by Chen et al. [10], suggesting that its 3D understanding of shape corresponds, to some extent, with human perception. + +# 5.2. Point Describability and Consistency + +During our experiments, we observed that certain keypoints were more difficult to retrieve than others. This was raised when performing our quantitative analysis on the Keypoint-Net dataset [54], which provides ground-truth keypoints for various 3D models. The assignment of a concise and descriptive text prompt to each ground truth point proved to be part of the challenge. We noticed a correlation: when a point was easy to name succinctly by a human observer, ZeroKey was able to retrieve it more effectively. Based on this intuition, we propose to study how describable different points are and how consistently ZeroKey can retrieve them. + +To quantify how "describable" a point is, we leverage the distributions reported in the Schelling points paper [10]. Points with high-density values in the distributions reported by [10] are considered more describable due to their inherent semantic significance. We refer to such points as salient points. The evaluation algorithm is described in the supplementary material. + +The qualitative results of our evaluation are showcased in Fig. 4. We ask Molmo to describe the green dot, using a simple prompt. We then give the prompt generated by Molmo to Zerokey to predict a 3D location of the original + +![](images/adf2bd9ed34b67b6d866109a0871de10d9e678406d04486e2a5cd6aee7925c34.jpg) +Figure 5. We show through a quantitative study that "salient" points are retrieved with higher accuracy than "non-salient" ones, regardless of the distance threshold used. + +point. For salient points, ZeroKey successfully leverages Molmo's prompt to relocate the queried point with high accuracy. However, for non-salient (random) points, retrieval is significantly more challenging. ZeroKey may identify similar object parts (e.g., an airplane wing) or even unrelated regions (e.g., an incorrect area on a teddy bear). Additionally, we observed a fail state where for some non-salient points Molmo was unable to give a good prompt, resulting in ZeroKey affirming that no points matching the prompt were present in the scene. + +Fig. 5 presents the quantitative results on the whole Schelling points dataset, showing that retrieval accuracy is significantly higher for semantically meaningful points than for arbitrary ones. This suggests that ZeroKey, guided by Molmo's descriptions, is better at reasoning about 3D structure when a point carries strong linguistic significance. These findings show that Molmo's 3D understanding is closely tied to the describability of keypoints. + +# 6. Experiments + +# 6.1. Setup and Dataset + +We evaluate our method using the KeypointNet dataset. Our evaluation strategy follows the approach of You et al. [55], which computes the Intersection over Union (IoU) between predicted keypoints and ground-truth keypoints from the KeypointNet dataset, using varying distance thresholds. We also stick to the same three classes of the dataset for evaluation: airplane, chair, and table. Please refer to our supplementary for the details. + +# 6.2. Zero-Shot 3D Keypoint Detection Baseline + +RedCircle. The Red Circle method [42] is one of the simplest and most straightforward ways to highlight points in the image based on a text query by randomly sampling points and highlighting them with a red circle and picking the one point with the highest CLIP similarity to the text + +
MethodIoU (in %)
Distance Thresholds0.0010.010.020.030.040.050.060.070.080.090.10
HARRIS-3D [43]0.150.762.193.966.168.8811.9115.1318.6322.1025.69
SIFT-3D [38]0.291.052.624.386.659.4212.3815.6519.3922.7126.15
ISS [57]0.491.102.694.677.2710.1413.2916.5620.0223.5827.20
USIP [26]0.831.703.255.248.0011.7615.9820.6325.5630.2033.67
D3FEAT [6]2.363.867.2710.3713.3717.1221.4126.1230.3934.8238.41
UKPGAN [55]3.956.5412.2717.8622.4226.5530.2834.3238.4642.6546.49
FSKD [5]7.007.9411.1716.7423.9931.1438.1443.9749.3053.6257.03
B2-3D [49]12.4120.2935.5546.2552.7457.7261.6164.1966.5668.5570.57
PaliGemma 2 [45]0.000.341.032.984.255.706.628.178.9211.4911.65
RedCircle [42]0.210.340.641.161.903.054.817.5511.0614.9218.50
GPT-4o0.180.481.202.614.196.048.4810.8914.0317.0320.73
CLIP-DINOiser [50]0.731.413.004.947.319.8012.6615.5218.5221.7625.56
StablePoints [17]3.665.809.9413.3216.5819.9123.4827.2631.2834.8938.22
ZeroKey (Ours)9.1213.1624.0137.4348.4356.6062.0667.5371.7475.7579.43
+ +Table 1. Comparison of IoU between the predicted and ground-truth keypoints from KeypointNet using different methods across various geodesic distance thresholds. The bold text indicates the best zero-shot methods, while the underlined text highlights the best methods. + +query. We lift the prediction of this method to 3D using the same lifting procedure used in our method to compare 3D Zero-shot keypoint detection. + +CLIP-DINOiser. The CLIP-DINOiser [50] combines the strengths of CLIP and DINO to perform zero-shot object localization and segmentation based on text query in images. To adapt the CLIP-DINOiser for zero-shot 3D keypoint detection, we first apply the method to each view of the 3D scene to obtain 2D keypoint predictions corresponding to the text prompt. For each image, the CLIP-DINOiser produces a heatmap indicating the regions of interest related to the query. We select the point with the highest activation in the heatmap as the 2D keypoint prediction in that view. We then lift these 2D keypoints to 3D using the same backprojection technique described in our method. + +StablePoint. In addition to the supervised and few-shot methods, we also compared our approach with an unsupervised baseline called StablePoint [17]. This method learns keypoints by optimizing text embeddings from latent diffusion models. The main idea is to identify text embeddings that guide the generative model to consistently focus on compact regions within images, which are then used as keypoints. Unlike our approach, this baseline does not retrieve keypoint according to a text prompt and operates fully unsupervised. However, it does necessitate specifying the exact number of keypoints we wish to detect. In our experiments, we set the number of keypoints to match the number of ground-truth points in the KeypointNet dataset. + +GPT-4o. In this Baseline, we replace our model with GPT-4o to observe variations in performance. Notably, we highlight that MLLMs trained solely on image-level tasks fail to provide useful signals for keypoint detection. In the GPT-4o version, we replace Molmo model in our method with the latest GPT-4o (2024-08-06) and evaluate its performance on point-level tasks. This comparison highlights the necessity of specialized training or architectural features that support + +fine-grained reasoning in visual contexts. + +PaliGemma 2. The 3B Mix 224 variant of PaliGemma 2 is a versatile and lightweight MLLM based on the Gemma language model [45]. It processes both images and text as input and generates text as output. Designed for a range of tasks—including short video captioning, visual question answering, text reading, object detection, and object segmentation—PaliGemma offers flexible multimodal capabilities. In our work, we adapt PaliGemma's object segmentation functionality to identify points of interest. Given an image and a text prompt, PaliGemma outputs a segmentation mask around the relevant region. We then select the area overlapped with the bounding box prediction as the detected points and lift these 2D points to 3D using our method. + +# 6.3. Quantitative and Qualitative Analysis + +Our KeypointNet evaluation shows (see Table.1) that our Zero-Shot method significantly outperforms MLLM-based baselines (PaliGemma 2[45], GPT-4o, CLIP-DINOiser [50]) across all distance thresholds. This provides strong evidence for our claim that the pixel-level annotations used to train MLLMs can be leveraged to both extract and name salient keypoints on 3D models without requiring any ground truth labels or supervision. Furthermore, our method achieves IoU levels comparable to those of reference-based Few-Shot and fully supervised methods tailored for this dataset, such as B2-3D [49]. Notably, after a certain distance threshold, our approach surpasses even few-shot methods, despite being zero-shot. The drop in performance at smaller thresholds is expected, given the inherently semantic nature of our approach, which prioritizes high-level understanding over fine-grained localization. + +Side-by-side comparisons between ground truth keypoints and our Zero-Shot predictions, a figure of GPT-4o fails to precisely locate the keypoint, and a comparison of our detected keypoints with those of other baselines is + +![](images/97cdf16d4c098119a94108d123e37ddae9f8b5dc71f12cb892dbcb04974d8196.jpg) +Figure 6. We compare against baselines CLIP-DINOiser and RedCircle. While both baselines identify some prominent regions, they fall in accurately localizing keypoints according to the text prompt. In contrast, our method precisely locates keypoints. + +![](images/1eaa762f4ddd98a782d0b5f73b80f3a3a630ce5021dbdfa1aad4aabc11be0ab7.jpg) + +![](images/96214c8a92fdac6793f99b4a2e1e730d14345978f52b7373adf4073d013bbaf5.jpg) + +![](images/032eb17c6c4c63ce99f825c2cc54bdffd7f7bf28e65b966f03f43404f8e8c99c.jpg) + +shown in Fig. 6. We observe that our method consistently produces reasonable predictions, whereas MLLM-based baselines often fail to generate adequate results. These findings show the need for an MLLM backbone trained with point-level tasks for precise keypoint detection, underscore the potential of point-level reasoning as a powerful task for MLLMs, suggesting that it may induce a more generalizable understanding of vision and geometry compared to conventional image-level tasks like detection and segmentation. + +# 6.4. Ablation Studies + +In this ablation study, we modify different components of our method to gain deeper insights into the point-level reasoning abilities of MLLMs. Specifically, we compare the effects of various text prompts, experiment with different MLLM backbones, and evaluate the impact of aggregating information from different numbers of views. + +Global Text prompt We evaluate the effectiveness of a point-specific text prompt. The original method utilizes prompts tailored to specific points in the image, providing precise guidance to the model. In the ablated version, we replace these point-specific text prompts with a global prompt: "Point to all salient points in this image." We then compare the detected set of keypoints with the ground truth to assess how this change affects the model's ability to localize specific points. This experiment helps us understand the importance of prompt specificity in guiding the MLLM's attention to particular regions of the image. + +Aggregation ablation Our original method aggregates keypoint information from 26 viewpoints to improve accuracy. In the 6-views variant, we reduce the number of viewpoints to 6. In another variant, we compare the performance of direct average against HDBSCAN clustering for aggregating detections. By evaluating these different aggregation strategies, we aim to determine how the number of views and the method of combining information influence the model's ability to accurately identify keypoints. + +The ablation results shown in Fig.7 illustrate the relative performance compared with ZeroKey. The ability to + +![](images/98ad0360ca1e87c73aa38bfa61ac823410170de39be52d576f2b03d5d244c251.jpg) +Figure 7. Comparison of the performance across different configurations: (blue) our original method; (red) results with a Global Text prompt; (orange, purple, brown) results using different rendering; (green) results without HDBSCAN clustering. This figure shows the impact of each modification on the overall performance. + +achieve $80\%$ performance with only 6 views indicates that our method efficiently aggregates points. The poor performance of direct averaging suggests the necessity of HDBSCAN clustering. Binary Back-projection applies a constant weight of 1 for all points, rather than utilizing the soft voting weights, which significantly hinders its performance. Additionally, Replacing the shading with pointmap color and mesh texture from the dataset does not improve performance. This is because pointmap might be out-of-distribution of Molmo's training data. The poor quality of texture from KeypointNet also limit its effectiveness. + +# 7. Conclusion and Future Work + +In this paper, we presented a novel zero-shot approach for 3D keypoint detection by harnessing the capabilities of recent MLLMs trained with enhanced localized reasoning capabilities. Unlike traditional methods that rely heavily on annotated 3D datasets and extensive supervised training, our method leverages the pixel-level annotations inherent in MLLMs to both extract and name salient keypoints without any ground truth labels or supervision. Our evaluations demonstrate the efficacy of our approach and suggest that point-level reasoning is an effective way to endow MLLMs with a robust understanding of vision and geometry. + +We believe that this work opens the door to entirely novel applications by providing a bridge between 3D data and MLLMs with localized reasoning. This can help solve downstream tasks such as shape manipulation, deformation, and processing, which typically require significant domain knowledge. Furthermore, this link can also help to enhance the power of MLLMs by providing feedback thanks to multiview consistency from 3D data, thus potentially helping to improve the robustness of their pixel-level understanding. + +Acknowledgments. Parts of this work were supported by the ERC Consolidator Grant 101087347 (VEGA). The authors also gratefully acknowledge gifts from Ansys and Adobe Inc. This work was also supported by the KAUST Center of Excellence for Generative AI, under award number 5940, and the KAUST Ibn Rushd Postdoc Fellowship program. + +# References + +[1] Ahmed Abdelreheem, Abdelrahman Eldesokey, Maks Ovsjanikov, and Peter Wonka. Zero-shot 3d shape correspondence. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3 +[2] Ahmed Abdelreehem, Ivan Skorokhodov, Maks Ovsjanikov, and Peter Wonka. Satr: Zero-shot semantic segmentation of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15166-15179, 2023. 3 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1, 2, 3 +[4] AI Anthropic. Claude 3.5 sonnet model card addendum. Claude-3.5 Model Card, 2024. 2 +[5] Souhaib Attaiki and Maks Ovsjanikov. Ncp: Neural correspondence prior for effective unsupervised shape matching. Advances in Neural Information Processing Systems, 35:28842-28857, 2022. 2, 7 +[6] Xuyang Bai, Zixin Luo, Lei Zhou, Hongbo Fu, Long Quan, and Chiew-Lan Tai. D3feat: Joint learning of dense detection and description of 3d local features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6359-6367, 2020. 7 +[7] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2 +[8] Declan Campbell, Sunayana Rane, Tyler Giallanza, Nicolò De Sabbata, Kia Ghods, Amogh Joshi, Alexander Ku, Steven M Frankland, Thomas L Griffiths, Jonathan D Cohen, et al. Understanding the limits of vision language models through the lens of the binding problem. arXiv preprint arXiv:2411.00238, 2024. 2 +[9] Nenglun Chen, Lingjie Liu, Zhiming Cui, Runnan Chen, Duygu Ceylan, Changhe Tu, and Wenping Wang. Unsupervised learning of intrinsic structural representation points. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9121-9130, 2020. 3 +[10] Xiaobai Chen, Abdulhair Saparov, Bill Pang, and Thomas Funkhouser. Schelling points on 3d surface meshes. ACM Transactions on Graphics (TOG), 31(4):1-12, 2012. 3, 6 +[11] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multiview prediction for 3d semantic scene segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 452-468, 2018. 3 + +[12] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 2, 4, 5 +[13] Clara Fernandez-Labrador, Ajad Chhatkuli, Danda Pani Paudel, Jose J Guerrero, Cédric Demonceaux, and Luc Van Gool. Unsupervised learning of category-specific symmetric 3d keypoints from point sets. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 546-563. Springer, 2020. 3 +[14] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1-11, 2021. 3 +[15] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Voint cloud: Multi-view point cloud representation for 3d understanding. In The Eleventh International Conference on Learning Representations, 2023. 3 +[16] Abdullah Hamdi, Faisal AlZahrani, Silvio Giancola, and Bernard Ghanem. Mvtn: Learning multi-view transformations for 3d understanding. International Journal of Computer Vision, 2024. 3 +[17] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Xingzhe He, Hossam Isack, Abhishek Kar, Helge Rhodin, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised keypoints from pretrained diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22820-22830, 2024. 3, 7 +[18] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 2, 3, 4 +[19] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4421-4430, 2019. 3 +[20] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3d: Learning fine-grained class-agnostic 3d segmentation without manual labels. European Conference on Computer Vision (ECCV), 2024. 3 +[21] Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3d keypoint discovery for shape control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12783-12792, 2021. 3 +[22] Maximilian Jaritz, Jiayuan Gu, and Hao Su. Multi-view pointnet for 3d scene understanding. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 0-0, 2019. 3 +[23] Evangelos Kalogerakis, Melinos Averkiou, Subhransu Maji, and Siddhartha Chaudhuri. 3d shape segmentation with + +projective convolutional networks. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 3779-3788, 2017. 3 +[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 3 +[25] Abhijit Kundu, Xiaqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In European Conference on Computer Vision (ECCV), pages 518-535. Springer, 2020. 3 +[26] Jiaxin Li and Gim Hee Lee. Usip: Unsupervised stable interest point detection from 3d point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 361-370, 2019. 7 +[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 2, 3 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3 +[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3 +[30] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. 3d-to-2d distillation for indoor scene parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4464–4474, 2021. 3 +[31] Yecheng Lyu, Xinming Huang, and Ziming Zhang. Learning to segment 3d point clouds in 2d image space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12255-12264, 2020. 3 +[32] Jinjie Mai, Abdullah Hamdi, Silvio Giancola, Chen Zhao, and Bernard Ghanem. Egoloc: Revisiting 3d object localization from egocentric videos with visual queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 45-57, 2023. 3 +[33] Jinjie Mai, Wenxuan Zhu, Sara Rojas, Jesus Zarzar, Abdullah Hamdi, Guocheng Qian, Bing Li, Silvio Giancola, and Bernard Ghanem. Tracknerf: Bundle adjusting nerf from sparse and noisy views via feature tracks, 2024. 3 +[34] OpenAI. Gpt-4 technical report, 2023. 1, 2, 4 +[35] Wujian Peng, Sicheng Xie, Zuyao You, Shiyi Lan, and Zuxuan Wu. Synthesize diagnose and optimize: Towards fine-grained vision-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13279-13288, 2024. 2 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2, 3 +[37] Pooyan Rahmanzadehgervi, Logan Bolton, Mohammad Reza Taesiri, and Anh Totti Nguyen. Vision language models are blind. arXiv preprint arXiv:2407.06581, 2024. 2 + +[38] Blaine Rister, Mark A Horowitz, and Daniel L Rubin. Volumetric image registration from invariant keypoints. IEEE Transactions on Image Processing, 26(10):4900-4910, 2017. 7 +[39] Thomas C Schelling. The Strategy of Conflict: with a new Preface by the Author. Harvard university press, 1980. 5 +[40] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In International Conference on Robotics and Automation (ICRA), 2023. 3 +[41] Ruoxi Shi, Zhengrong Xue, Yang You, and Cewu Lu. Skeleton merger: an unsupervised aligned keypoint detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 43-52, 2021. 3 +[42] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11987-11997, 2023. 3, 6, 7 +[43] Ivan Sipiran and Benjamin Bustos. Harris 3d: a robust extension of the harris operator for interest point detection on 3d meshes. The Visual Computer, 27:963-976, 2011. 7 +[44] Supasorn Suwajanakorn, Noah Snavely, Jonathan J Tompson, and Mohammad Norouzi. Discovery of latent 3d keypoints via end-to-end geometric reasoning. Advances in neural information processing systems, 31, 2018. 3 +[45] Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Machel Reid, Maciej Mikula, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. + +Gemma: Open models based on gemini research and technology, 2024. 7 +[46] Brian H Wang, Wei-Lun Chao, Yan Wang, Bharath Hariharan, Kilian Q Weinberger, and Mark Campbell. Ldls: 3-d object segmentation through label diffusion from 2-d images. IEEE Robotics and Automation Letters, 4(3):2902-2909, 2019. 3 +[47] Hanyu Wang, Jianwei Guo, Dong-Ming Yan, Weize Quan, and Xiaopeng Zhang. Learning 3d keypoint descriptors for non-rigid shape matching. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018. 3 +[48] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International conference on machine learning, pages 23318-23340. PMLR, 2022. 1 +[49] Thomas Wimmer, Peter Wonka, and Maks Ovsjanikov. Back to 3d: Few-shot 3d keypoint detection with back-projected 2d features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4154-4164, 2024. 2, 4, 7 +[50] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzcinski, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation. In Proc. ECCV, 2024. 1, 7 +[51] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European Conference on Computer Vision, pages 131-147. Springer, 2024. 2, 4 +[52] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1179-1189, 2023. 3 +[53] Zi Jian Yew and Gim Hee Lee. 3dfeat-net: Weakly supervised local 3d features for point cloud registration. In Proceedings of the European conference on computer vision (ECCV), pages 607-623, 2018. 3 +[54] Yang You, Yujing Lou, Chengkun Li, Zhoujun Cheng, Liangwei Li, Lizhuang Ma, Cewu Lu, and Weiming Wang. Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. arXiv preprint arXiv:2002.12687, 2020. 3, 4, 6 +[55] Yang You, Wenhai Liu, Yanjie Ze, Yong-Lu Li, Weiming Wang, and Cewu Lu. Ukpgan: A general self-supervised keypoint detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 6, 7 +[56] Chenhui Zhang and Sherrie Wang. Good at captioning bad at counting: Benchmarking gpt-4v on earth observation data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7839–7849, 2024. 2 + +[57] Yu Zhong. Intrinsic shape signatures: A shape descriptor for 3d object recognition. In 2009 IEEE 12th international conference on computer vision workshops, ICCV Workshops, pages 689-696. IEEE, 2009. 7 +[58] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. arXiv preprint arXiv:2310.06773, 2023. 2, 4 +[59] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 3 \ No newline at end of file diff --git a/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/images.zip b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..252a009d61ebad49ac1aed20561971c220611ebd --- /dev/null +++ b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c912fbf04adb41c321969662b1f6c84503687a6f2b143615bb5c4445040949da +size 422077 diff --git a/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/layout.json b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..75cc602b1f67f8db941ef6ac4ea629fc0250ca7f --- /dev/null +++ b/zerokeypointlevelreasoningandzeroshot3dkeypointdetectionfromlargelanguagemodels/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79bae3c8c6e1e531a44587664c53b0023b75d93f7d2fe10802a75caf1e55c232 +size 368627 diff --git a/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/21e6de6e-abc8-43a4-bcfb-a4a84f3eeee1_content_list.json b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/21e6de6e-abc8-43a4-bcfb-a4a84f3eeee1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40a8f5aa86fd0791403d2a0d4d2cccea85bfadd7 --- /dev/null +++ b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/21e6de6e-abc8-43a4-bcfb-a4a84f3eeee1_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d2ba5e5a1da3f6bac3602e9c9d6ef366aeffe902e6e38b275b16a16c010631e +size 83761 diff --git a/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/21e6de6e-abc8-43a4-bcfb-a4a84f3eeee1_model.json b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/21e6de6e-abc8-43a4-bcfb-a4a84f3eeee1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8606ca16a1e9a9bfc8a2bc6a6b6ac4caa7e51728 --- /dev/null +++ b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/21e6de6e-abc8-43a4-bcfb-a4a84f3eeee1_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55ff763ab1fa2d0bd27a564e672534cefd5d55b9096c15156d05037c9e8e39d3 +size 103072 diff --git a/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/21e6de6e-abc8-43a4-bcfb-a4a84f3eeee1_origin.pdf b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/21e6de6e-abc8-43a4-bcfb-a4a84f3eeee1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d7d05b61dbcd72e3bb09fcad78a86f002650189a --- /dev/null +++ b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/21e6de6e-abc8-43a4-bcfb-a4a84f3eeee1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3e0714992eabb67c22989c390991bd2da95e278c3adff9684050b67432c6c5a +size 2091227 diff --git a/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/full.md b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..be9af16cbc2ced4eaaedfc1be2c26522cdeecf79 --- /dev/null +++ b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/full.md @@ -0,0 +1,315 @@ +# Zero-Shot Composed Image Retrieval via Dual-Stream Instruction-Aware Distillation + +Wenliang Zhong $^{1}$ , Rob Barton $^{2}$ , Weizhi An $^{1}$ , Feng Jiang $^{1}$ , Hehuan Ma $^{1}$ , Yuzhi Guo $^{1}$ , Abhishek Dan $^{2}$ , Shioulin Sam $^{2}$ , Karim Bouyarmane $^{2}$ , and Junzhou Huang $^{1}$ + +1The University of Texas at Arlington, 2Amazon + +{wxz9204, weizhi.an, fxj8843, hehuan.ma, yuzhi.guo}@mavs.uta.edu {rab, abhidan, shioulin, bouykari}@amazon.com jzhuang@uta.edu + +# Abstract + +Composed Image Retrieval (CIR) targets the retrieval of images conditioned on a reference image and a textual modification, but constructing labeled triplets (reference image, textual modification, target image) is inherently challenging. Existing Zero-Shot CIR (ZS-CIR) approaches often rely on well-aligned vision-language models (VLMs) to combine visual and textual inputs, or use large language models (LLMs) for richer modification understanding. While LLM-based methods excel in capturing textual details, they are computationally costly, slow to infer, and often restricted by proprietary constraints. In this paper, we argue that the superior performance of LLM-based ZS-CIR methods primarily stems from their capacity to follow instructions, an aspect largely missing in more efficient projection-based models built upon VLMs. To bridge this gap, we introduce DistillCIR, a dual-stream distillation framework that transfers LLMs' instruction-following capability into compact, projection-based architectures. By synthesizing triplet data with an LLM and incorporating a novel reasoning process, DistillCIR learns both composed retrieval and instruction awareness. In addition, we train an open-source multimodal LLM on the generated data, and further distill its instruction-aware embeddings into the projection-based model. Without any reliance on LLMs at inference, DistillCIR significantly surpasses state-of-the-art ZS-CIR methods in both performance and efficiency, offering a promising direction for instruction-aware, lightweight CIR. Project page is at: https://distillcir.github.io/. + +# 1. Introduction + +Composed Image Retrieval (CIR) is a retrieval task that refers to finding target images with a composition of ref- + +![](images/0903b17517267ea0362a177169f7ac9afb81ec3edce986c4d5db3bc5ac6968fe.jpg) +Figure 1. Comparison with DistillCIR and Existing Methods. Pic2Word relies solely on CLIP for both training and inference, offering efficiency but limited comprehension of textual modifications. In contrast, CIREVL employs an LLM to parse instructions during inference, making the process slower and less flexible. DistillCIR bridges this gap by distilling the LLM's instruction-following capabilities into a CLIP-based CIR model, achieving both superior performance and efficient inference. + +erence images and a textual modification requirement [2, 12, 36]. Such a task is increasingly attractive in real-world scenarios, such as e-commerce platforms where users might want to find similar products with specific alterations [5, 62]. However, CIR often faces significant data acquisition challenges, as collecting training samples typically requires substantial human effort. In response, recent works explore Zero-Shot CIR (ZS-CIR) [4, 42], where models are pretrained on specific datasets and then applied directly to downstream tasks without additional tuning. + +Most ZS-CIR methods are projection-based models [2, 12, 36] founded on well-aligned Vision-Language Models (VLMs) such as CLIP [35]. Specifically, they generate + +composed embeddings by projecting the image embedding from the VLM into the text embedding space and combining it with the textual modification for the output embedding. Despite their staged progress, concerns remain about whether they truly understand the textual modification or just simply combine multimodal information [47]. More recently, some methods [6, 19, 21, 22, 27] explore the involvement of Large Language Models (LLMs) [57] to enhance models' comprehension of modifications. These methods either process the textual modification and reference images with LLMs or directly leverage LLMs for embedding generation. Though qualitative experiments are impressive, these models are expensive to deploy and slow in inference due to the significantly larger number of parameters in LLMs. Worse still, some methods [21] inherently rely on proprietary LLMs in downstream tasks, which may be prohibited in commercial scenarios due to privacy concerns. + +Witnessing the rise of LLM-based ZS-CIR models and their shortcomings, a question naturally arises: what internal capabilities make these LLM-based models more powerful than most projection-based models? In this paper, we argue that projection-based models based on pretrained VLMs excel at multimodal retrieval; however, they struggle to understand textual modifications. It is the instruction-following capability [46, 60] within LLMs that makes LLM-based models outperform projection-based models. Motivated by this argument, we propose to distill the instruction awareness of LLMs into projection-based models, making them small in size yet powerful in performance. + +Nevertheless, distillation from LLMs to CIR models based on CLIP is challenging, primarily because of the inherent task difference: LLMs are trained to generate content, while CLIP is designed for multimodal representation alignment. To address this challenge, we propose a dual-stream distillation strategy. Our intuition is to leverage a powerful LLM [1, 6] to enhance existing data with its internal knowledge to train CIR models. Specifically, we prompt the LLM to extend an image-caption pair to a triplet similar to the CIR format, which contains an image, a textual modification, and a modified caption. While the triplet data can teach the projection-based model the basics of composed retrieval, it does not fully reflect the instruction-following capability of LLMs. To elicit instruction awareness, we add a reasoning process [48] during generation, where the LLM explicitly describes how it derives the modification and how the modification affects the final caption. A novel reasoning loss is then proposed to guide the projection-based model in learning from this detailed generation process. Besides learning from reasoning, it is also crucial to learn instruction-aware embeddings. To achieve this, we propose training an open-source Multimodal LLM (MLLM) [28, 29] using the triplet data and distilling its instruction-following capability into the + +projection-based model through another feature distillation loss. We term our dual-stream distilled model DistillCIR, which not only achieves superior performance compared to existing LLM-based ZS-CIR models but also remains efficient in model size. Furthermore, DistillCIR demonstrates extraordinary instruction-following capabilities without requiring any LLMs during inference. + +To summarize, our contributions are threefold: (1) We propose a novel insight for enhancing projection-based ZS-CIR models through instruction awareness, which originally resides in LLMs but missing in retrieval VLMs. (2) We introduce DistillCIR, a novel method that learns the instruction-following capability from LLMs through dual-stream distillation. (3) Extensive experiments on four public datasets are conducted, and the results reflect the superiority of DistillCIR. Moreover, DistillCIR is instruction-aware without requiring any LLMs during inference. + +# 2. Related Works + +# 2.1. Composed Image Retrieval + +Due to the scarcity of labeled data, recent research in Zero-Shot Composed Image Retrieval (ZS-CIR) has gained traction. Existing ZS-CIR methods generally fall into two categories: (1) Projection-based Models [2, 12, 36], which often build on established multimodal retrieval frameworks [25, 35]. These methods generate a composed embedding by projecting image embeddings into text space and combining them with modification embeddings [4, 36] or by jointly projecting both image and text embeddings into a shared interpolated space [18]. Although effective, they frequently struggle to interpret textual modifications [47]. (2) LLM-based Models [21, 27], which harness the capabilities of LLMs. Some leverage powerful LLMs at inference time to parse textual modifications [21], while others use LLMs primarily as text encoders [22, 27]. However, applying LLMs directly for downstream tasks may lead to slower, less efficient inference, and privacy issue in many commercial contexts. Although some approaches [19] sidestep this by using LLMs to synthesize training data, they generally rely on a specific data gallery, limiting broader ZS-CIR applicability. To address these challenges, we propose distilling the LLM's instruction-following ability into projection-based CIR models, thus enabling robust textual understanding without the computational overhead of LLM-based inference. + +# 2.2. Vision-Language Model Distillation + +Model distillation [11, 14, 34, 44] aims to transfer knowledge from a teacher model to a student model. Most existing approaches concentrate on feature-level alignment [8, 9, 33, 45, 50, 52], matching representations between teacher and student networks. With the rise of large language models (LLMs), recent work has also begun exploring how to + +distill from a larger LLM into a smaller one [15, 32, 38, 46, 51]. One popular strategy prompts the teacher to synthesize data, which the student then uses for training to mimic the teacher's outputs. Beyond LLMs, some studies also investigate the distillation of retrieval models, but they primarily focus on feature alignment. Crucially, few endeavors address distilling instruction-awareness from LLMs into retrieval models. In this paper, we tackle this issue with a novel dual-stream distillation strategy. + +# 3. Methodology + +The overview of DistillCIR is shown in Figure 2. The subsequent sections provide details on the model architecture and the three training objectives: the composed loss, the reasoning distillation loss, and the feature distillation loss. + +# 3.1. Model Architecture + +DistillCIR uses a projection-based CIR model $f$ based on the pretrained Pic2Word [35, 36], providing basic multimodal retrieval capability and efficiency. The image embedding $h_i$ is obtained by passing the reference image $i$ through the CLIP visual tower $f_V$ (Equation 1). Subsequently, a projection module $f_{\phi}$ , consisting of a three-layer MLP with ReLU activations, maps the image embedding into the textual token embedding space (Equation 2). The textual modification $t$ , represented by a sentence of instruction, is tokenized and combined with the projected image embedding. The composed input sequence is then fed to the CLIP text tower $f_T$ to obtain the final composed embedding $h_{(i,t)}$ (Equation 3). $D$ in all equations refers to the model embedding dimension and the subscript refers to the corresponding model. An overview of this composed embedding generation process is shown in Figure 3. + +$$ +h _ {i} = f _ {V} (i), h _ {i} \in \mathbb {R} ^ {D _ {V}} \tag {1} +$$ + +$$ +h _ {i} ^ {\prime} = f _ {\phi} \left(h _ {i}\right), h _ {i} ^ {\prime} \in \mathbb {R} ^ {D _ {T}} \tag {2} +$$ + +$$ +h _ {(i, t)} = f _ {T} \left(h _ {i} ^ {\prime}, t\right), h _ {(i, t)} \in \mathbb {R} ^ {D _ {T}} \tag {3} +$$ + +In inference, all target images can be directly encoded by the visual tower $f_{V}$ and cached in the database. Cosine similarity is then used to match the composed embedding with all target image embeddings. + +# 3.2. Eliciting LLM's Knowledge via Data Extension + +Obtaining large-scale triplet data (reference image, textual modification, target image) for CIR is challenging. In addition, the difference in tasks between LLMs and CIR models hinders the application of distillation techniques used for transferring larger language models to smaller ones in CIR. To address this challenge, we propose using publicly available image-text datasets and LLMs to derive triplet data with a format similar to CIR. This triplet data provides key + +benefits including knowledge enhancement from LLMs and training objectives akin to CIR. + +Specifically, we leverage the CC3M [37] dataset, which is commonly used in existing ZS-CIR models for consistency. Given an image-caption pair $(i, c)$ from the dataset, $i$ and $c$ represent the image and caption, respectively. We use a powerful frontier LLM in the class of Claude 3.5 Sonnet [3] or GPT-4o [1] to generate a triplet $(i, t, m)$ , where $t$ is the textual modification, and $m$ is the modified caption based on the image and modification. In detail, the caption $c$ is fed to the LLM, which leverages its internal knowledge to brainstorm a suitable modification and generate the modified caption. By generating triplet data in a CIR-like format, we distill the LLM's internal knowledge for training CIR models. We then establish a composed training objective as in Equation 6, where we obtain the composed embedding $h_{it}$ by forwarding $(i, t)$ through the CIR model $f$ and target embedding $h_m$ by passing the modified caption through the text tower $f_T$ . + +$$ +h _ {(i, t)} = f (i, t), h _ {i t} \in \mathbb {R} ^ {D _ {T}} \tag {4} +$$ + +$$ +h _ {m} = f _ {T} (m), h _ {m} \in \mathbb {R} ^ {D _ {T}} \tag {5} +$$ + +$$ +\mathcal {L} _ {c o m} = - \log \frac {\operatorname {s i m} \left(h _ {(i , t)} , h _ {m}\right)}{\operatorname {s i m} \left(h _ {(i , t)} , h _ {m}\right) + \sum_ {m ^ {\prime} \in \mathbb {N}} \operatorname {s i m} \left(h _ {(i , t)}, h _ {m ^ {\prime}}\right)} \tag {6} +$$ + +$\mathrm{sim}(\cdot ,\cdot)$ represents the cosine similarity with temperature and $\mathbb{N}$ represents the negative set, which is from other samples in the same batch. + +# 3.3. Distillation from Reasoning Enhancement + +Directly generating triplets from pairs using LLMs poses some potential issues. The generated triplet may be undermined by hallucinations [17, 53] within the LLM. More critically, the triplet can only teach the CIR model what targets to retrieve rather than how to retrieve the target, which we believe is the essence of instruction awareness. To learn the instruction-following capability, the reasoning process [43, 48] of how a triplet is generated by the LLM should be detailed and taught to the CIR model. + +To achieve this goal, the triplet generation should not be free-form. Instead, a step-by-step reasoning process [48] should be incorporated into the generation. In the generation prompt, we first ask the LLM to identify objects or environments in the caption that are possibly changeable. Then, we instruct the LLM to brainstorm a plausible modification for the identified term. For example, given the caption "A student is playing basketball," one possibly changeable object is basketball, and the corresponding modification is "changing the sport to football." The precise generation process ensures that the modification is feasible, thereby mitigating issues caused by hallucinations within the LLM. The LLM is subsequently prompted to reason + +![](images/1c4d09a7f0d435a7bad8a3ed82b9f278c61b4a9b507989df6ac727c557291860.jpg) +Figure 2. Overview of DistillCIR. We begin by augmenting the CC3M dataset via LLM-based reasoning, where the LLM produces plausible modifications and corresponding modified captions step by step, revealing how the modification is derived. We then train a projection-based CIR model on this LLM-enhanced data using three contrastive losses. The composed loss $\mathcal{L}_{com}$ forms the backbone of composed retrieval by mapping the reference image and its modification to the modified caption. The reasoning distillation loss $\mathcal{L}_{rea}$ contrasts the reference image and the modified caption against the LLM's reasoning process, employing learnable prompt tokens to differentiate it from the composed loss. Lastly, the feature distillation loss $\mathcal{L}_{fea}$ aligns the projection model's embeddings with those of a separately trained teacher MLLM. By combining these losses, our framework equips the projection model with strong instruction-following capabilities for composed image retrieval. + +![](images/db8c26688f6adbf16f1f29f223f905a2636377b7d7124b474398986498d42593.jpg) +Figure 3. Model Architecture. The reference image is first processed by the visual tower $f_V$ and then projected via $f_{\phi}$ . We then insert the projected image embedding into the text sequence by replacing the $ token embedding. This combined input is fed into $f_T$ , producing the composed embedding. At inference time, the target image is passed directly to $f_V$ , and we measure the similarity between the composed embedding and the target embedding. + +how the change will modify the original caption and ultimately derive a modified caption. Due to space limitations, the detailed prompt is shown in Appendix A. In a nutshell, the reasoning process consists of three parts: (1) identifying plausible terms to be changed; (2) brainstorming an applicable change for the identified terms; (3) analyzing how the modification affects the original caption and deriving the final modified caption. We denote the entire reasoning process as $r$ , which will be used for our reasoning distillation to teach the CIR model instruction awareness. + +To train the CIR model with the reasoning $r$ , we pro + +pose a reasoning distillation training objective, which aims to align the image and modified caption $(i,m)$ with its reasoning process. Specifically, the objective is a contrastive loss matching the consequential embedding $h_{(i,m)}$ and the reasoning embedding $h_r$ as shown in Equation 9. However, in our experiments, we find that this loss can interfere with $\mathcal{L}_{com}$ because they share a similar format but differ in semantics. To distinguish it from $\mathcal{L}_{com}$ , we introduce a set of learnable prompt tokens $p\in \mathbb{R}^{K\times D_T}$ , where $K$ is the number of tokens [23, 61]. These learnable tokens are pretended to $m$ to capture specific reasoning semantics. + +$$ +h _ {(i, m)} = f (i, m, p), h _ {(i, m)} \in \mathbb {R} ^ {D _ {T}} \tag {7} +$$ + +$$ +h _ {r} = f _ {T} (r), h _ {r} \in \mathbb {R} ^ {D _ {T}} \tag {8} +$$ + +$$ +\mathcal {L} _ {r e a} = - \log \frac {\operatorname {s i m} \left(h _ {(i , m)} , h _ {r}\right)}{\operatorname {s i m} \left(h _ {(i , m)} , h _ {r}\right) + \sum_ {r ^ {\prime} \in \mathbb {N}} \operatorname {s i m} \left(h _ {(i , m)}, h _ {r ^ {\prime}}\right)} \tag {9} +$$ + +# 3.4. Distillation from Feature Alignment + +So far, $\mathcal{L}_{com}$ and $\mathcal{L}_{rea}$ learn the instruction-following capability at the semantic level. However, to generate instruction-aware embeddings, it is also important to learn at the feature level. Since most state-of-the-art LLMs are large or proprietary, we train an open-source Multimodal LLM (MLLM [26, 29]) $f_{M}$ for this purpose. + +Instruction-tuned MLLMs provide an aligned vision-language space. To distill knowledge at the feature level, one remaining challenge is extracting instruction-aware embeddings from the MLLM. Similar to [20, 58], we propose + +
Method# Params.CIRRCIRCO
R@1R@5R@10Rs@1Rs@2Rs@3mAP@5mAP@10mAP@25mAP@50
ViT-BSEARLE165M24.0053.4266.8254.8976.6088.199.359.9411.1311.84
Slerp88M28.1955.8868.7761.1380.6390.689.3410.2611.6512.33
CIREVL12.3B*23.9452.5166.0060.1780.0590.1914.9415.4217.0017.82
DistillCIR153M29.8858.7570.2662.3382.4690.7516.3116.9018.5419.37
ViT-LPic2Word [36]429M23.9051.7065.3053.7674.4687.088.729.5110.6411.29
SEARLE [4]442M24.2452.4866.2953.7675.0188.1911.6812.7314.3315.12
KEDs [39]429M26.4054.8067.20-------
Context-I2W [40]496M25.6055.1068.50-------
LinCIR [12]442M25.0453.2566.6857.1177.3788.8912.5913.5815.0015.85
Slerp [18]428M30.9459.4070.9464.7082.9292.3118.4619.4121.4322.41
CIREVL [21]12.5B*24.5552.3164.9259.5479.8889.6918.5719.0120.8921.80
MCL [27]7.4B26.2256.84-61.45--17.6718.8620.8021.68
DistillCIR429M32.3463.5874.2966.7584.0291.8318.6720.3322.8223.86
+ +Table 1. Experiment Results on the CIRCO and CIRR Test Sets. Bold values indicate the highest scores. *CIREVL includes ChatGPT as a part of components for retrieval. Only parameters of components with known sizes are reported. Our model achieves significant performance gains over the baselines while retaining a compact model size. + +training a commonly used MLLM (LLaVA [29]) as an embedding model. Specifically, we use the triplet $(i,t,m)$ and $\mathcal{L}_{com}$ to train the model. To acquire embeddings, the [EOS] token is appended to the textual input, and its corresponding output embedding is used as the representation. In training, the image and textual modification are fed to the MLLM for the composed embedding, while the modified caption is fed to the MLLM for the modified target embedding. The MLLM is then optimized by $\mathcal{L}_{com}$ . Detailed training procedures are described in Appendix B. We then use the trained MLLM as the teacher model for feature distillation. Specifically, the CIR model generates the composed embedding for $(i,t)$ , which is projected and used to match the modified caption embedding generated by the MLLM in Equation 12. + +$$ +h _ {(i, t)} ^ {\prime} = \operatorname {L i n e a r} \left(h _ {(i, t)}\right), h _ {(i, t)} ^ {\prime} \in \mathbb {R} ^ {D _ {M}} \tag {10} +$$ + +$$ +h _ {m} ^ {M} = f _ {M} (m), h _ {m} ^ {M} \in \mathbb {R} ^ {D _ {M}} \tag {11} +$$ + +$$ +\mathcal {L} _ {f e a} = - \log \frac {\operatorname {s i m} \left(h _ {(i , m)} ^ {\prime} , h _ {m} ^ {M}\right)}{\operatorname {s i m} \left(h _ {(i , m)} ^ {\prime} , h _ {m} ^ {M}\right) + \sum_ {m ^ {\prime} \in \mathbb {N}} \operatorname {s i m} \left(h _ {(i , m)} ^ {\prime} , h _ {m ^ {\prime}} ^ {M}\right)} \tag {12} +$$ + +Finally, we combine three losses to train the model for the composed retrieval, reasoning distillation, and feature distillation. The combined loss is presented in Equation 13. + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o m}} + \alpha \cdot \mathcal {L} _ {\text {r e a}} + \beta \cdot \mathcal {L} _ {\text {f e a}} \tag {13} +$$ + +The final loss ensures that the CIR model not only learns to perform composed retrieval using the LLM's knowledge but also acquires instruction-following capabilities through the reasoning and feature distillation processes. + +# 4. Experiment + +# 4.1. Training Settings + +For data processing, we use the frontier LLM to process the CC3M dataset and save all distilled tuples $(i,t,m,r)$ as + +described in Section 3, which will be used for the model training. + +We experiment with CLIP [35] ViT-B and ViT-L as the backbones of DistillCIR. In alignment with common baselines, all data processing and model training are conducted on the CC3M [37] dataset. For the teacher MLLM, we use xtuner/llava-phi-3-mini-hf [10]. After training the MLLM, we extract and cached modified caption embeddings from it. The training is conducted on a cluster of six H100 GPUs, with a learning rate of $2 \times 10^{-5}$ and the batch size is set to 768. For parameter efficiency, we train the entire Pic2Word projection module while applying LoRA [16] on linear layers within the CLIP backbone. Appendix D includes further details and other hyperparameters. + +# 4.2. Benchmarks and Baselines + +We conduct our experiments on four prominent zero-shot CIR benchmarks: FashionIQ [49], CIRR [31], CIRCO [4], and GeneCIS [41]. As one of the earlier datasets, FashionIQ concentrates on fashion e-commerce imagery. By contrast, CIRR and CIRCO revolve around broader, real-world scenes. While CIRR pioneered CIR research on natural images, its design of assigning only a single target image per query may lead to overlooked positives. CIRCO addresses this concern by enabling multiple ground-truth targets per query, offering a more robust measure of retrieval performance. Meanwhile, GeneCIS specializes in conditional retrieval through four types of attribute and object modifications. For evaluation, we follow established practice by reporting Recall@k ( $R@k$ ) for FashionIQ, CIRR, and GeneCIS, with an additional subset metric ( $R_s@k$ ) for CIRR. Because CIRCO can include multiple correct targets for a single query, we employ mean Average Precision ( $mAP@k$ ) to balance both precision and recall across various ranking positions. + +We compare DistillCIR against two categories of state- + +
MethodFocus AttributeChange AttributeFocus ObjectChange ObjectAverage
R@1R@2R@3R@1R@2R@3R@1R@2R@3R@1R@2R@3R@1
ViT-BSEARLE18.9030.6041.2013.0023.8033.7012.2023.0033.3013.6023.8033.3014.40
CIRReLU17.9029.4040.4014.8025.8035.8014.6024.3033.3016.1027.8037.6015.90
DistillCIR20.0130.9942.5015.1526.8935.8513.1723.7732.9616.6728.1937.1016.25
ViT-LPic2Word15.6528.1638.6513.8724.6733.058.4218.0125.776.6815.0524.0311.15
SEARLE17.1029.6040.7016.3025.2034.2012.0022.2030.9012.0024.1033.9014.35
LinCIR16.9029.9541.4516.1927.9836.848.2717.4026.227.4015.7125.0012.19
CIRReLU19.5031.8042.0014.4026.0035.2012.3021.8030.5017.2028.9037.6015.85
DistillCIR20.6532.5544.1816.0627.5837.6613.2824.3533.6017.3428.8437.6516.83
+ +Table 2. Experiment Results on the GeneCIS Test Set. Bold indicates the highest score. + +of-the-art methods: (1) Projection-based models such as Pic2Word [36], SEARLE [4], Context-I2W [40], KEDs [39], Slerp [18], and LinCIR [12]; (2) LLM-based models such as CIREVL [21] and MCL [27]. To ensure fair comparisons, we mainly consider baselines trained on CC3M and exclude those [13, 19, 24, 56] relying on different training corpora. + +# 4.3. Result Analysis + +Table 1 presents our results on the hidden test sets for both CIRCO and CIRR, obtained from their respective submission servers. On the CIRCO benchmark, a dataset noted for its comprehensive annotations and allowance of multiple valid targets per query, our method achieves an $mAP@5$ of $18.67\%$ surpassing LinCIR and Slerp by $6.08\%$ and $0.21\%$ , respectively. It also provides a $0.1\%$ boost over CIREVL, which relies on two MLLMs, BLIP-2 [26] and ChatGPT. Turning to the CIRR dataset, which often features minimal overlap between reference and target images, our approach excels in handling text-driven retrieval. We obtain a $5.94\%$ improvement over KEDs and a $6.74\%$ improvement over Context-I2W in $R@1$ . Additionally, we exceed the LLM-based baseline MCL by $6.12\%$ in $R@1$ and $5.30\%$ in $R_s@1$ . These results underscore the strength and adaptability of our framework in managing the nuanced challenges posed by both CIRCO's multi-target scenarios and CIRR's instruction-dominated retrieval tasks. Notably, besides the impressive performance, DistillCIR also has a compact size comparable to most projection-based baselines. Theses findings validate the efficiency and superiority of Distill-CIR. + +Table 2 compares our results on GeneCIS, a dataset distinguished by condition statements that may convey opposing meanings, which necessitates a model capable of nuanced interpretation. Our DistillCIR framework outstrips Pic2Word and CIREVL by $5.68\%$ and $0.98\%$ in average $R@1$ , respectively. The superiority in this benchmark highlights the robustness of DistillCIR and its comprehension capability in following instruction. + +Table 3 compares our approach against existing zero-shot baselines on FashionIQ revealing substantial gains in + +![](images/3d7e19ea9936c5063a44ed7716b0861c7b58a6cecb8d02e2760596785d7464d4.jpg) +Figure 4. Qualitative Comparison between DistillCIR and Pic2Word on the CIRCO (Top) and FashionIQ (Bottom) Validation Set. Green boxes indicate ground truths. + +average $R@10$ . Specifically, our model surpasses Pic2Word and KEDs by $5.98\%$ and $3.98\%$ , respectively. Although our training primarily utilizes natural imagery, FashionIQ is highly specialized in fashion e-commerce, making our strong performance there indicative of the model's robust generalization capabilities. Effectively transferring knowledge from general-purpose data to more domain-specific tasks underscores the versatility of our method across both fashion-centric and natural-image CIR settings. + +Figure 4 presents qualitative comparisons from both CIRCO and FashionIQ, demonstrating the adaptability of our method across diverse tasks. In particular, we observe Pic2Word correctly captures alterations in color within FashionIQ but struggles with shifts in style. By contrast, our approach successfully handles both color and style modifications, showcasing its superior ability to interpret and apply textual guidance. + +
MethodShirtDressToptee
R@10R@50R@10R@50R@10R@50
ViT-BSEARLE24.4441.6118.5439.5125.7046.46
Slerp23.0641.9519.2442.1426.5747.78
CIREVL28.3647.8425.2946.3631.2153.85
DistillCIR29.3449.0525.1445.7633.8753.06
ViT-LPic2Word26.2043.6020.0040.2027.9047.40
SEARLE26.8945.5820.4843.1329.3249.97
KEDs28.9048.0021.7043.8029.9051.90
Context-I2W29.7048.6023.1045.3030.6052.90
LinCIR29.1046.8120.9242.4428.8150.18
Slerp29.6446.4723.3545.1231.9751.20
CIREVL29.4947.4024.7944.7631.3653.65
DistillCIR30.7350.1225.5146.6433.8853.74
+ +# 5. Ablation Study + +In this section, we conduct an in-depth analysis of each part of DistillCIR and examine its instruction awareness learned from LLMs. Specifically, we aim to answer the following questions: (1) How does reasoning distillation affect the model performance, and what design choices should we make for the training objective? (2) What is the impact of feature distillation, and how do different MLLMs influence the results? (3) Does DistillCIR actually learn the instruction-following capability, and how should we evaluate it? We address these questions in each subsection. Unless otherwise specified, we use the ViT-L backbone and the validation sets in CIRR and CIRCO to conduct ablations. + +# 5.1. Effect of Reasoning Distillation + +While $\mathcal{L}_{com}$ focuses on composed retrieval, $\mathcal{L}_{rea}$ and $\mathcal{L}_{fea}$ further enhance instruction awareness. Unlike conventional distillation approaches, $\mathcal{L}_{rea}$ aims to enable CIR models to acquire instruction-following capabilities from LLMs at the semantic level by performing cross-task distillation from text generation to retrieval. This introduces a key challenge in designing the distillation loss, since the teacher and student models optimize fundamentally different objectives. We suggest two potential solutions: (1) aligning with the teacher (LLM) through text generation with a Negative Log Likelihood loss, or (2) aligning with the student (CIR) through retrieval via a contrastive loss. As shown in Cases 1-3 of Table 4, despite prior work [25, 26] suggesting that joint training for text generation and contrastive learning can be mutually beneficial, we observe that this approach actually underperforms compared to contrastive-only training. We attribute this discrepancy to differences in inputs and outputs: in contrast to the multitask setting described in [25] where both tasks share the same inputs and outputs, our $\mathcal{L}_{com}$ (composed retrieval) and $\mathcal{L}_{rea}$ (instruction reasoning) target distinct inputs and outputs, complicating training when the losses themselves also diverge. + +Hence, we recommend using contrastive learning for $\mathcal{L}_{rea}$ . However, Cases 3-4 shows that contrastive-only opti + +Table 3. Experiment Results on the FashionIQ Validation Set. Bold indicates the highest score. + +
CasesCIRR R@1CIRCO mAP@10GeneCIS R@1
1. Lcom Only30.3819.7516.04
2. + Lrea (NLL)28.7518.9615.48
3. + Lrea (Contrast) w/o p31.4020.0016.06
4. + Lrea (Contrast) w/ p31.9620.3316.21
+ +Table 4. Ablations on the Reasoning Distillation. $\mathcal{L}_{rea}$ (NLL) and $\mathcal{L}_{rea}$ (Contrast) refer to using the negative log likelihood for text generation and using the contrastive loss for retrieval as the reasoning distillation loss together with the composed loss, respectively. $p$ are learnable prompt tokens. + +
CasesCIRR R@1CIRCO mAP@10GeneCIS R@1
1. Lcom Only30.3819.7516.04
2. Lcom + Lfea (TinyLLaVA)30.6219.7816.02
3. Lcom + Lfea (LLaVA-LLaMA3)31.4520.5516.63
4. Teacher Model (TinyLLaVA)31.9321.4216.95
5. Teacher Model (LLaVA-Phi3)33.7922.6816.99
6. Teacher Model (LLaVA-LLaMA3)34.1723.2017.15
7. Lcom + Lfea (LLaVA-Phi3)31.4920.4116.46
+ +Table 5. Ablations on the Feature Distillation. Case 1-3 show results of feature distillation with different MLLMs together with the composed loss. Case 4-6 show result of directly using teacher models for CIR, which is large in parameters and slow in inference compared with DistillCIR. + +mization does not adequately capture instruction reasoning. As discussed in Section 3, we attribute this to the semantic discrepancy between $\mathcal{L}_{\text {com }}$ and $\mathcal{L}_{\text {rea }}$ ; forcing them to share an identical loss complicates training. To address this, we propose a prompt-based approach. By introducing learnable prompt tokens into $\mathcal{L}_{\text {rea }}$ , the model can better acquire instruction-awareness while maintaining a clear distinction from the basic composed retrieval task. + +# 5.2. Effect of Feature Distillation + +While the reasoning distillation exploits LLM outputs at the semantic level, the feature distillation aligns representations from the CIR model with those of an open-source MLLM. As shown in Cases 1 and 7 of Table 5, incorporating the feature distillation loss $\mathcal{L}_{fea}$ improves performance beyond the base $\mathcal{L}_{com}$ . Because open-source MLLMs vary in scale, we investigate different model sizes of them. In addition to xtuner/llava-phi-3-mini-hf (4.2B), we examine bczhou/tiny-llava-v1-hf (1.4B) [59] and xtuner/llava-llama-3-8b (8B) [10]. The results in Cases 2, 3, and 7 show that the smallest model yields the lowest performance, while the medium and larger models perform comparably. Given evaluations on public MLLM benchmarks [30, 55], this suggests that distillation performance is influenced more by the LLM's overall capability than its size. Since llava-phi-3 matches llava-llama-3 in accuracy but is more compact, we adopt llava-phi-3 to balance performance and efficiency. + +# 5.3. Learning Instruction Awareness from LLMs + +In this section, we examine how DistillCIR acquires instruction awareness. To assess its understanding of instructions during composed retrieval, we propose a specific criterion: whether the model can selectively focus on particular regions of an image according to textual modifications. Pretrained CLIP models often emphasize dominant objects in an image, but the composed embedding should instead shift attention to the details specified by the textual cues. Drawing inspiration from [40, 54], we visualize patch-level attention maps to investigate whether DistillCIR indeed adjusts its focus based on the modifications. + +Specifically, let $i \in \mathbb{R}^{H \times W \times C}$ be an input image, where $H, W$ , and $C$ denote its height, width, and channel dimensions, respectively. The visual encoder $f_{V}$ produces patch embeddings $P \in \mathbb{R}^{L^2 \times D_V}$ , with $L^2$ representing the total number of patches. Next, we project these patch embeddings into the text embedding space using $f_{\phi}$ and $f_{T}$ . Finally, we compute an attention map between the projected patch embeddings and the composed embedding, as illustrated in Equation 19. The attention map $A$ is resized to a matrix and which is interpolated to the size of the image. The detailed computation process is shown in Appendix C and visualized examples are shown in Figure 5. + +![](images/4ab925e69aadd514c6206876bbe1210b623bf7e549b65381d7dfe0c23eabad42.jpg) +Figure 5. Attention Map Comparison between DistillCIR and Pic2Word on the CIRCO Validation Set with Color Map JET. Warmer colors indicate higher attention than cooler colors. + +From these visualizations, it is evident that the earlier CIR model, Pic2Word, fails to highlight the specific regions indicated by textual modifications. We conjecture that previous projection-based models, trained solely on image-caption pairs, primarily merge image and text features without fully addressing the discrepancy between standard image-text retrieval and composed image retrieval [7]. In contrast, DistillCIR selectively attends to the relevant areas, reflecting its instruction awareness acquired via LLM-based distillation. While there has been limited research on + +
CasesCIRR R@1CIRCO mAP@10GeneCIS R@1
1. LoRA on fV32.3820.9416.27
2. LoRA on fT31.4021.3616.05
3. Train fφ only29.6520.0715.85
4. MLP layers in fφ = 130.3820.1916.44
5. MLP layers in fφ = 232.8921.4716.59
LoRA on (fV, fT), Full on fφ (Ours)33.4521.5216.83
+ +Table 6. Ablations on Other Training Settings. Case 1-3 show results of training different components only and freeze others. Case 4-5 show results of training different MLP layers in $f_{\phi}$ under the same training and freezing strategy. Ours indicate train $f_{V}$ and $f_{T}$ with LoRA and full $f_{\phi}$ . + +instruction-following metrics for retrieval, the attention map serves as an initial quantitative metric. We leave the development of qualitative evaluation metrics for future work. + +# 5.4. Analysis of Other Training Settings + +In this section, we explore additional training setups. Table 6 reports the performance of DistillCIR under varying component configurations. As shown in Cases 1, 2, 3, and 6, applying LoRA to $f_{V}$ and $f_{T}$ while training full $f_{\phi}$ achieves the highest scores. This suggests that training more components generally enhances the model's learning capacity. Specifically, Applying LoRA on $f_{V}$ yields better performance than $f_{T}$ , indicating that the visual tower potentially acts a more important role. In addition, Cases 4-5 examines the effect of different MLP depths in $f_{\phi}$ . We observe continuous gains with increasing layers, reaching a maximum at three layers. Overall, we find that applying LoRA to both the visual and text towers, while restricting MLP layers in $f_{\phi}$ , achieves the best performance. + +# 6. Conclusion + +In this paper, we address instruction-aware Zero-Shot Composed Image Retrieval by distilling instruction-following capabilities from large language models into compact projection-based architectures. Our DistillCIR framework adopts a dual-stream strategy: it synthesizes pseudo triplet data with rationales to guide learning, and further leverages an open-source MLLM to embed instruction-aligned knowledge. This approach achieves strong performance while removing the need for LLMs during inference. DistillCIR highlights a practical path to instruction-aware retrieval, bridging projection-based and LLM-based models. + +# 7. Acknowledgments + +This work was partially supported by US National Science Foundation IIS-2412195, Microsoft Accelerate Foundation Models Research, 2024, CCF-2400785, the Cancer Prevention and Research Institute of Texas (CPRIT) award (RP230363) and the National Institutes of Health (NIH) R01 award (1R01AI190103-01). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2.3 +[2] Lorenzo Agnolucci, Alberto Baldrati, Marco Bertini, and Alberto Del Bimbo. isearle: Improving textual inversion for zero-shot composed image retrieval. arXiv preprint arXiv:2405.02951, 2024. 1, 2 +[3] Anthropic. The claude 3 model family: Opus, sonnet, haiku. 3 +[4] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1, 2, 5, 6 +[5] Oriol Barbany, Michael Huang, Xinliang Zhu, and Arnab Dhua. Leveraging large language models for multimodal search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1201-1210, 2024. 1 +[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2 +[7] Jaeseok Byun, Seokhyeon Jeong, Wonjae Kim, Sanghyuk Chun, and Taesup Moon. Reducing task discrepancy of text encoders for zero-shot composed image retrieval. arXiv preprint arXiv:2406.09188, 2024. 8 +[8] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer distillation with semantic calibration. In Proceedings of the AAAI conference on artificial intelligence, pages 7028-7036, 2021. 2 +[9] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4794-4802, 2019. 2 +[10] XTuner Contributors. Xtuner: A toolkit for efficiently fine-tuning llm. https://github.com/InternLM/xtuner, 2023.5.7 +[11] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 2 +[12] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangwoo Yun. Language-only efficient training of zero-shot composed image retrieval--. 1, 2, 5, 6 +[13] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 6 +[14] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[15] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, + +Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301, 2023. 3 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5 +[17] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1-55, 2025. 3 +[18] Young Kyun Jang, Dat Huynh, Ashish Shah, Wen-Kai Chen, and Ser-Nam Lim. Spherical linear interpolation and text-anchoring for zero-shot composed image retrieval. arXiv preprint arXiv:2405.00571, 2024. 2, 5, 6 +[19] Young Kyun Jang, Donghyun Kim, Zihang Meng, Dat Huynh, and Ser-Nam Lim. Visual delta generator with large multi-modal models for semi-supervised composed image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16805-16814, 2024. 2, 6 +[20] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-v: Universal embeddings with multimodal large language models. arXiv preprint arXiv:2407.12580, 2024. 4, 1 +[21] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. arXiv preprint arXiv:2310.09291, 2023. 2, 5, 6 +[22] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. 2023. 2 +[23] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 4 +[24] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and quality assessment for composed image retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2991-2999, 2024. 6 +[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 2, 7 +[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 4, 6, 7 +[27] Wei Li, Hehe Fan, Yongkang Wong, Yi Yang, and Mohan Kankanhalli. Improving context understanding in multimodal large language models via multimodal composition learning. In *Forty-first International Conference on Machine Learning*. 2, 5, 6 + +[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 4, 5 +[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 7 +[31] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2125–2134, 2021. 5 +[32] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36:46534-46594, 2023. 3 +[33] Peyman Passban, Yimeng Wu, Mehdi Rezagholizadeh, and Qun Liu. Alp-kd: Attention-based layer projection for knowledge distillation. In Proceedings of the AAAI Conference on artificial intelligence, pages 13657-13665, 2021. 2 +[34] Antonio Polino, Razvan Pascanu, and Dan Alistarh. Model compression via distillation and quantization. arXiv preprint arXiv:1802.05668, 2018. 2 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5 +[36] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19305-19314, 2023. 1, 2, 3, 5, 6 +[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, Melbourne, Australia, 2018. Association for Computational Linguistics. 3, 5 +[38] Krishna Srinivasan, Karthik Raman, Anupam Samanta, Lingrui Liao, Luca Bertelli, and Michael Bendersky. QUILL: Query intent with large language models using retrieval augmentation and multi-stage distillation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 492-501, Abu Dhabi, UAE, 2022. Association for Computational Linguistics. 3 +[39] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 26951-26962, 2024. 5, 6 +[40] Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Yue Hu, and Qi Wu. Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5180-5188, 2024. 5, 6, 8 +[41] Sagar Vaze, Nicolas Carion, and Ishan Misra. Genecis: A benchmark for general conditional image similarity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6862-6872, 2023. 5 +[42] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Güll Varol. Covr: Learning composed video retrieval from web video captions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5270-5279, 2024. 1 +[43] Boshi Wang, Xiang Yue, and Huan Sun. Can chatgpt defend its belief in truth? evaluating lIm reasoning via debate. arXiv preprint arXiv:2305.13160, 2023. 3 +[44] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 2 +[45] Xiaobo Wang, Tianyu Fu, Shengcai Liao, Shuo Wang, Zhen Lei, and Tao Mei. Exclusivity-consistency regularized knowledge distillation for face recognition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 325-342. Springer, 2020. 2 +[46] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022. 2, 3 +[47] Cong Wei, Yang Chen, Haonan Chen, Hexiang Hu, Ge Zhang, Jie Fu, Alan Ritter, and Wenhu Chen. Uniir: Training and benchmarking universal multimodal information retrievers, 2023. 2 +[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 3 +[49] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 5 +[50] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In European conference on computer vision, pages 664-680. Springer, 2020. 2 +[51] Xiaohan Xu, Ming Li, Chongyang Tao, Tao Shen, Reynold Cheng, Jinyang Li, Can Xu, Dacheng Tao, and Tianyi Zhou. A survey on knowledge distillation of large language models. arXiv preprint arXiv:2402.13116, 2024. 3 +[52] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. Clip-kd: An empirical study of clip model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15952-15962, 2024. 2 + +[53] Jia-Yu Yao, Kun-Peng Ning, Zhen-Hui Liu, Mu-Nan Ning, Yu-Yang Liu, and Li Yuan. Llm lies: Hallucinations are not bugs, but features as adversarial examples. arXiv preprint arXiv:2310.01469, 2023. 3 +[54] Runpeng Yu, Weihao Yu, and Xinchao Wang. Attention prompting on image for large vision-language models. arXiv preprint arXiv:2409.17143, 2024. 8 +[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 7 +[56] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions. arXiv preprint arXiv:2403.19651, 2024. 6 +[57] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 2 +[58] Wenliang Zhong, Weizhi An, Feng Jiang, Hehuan Ma, Yuzhi Guo, and Junzhou Huang. Compositional image retrieval via instruction-aware contrastive learning. arXiv preprint arXiv:2412.05756, 2024. 4, 1 +[59] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289, 2024. 7 +[60] Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. Instruction-following evaluation for large language models. arXiv preprint arXiv:2311.07911, 2023. 2 +[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4 +[62] Xinliang Zhu, Sheng-Wei Huang, Han Ding, Jinyu Yang, Kelvin Chen, Tao Zhou, Tal Neiman, Ouye Xie, Son Tran, Benjamin Yao, et al. Bringing multimodality to amazon visual search system. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6390–6399, 2024. 1 \ No newline at end of file diff --git a/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/images.zip b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..50a1a3efef1646884cd566441bfe5c7c5ff7f6d9 --- /dev/null +++ b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19b9dc4ae7137fcbe100c251cd39048d7078b30584344e73105ebc979d69dbdf +size 582192 diff --git a/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/layout.json b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38197c3a2d6fb3e4d60c55b6d0d75237f010bf92 --- /dev/null +++ b/zeroshotcomposedimageretrievalviadualstreaminstructionawaredistillation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:134e486b485dbab327ba57193c01d0f8a605ecda24c53e9acdc5c1b81f549b0b +size 407605 diff --git a/zeroshotcompositionalvideolearningwithcodingratereduction/28edbc35-c9cb-4584-bcf5-27480c4c0924_content_list.json b/zeroshotcompositionalvideolearningwithcodingratereduction/28edbc35-c9cb-4584-bcf5-27480c4c0924_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..835bacfbfd5235e371b6bdb6318a6be5c4b2eaa9 --- /dev/null +++ b/zeroshotcompositionalvideolearningwithcodingratereduction/28edbc35-c9cb-4584-bcf5-27480c4c0924_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f2b98d21b9aaeeea0d91db67c76851645272addeb8ec5a595112c0746352beb +size 93271 diff --git a/zeroshotcompositionalvideolearningwithcodingratereduction/28edbc35-c9cb-4584-bcf5-27480c4c0924_model.json b/zeroshotcompositionalvideolearningwithcodingratereduction/28edbc35-c9cb-4584-bcf5-27480c4c0924_model.json new file mode 100644 index 0000000000000000000000000000000000000000..034076ef036b5e0079854db5e9b3fbd07808e5d5 --- /dev/null +++ b/zeroshotcompositionalvideolearningwithcodingratereduction/28edbc35-c9cb-4584-bcf5-27480c4c0924_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:641740540884d315f6ad75eb81cfb49f22afedbac831ac617f1c502545ec4962 +size 113675 diff --git a/zeroshotcompositionalvideolearningwithcodingratereduction/28edbc35-c9cb-4584-bcf5-27480c4c0924_origin.pdf b/zeroshotcompositionalvideolearningwithcodingratereduction/28edbc35-c9cb-4584-bcf5-27480c4c0924_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dad727e9cd07d9754b8928cee98b003e39ecfce8 --- /dev/null +++ b/zeroshotcompositionalvideolearningwithcodingratereduction/28edbc35-c9cb-4584-bcf5-27480c4c0924_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80bc8ec7ecd3b4ae35a44994b06db65f99dc84036ddfc9eb9705babf96eefaf7 +size 8939750 diff --git a/zeroshotcompositionalvideolearningwithcodingratereduction/full.md b/zeroshotcompositionalvideolearningwithcodingratereduction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..976b1a8d0f24bb152f42d3f9d61533d463dda06c --- /dev/null +++ b/zeroshotcompositionalvideolearningwithcodingratereduction/full.md @@ -0,0 +1,400 @@ +# Zero-Shot Compositional Video Learning with Coding Rate Reduction + +Heeseok Jung $^{1}$ Jun-Hyeon Bak $^{1}$ Yujin Jeong $^{2}$ Gyugeun Lee $^{1}$ Jinwoo Ahn $^{3}$ Eun-Sol Kim $^{1,2,3*}$ + +$^{1}$ Department of Computer Science, Hanyang University + +$^{2}$ Department of Artificial Intelligence, Hanyang University + +$^{3}$ Department of Artificial Intelligence Application, Hanyang University + +{heeseokjung, gillma, itswarm, ggee, jinwooahn, eunsolkim}@hanyang.ac.kr + +# Abstract + +In this paper, we propose a novel zero-shot compositional video understanding method inspired by how young children efficiently learn new concepts and flexibly expand their existing knowledge framework. While recent large-scale visual language models (VLMs) have achieved remarkable advancements and demonstrated impressive performance improvements across various tasks, they require massive amounts of data and computational resources. However, despite their strong benchmark performance, they often fail to solve simple zero-shot composition tasks. Moreover, VLMs designed for video data demand even greater computational resources. We introduce a new video representation learning method inspired by human compositional learning to address these challenges. Specifically, we demonstrate that achieving zero-shot compositional learning requires effective representation learning that disentangles given data into meaningful semantic units. We propose a novel method that learns such disentangled representations based on an information-theoretic measure. By optimizing coding rate reduction, we successfully learn spatio-temporally disentangled features from videos, one of the most challenging data. Our approach significantly enhances compositional generalizability, demonstrating its effectiveness in zero-shot learning scenarios. Codes will be available at https://github.com/heeseokjung/zs-comp-mcr2. + +# 1. Introduction + +Recently, various types of large-scale visual language models (VLMs) have been proposed, demonstrating strong performance on various multimodal tasks such as visual question answering [29, 36, 49], image retrieval [47], and image segmentation [10, 22, 28, 33]. These VLMs have attracted significant attention not only for their strong performance + +![](images/f6dcbc5dfa5a80cf1634a609c5f30c5e9a24182539d685371ff18de65639311b.jpg) +(a) + +![](images/ccab8b684e65613fe63430c42df5b90f75c267439e1b50b286fae3d1d6dd037c.jpg) +(b) +Figure 1. The illustration of compositional generalization problem in (a) Sth-Com [30] and (b) CATER [11] datasets. The action labels are defined as a composition of primitive spatial and temporal categories, and the model is required to predict novel action labels that are unseen during training by recombining learned primitive concepts. + +across various downstream tasks but also for their ability to generate reasonable answers to unseen (novel) problems [1, 3, 20, 34, 47]. However, the training of VLMs relies heavily on tremendous amounts of data and computing resources, and it is not possible to clearly explain the principles of the zero-shot capability. + +In contrast, the way humans learn new problems or concepts can be clearly explained based on compositional learning. Humans can easily learn new concepts by recombining previously acquired knowledge, efficiently learning + +numerous concepts with minimal experience through such extrapolation [26]. For example, as can be seen in Figure 1, when we watch videos depicting the actions of (sphere, slide) and (cube, pick-place), we can easily imagine scenarios involving (sphere, pick-place) and (cube, slide), even if we never encountered such videos. This zero-shot compositional learning problem has garnered significant attention from computer vision researchers in both the image [42] and video [30, 58] domains, but there remain many limitations to achieving compositional generalization with machines. + +The most important point of zero-shot compositional learning is to disentangle data representations into primitive semantic concepts [16] and to make predictions for their novel compositions by combining the learned concepts [26]. Here, we present a novel information-theoretic video representation learning method for the zero-shot compositional learning problem, where coding rate reduction [6, 39, 56, 57] is used to measure the disentanglement quality of representations. Specifically, the proposed method disentangles spatiotemporal information in videos into multiple hierarchical subspaces (i.e., spatial, temporal) by optimizing the coding rate reduction. Our method is implemented within a resampler-like transformer decoder architecture where consecutive layers can be interpreted as an unrolled optimization of coding rate reduction. Through experiments on three challenging benchmarks for zero-shot compositional learning, we demonstrate that the suggested method exhibits strong zero-shot capability and learns interpretable representations. + +The main contributions of our work are summarized as follows: + +- We propose an information-theoretic approach for video representation learning by disentangling spatiotemporal information into multiple subspaces. +- The proposed method achieves a new state-of-the-art performance on three zero-shot compositional learning benchmarks across image and video. +- Rigorous qualitative analysis underscores the interpretability of the proposed method. + +# 2. Related Work + +# 2.1. Compositional Zero-Shot Learning + +Compositional zero-shot learning (CZSL) focuses on recognizing novel combinations of previously learned primitive concepts so that the model can generalize beyond its training data. Misra et al. [42] first proposed the CZSL problem in image understanding where the models are required to recognize images with novel object-attribute compositions by combining learned object and attribute concepts in training data. Subsequent studies have explored various approaches, including disentangling attribute and object features [14, 21, 48], modular networks [46], and + +leveraging graphs [23, 43, 52]. In addition, [24, 40] investigates the CZSL problem without relying on prior knowledge about object-attribute compositions, so called open-world setting. More recent works leverage the generalizable semantic understanding of CLIP [47] by using prompt learning [18, 32, 37, 44] or adaptation [32, 60]. + +The compositional learning for video understanding introduces additional challenges due to the increased complexity of temporal variations. Materzynska et al. [41] presented the Something-else (Sth-else) dataset, based on the Something-Something V2 dataset [12], to evaluate models on unseen spatial-temporal (i.e., object-verb) concept combinations in compositional action recognition. However, Sth-else only requires models to recognize unseen temporal categories (verb) without considering their combination between objects, so the evaluation is limited just to the spatial bias test. To address this limitation, Li et al. [30] recently introduced a novel Something-Composition (SthCom) dataset, evaluating the ability to generalize unseen combinations between spatial (object) and temporal (verb) categories. Moreover, Yun et al. [58] introduced a novel compositional generalization benchmark derived from the CATER [11] dataset with a strong baseline method that leverages spatiotemporal graphs. + +# 2.2. Representation Learning with Rate Reduction + +The representation learning with rate reduction framework [6] primarily aims to find the compressed and structured representations from high-dimensional data having complex distributions. Previous studies have shown that the learning objective of this framework can be described by the information-theoretic measure regarding data compression, known as coding rate reduction. In the rate-distortion theory [8], (lossy) coding rate measures the compactness of random distribution by counting the minimal binary bits required to encode the data within a given decoding error $\epsilon$ , which can be considered as the estimation of the volume of the space spanned by the data. Based on this idea, the rate reduction framework assumes high-dimensional data can be represented as a mixture of low-dimensional orthogonal subspaces, and considers the optimization problem minimizing the coding rate computed within the same subspace while maximizing the coding rate of the entire space. + +Several studies have been proposed within this framework. Ma et al. [39] presented a tight approximation formula for the rate reduction of finite data samples from an unknown distribution, which promoted subsequent research. Yu, Chan et al. [56] proposed a unified learning framework by using rate reduction as a training objective. Yu et al. [57] introduced a transformer-like mathematically interpretable architecture by unrolling iterative optimization over rate reduction as a model forward pass. Lu et al. [38] leveraged the rate reduction framework as a feature denoiser + +to discover informative subspaces, successfully applying it to a wide range of motion learning tasks. Recently, Wu et al. [53] derived a variational form of rate reduction objective to propose a new attention operator having linear complexity. Furthermore, Wu et al. [54] showed that using coding rate maximization as a regularizer can mitigate feature collapse in self-supervised learning [4, 45]. In this paper, we show that the rate reduction framework can also be extended to spatiotemporally disentangled representation learning, allowing us to consider the highly challenging problem of compositional generalization in video action recognition. + +# 3. Method + +The most important point of compositional learning, where new concepts are learned by recombining previously acquired knowledge, is to disentangle data representations into primitive semantic units. In this section, we introduce a novel information theory-based representation learning method for efficiently disentangling spatiotemporal video representations. The proposed method is based on a transformer decoder architecture compatible with any kind of video encoder. The model learns disentangled representations by denoising the spatiotemporally entangled feature map from the video encoder into multiple subspaces of learned latent queries. The illustration of the overall framework can be found in Figure 2. We first introduce preliminary information about coding rate reduction and then elaborate on the architectural details of the proposed method. + +# 3.1. Coding Rate Reduction + +Measuring the quality of feature vectors is the most fundamental topic of machine learning. The most important criterion for the feature quality is the ability to retain (reconstruct) as much of the original data as possible using the fewest possible bits. From information theory [8], many studies have suggested the measures, and recently, a coding rate-based measure [6, 39, 56, 57] has been introduced that can robustly measure the quality of features in high-dimensional continuous variables. + +From rate-distortion theory [8], the (lossy) coding rate can be defined to measure the quality of the representation in terms of data compression. For a given set of $N$ data vectors $\mathbf{Z} = \{\mathbf{z}_i\}_{i=1}^N \in \mathbb{R}^{D \times N}$ and their reconstructions $\hat{\mathbf{Z}} = \{\hat{\mathbf{z}}_i\}_{i=1}^N$ , coding rate measures the number of binary bits required to reconstruct within a specified decoding error $\epsilon$ , where $\mathbb{E}[\| \mathbf{z}_i - \hat{\mathbf{z}}_i \|^2] \leq \epsilon^2$ . Ma et al. [39] presented a tight approximation formula of the coding rate for data of arbitrary distribution as + +$$ +R (\mathbf {Z}) = \frac {1}{2} \log \det \left(\mathbf {I} + \frac {D}{N \epsilon^ {2}} \mathbf {Z} \mathbf {Z} ^ {*}\right). \tag {1} +$$ + +It can be assumed that high-dimensional data $\mathbf{Z}$ can be represented as a mixture of $K$ number of lower-dimensional + +subspaces, so we can compute the coding rate more accurately by summing up the respective coding rate of each subspace as + +$$ +R ^ {c} (\mathbf {Z} | \boldsymbol {\Pi}) = \sum_ {j = 1} ^ {K} \frac {\operatorname {t r} \left(\boldsymbol {\Pi} _ {j}\right)}{2 N} \log \det (\mathbf {I} + \frac {D}{\operatorname {t r} \left(\boldsymbol {\Pi} _ {j}\right) \epsilon^ {2}} \mathbf {Z} \boldsymbol {\Pi} _ {j} \mathbf {Z} ^ {*}), \tag {2} +$$ + +where $\Pi = \{\Pi_{j} \in \mathbb{R}^{N \times N}\}_{j=1}^{K}$ is a set of diagonal membership matrices whose $i^{\text{th}}$ diagonal entry denotes the probability of $\mathbf{z}_i$ belongs to the $j^{\text{th}}$ subspace. The primary goal of representation learning with rate reduction is to find a mapping $f$ maximizing the rate reduction as + +$$ +\max _ {f \in \mathcal {F}} \mathbb {E} _ {\mathbf {Z}} [ \Delta R (\mathbf {Z}) ] = \max _ {f \in \mathcal {F}} \mathbb {E} _ {\mathbf {Z}} \left[ R (\mathbf {Z}) - R ^ {c} (\mathbf {Z} | \boldsymbol {\Pi}) \right]. \tag {3} +$$ + +This objective drives representation to be highly compressed in the same subspace while being incoherent between different subspaces, resulting in maximally discriminative representations. There are several methods for optimizing the rate reduction objective, and we use a simple gradient ascent method [6] with closed-form derivatives of Eq. 1 and 2 as + +$$ +\nabla R (\mathbf {Z}) = \alpha \left(\mathbf {I} + \alpha \mathbf {Z} \mathbf {Z} ^ {*}\right) ^ {- 1} \mathbf {Z}, \tag {4} +$$ + +$$ +\nabla R ^ {c} (\mathbf {Z} | \boldsymbol {\Pi} _ {j}) = \gamma_ {j} \alpha_ {j} \left(\mathbf {I} + \alpha_ {j} \mathbf {Z} \boldsymbol {\Pi} _ {j} \mathbf {Z} ^ {*}\right) ^ {- 1} \mathbf {Z} \boldsymbol {\Pi} _ {j}, \tag {5} +$$ + +where $\alpha = \frac{D}{N\epsilon^2}$ , $\alpha_{j} = \frac{D}{\mathrm{tr}(\Pi_{j})\epsilon^{2}}$ , and $\gamma_{j} = \frac{\mathrm{tr}(\Pi_{j})}{N}$ for $j = 1,\dots ,K$ + +# 3.2. Resampler with Rate Reduction + +We aim to learn the $M$ -representative feature vectors that can compactly represent the spatiotemporal semantics of an input video. To find these $M$ feature vectors, we use a resampler-based approach and apply the coding rate reduction optimization method introduced in Section 3.1 to ensure effective compression of these vectors. + +For a video $\mathcal{V} \in \mathbb{R}^{T \times C \times H \times W}$ given by $T$ frames with spatial resolution of $H \times W$ and $C$ channels, we first obtain a feature map $\mathcal{X} \in \mathbb{R}^{T \times D \times h \times w}$ with pre-trained video encoder $f: \mathcal{V} \to \mathcal{X}$ . The feature map of the encoder output is then flattened to $\mathcal{X} \in \mathbb{R}^{TN \times D}$ where $N$ denotes the number of non-overlapping patches in each frame ( $N = hw$ ). To preserve the spatial location and time information of each patch, learnable spatiotemporal position embeddings $\mathbf{E}_{\mathrm{spatial}} \in \mathbb{R}^{N \times D}$ , $\mathbf{E}_{\mathrm{temporal}} \in \mathbb{R}^{T \times D}$ are added by broadcasting as + +$$ +\hat {\mathcal {X}} = \mathcal {X} + \mathbf {E} _ {\text {s p a t i a l}} + \mathbf {E} _ {\text {t e m p o r a l}}. \tag {6} +$$ + +We assume that $\hat{\mathcal{X}}$ is a spatiotemporally entangled video feature and present an algorithm below that can disentangle this feature along spatial and temporal semantic axes. + +From the $\hat{\mathcal{X}}$ , $M$ ( $M \ll TN$ ) number of disentangled video representations $\mathcal{Z}^L \in \mathbb{R}^{M \times D}$ are learned by itera + +![](images/d69a6b29227b26b9b3975fcdaac60fec88e995d7084b3f348a5321c6d4fc43a7.jpg) + +![](images/ee358c724510076d96e1efff23941b6cf15dfe41d8a80f37efe43a3ffd7d1b67.jpg) + +![](images/7e96a29c6ccc1e0c291150d64625aa2803beb4827df2fb687334b77c243c8ac9.jpg) +Figure 2. (a) Overall framework of the proposed method. The input video is first passed through the pre-trained video encoder. The encoder output is then augmented with spatiotemporal position embeddings and flattened to obtain the input feature map $\hat{\mathcal{X}}$ . This feature map is entangled in both space and time, so the resampler-based architecture maps this input feature map to the compressed latent embeddings $\mathcal{Z}^L$ by iteratively updating with cross-attention to the direction of maximizing coding rate reduction. (b) The architectural details of the implementation of rate reduction optimization in the resampler module. (c) The conceptual illustration of hierarchical multiple subspaces, where the features are disentangled in our approach. + +![](images/6f136b29fbae6dc49feb842cb557ab924bb46b63aad3da7b1a3b835c0c2c854d.jpg) + +![](images/965578cdc8ef6e0d2254d8d071f9167b05454297a6c8fc57924ccc71f9127a2f.jpg) + +tively updating initial latent embeddings $\mathcal{Z}^0\in \mathbb{R}^{M\times D}$ according to the coding rate reduction. Let $\mathcal{Z}^{\ell}$ denote the latent embeddings at iteration step $\ell$ . In each iterative step, we define an update function $g^{\ell}:\mathcal{Z}^{\ell}\to \mathcal{Z}^{\ell +1}$ . This function $g$ first maps the information of the encoder output into latent embeddings using a cross-attention block, followed by a feed-forward network, each with a skip-connection as + +$$ +\bar {\mathcal {Z}} ^ {\ell} = \operatorname {M C A} \left(\mathcal {Z} ^ {\ell}, \hat {\mathcal {X}}, \hat {\mathcal {X}}\right) + \mathcal {Z} ^ {\ell}, \tag {7} +$$ + +$$ +\hat {\mathcal {Z}} ^ {\ell} = \operatorname {F F N} \left(\bar {\mathcal {Z}} ^ {\ell}\right) + \bar {\mathcal {Z}} ^ {\ell}, \tag {8} +$$ + +where MCA denotes multi-head cross-attention block with the query is given by $\mathcal{Z}^{\ell}$ and the key and value are given by $\hat{\mathcal{X}}$ , and FFN denotes feed-forward network. + +Now, we introduce an information-theoretic approach to disentangle the $M$ resampled feature vectors. First, we assume that resampled representation $\hat{\mathcal{Z}}^{\ell}$ from $\hat{\mathcal{X}}$ can be represented as a mixture of lower-dimensional subspaces $\{\hat{z}_k^\ell\}_{k=1}^K$ where $\hat{z}_k^\ell \in \mathbb{R}^{M \times d}$ and $d \cdot K = D$ . Let $\mathbf{P} = \{\mathbf{P}_k \in \mathbb{R}^{D \times d}\}_{k=1}^K$ to be a set of sparse partitioning matrices where rows of $\mathbf{P}_k$ are divided into $K$ groups and only the elements in $k^{\mathrm{th}}$ group are 1 and otherwise 0. Then we define the following two operators to compute the gradient of the coding rate reduction as described in Section 3.1 + +$$ +\mathrm {G E} (\mathbf {Z}) = \alpha (\mathbf {I} + \alpha \mathbf {Z} \mathbf {Z} ^ {*}) ^ {- 1} \mathbf {Z} \tag {9} +$$ + +$$ +\mathbb {G C} (\mathbf {Z} \mid \mathbf {P}) = \sum_ {k = 1} ^ {K} \gamma_ {k} \alpha_ {k} \left(\mathbf {I} + \alpha_ {k} \mathbf {Z P} _ {k} \mathbf {P} _ {k} ^ {*} \mathbf {Z} ^ {*}\right) ^ {- 1} \mathbf {Z P} _ {k} \tag {10} +$$ + +Using these operators, we perform a gradient ascent update step at the end of each resampler layer for maximizing the coding rate reduction as + +$$ +\mathcal {Z} ^ {\ell + 1} = \hat {\mathcal {Z}} ^ {\ell} + \eta \cdot \left(\mathrm {G E} \left(\hat {\mathcal {Z}} ^ {\ell}\right) - \mathrm {G C} \left(\hat {\mathcal {Z}} ^ {\ell} \mid \mathbf {P}\right)\right), \tag {11} +$$ + +where $\eta$ is a fixed learning rate. By repeatedly updating $\mathcal{Z}^0$ for $L$ steps with $g: \mathcal{Z}^\ell \to \mathcal{Z}^{\ell+1}$ , we obtain resampled $\mathcal{Z}^L$ that are disentangled in both space and time. Finally, we feed average-pooled $\mathcal{Z}^L$ into two fully-connected layers $\psi_s$ , $\psi_t$ to get the final representations $\mathbf{z}_s \in \mathbb{R}^D$ and $\mathbf{z}_t \in \mathbb{R}^D$ for spatial and temporal information respectively. These two heads are expected to select informative subspaces for $\mathbf{z}_s$ , $\mathbf{z}_t$ from the resampled representation $\mathcal{Z}^L$ . + +# 3.3. Training for CZSL problem + +Following previous works [30, 41], action labels are composed of primitive spatial and temporal categories as $\mathbf{y}_{\mathrm{action}} = (\mathbf{y}_s,\mathbf{y}_t)$ . For compositional action recognition, we first encode primitive spatial and temporal categories with learnable embeddings as $\mathbf{y}_s = \{e_i^s\in \mathbb{R}^D\}_{i = 1}^{N_s}$ and $\mathbf{y}_t = \{e_i^t\in \mathbb{R}^D\}_{i = 1}^{N_t}$ . The probability score estimation for spatial and temporal categories is the confidence score computed by cosine similarity between $\mathbf{z}_s$ , $\mathbf{z}_t$ , and label embeddings $\mathbf{y}_s$ , $\mathbf{y}_t$ with proper normalization as + +$$ +p \left(\mathbf {y} _ {\mathbf {s}} ^ {(i)} \mid \mathcal {V}\right) = 0. 5 \cdot \cos \left(\mathbf {z} _ {s}, e _ {i} ^ {s}\right) + 0. 5, \tag {12} +$$ + +$$ +p \left(\mathbf {y} _ {\mathbf {t}} ^ {(i)} \mid \mathcal {V}\right) = 0. 5 \cdot \cos \left(\mathbf {z} _ {t}, e _ {i} ^ {t}\right) + 0. 5. \tag {13} +$$ + +We assume conditional independence between spatial and temporal characteristics of the given video, so the prediction for action labels is computed as a product of these two probability scores: + +$$ +\begin{array}{l} p \left(\mathbf {y} _ {\text {a c t i o n}} ^ {(i, j)} \mid \mathcal {V}\right) = p \left(\mathbf {y} _ {s} ^ {(i)}, \mathbf {y} _ {t} ^ {(j)} \mid \mathcal {V}\right) \\ = p \left(\mathbf {y} _ {\mathbf {s}} ^ {(i)} \mid \mathcal {V}\right) \cdot p \left(\mathbf {y} _ {\mathbf {t}} ^ {(j)} \mid \mathcal {V}\right). \tag {14} \\ \end{array} +$$ + +As a result, the model is trained with the following objective: $\mathcal{L}_{\mathrm{action}} + \gamma_1\mathcal{L}_{\mathrm{spatial}} + \gamma_2\mathcal{L}_{\mathrm{temporal}}$ , where each term denotes cross-entropy loss for the composite action, primitive spatial and temporal categories, with scaling hyperparameters $\gamma_{1},\gamma_{2}$ . + +# 4. Experiments + +# 4.1. Dataset and Task Definition + +The main goal of this paper is to learn disentangled informative representations of videos so that it can be applied to various downstream tasks in the video domain. However, quantitative evaluation of compositional generalizability is currently limited due to the lack of well-defined benchmarks. Therefore, we focus on the action recognition problem in two datasets where zero-shot composite action labels can be clearly defined. Furthermore, we also present experimental results on image data to demonstrate that our approach is not limited to the video domain. We first describe benchmark datasets and their task definition of zero-shot compositional learning in the following sections to make further discussions clear. + +# 4.1.1. Sth-Com dataset + +The Sth-Com dataset was recently introduced by Li et al. [30] to evaluate the zero-shot compositional action recognition problem. The action labels are defined as a composition of verb and object labels included in the Something-Something V2 (SSv2) dataset [12]. Specifically, Sth-Com requires the model to predict unseen action classes based on the known concept of verbs and objects from seen action classes as can be seen in Figure 1 (a). As a result, the dataset consists of 79K videos from the SSv2 dataset, currently the most large-scale benchmark for evaluating zero-shot compositional action recognition. The performance is measured with the following 6 different aspects based on the top-1 classification accuracy: (i) seen/unseen composite action class (ii) verb/object atomic concept class (iii) harmonic mean (HM) of seen/unseen accuracies (iv) area under the curve (AUC) of seen-unseen accuracy curve. + +# 4.1.2. CATER dataset + +The CATER dataset is one of the most challenging benchmarks for evaluating the recognition of compositional actions and temporal reasoning, as introduced by Girdhar et al. [11]. It contains 5.5k synthetic videos, each featuring + +multiple objects performing actions simultaneously while interacting. Recently, Yun et al. [58] proposed a new benchmark based on the CATER dataset for zero-shot compositional action recognition. The action classes are defined as a composition of five primitive objects (e.g., sphere) and four primitive temporal movements (e.g., pick-place), resulting in 14 distinct action categories. As illustrated in Figure 1 (b), videos with some classes are not used in the training to evaluate the compositional generalizability. More detailed problem setups are described in the supplementary material. It can be considered that this task is quite similar to the task of the Sth-Com dataset; however, it is worth noting that the setting in this benchmark is more challenging due to the multiple objects and their interactions (i.e. multi-label classification). The performance is measured based on the mAP for seen/unseen classes separately. We also report a normalized performance gap between seen and unseen splits, computed as $(\mathrm{mAP}_{\mathrm{seen}} - \mathrm{mAP}_{\mathrm{unseen}}) / \mathrm{mAP}_{\mathrm{seen}}$ , to examine the memorization-generalization trade-off. + +# 4.1.3. C-GQA dataset + +Similar to the problem setup of zero-shot compositional action recognition, C-GQA focuses on the compositional generalization problem in the image domain. The C-GQA is curated from the Stanford GQA [19] dataset, and it contains a vocabulary of 457 and 870 primitive attributes and object categories, where each selected bounding box in the images is annotated with both the object and its attribute. Some compositions of object-attribute pairs are defined as novel and never presented in the training data, so it is required to recognize unseen compositions of object-attribute pairs by combining learned primitive concepts of objects and attributes. The performance is similarly measured as in the Sth-Com benchmark, based on the top-1 classification accuracy of seen/unseen classes with harmonic mean (HM) and area under the curve (AUC). + +# 4.2. Implementation Details + +For fair comparisons with previous approaches, a pretrained VideoSwin-T [35] is used as the video encoder for both the Sth-Com and CATER datasets. Following the recently suggested method [30], we also report results on Sth-Com using CLIP [47] visual encoder adapted with AIM [55] as video encoder. The resampler module consists of 6 layers, each containing a cross-attention block followed by a feed-forward network, and the single rate reduction optimization step. We used 128 latent queries and 24 heads, which were empirically found to be the best-performing configuration. For gradient ascent update with rate reduction objective, we used $\eta = 0.1$ for the learning rate. The model was trained for 50 and 100 epochs on the Sth-Com and CATER datasets. We used the maximum learning rate of $2.5\mathrm{e - }4$ with linear warmup and cosine decay scheduling. + +Table 1. Zero-shot compositional action recognition results on the Sth-Com [30] dataset with varying the backbone video encoder and label encoding scheme. + +
Methodverbobj.seenunseenHMAUC
VideoSwin-T + fastText
TMN [46]33.445.631.030.223.07.9
Compcos [40]36.750.534.237.827.210.9
CGE [43]31.848.131.230.323.17.9
OADis [48]41.648.938.138.330.212.6
ADE [14]35.349.133.836.226.810.3
CAN [51]32.649.533.033.325.09.2
C2C (Vanilla) [30]47.950.942.143.634.916.4
C2C (Enhanced) [30]48.852.843.444.136.717.4
Ours50.050.644.645.437.318.3
CLIP
Coop [61]48.455.143.946.436.618.1
CSP [44]49.054.843.647.436.018.0
SPM [37]48.954.544.046.335.517.7
+DFM [37]47.455.443.147.335.817.9
C2C [30]53.856.247.250.241.021.7
Ours56.354.749.652.442.623.5
+ +# 4.3.Quantitative Results + +# 4.3.1. Comparisons on the Sth-Com + +For experiments on the Sth-Com dataset, we present experimental results in two different settings by varying the video encoder and label encoding schemes. For the first evaluation setting, we used VideoSwin-T [35] as a video encoder and fastText [2] word embeddings for encoding labels. For the second, CLIP [47] visual and text encoders are used for video and label encoders, respectively, and the visual encoder is trained with AIM [55] adapter. + +Zero-shot compositional action recognition results on the Sth-Com dataset are summarized in Table 1. The proposed method outperforms state-of-the-art baseline methods with a meaningful margin. It is worth noting that even if C2C (Vanilla) explicitly leverages label encodings for combining spatial and temporal representation by concatenating label encodings to model outputs, our method shows better performance without relying on label encodings. Furthermore, we also emphasize that our method outperforms C2C (Enhanced), which is trained with a more sophisticated training strategy compared to ours, including auxiliary losses [13] and data augmentation [59]. For experiments with CLIP backbone in the same table, our method also outperforms text prompting-based baseline compositional learning methods [37, 44, 61] without any learnable prompt in label encoding. In comparison with C2C, the proposed method shows a more significant improvement by using the VLM backbone, resulting in a larger performance gap compared to the first evaluation setting. + +# 4.3.2. Comparisons on the CATER + +To further challenge our method, we evaluate it for multi-label zero-shot compositional action recognition on the + +Table 2. Zero-shot compositional action recognition results on the CATER [11] dataset. + +
Methodseen ↑unseen ↑seen-unseen gap ↓
Using GT graph input
ST-GT99.171.50.27
Using RGB video input
R3D93.025.10.73
C2C82.656.40.31
Ours88.574.00.16
+ +Table 3. Comparison with VLMs on the CATER [11] dataset. + +
Methodsseenunseenseen-unseen gap ↓
InternVL2-8B (Chen et al.)57.237.00.35
Qwen2-VL-8B (Bai et al.)66.136.70.44
Ours88.574.00.16
+ +CATER dataset [11]. Table 2 summarizes the comparative results of experiments on the CATER dataset. Although all presented methods achieve comparable performance on the seen split, their score varies significantly on the unseen split. The compositional learning approach C2C [30] demonstrates substantially improved generalization on the unseen split compared to the monolithic R3D [5, 15] architecture. While ST-GT [58] demonstrates decent scores on both seen and unseen splits, note that this method leverages graph inputs constructed with ground-truth underlying parameters in the data-generated environment. On the other hand, the proposed method achieves the highest score on the unseen split with the closest gap between scores on seen/unseen splits, only consuming the RGB input. + +To provide a more robust evaluation, we further compare the proposed method with the most recent VLMs, InternVL2-8B [7] and Qwen2-VL-8B [50]. We provide all possible primitive objects and temporal action categories along with videos in the input sequence and train them with LoRA adapters [17] to predict action classes by combining them. More implementation details for training VLMs for the zero-shot compositional action recognition task are described in the supplementary material. As shown in Table 3, the proposed method outperforms strong VLM baselines by a large margin. From these results, we can infer that VLMs struggle to generalize beyond training data distribution by combining previously learned concepts, so they require a massive amount of training data and parameters. + +# 4.3.3. Comparisons on the C-GQA + +To implement the proposed method for the experiments on the C-GQA, we simply replace the spatial and temporal primitive classes with object and state classes defined in the C-GQA, without modifying the model architecture. Following previous approaches, we use the ViT-B [9] as a visual backbone encoder and fastText [2] to encode labels, and most of the hyperparameter settings regarding + +![](images/fa235f613461e6aeeccadf45df20e31933468650af236ec8374663c9ec2d4ad1.jpg) +(a) + +![](images/441312d2b0ee8856b06ffa9f819701a91c6447c9ca2ed3c3f10da3245de764fd.jpg) + +![](images/a94b540ba437df84883f32db9cdcd7bff3e6c0595db587616549bbf94e8df879.jpg) + +![](images/b80374cacff4b2469ec1d7cebfb75fa4658ef87b3565d755cc39b9a285ec328f.jpg) + +![](images/b2a63ac6536d1ed8a95b8fc192172ab8ab63cbef839852b56fc287a1fdbe45f5.jpg) + +![](images/217b1990c141beb0ac3b404a17f23bfd43719e2ccf3120244a92307096fcf60c.jpg) + +![](images/a2bbb7db56dd456e90eec926291c4359c27748f6548fb6798bfbbdb9acbf2f18.jpg) + +![](images/0d742dbd8a90671c6b54ce963263dc17e14ff7896be289d07a464256a1018c0e.jpg) + +![](images/72ca4d57bd36a3e5df9a3f4e70f5b411dc086f8776e5a54d795c17933f082125.jpg) +(b) + +![](images/1713137a6fd8a404616284ee17d8068a282262ba7406b7a1d48c87c56ae9676f.jpg) + +![](images/72bee71ec9c98160023be3fd2b6f1ee7682588d0432684e3b5ab0bdcfb520876.jpg) + +![](images/1c19dc455bd94aa470840693e8026d5e58e89e86872760f791b8d3bd0c870957.jpg) + +![](images/b10988af1f310a5218c874ca0a472d10118a1186f931565b107b14f0fd54bf1d.jpg) + +![](images/4b198be26480864bafd95a23d6580caa6a12b90990790fed4c15ffe0e75cf922.jpg) + +![](images/03fc40f978f5c250d96df4534a2b4bc61c0c2c2270eee75c09b21af48f73e341.jpg) + +![](images/bc64afc2cbdf4c3e1f7924992f19e2de0cfd6e27b77c4e1db99a7a29495d3594.jpg) + +![](images/95195c85f4808c22bce63e49c3e952f8eb2e7a57935ccafcb38fb86101e84764.jpg) +(c) + +![](images/6db44ce001b9edd6e128d4db65a7f7b897c59913052aceac2b3c33b0c32f5cb4.jpg) +Figure 3. The visualization of cross-attention maps in the last layer of the resampler module. Example videos are selected in the Sth-Com [30] dataset, having three different action labels: (a) Tilting a lid with a coin on it until it falls off (b) Putting usb next to mug (c) Keys falling like a rock. For each video, we select a representative head and visualize its attention map on the input frames across time. It can be observed that each head captures the informative object and its temporal variations, even trained without dense labels of object positions such as bounding boxes. + +![](images/e819e5d9200fdf799524416ee980542a6bf4e71b16a80c5362607c9f98c5d497.jpg) + +![](images/7a38e1b4ed91d3c0669eb4d5658550e129fb5e5df633f5d9e6d0941b38487a1f.jpg) + +![](images/2e550baffa047c8591db0ccf8c2d42e87bd4d4b78949e69824864c04c85f77c9.jpg) + +![](images/a15b131915753053e8dd17d2cff2f092bf7ba8c5856ccd7301b8f4d769973fba.jpg) + +![](images/674e481543ef88e9dfd935376eba0bd802598e11ba12287dc9cfc1412a20dcbd.jpg) + +![](images/5d276d793d403c3f85fade0fe4763a6b3ccf744f38c11564a132045bd0227471.jpg) + +Table 4. Zero-shot compositional image recognition results on the C-GQA [43] dataset. + +
MethodseenunseenHMAUC
CGE [43]38.017.118.55.4
OADis [48]38.319.820.17.0
CoT [25]39.222.722.17.4
C2C [30]39.123.023.07.7
Ours40.323.823.68.2
+ +model architecture are borrowed from the experiments on the video data. As shown in Table 4, we compare the proposed method with state-of-the-art zero-shot compositional learning approaches, and our method outperforms all strong baselines. These results highlight the strength of our work as it can be naturally extended to a different type of data modality. + +# 4.3.4. Ablation Studies + +Table 5 shows the ablation experiment results on the SthCom dataset. We first set the simplest baseline of our method, where spatiotemporal position embedding and rate reduction optimization are removed. It can be observed that this baseline makes spatially biased predictions, resulting in significantly lower performance in capturing temporal variations. We hypothesize that this is due to the static appearance bias in most of the datasets used for pre-training the video encoder, as discovered in previous studies [27, 31]. On the other hand, adding spatiotemporal position information and disentangling resampled representation by coding rate reduction significantly improves the performance, especially for capturing temporal variations. For experiments under different hyperparameter settings, we investigate the effect of the number of latent queries $(M)$ and subspaces $(K)$ . We found that using too many latent queries rather + +Table 5. Ablation experiment results on the Sth-Com [30] dataset. + +
Methodverbobj.seenunseenHMAUC
baseline37.450.735.437.128.111.3
+ position embedding50.450.444.445.036.818.0
+ rate reduction update50.050.644.645.437.318.3
M=3249.950.043.544.636.317.5
M=6450.250.143.845.236.717.9
M=12850.050.644.645.437.318.3
M=25649.850.443.845.036.617.8
K=649.550.444.344.936.517.9
K=1250.150.143.945.136.517.7
K=2450.050.644.645.437.318.3
K=4850.049.743.945.036.617.8
+ +decreases the performance, and the optimal number of subspaces is 24, which is the number of heads in the backbone video encoder. + +# 4.4. Qualitative Results + +Attention map Figure 3 illustrates the attention map from the cross-attention block in the last update step of the resampler module for three example videos. In three examples, our method accurately captures the informative object and its movement across time steps despite being trained without dense labels of object position, such as bounding boxes. + +Measuring disentanglement For Sth-Com videos in the test split, we aggregated the learned representations for two groups: spatial $\mathbf{Z}_s = \{\mathbf{z}_s^i\}_{i=1}^N \in \mathbb{R}^{D \times N}$ and temporal $\mathbf{Z}_t = \{\mathbf{z}_t^i\}_{i=1}^N \in \mathbb{R}^{D \times N}$ representations. Then, we compute the coding rate of each representation group, denoted as $R(\mathbf{Z}_s)$ and $R(\mathbf{Z}_t)$ . The degree of entanglement between two representation groups can be measured by the coding rate of the overlapping regions, which estimate the volume of the space spanned by representations, computed as $R(\mathbf{Z}_s) + R(\mathbf{Z}_t) - R(\mathbf{Z}_s \cup \mathbf{Z}_t)$ . Based on these three aspects, + +![](images/a1d1446b9b79de93e6c0e13b829f74c4ba3d02d2666cfbf9c687b89495d51843.jpg) +(a) w/ rate reduction optimization (ours) + +(b) w/o rate reduction optimization +Figure 4. PCA results of resampled representations from the proposed method (a) w/ rate reduction optimization and (b) w/o rate reduction optimization. The first two principal components of the resampled representations from videos in the Sth-Com [30] dataset belong to 4 possible action labels, which are the combination of two pre-defined categories for each spatial and temporal attribute. It can be observed that the principal component axes of learned representations are closely aligned with the semantic axes of attributes. +![](images/4fdc32928a7b800aa191e41999f8c45150760d55d1690771026797e75c257566.jpg) +- Approaching [pendrive] with your camera +- Moving away from [pendrive] with your camera +- Approaching [chair] with your camera +- Moving away from [chair] with your camera + +the proposed method with and without the rate reduction optimization is compared to the C2C [30]. As shown in Table 6, the proposed method shows the lowest coding rate of $R(\mathbf{Z}_s)$ and $R(\mathbf{Z}_t)$ , which means that learned representations are more compressed in each spatial and temporal subspace. Moreover, as seen with the lowest entanglement score (coding rate of the overlapping region), we also argue that the proposed method, equipped with the rate reduction optimization, learns the most disentangled representation in both space and time. + +Table 6. Comparison of representations learned by C2C [30] and our method with and without the rate reduction optimization based on the coding rate. For every video in the test split of the Sth-Com dataset, we aggregate the spatial and temporal learned representations denoted as $\mathbf{Z}_s = \{\mathbf{z}_s^i\}_{i=1}^N \in \mathbb{R}^{D \times N}$ and $\mathbf{Z}_t = \{\mathbf{z}_t^i\}_{i=1}^N \in \mathbb{R}^{D \times N}$ . Here, $R(\mathbf{Z}_s)$ and $R(\mathbf{Z}_t)$ denote the coding rate computed on $\mathbf{Z}_s$ and $\mathbf{Z}_t$ respectively. We also evaluate the degree of entanglement between $\mathbf{Z}_s$ and $\mathbf{Z}_t$ by measuring the coding rate of the overlapping area of the space they span computed as $R(\mathbf{Z}_s) + R(\mathbf{Z}_t) - R(\mathbf{Z}_s \cup \mathbf{Z}_t)$ . + +
MethodR(Zs)↓R(Zt)↓R(Zs∩Zt)↓
C2C [30]1125.2870.1827.2
Ours w/o rate reduction791.8742.9508.4
Ours w/ rate reduction788.3740.9504.3
+ +PCA results of resampled representations We explore the characteristics of subspaces of the learned representations with principal component analysis (PCA). Based on two distinct categories predefined for each of the spatial and temporal attributes, we select example videos belonging to the 4 possible action labels, which are the combination of spatial and temporal categories. For example, as seen in Figure 4, if we set (pendrive, chair) as spatial categories and (Approaching [something] with your camera, Moving away from [something] with your camera) as temporal categories, we select videos having action labels of (Approaching pendrive with your camera, Approaching chair with your camera, Moving away from pendrive with your camera, Moving away from chair with your camera). Then, we use the PCA to visualize the subspaces of the learned representations of the selected videos. As shown in Figure 4 (a), the first two principal components of the latent embedding space are closely aligned with the semantic axes of spatial and temporal attributes. Furthermore, in comparison with Figure 4 (b), the optimization with the rate reduction promotes each subspace to be highly compressed while being incoherent between different subspaces. Similar examples can be found in the supplementary material. + +# 5. Conclusion + +In this paper, we have presented a novel information-theoretic approach for video representation learning to achieve compositional video understanding. To this end, we proposed a resampler-based transformer decoder architecture that projects the input feature map into latent embeddings, disentangling spatial and temporal factors into multiple subspaces by maximizing the coding rate reduction. The proposed method achieved new state-of-the-art results on challenging zero-shot compositional action recognition and visual question answering problems. We hope our approach can suggest a new direction for the challenging problem of compositional video understanding. + +# 6. Acknowledgements + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2022-II220264/50%; RS-2022-II220951/20%; RS-2022-II220612/20%; RS-2020-II201373/10%, Artificial Intelligence Graduate School Program(Hanyang University)). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. +[2] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. +[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. +[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. +[6] Kwan Ho Ryan Chan, Yaodong Yu, Chong You, Haozhi Qi, John Wright, and Yi Ma. Redunet: A white-box deep network from the principle of maximizing rate reduction. Journal of machine learning research, 23(114):1-103, 2022. +[7] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. +[8] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[10] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. +[11] Rohit Girdhar and Deva Ramanan. CATER: A diagnostic dataset for Compositional Actions and TEmporal Reasoning. In ICLR, 2020. +[12] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, + +Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. "The" something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. +[13] Arthur Gretton, Olivier Bousquet, Alex Smola, and Bernhard Schölkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory, pages 63-77. Springer, 2005. +[14] Shaozhe Hao, Kai Han, and Kwan-Yee K Wong. Learning attention as disentangler for compositional zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15315-15324, 2023. +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[16] Irina Higgins, Loic Matthey, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR (Poster), 3, 2017. +[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022. +[18] Siteng Huang, Biao Gong, Yutong Feng, Min Zhang, Yiliang Lv, and Donglin Wang. Troika: Multi-path cross-modal traction for compositional zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24005-24014, 2024. +[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. +[20] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. +[21] Dat Huynh and Ehsan Elhamifar. Compositional zero-shot learning via fine-grained dense feature composition. Advances in Neural Information Processing Systems, 33: 19849-19860, 2020. +[22] Li Jiang, Shaoshuai Shi, and Bernt Schiele. Open-vocabulary 3d semantic segmentation with foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21284-21294, 2024. +[23] Michael Kampffmeyer, Yinbo Chen, Xiaodan Liang, Hao Wang, Yujia Zhang, and Eric P Xing. Rethinking knowledge graph propagation for zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11487-11496, 2019. +[24] Shyamgopal Karthik, Massimiliano Mancini, and Zeynep Akata. Kg-sp: Knowledge guided simple primitives for open world compositional zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9336-9345, 2022. + +[25] Hanjae Kim, Jiyoung Lee, Seongheon Park, and Kwanghoon Sohn. Hierarchical visual primitive experts for compositional zero-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5675-5685, 2023. +[26] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 40:e253, 2017. +[27] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428, 2022. +[28] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. +[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. +[30] Rongchang Li, Zhenhua Feng, Tianyang Xu, Linze Li, Xiao-Jun Wu, Muhammad Awais, Sara Atito, and Josef Kittler. C2c: Component-to-composition learning for zero-shot compositional action recognition. In European Conference on Computer Vision, pages 369-388. Springer, 2025. +[31] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. +[32] Yun Li, Zhe Liu, Hang Chen, and Lina Yao. Context-based and diversity-driven specificity in compositional zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17037-17046, 2024. +[33] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. +[35] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. +[36] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. +[37] Xiaocheng Lu, Song Guo, Ziming Liu, and Jingcai Guo. Decomposed soft prompt guided fusion enhancing for compositional zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23560-23569, 2023. + +[38] Yawen Lu, Dongfang Liu, Qifan Wang, Cheng Han, Yiming Cui, Zhiwen Cao, Xueling Zhang, Yingjie Victor Chen, and Heng Fan. Promotion: Prototypes as motion learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28109-28119, 2024. +[39] Yi Ma, Harm Derksen, Wei Hong, and John Wright. Segmentation of multivariate mixed data via lossy data coding and compression. IEEE transactions on pattern analysis and machine intelligence, 29(9):1546-1562, 2007. +[40] Massimiliano Mancini, Muhammad Ferjad Naeem, Yongqin Xian, and Zeynep Akata. Open world compositional zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5222-5230, 2021. +[41] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1049–1059, 2020. +[42] Ishan Misra, Abhinav Gupta, and Martial Hebert. From red wine to red tomato: Composition with context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1792-1801, 2017. +[43] Muhammad Ferjad Naeem, Yongqin Xian, Federico Tombari, and Zeynep Akata. Learning graph embeddings for compositional zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 953-962, 2021. +[44] Nihal V Nayak, Peilin Yu, and Stephen H Bach. Learning to compose soft prompts for compositional zero-shot learning. arXiv preprint arXiv:2204.03574, 2022. +[45] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. +[46] Senthil Purushwalkam, Maximilian Nickel, Abhinav Gupta, and Marc'Aurelio Ranzato. Task-driven modular networks for zero-shot compositional learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3593-3602, 2019. +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[48] Nirat Saini, Khoi Pham, and Abhinav Shrivastava. Disentangling visual embeddings for attributes and objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13658-13667, 2022. +[49] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. +[50] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin + +Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. +[51] Qingsheng Wang, Lingqiao Liu, Chenchen Jing, Hao Chen, Guoqiang Liang, Peng Wang, and Chunhua Shen. Learning conditional attributes for compositional zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11197-11206, 2023. +[52] Xiaolong Wang, Yufei Ye, and Abhinav Gupta. Zero-shot recognition via semantic embeddings and knowledge graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6857-6866, 2018. +[53] Ziyang Wu, Tianjiao Ding, Yifu Lu, Druv Pai, Jingyuan Zhang, Weida Wang, Yaodong Yu, Yi Ma, and Benjamin D Haeffele. Token statistics transformer: Lineartime attention via variational rate reduction. arXiv preprint arXiv:2412.17810, 2024. +[54] Ziyang Wu, Jingyuan Zhang, Druv Pai, XuDong Wang, Chandan Singh, Jianwei Yang, Jianfeng Gao, and Yi Ma. Simplifying dino via coding rate regularization. arXiv preprint arXiv:2502.10385, 2025. +[55] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. arXiv preprint arXiv:2302.03024, 2023. +[56] Yaodong Yu, Kwan Ho Ryan Chan, Chong You, Chaobing Song, and Yi Ma. Learning diverse and discriminative representations via the principle of maximal coding rate reduction. Advances in neural information processing systems, 33: 9422-9434, 2020. +[57] Yaodong Yu, Sam Buchanan, Druv Pai, Tianzhe Chu, Ziyang Wu, Shengbang Tong, Benjamin Haeffele, and Yi Ma. White-box transformers via sparse rate reduction. Advances in Neural Information Processing Systems, 36:9422-9457, 2023. +[58] Hoyeoung Yun, Jinwoo Ahn, Minseo Kim, and Eun-Sol Kim. Compositional video understanding with spatiotemporal structure-based transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18751-18760, 2024. +[59] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. +[60] Zhaoheng Zheng, Haidong Zhu, and Ram Nevatia. Caila: concept-aware intra-layer adapters for compositional zero-shot learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1721-1731, 2024. +[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. \ No newline at end of file diff --git a/zeroshotcompositionalvideolearningwithcodingratereduction/images.zip b/zeroshotcompositionalvideolearningwithcodingratereduction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5c1be3bb346623c6a8d78191f686471cbefead5f --- /dev/null +++ b/zeroshotcompositionalvideolearningwithcodingratereduction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a45a08b61c8235a8a91abf4fc744f889b929e078010e8f159f4f405de949e97 +size 492303 diff --git a/zeroshotcompositionalvideolearningwithcodingratereduction/layout.json b/zeroshotcompositionalvideolearningwithcodingratereduction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..93d1a721177b33de61223fd59a403f4dd6861cd4 --- /dev/null +++ b/zeroshotcompositionalvideolearningwithcodingratereduction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98c773ae8a2808dfe682c95d613359b28536d6db6c64774eab350d82354bfbdc +size 473085 diff --git a/zeroshotdepthawareimageeditingwithdiffusionmodels/073d8c0b-4053-48f6-972c-264031ab2e86_content_list.json b/zeroshotdepthawareimageeditingwithdiffusionmodels/073d8c0b-4053-48f6-972c-264031ab2e86_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9da89677157dfce77a7b20b5a40287b7f213bff7 --- /dev/null +++ b/zeroshotdepthawareimageeditingwithdiffusionmodels/073d8c0b-4053-48f6-972c-264031ab2e86_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a3272d6c0ab3e987104f262c117eb3a165310cf994408a27fcc8e524bdc9a12 +size 103328 diff --git a/zeroshotdepthawareimageeditingwithdiffusionmodels/073d8c0b-4053-48f6-972c-264031ab2e86_model.json b/zeroshotdepthawareimageeditingwithdiffusionmodels/073d8c0b-4053-48f6-972c-264031ab2e86_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4020e06db550d52c1a374c255c247d6346868fb8 --- /dev/null +++ b/zeroshotdepthawareimageeditingwithdiffusionmodels/073d8c0b-4053-48f6-972c-264031ab2e86_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d80a0189fb41da49e532cb49ef196a23b61955b5b83d92ae0847c15a290d8bf7 +size 132172 diff --git a/zeroshotdepthawareimageeditingwithdiffusionmodels/073d8c0b-4053-48f6-972c-264031ab2e86_origin.pdf b/zeroshotdepthawareimageeditingwithdiffusionmodels/073d8c0b-4053-48f6-972c-264031ab2e86_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6365575c0f15024b4e6fc89c2973488c5271489e --- /dev/null +++ b/zeroshotdepthawareimageeditingwithdiffusionmodels/073d8c0b-4053-48f6-972c-264031ab2e86_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04baa8045748372f13c957b4fb1bedea70462bfc2c67c60fa66126675fc5e9fb +size 25356059 diff --git a/zeroshotdepthawareimageeditingwithdiffusionmodels/full.md b/zeroshotdepthawareimageeditingwithdiffusionmodels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..458f850601c5213c24f00b88a57e8bb1ffca2dfe --- /dev/null +++ b/zeroshotdepthawareimageeditingwithdiffusionmodels/full.md @@ -0,0 +1,441 @@ +# Zero-Shot Depth Aware Image Editing with Diffusion Models + +Rishubh Parihar* Sachidanand VS* R. Venkatesh Babu +IISc Bangalore + +![](images/e73b401e7ecaf4dc36adc83c8150f03eae5c4dc3fb2748e1c764178a20fc6c48.jpg) +(a) + +![](images/80f8d059eda7539960506df41b7b11bb78e15897d7b4d12b61da34967a860aa7.jpg) +Depth Aware Scene Composition +Composition at d1 + +![](images/1aee190dfb76e3bf02aa7e6c1c21dbb1c4769121dce5f52b38331895fafbae05.jpg) +Composition at d2 + +![](images/49edf767ed36e739ded04f5580eda0829b7475e6225e3cc9bca5a65720bf93b8.jpg) +Figure 1. Our method performs precise depth-aware image editing at user-specified depth d. a) Given two input scenes and specified d, our method seamlessly composite the foreground (depth < d) of one scene with the background of another. b) Given a background image, an object image and a 2D bounding box, our method can realistically place the object at depth d, with appropriate scene occlusions. + +![](images/361628c7fa83b57ec3d09ddf38cfe0f12789bed76bf878d7b9e4d5e2853b5dfe.jpg) + +![](images/bbd750d9070531191c58d0d53568b56960b1e27cd0134e3eb33b6973594fbd3d.jpg) + +![](images/90b713860bdfea324ccf4c89307817f175ab293f9fe1b2da1d2ad7d5f5ce1aae.jpg) + +![](images/f53ad0191e9ddbd2f18d5204a406084ce1606155a41a61bbb2dfaf7f037d1ae2.jpg) +(b) +Input scene + +![](images/08929967cf54f8423f574e8357f60773232a6437c2dcb7a0baf098ec24fd562b.jpg) +Object + +![](images/640751b58146f6076d931f5e270e0d47c8692cfdf54ea772d456332e9f9a0c75.jpg) +Insertion at depth d1 + +![](images/4c575a553ce7511baddf28a91a118972b1d1a38b19e6eae48cbeede577b64a8b.jpg) +Insertion at depth d2 + +![](images/2f9f0291c5f744851261fb10ef0d2b086d21ee2cc34e29e8237f92edb0e3a438.jpg) + +![](images/0e2ba25036d00e6c92c7aea64a5db34deb3d9e384bc16efc98f8c1c289affe61.jpg) + +![](images/46a02d7d76bdb44c7955c1e1e944545cd910233e326cfbe70cd584af51b196bd.jpg) + +![](images/fdc1fb5a39a2ebc9dd1fe14fe10f69b9d5d891790fa1babbc13a5d50970c7997.jpg) + +![](images/3c6919c3c9d0d78322da00d9a7e225200d4e96a232b16bdaa5c18e71836c4d0c.jpg) + +![](images/e2f9d1ca16fc7783c8b715ed491449badb40381d877af8eedd03bb5d0a450142.jpg) + +# Abstract + +Diffusion models have transformed image editing but struggle with precise depth-aware control, such as placing objects at a specified depth. Layered representations offer fine-grained control by decomposing an image into separate editable layers. However, existing methods simplistically represent a scene via a set of background and transparent foreground layers while ignoring the scene geometry - limiting their effectiveness for depth-aware editing. We propose Depth-Guided Layer Decomposition - a layering method that decomposes an image into foreground and background layers based on a user-specified depth value, enabling precise depth-aware edits. We further propose Feature Guided Layer Compositing - a zero-shot approach for realistic layer compositing by leveraging generative priors from pretrained diffusion models. Specifically, we guide the internal U-Net features to progressively fuse individual layers into a composite latent at each denoising step. This preserves the structure of individual layers while generating realistic outputs with appropriate color and lighting adjustments without a need for post-hoc harmonization + +models. We demonstrate our method on two key depth-aware editing tasks: 1) scene compositing by blending the foreground of one scene with the background of another at a specified depth, and; 2) object insertion at a user-defined depth. Our zero-shot approach achieves precise depth ordering and high-quality edits, surpassing specialized scene compositing and object placement baselines, as validated across benchmarks and user studies. webpage - https://rishubhpar.github.io/DAEdit/ + +# 1. Introduction + +Recent advancements in diffusion models [1, 2] have significantly improved image editing [3-10]. Though these approaches work well for coarse image modifications, such as altering object appearance, adding attributes or changing image style via text prompts, they lack the precise control over image content that artists and designers require. Layered image representation offers finer control by decomposing an image into editable layers, a widely used technique in visual content editing workflows. While recent works have explored layered generation with diffusion models [11-13], their use in layered image editing is largely unexplored. + +Current layering approaches [11, 14, 15] decompose + +images into a background layer and multiple transparent foreground layers each corresponding to a distinct object (Fig. 2). This enables precise editing of existing objects, such as removal, resizing, and translation within the image plane via editing the individual layers. However, this object-centric layering overlooks the spatial geometry of the scene, including the depth of individual objects and their arrangement in 3D space. As a result, they are incapable of performing depth-aware editing, such as composing foreground from a scene with background from another at a specified depth (Fig. 1a)). Moreover, when composing layers from two different scenes, these methods require additional image harmonization models [16, 17] to adjust lighting and color for photorealistic outputs. + +In this work, we propose a novel zero-shot depth-aware editing framework that introduces Depth-Guided Layer Decomposition (DeGLaD). Given an input image, its depth map (from an off-the-shelf predictor [18]), and a user-specified depth $\mathbf{d}$ , DeGLaD decomposes the image into foreground (depth $< \mathbf{d}$ ) and background (depth $> \mathbf{d}$ ) layers (Fig. 2) based on the scene depth (Fig. 2). This decomposition enables precise depth-aware editing via independent editing of each layer. For example, to composite two scenes at specified depth $\mathbf{d}$ , the background layer from one scene can be replaced with another. Similarly, a novel object can be inserted at depth $\mathbf{d}$ by inpainting the object in the background layer using off-the-shelf inpainting model [19] and composite with the unedited foreground layer, ensuring placement at intended depth. As providing a scalar depth value can be challenging for the user, we offer an intuitive top-view interface that allows users to specify $\mathbf{d}$ with a single click (Suppl.Sec.B). + +Directly compositing edited layers in image space leads to unrealistic results, lacking proper lighting and color consistency. To address this, we integrate DeGLaD in the latent space of pretrained diffusion models, leveraging their rich generative priors for photorealistic compositing. First, we invert the input images in the diffusion latent space and then apply DeGLaD to obtain latent depth layers. For the seamless compositing of these layers, we propose Feature-Guided Layer Compositing (FeatGLaC) - a training-free method that gradually composes the edited layers by guiding the diffusion features towards the target composition at each denoising step, similar to classifier guidance [20]. This progressive compositing approach preserves the structure of individual layers while ensuring realistic compositing with natural lighting and color consistency. + +We evaluate our method on two novel depth-aware editing tasks: $\pmb{a}$ photorealistic compositing of scenes at a specified depth with appropriate relighting, and $\pmb{b}$ inserting a novel object at a precise depth. To benchmark these new tasks, we introduce a dataset featuring diverse objects and background scenes. Our zero-shot method outper + +![](images/383fa342d2e370a866dd7472dee87c79963495ee61ed4b704be9ceaef8d1c6b0.jpg) +Figure 2. Existing layering method [14] decompose an image into background and foreground object layers but disregard spatial scene geometry, limiting their applicability for depth-aware editing. In contrast, our Depth-Guided Layer Decomposition (DeGLaD) decomposes a scene based on the scene depth and a user-specified depth $\mathbf{d}$ , enabling precise depth-aware editing such as inserting objects at a precise depth. + +forms specialized baselines trained for object insertion and scene compositing, demonstrating superior depth awareness and photorealistic compositing supported quantitatively and with user studies. In summary, our key contributions are: + +- Depth-Guided Layer Decomposition - a novel layering method to decompose a given image into editable layers using its depth map and a user-specified depth value. +- Feature-Guided Layer Compositing — a zero-shot method that leverages pretrained diffusion models to progressively blend multiple layers with feature guidance, ensuring photorealistic lighting and color harmonization. +- Downstream applications of depth-based layering in novel depth-aware editing tasks - object insertion and scene compositing at a user-specified depth. +- Depth Edit Benchmark - A benchmark dataset consisting of diverse images with depth-aware editing annotations for evaluation of the two depth-aware editing tasks. + +# 2. Related Works + +Layered Image Generation. Recent methods perform large-scale diffusion model training on transparent layer dataset to perform layered generation [11, 12, 21], facilitating transparent content for editing workflows. Other methods [13, 14, 22, 23] decompose images into layered foreground and background representations or generate only a transparent foreground [24] for compositing-based edits. However, they lack depth-aware editing and are limited to individual object layers. PAIR-Diffusion [25] learns object-centric features to compose multiple images using scene segmentation maps, while [26] leverages a layered latent representation for object movement within a scene. Additionally, works on harmonization [17] and relighting [16, 27, 28] use layered representations for scene compositing but rely on large-scale paired datasets. + +Image Editing with Diffusion Models. Text-to-Image + +models [1, 2, 29] are extensively used for image editing and controlled image synthesis [3, 4, 30, 31]. A set of existing methods manipulate the cross-attention maps [3, 6, 32] during inference to control the image layouts. These include swapping the attention maps [3, 32], or taking attention across the batch of images [33, 34]. Another set of works explores the text conditioning space of the T2I model to achieve more control [6, 35-37]. Others aim to find semantic direction in latent space or the text space [38-42] for editing. However, these approaches focus primarily on appearance-based edits and lack precise 3D control. + +3D editing with Generative Models. While diffusion models excel at generating realistic images, they struggle with consistent 3D effects [43, 44]. To introduce 3D control, some methods condition diffusion models on scene normals or depth maps [1, 45], while others train on large-scale 3D-annotated datasets, using 3D bounding boxes [46] or geometric properties [47]. More recent approaches leverage generative priors from pretrained diffusion models for 3D-aware editing [31, 48-51]. Some lift 2D diffusion features to 3D space via depth maps for direct 3D editing [31, 48], while others edit the inferred mesh [52] or point cloud [49, 53] from a single image and refine the rendered image with pretrained diffusion models as postprocessing. Another set of methods [51, 54, 55], personalize the diffusion models on multi-view images to achieve 3D editing and view control for personalized object. + +Object Insertion. Existing methods formulate object insertion from a single image as an object inpainting task. A widely used approach is to condition the diffusion models on object features extracted from an additional image encoder, enabling object insertion within a specified 2D box [19, 56-60]. Further, some approaches enhance realism of the inserted object by implicitly modeling lighting and shading via curating high-quality datasets [61]. However, these methods lack control over object placement at a specific depth and always generate complete objects without considering occlusions from the scene. In contrast, 3D-based methods enable realistic object insertion [62] and 3D-aware edits [63] but require multiple images to construct accurate 3D representations such as NeRFs. Another approach for object insertion estimates floor planes and scene lighting to place synthetic 3D assets [64], but obtaining realistic 3D assets from a single image remains challenging. Unlike these methods, our approach enables realistic, depth-aware object insertion using only a single object and background image while considering occlusions. + +# 3. Method + +# 3.1. Preliminaries + +Diffusion models generate images by iteratively denoising a random noise sample. In the forward diffusion process, image $x_0$ is corrupted by sequentially adding stan + +dard Gaussian noise $\epsilon$ to obtain $x_{t}$ . A denoiser network $\epsilon_{\theta}$ is trained to estimate the added noise, conditioned on the timestep and optional conditioning such as text. For generating images, the reverse diffusion process denoises the random noise $x_{T}$ , with multiple passes through denoising network $\epsilon_{\theta}$ . To accelerate the diffusion models, Latent Diffusion Models [1] take a two-stage approach where the input image is first encoded into a lower dimensional latent space of a pretrained variational autoencoder, and the diffusion process is applied in the compressed latent space, reducing the computational demands. + +Guidance. Diffusion models generate images by iteratively denoising a random noise sample. Classifier guidance [20] provides a mechanism to steer this iterative sampling process using a predefined energy function $\mathcal{G}$ . This enables the ability to condition the generation during inference time without a need for model retraining. For example, to generate class-conditioned images, the energy function as the cross-entropy loss $\mathcal{L}$ between the pretrained classifier's prediction $f(x_{t})$ and the given class $y$ as $\mathcal{G} = \mathcal{L}(f(x_t,y))$ . During generation, the predicted noise $\epsilon_{\theta}$ is adjusted to minimize the classifier loss $\mathcal{L}$ by taking the loss gradient with respect to $x_{t}$ , with $\lambda$ as the classifier guidance weight as: + +$$ +\tilde {\epsilon} _ {\theta} \left(x _ {t}, t\right) = \epsilon_ {\theta} \left(x _ {t}, t\right) + \lambda \nabla_ {x _ {t}} \mathcal {L} (f (x _ {t}), y) \tag {1} +$$ + +Several recent guidance approaches achieve inference time conditioning on sketch [65], layout [30, 66], features [31], optical flow [67] and human skeleton [68]. + +3.2. Depth-Guided Layer Decomposition (DeGLaD) Scene depth serves as an effective representation to model the underlying scene geometry and enables enhanced control over 3D scene structure [31, 45]. Motivated by this, we propose Depth-Guided Layer Decomposition - a depth-based layering approach for precise depth-aware editing. The requirements for DeGLaD are an input image $\mathbf{x}$ , its corresponding depth map $\mathbf{D}$ (can be obtained from off-the-shelf depth predictor [18]). Additionally, the user has to specify a scalar depth value $\mathbf{d}$ , where the edit needs to be performed. Given these inputs, we decompose the image into foreground and background layers with corresponding binary masks $\mathbf{M}_{\mathrm{fg}}$ and $\mathbf{M}_{\mathrm{bg}}$ , computed as follows: + +$\mathbf{M}_{\mathrm{fg}}(\mathbf{i}, \mathbf{j}) = \mathbb{I}(\mathbf{D}(\mathbf{i}, \mathbf{j}) < \mathbf{d}), \quad \mathbf{M}_{\mathrm{bg}}(\mathbf{i}, \mathbf{j}) = \mathbb{I}(\mathbf{D}(\mathbf{i}, \mathbf{j}) \geq \mathbf{d})$ , (2) where $\mathbf{i}$ and $\mathbf{j}$ denotes the pixel coordinates, and $\mathbb{I}(\cdot)$ is the indicator function. The obtained layers can be edited independently and recomposed for precise depth-aware editing. While we illustrate decomposition at a single depth, our method naturally extends to multiple depths by providing a set of depth values $\{\mathbf{d}_1,.., \mathbf{d}_k\}$ , enabling multi-depth editing, such as composing multiple scenes or iteratively inserting objects (Fig. 1). + +3.3. Feature-Guided Layer Composition (FeatGLaC) Directly compositing the obtained layers in the image space leads to unnatural results, lacking color harmonization and proper lighting (Fig. 3). To address this, we integrate DeGLaD in the latent + +![](images/34b89e0aba36396c84696bba818bd446e2e384554d151857a6d73223f314aac3.jpg) +Scene 1 + +![](images/a5372eab879c683f308871917e8007c2890a29beffde48fa13719f42ac1d959e.jpg) +Scene 2 + +![](images/49d3c52a524c7a36cafaed3de89b0662951e89132b3c36894c765cbb56cd316e.jpg) +Masked FG + +![](images/ea276121a457a85ea668d4532215c3a98805881faa49f62c259d9e68dce0d518.jpg) +Masked BG + +![](images/46584b0c394cc232c60e6a6ee06556d4d0193aa67e27a0b2523e8c134bbf0e1a.jpg) +Image $\alpha$ -comp + +![](images/dcf202e2d769ded392250f6894014cb5d6f527841feb40396b6e2feef7e705ed.jpg) +Latent $\alpha$ -comp +T = 10 +Figure 3. Ablation for compositing layers: a) Naively $\alpha$ compositing the layers in the image space (Image $\alpha$ -comp) results in unnatural 'cut-paste' artifacts without adjusting color and scene lighting. b) Performing $\alpha$ compositing in the diffusion latent space at $t = \tau$ (Eq.3) followed by denoising (Latent $\alpha$ -comp $\tau$ ) has an inherent tradeoff. Compositing at an early stage of diffusion ( $\tau = 40$ ) results in identity loss due to excessive denoising, and compositing at a later stage ( $\tau = 10$ ) results in unnatural blending similar to Image $\alpha$ -comp. Our diffusion feature guidance-based layer compositing generates photorealistic composition while preserving the structure and identity of individual images. + +![](images/e42f8ca4ae2930223491e2a6f70f48861d0ea8b89f0d22ba09fdb9b51c90e6c6.jpg) +Latent $\alpha$ -comp +T=20 + +![](images/78b5ab272972ececaa65a40a6feeb06848b4e8e00ff2dd20d9c4cdb6cc90c541.jpg) +Latent $\alpha$ -comp +T=30 + +![](images/d67f5720378e8603f47611baa8157a0e6a5c3e2df44df61d4adea974a21f99c2.jpg) +Latent $\alpha$ -comp +T=40 + +![](images/a515dd609ee675f25ce7667497df7a0f8b6293a0c9524409c44e8c2476233952.jpg) +Feature Guidance +(Ours) + +space of pretrained diffusion models and progressively compose the layers during denoising to generate realistic outputs. We explain this approach using an example where the foreground layer from scene $\mathbf{x}_{\mathrm{a}}$ (extracted by binary mask $\mathbf{M}_{\mathrm{fg}}^{\mathrm{a}}$ ) is composed with the background from scene $\mathbf{x}_{\mathrm{b}}$ . For composing foreground from $\mathbf{x}_{\mathrm{a}}$ with background of $\mathbf{x}_{\mathrm{b}}$ , we define the background mask for $\mathbf{x}_{\mathrm{b}}$ to be the same as that of $\mathbf{x}_{\mathrm{a}}$ , i.e., $\mathbf{M}_{\mathrm{bg}}^{\mathrm{b}} = \mathbf{M}_{\mathrm{bg}}^{\mathrm{a}}$ . We start by inverting $\mathbf{x}_{\mathrm{a}}$ and $\mathbf{x}_{\mathrm{b}}$ with null-text inversion [69] to obtain the corresponding latents $\mathbf{z}_{0:T}^{\mathrm{a}}$ and $\mathbf{z}_{0:T}^{\mathrm{b}}$ . + +Baseline. One straightforward approach is to first $\alpha$ -composite the latents $\mathbf{z}_{\tau}^{\mathbf{a}}$ and $\mathbf{z}_{\tau}^{\mathbf{b}}$ at intermediate timestep $\tau$ to obtain a composite intermediate latent $\mathbf{z}_{\tau}^{\mathbf{c}}$ : + +$$ +\mathbf {z} _ {\tau} ^ {\overline {{\mathbf {c}}}} = \mathbf {M} _ {\mathbf {f g}} ^ {\mathbf {a}} * \mathbf {z} _ {\tau} ^ {\mathbf {a}} + \mathbf {M} _ {\mathbf {b g}} ^ {\mathbf {b}} * \mathbf {z} _ {\tau} ^ {\mathbf {b}} \tag {3} +$$ + +where $\mathbf{M}_{\mathrm{fg}}^{\mathrm{a}}$ is downsampled to match the dimension of latent $\mathbf{z_t}$ . The composed latent $\mathbf{z}_{\tau}^{\mathrm{c}}$ is then denoised with the diffusion model for the remaining $\mathbf{T} - \tau$ timesteps for realistic blending of the two layers [5]. Though this framework seems promising, it has an inherent tradeoff between realistic blending with complex scene effects and preserving layer identity, as shown in Fig. 3. A large $\tau$ (close to clean image) does not provide enough freedom to recover the complex scene effects with denoising, and a small $\tau$ (close to noisy image) generates plausible composition but changes the scene contents significantly. + +Composition with guidance. Rather than directly $\alpha$ -compositing the inverted latents, we introduce a more gradual fusion strategy that incorporates feature guidance at each denoising step. We call this approach Feature-Guided Layer Composition, FeatGLaC in short. We start by initializing the composite latent $\mathbf{z}_{\mathrm{T}}^{\mathrm{c}}$ as the background latent $\mathbf{z}_{\mathrm{T}}^{\mathrm{b}}$ and iteratively denoise it with feature guidance, similar to classifier guidance [20]. Following prior works [31, 65], which demonstrate that the internal features of the denoising U-Net are highly expressive and enable fine-grained control over generation, we guide these features towards the target composition to achieve seamless blending. We denote the U-Net features as $\Psi_{i,t}$ where i is the diffusion model layer index and t is the diffusion timestep. At each timestep, we extract the features $\Psi_{i,t}^{a}$ , $\Psi_{i,t}^{b}$ and + +![](images/df97508626dad8637dfde688fa49a92cfc34425ef611cf27b3817755ee2dffb5.jpg) +a +DeGLaD +Figure 4. Overall framework for scene compositing: a) Given an input foreground image $\mathbf{x}^{\mathrm{a}}$ and a user specified depth $\mathbf{d}$ , we decompose the image with Depth-Guided Layer Decomposition (DeGLaD) and obtain foreground $\mathbf{M}_{\mathrm{fg}}^{\mathrm{a}}$ and background $\mathbf{M}_{\mathrm{bg}}^{\mathrm{a}}$ masks. b) We compose the foreground latent $\mathbf{z}_{\mathbf{t}}^{\mathbf{a}}$ with background latent $\mathbf{z}_{\mathbf{t}}^{\mathbf{b}}$ using the obtained mask with diffusion feature guidance. We guide the features of the composite latent $\Psi_{i,t}^{c}$ using the activations for foreground $\Psi_{i,t}^{a}$ and background features $\Psi_{i,t}^{b}$ and update the composite latent $\mathbf{z}_{\mathbf{t}}^{\mathbf{c}}$ for $K$ iterations at each denoising step. + +![](images/38ff9f4bc09c062c91b6e3d73b4ddd0424582ad03010c234e676636fdc6c0363.jpg) +b) +Feature-Guided Layer Composition (FeatGLaC) + +$\Psi_{i,t}^{c}$ from $\mathbf{z}_{\mathbf{t}}^{\mathbf{a}}$ , $\mathbf{z}_{\mathbf{t}}^{\mathbf{b}}$ and composite latent $\mathbf{z}_{\mathbf{t}}^{\mathbf{c}}$ respectively. Next, we define the diffusion guidance energy $\mathcal{G}$ for progressive composition of these layers. + +Intuition: We force the foreground layer $(fg)$ of $\Psi_{i,t}^{c}$ to be close to foreground layer of $\Psi_{i,t}^{a}$ and the background layer $(bg)$ of $\Psi_{i,t}^{c}$ to be close to the background layer of $\Psi_{i,t}^{b}$ as shown in Fig. 4. This is implemented by defining $\mathcal{G} =$ + +$$ +\sum_ {i} | | \mathbf {M} _ {\mathrm {f g}} ^ {\mathbf {a}} * (\boldsymbol {\Psi} _ {\mathbf {i}, \mathbf {t}} ^ {\mathbf {a}} - \boldsymbol {\Psi} _ {\mathbf {i}, \mathbf {t}} ^ {\mathbf {c}}) | | ^ {2} + | | \mathbf {M} _ {\mathrm {b g}} ^ {\mathbf {b}} * (\boldsymbol {\Psi} _ {\mathbf {i}, \mathbf {t}} ^ {\mathbf {b}} - \boldsymbol {\Psi} _ {\mathbf {i}, \mathbf {t}} ^ {\mathbf {c}}) | | ^ {2} (4) +$$ + +We compute the gradients of guidance energy $\mathcal{G}$ with respect to composite latent $\mathbf{z}_{\mathbf{t}}^{\mathbf{c}}$ and backpropagate them to update the next sample prediction as $\tilde{\mathbf{z}}_{\mathbf{t}}^{\mathbf{c}} = \mathbf{z}_{\mathbf{t}}^{\mathbf{c}} - \nabla_{\mathbf{z}_{\mathbf{t}}^{\mathbf{c}}} \mathcal{G}$ for $\mathbf{K}$ iterations at each denoising timestep. This gradual layer composition approach strikes a good tradeoff in preserving layer identity and generating photorealistic compositions. Fig. 1 & 3. + +Notably, the proposed FeatGLaC framework is model-agnostic and can be integrated into any pretrained diffusion model. We demonstrate its flexibility by applying it to a depth-conditioned diffusion model for scene composition and an inpainting diffusion model for object insertion, leveraging their specialized editing capabilities for depth-aware tasks. + +# 3.4. Application in Depth-Aware Editing + +We implement layer decomposition with DeGLaD and composition with FeatGLaC to perform depth-aware scene composition in Sec. 3.4.1 and object insertion in Sec. 3.4.2. A detailed algorithm of our method for both these tasks is provided in Suppl.Sec.A. requires users to specify a depth value $\mathbf{d}$ for the scene where the edit needs to be performed, which may not be user-friendly. To address this, we alternatively provide an intuitive interface, as discussed in Suppl.Sec.B that visualizes segmented scene from a top-down view, allowing users to easily specify $\mathbf{d}$ with a single user click. + +# 3.4.1. Depth Aware Scene Composition + +The goal of this task is to seamlessly compose the foreground of one scene, $\mathbf{x_a}$ , with the background of another, $\mathbf{x_b}$ , at the user-specified depth $\mathbf{d}$ . We first decompose the images into depth-based layers using DeGLaD and compose the layers with FeatGLaC as discussed in Sec. 3.2 & 3.3 and Fig. 2. To ensure structure preservation of individual layers during composition, we incorporate a pretrained depth-conditioned diffusion model [1] for guidance. Specifically, we condition the model using an $\alpha$ -composited depth map, obtained by blending $\mathbf{D_a}$ (depth of $\mathbf{x_a}$ ) and $\mathbf{D_b}$ (depth of $\mathbf{x_b}$ ) with their respective foreground and background masks ( $\mathbf{M_{fg}^a}$ , $\mathbf{M_{bg}^a}$ ). This composite depth input allows the model to maintain the structure of individual layers during the composition. + +# 3.4.2. Depth Aware Object Insertion + +We introduce a novel task of realistically inserting a given object $\mathbf{x}_0$ in an input scene $\mathbf{x}$ at a user-specified depth $\mathbf{d}$ and inside a 2D bounding box $\mathbf{b}$ . Existing approaches [19, 56, 57] can insert a given object in the specified 2D bounding box but do not provide explicit depth-aware control during object insertion. To this end, we lift an object insertion model $\mathcal{H}$ [19] and make it depth-aware. Note that our method is generalized and can also work without an inpainting model (Fig. 5c and Suppl.Sec.F3). We first perform null-text inversion [69] on $\mathbf{x}$ to obtain latent $\mathbf{z}_{\mathbf{0:T}}$ and store corresponding diffusion U-Net features $\Psi_{i,t}$ for guidance. Next, we obtain the foreground $(\mathbf{M}_{\mathrm{fg}})$ and background $(\mathbf{M}_{\mathrm{bg}} = 1 - \mathbf{M}_{\mathrm{fg}})$ layers for the input scene $\mathbf{x}$ using DeGLaD at specified depth $\mathbf{d}$ . The object insertion model $\mathcal{H}$ takes input scene $\mathbf{x}$ , object image $\mathbf{x_o}$ and 2D bounding box $\mathbf{b}$ as input and iteratively denoises a edit latent $\mathbf{z}_{\mathbf{T}}^{\mathrm{e}}$ initialized from $\mathcal{N}(0,I)$ to inpaint the object in the bounding box. To perform depth-aware object insertion, we extract the features $\Psi_{i,t}^{e}$ of edit latent $\mathbf{z}_{\mathbf{t}}^{\mathrm{e}}$ from $\mathcal{H}$ at each timestep $\mathbf{t}$ and apply FeatGLaC to compose with only unedited foreground allowing the background layer to change. + +Intuition. We force the foreground (depth $< \mathbf{d}$ ) features $\Psi_{i,t}^{e}$ of the edit latent $\mathbf{z}_{\mathbf{t}}^{\mathbf{e}}$ to be close to the foreground features $\Psi^{i,t}$ of the inverted image latent $\mathbf{z}_{\mathbf{t}}$ at each denoising step. This is implemented by defining the guidance energy $\mathcal{G}$ as: + +$$ +\mathcal {G} = \sum_ {i} \left| \left| \mathbf {M} _ {\mathbf {f g}} * \left(\boldsymbol {\Psi} _ {\mathbf {i}, \mathbf {t}} - \boldsymbol {\Psi} _ {\mathbf {i}, \mathbf {t}} ^ {\mathrm {e}}\right) \right| \right| ^ {2} \tag {5} +$$ + +We use the above guidance loss to refine the intermediate edit latent $\mathbf{z}_{\mathbf{t}}^{\mathrm{e}}$ for $\mathbf{K}$ iterations at each denoising step. Empirically, replacing the foreground layer of $\mathbf{z}_{\mathbf{t}}^{\mathrm{e}}$ with the foreground layer of $\mathbf{z}_{\mathbf{t}}$ at an intermediate timestep $\tau$ , followed by the guidance update for the remaining timesteps, yields more accurate object insertion results. We believe this improves $\mathcal{H}$ 's ability to interpret object depth, especially scene occlusions, leading to seamless compositions (Fig. 6). + +# 4. Experiments + +We perform extensive experiments to evaluate our method for depth-aware editing. In this section, we first discuss the implementation and dataset details, followed by experiments on scene composition, object insertion, and ablation studies. Additional experiment and dataset details are provided in the supp. document. + +![](images/770060b580e11d7e9bb181e5445ca10a0eeeeaa5c192e469cb5f67ed361384cc.jpg) + +![](images/7b2a52933fc4c6a2be4e6f06ba17b0e127e6189beb07232471184fdc48a52bc4.jpg) +Figure 5. Depth-aware object insertion. Our method enables realistically placing multiple objects at precise user specified depths. b) It can also place the same object at multiple locations and depths in a given scene. c) Further, our method also works with depth-conditioned T2I model [1] by first inverting the background and object images and then compositing them using FeatGLaC. + +# 4.1. Implementation Details + +For scene composition, we use the depth-conditioned Stable Diffusion v2-depth [1], and for object insertion, we use the Any-door inpainting model [19] and depth conditioned text-to-image diffusion models [1]. The guidance is applied to features from the last and penultimate layers of the diffusion U-Net, which enhances edit plausibility, which is also shown in our ablations. For scene composition, guidance is applied from timesteps 0 to 38, updating the latent $\mathbf{z}_{\mathrm{t}}^{\mathrm{c}}$ for $K = 5$ iterations per step. For object insertion, guidance is given from timesteps 30 to 50, with the latent $\mathbf{z}_{\mathrm{t}}^{\mathrm{e}}$ updated for $K = 3$ iterations per step. Additionally, we use Depth Anything [18] to obtain the depth for object insertion and Zoedepth [70] for scene composition, as the depth-conditioned model SD v2-depth is pretrained with a metric depth map. We extracted captions of input scenes using a captioning model [71] to perform null-text inversion. Also, for scene composition, we compose the captions of two input scenes to condition the diffusion model during guided generation. + +# 4.2. Dataset + +Since we are the first to introduce the two depth-aware editing tasks, no public dataset exists for their evaluation. To address this, we curated the Depth Edit Benchmark - a dataset comprising two subsets tailored for the extensive evaluation of depth-aware scene composition and object insertion. + +Object insertion. We gathered a collection of 490 scene-object image pairs from the web. Each pair includes annotations for the corresponding scene's depth map, depth value $\mathbf{d}$ , where the object + +![](images/12cd0881ff5e4e0c4d671b00a4fb0387ec8de93b89234af68d059b9f0e1e5e14.jpg) +Inputs + +![](images/d06d649d772c3a5596fe05a95aab7b2f9a6655eab02905b8f746db38f362fbcb.jpg) +IP-Adapter + +![](images/bd774c8344fa53476b9a4b3cdbce760b67993ef8e61597c28017dd142a3f9a62.jpg) +PbE + +![](images/9d95f2aba110c9fab682dc77330e96b6f1832145e6319dc882397fe46675fc70.jpg) +Anydoor + +![](images/e83d543040e342d74e81ee10d56f96d4eef8e8138e12ae3c44cbc79f38e46561.jpg) +Ours + +![](images/c84cd59e6d987d3901b9848ba95423e32fff068250a6d7e2d543c21e515819dc.jpg) + +![](images/980ad624608408fef089d15f79781b74e12a541c9d2a471cc8c4c9136b4ac309.jpg) + +![](images/71a942f3070505214781b716da06f03b4273c960280d22aafcb42ac34aa1f0f4.jpg) + +![](images/6057d0d8af97a8e6ee6b68a2d94c6e20511d3fb1808033f7bc9af73d3b475334.jpg) + +![](images/aef580d3b4ca78620a61e37a86f8b6924e2feb12751ee72709ea24e9429231d4.jpg) + +![](images/9c9013b0792d963d9307dfd2644f5c9e5e6986a5b4a66b41b3f11318e4a51112.jpg) +Figure 6. Comparison for depth-aware object insertion: IP-Adapter and PbE struggles to insert objects with consistent identity within an amodal bounding box. Anydoor achieves plausible placement but generates artifacts along the mask border (marked in red). Our method enables realistic object insertion while preserving both object identity and scene consistency. + +![](images/32de00daa31b001812836d27bbf03412a14dbbd463c30506cfcd5b04a42e3f03.jpg) + +![](images/fcf6a1811abd4c8362403426ade5686e94e0a84d3b6401c7f9b3b27d89c8e422.jpg) + +![](images/50a03a265215eafb97cd2c8bdecb549b38b41a45c6bd6a813c5cca27d4697b9b.jpg) + +![](images/78b18a3902269dcbb772a0ae65696efce0d2d46cda3b9a4e53be917de599df45.jpg) + +can be plausibly placed, and a corresponding 2D bounding box for object insertion. This dataset includes diverse objects from indoor and outdoor environments, with possible occlusion for inserted objects to effectively assess depth-aware object insertion. + +Scene composition. We curated a dataset with 2,844 image pairs with diverse foreground and background scenes, sourced from the SSHarmonization dataset [72] and the web. The dataset covers a broad range of indoor and outdoor environments with varying lighting, composition, and appearance. Each image is annotated with depth maps (extracted from [70]) and the depth value $\mathbf{d}$ for each foreground scene for plausible scene composition. Additionally, we generate text prompts using an off-the-shelf image captioning model [71]. + +# 4.3. Object Insertion + +To our knowledge, we are the first to perform depth-aware object insertion using only a single object and background image; hence, we compare with reference-based inpainting baselines. We use state-of-the-art reference conditioned inpainting methods IPAdapter [56], Paint by example (PbE) [57], and Anydoor [19] to inpaint the given object in a scene in a specified bounding box. These methods take a bounding box as input and place the object without accounting for occlusions. For a fair comparison, we adapt them for depth-aware placement by using the foreground layer mask to occlude the bounding box with overlapping objects (Fig. 6), resulting in an amodal bounding box mask. This will preserve the foreground regions during inpainting and give us the illusion that the object is placed behind other objects. We have provided a comparison with additional inpainting methods in Suppl.Sec.F and a 3D-aware object insertion method (3D + +edit [47]) in Suppl. Sec.H. Metrics. We evaluate object insertion method for object identity, realism of the output, correctness of the inserted object, and depth consistency for the inserted object. We use DINO [73] feature similarity (DINO-sim) between the generated object in the bounding box and the reference object to measure identity preservation. To measure image realism, we compute KID [74] against COCO [75] as our evaluation set is relatively smaller to compute FID. To evaluate whether the object is actually placed, we use CLIP [76] + +
MethodDINO-sim ↑KID ↓Δ depth ↓Clip-sim ↑
IP-Adapter [56]0.2445.39.36627.81
PbE [57]0.2734.96.73360.12
Anydoor [19]0.5074.93.17683.23
Ours0.5454.82.98984.86
+ +Table 1. Depth-Aware object insertion comparison. KID and $\Delta$ depth are reported in $\times {10}^{2}$ units. + +similarity (CLIP-sim) between 'a photo of object-name' and the cropped image from the generated image. If the object is correctly generated, the CLIP score should be higher. To assess depth consistency, we compute the discrepancy between the predicted object depth and the user-specified input depth. We estimate the depth of the generated image (using [18]), and compute the mean object depth of the object segment (obtained with SAM [77]). We report normalized $\Delta$ depth across the dataset, where lower values indicate more consistent depth-aware placement. + +Analysis. We present the results for depth-aware object insertion in Fig. 6, and Tab. 1. The reference-conditioned inpainting models, such as the IP-adapter and Paint-by-example (PbE), struggle to generate accurate objects in the amodal mask as they have been trained to primarily inpaint unoccluded objects with 2D bounding boxes. This is quantified with a poor CLIP-sim metric in Tab. 1. Anydoor is able to generate consistent objects; however, it generates significant border artifacts (marked in red), resulting in an unnatural composition. Our approach generates realistic compositions with accurate object insertion (highest Clip-sim) and superior identity preservation (highest Dino-sim) as compared to all the object insertion baselines. Further, the object is naturally placed at an accurate depth, as evident with lower $\Delta$ depth scores. Note that the identity preservation of the object is limited by the base inpainting model and is not a limitation of our guidance method. However, we can improve the identity by doing a post-processing step; experiments are in the Supp.Sec.F1 document. + +![](images/fcc323b8b276359e88093301cca4a6fb68241fef7d0f5e6c78f962515f28cdd4.jpg) +Figure 7. Depth consistency in object insertion: The depth map after object insertion appears visually consistent, confirming the object's placement accurately within the scene geometry. + +Inserted objects are consistent with input depth. To analyze the depth consistency of the inserted object, we visualize the depth map from [78] after object insertion in Fig. 7. The visualization confirms the object is places at accurate scene depth between the foreground and background scene objects. + +# 4.4. Scene Composition + +We compare our method with the following baselines: a) DeGLaD Image - We perform DeGLaD in the image space and compose + +![](images/066e50c000110d08fd405de96ebd215ddce9e51d5b4156929c6ff163bb4ba348.jpg) +Figure 8. Depth-aware scene compositing. a) Our method seamlessly blends input scenes at a specified depth with accurate color and lighting adjustments. b) Additionally, our method enables scene relighting by compositing a foreground object with a plain background featuring strong lighting effects. + +the edited image layers with $\alpha$ compositing. Further, we perform image harmonization using [17] as post-processing for realistic blending. b) DeGLaD + SDEdit [5] - We perform SDEdit (noise and denoise with diffusion model) on the output of DeGLaD composition in the image space for generating realistic compositions. c) PAIR Diffusion [25] allows for localized control during generation by copying content from a masked reference image. We use the layer masks $(\mathbf{M}_{\mathrm{fg}}$ and $\mathbf{M}_{\mathrm{bg}})$ to segment out the foreground and background regions and then pass desired reference scene images to PAIR Diffusion for generating the composed scene. + +Metrics. We measure the visual quality of the composed scene and identity (structure and appearance) preservation of the foreground and background. We report FID with the COCO dataset to quantify the realism of the generated image. To evaluate + +identity preservation, we report the average LPIPS distance between the background and the fore + +
MethodLPIPS ↓FID ↓
Pair-Diffusion0.45140.54
DeGLAD Image0.036132.6
DeGLaD + SDEdit0.395106.24
DeGLaD Diff0.263123.32
+ +Table 2. Scene compositing comparison + +ground region of the composite image with the input images. For realistic scene composition, both LPIPS and FID should be low, indicating superior identity preservation and realism. + +Analysis. We present our results and comparisons in Fig. 9 and Tab. 2. DeGLaD Image achieves harmonization of the foreground to improve blending; however, it still struggles with cut-pasting appearance at the layer borders, leading to unrealistic generation (inferior FID score). DeGLaD + SDEdit and PAIR-diffusion change the scene structure while generating consistent images in + +some examples. Our method generates photorealistic depth-aware scene composition with accurate scene illumination while preserving the scene structure. Notably, our method is robust to minor errors in the layered mask, as the masks are only used in guidance loss computation in a smaller dimensional latent space. Additionally, our method can realistically relight the scene by providing different sky backgrounds. + +![](images/4f8b6fe88e99f23568ef7023519d84217defb57602e598389482f27e1973ad61.jpg) +Figure 9. Comparison for scene compositing. DeGLaD Image result in cut-pasting artifacts leading to unnatural outputs. DeGLaD+SDEdit and Pair Diffusion generate unnatural compositions and distort the identity in some cases. Our method realistically composite the two scenes in a depth-aware manner with consistent intra-scene illumination. + +Composed scene follows accurate depth ordering. To analyze the depth consistency, we visualize the histogram of the input and the output scene in Fig. 10, which shows our method preserves the distribution of depth present in the foreground and background scene even during composition. + +![](images/5b3629c0de7899374f07933406fa5be4e9bb99ab2a057f40e082999fa2d242f0.jpg) +Figure 10. Depth consistency in scene editing: The initial depth regions in the composite image align with the foreground depth maps, while the later regions correspond to the background depth maps, indicating the preserved depth distribution. + +# 4.5. User study + +Due to the unavailability of well-established benchmarks for the task, we perform a user study to evaluate our approach across multiple aspects. We perform a user study to evaluate our method for depth-aware scene editing. We evaluate object insertion for the realism of the placement, identity preservation, and depth consistency. For the task of scene composition, we evaluate for the realism of the composition and depth consistency. The study was performed on 15 source images for each task, and 40 volunteers participated with varied expertise in image editing. We created 60 image pairs for object insertion and 40 pairs for scene composition, with each pair consisting of our generated output and a + +randomly sampled baseline. We divide this dataset into groups of 20 image pairs for separate analysis on each editing goal. Each user compared 20 pairs for each of the goals for the two tasks. The order of image pairs and the methods within each pair were randomized. The results of the user study are present in Fig. 11 + +![](images/8927f3e41c5bb9a5b4b2a2d97b100f904c8ac15b902cd6290e120bb684602098.jpg) +Figure 11. User study compiled from 40 responses. Our object insertion method is better than the baselines in realism, identity preservation and depth consistency. Similarly, for scene compositing, our approach surpasses the baselines in both realism and depth consistency, with image DeGLaD Image yielding comparable results but it suffers from image cut-pasting artifacts (Fig. 9). + +Object insertion. Our method significantly outperforms all baselines in terms of realism, identity preservation, and depth consistency. PbE and IP-adapter perform poorly across all three goals, indicating the challenge of depth-aware placement task. Our approach excels in depth consistency metrics, indicating that our method effectively performs depth-aware editing while producing highly realistic images. This can also be seen while visualising the generated image depth map Fig. 11 where the object depth map is consistent with the surrounding. + +Scene composition. As indicated in the user study, DeGLaD Image performs comparably to our approach for both goals. However, the harmonization model used in DeGLaD baseline, is specifically trained on a large-scale dataset for the task of blending objects in the background scene whereas our method is zero-shot. Further, applying DeGLaD in the image space suffers with cut-paste artifacts as shown in Fig. 9. As compared to all the other scene compositing baselines, our approach achieves significantly better performance. + +# 4.6. Ablations + +We ablate over the design choices for scene compositing in Fig. 12. We follow the same guidance parameters for the object insertion task as well. Additional quantitative ablations are provided in the Supp.Sec.C - Tab.1 & 2. + +Guidance Timesteps. We ablate over the timestep range from $0 - 50$ for applying the FeatGLaC guidance. Guiding only for small timesteps $(0 - 20)$ results in significant structure changes for the foreground and background scenes. On the contrary, providing guidance for all the timesteps preserves the structure but leads to unnatural composition (lighting mismatch). We found that guiding until an intermediate range of timesteps (0-38) and allowing the image to denoise freely for the remaining steps strikes a good balance, resulting in realistic compositions. + +Guidance Layers. We ablate over the U-Net decoder features used to calculate FeatGLaC guidance loss, and using all the decoder layers for guidance results in significant artifacts. We observe that guidance with the first decoder layers can significantly + +![](images/e21dabd03e4fb5728e7785d8cb170b6fbf106996994fc62d45e15e153b467b04.jpg) +Figure 12. Ablation for scene compositing guidance parameters + +hurt the generation. Finally, we achieve a combination of layer 3 (weight 8.5) and layer 2 (weight 0.2) works well in most cases. Using only one of these layers resulted in subpar compositions. + +Guidance weight. After finalizing the layers to be used for FeatGLaC guidance, we tried different weights for the guidance factor. Specifically, we ablate over a guidance multiplier $\lambda$ for foreground guidance. Having a smaller $\lambda$ results in generating only a background region, we achieve a good composition with $\lambda = 1$ . Notably, $\lambda$ is also a control parameter that a user uses to control the effect of the background on the foreground scene. + +# 5. Conclusion and Discussion + +Limitations. While our approach is highly effective, it has some limitations. Since our method builds on pretrained diffusion models, it inherits their biases. For object insertion, we rely on an inpainting model that may distort object identity in complex cases, such as objects with intricate textures (Fig. 13). Integrating the advancements in recent inpainting models can improve the identity in such cases. Additionally, our guidance mechanism involves optimization at each denoising step, increasing computational cost. + +Conclusion. In this work, we propose a zero-shot framework for depth-aware image editing. We introduce a novel depth-based layering approach that + +![](images/c07656e2b07637142548b6995cc77fb44246efb6a2a084ebffc7fc9d364d2109.jpg) +Figure 13. Failure cases. + +decomposes an image based on a user-specified depth value, enabling precise depth control. Additionally, we present a layer composition method that progressively blends layers using diffusion feature guidance at each denoising step, ensuring realistic layer composition. We demonstrate the effectiveness of our approach on two novel tasks: depth-aware object insertion and scene composition, achieving highly plausible edits with accurate depth control. Our work offers a fresh perspective on image layering and its applications in depth-aware editing. + +Acknowledgements. We thank Vaibhav Agarwal, Tejan Karmali and Jogendra Kundu for reviewing the manuscript and providing insightful feedback. This work is supported with PMRF from the Government of India. + +# References + +[1] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3, 5 +[2] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3 +[3] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 3 +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 3 +[5] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 4, 7 +[6] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 3 +[7] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8861-8870, 2024. +[8] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. In European Conference on Computer Vision, pages 365-381. Springer, 2024. +[9] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024. +[10] Yusuf Dalva, Kavana Venkatesh, and Pinar Yanardag. Fluxspace: Disentangled semantic editing in rectified flow transformers. arXiv preprint arXiv:2412.09611, 2024. 1 +[11] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 1, 2 +[12] Runhui Huang, Kaixin Cai, Jianhua Han, Xiaodan Liang, Renjing Pei, Guansong Lu, Songcen Xu, Wei Zhang, and Hang Xu. Layerdiff: Exploring text-guided multi-layered composable image synthesis via layer-collaborative diffusion model. In European Conference on Computer Vision, pages 144-160. Springer, 2024. 2 + +[13] Yusuf Dalva, Yijun Li, Qing Liu, Nanxuan Zhao, Jianming Zhang, Zhe Lin, and Pinar Yanardag. Layerfusion: Harmonized multi-layer text-to-image generation with generative priors. arXiv preprint arXiv:2412.04460, 2024. 1, 2 +[14] Jinrui Yang, Qing Liu, Yijun Li, Soo Ye Kim, Daniil Pakhoomov, Mengwei Ren, Jianming Zhang, Zhe Lin, Cihang Xie, and Yuyin Zhou. Generative image layer decomposition with visual effects. arXiv preprint arXiv:2411.17864, 2024. 1, 2 +[15] Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with diffusion models. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 1 +[16] Mengwei Ren, Wei Xiong, Jae Shin Yoon, Zhixin Shu, Jianming Zhang, HyunJoon Jung, Guido Gerig, and He Zhang. Relightful harmonization: Lighting-aware portrait background replacement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6452-6462, 2024. 2 +[17] Zhanghan Ke, Chunyi Sun, Lei Zhu, Ke Xu, and Rynson W.H. Lau. Harmonizer: Learning to perform white-box image and video harmonization. In European Conference on Computer Vision (ECCV), 2022. 2, 7 +[18] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 3, 5, 6 +[19] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2, 3, 5, 6 +[20] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2, 3, 4 +[21] Zipeng Qi, Guoxi Huang, Zebin Huang, Qin Guo, Jinwen Chen, Junyu Han, Jian Wang, Gang Zhang, Lufei Liu, Errui Ding, et al. Layered rendering diffusion model for zero-shot guided image synthesis. arXiv preprint arXiv:2311.18435, 2023. 2 +[22] Jinrui Yang, Qing Liu, Yijun Li, Soo Ye Kim, Daniil Pakhoomov, Mengwei Ren, Jianming Zhang, Zhe Lin, Cihang Xie, and Yuyin Zhou. Generative image layer decomposition with visual effects. arXiv preprint arXiv:2411.17864, 2024. 2 +[23] Petru-Daniel Tudosiu, Yongxin Yang, Shifeng Zhang, Fei Chen, Steven McDonagh, Gerasimos Lampouras, Ignacio Iacobacci, and Sarah Parisot. Mulan: A multi layer annotated dataset for controllable text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22413-22422, 2024. 2 +[24] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kassen, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 2 +[25] Vedit Goel, Elia Peruzzo, Yifan Jiang, Dejia Xu, Xingqian Xu, Nicu Sebe, Trevor Darrell, Zhangyang Wang, and Humphrey Shi. Pair diffusion: A comprehensive multimodal + +object-level image editor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8609-8618, 2024. 2, 7 +[26] Jiawei Ren, Mengmeng Xu, Jui-Chieh Wu, Ziwei Liu, Tao Xiang, and Antoine Toisoul. Move anything with layered scene diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6380-6389, 2024. 2 +[27] Junuk Cha, Mengwei Ren, Krishna Kumar Singh, He Zhang, Yannick Hold-Geoffroy, Seunghyun Yoon, HyunJoon Jung, Jae Shin Yoon, and Seungryul Baek. Text2relight: Creative portrait relighting with text guidance. arXiv preprint arXiv:2412.13734, 2024. 2 +[28] Sumit Chaturvedi, Mengwei Ren, Yannick Hold-Geoffroy, Jingyuan Liu, Julie Dorsey, and Zhixin Shu. Synthlight: Portrait relighting with diffusion model by learning to re-render synthetic faces. arXiv preprint arXiv:2501.09756, 2025. 2 +[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 3 +[30] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3 +[31] Karran Pandey, Paul Guerrero, Matheus Gadelha, Yannick Hold-Geoffroy, Karan Singh, and Niloy J Mitra. Diffusion handles enabling 3d edits for diffusion models by lifting activations to 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7695–7704, 2024. 3, 4 +[32] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 3 +[33] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. arXiv preprint arXiv:2405.01434, 2024. 3 +[34] Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation. ACM Transactions on Graphics (TOG), 43(4):1-18, 2024. 3 +[35] Yuxin Zhang, Weiming Dong, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Tong-Yee Lee, Oliver Deussen, and Changsheng Xu. Prospect: Prompt spectrum for attribute-aware personalization of diffusion models. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 3 +[36] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $\mathfrak{p}+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. +[37] Yuval Alaluf, Elad Richardson, Gal Metzer, and Daniel Cohen-Or. A neural space-time representation for text-to-image personalization. ACM Transactions on Graphics (TOG), 42(6):1-10, 2023. 3 + +[38] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 3 +[39] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting.Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36:25365-25389, 2023. +[40] Stefan Andreas Baumann, Felix Krause, Michael Neumayr, Nick Stracke, Vincent Tao Hu, and Björn Ommer. Continuous, subject-specific attribute control in t2i models by identifying semantic directions. arXiv preprint arXiv:2403.17064, 2024. +[41] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469-487. Springer, 2024. +[42] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6678, 2024. 3 +[43] Ayush Sarkar, Hanlin Mai, Amitabh Mahapatra, Svetlana Lazebnik, David A Forsyth, and Anand Bhattachad. Shadows don't lie and lines can't bend! generative models don't know projective geometry... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28140-28149, 2024. 3 +[44] Rishi Upadhyay. Improving Projective Geometry in Diffusion Models. PhD thesis, UCLA, 2024. 3 +[45] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3 +[46] Shariq Farooq Bhat, Niloy Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[47] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. Advances in Neural Information Processing Systems, 36, 2024. 3, 6 +[48] Rahul Sajnani, Jeroen Vanbaar, Jie Min, Kapil Katyal, and Srinath Sridhar. Geodiffuser: Geometry-based image editing with diffusion models. arXiv preprint arXiv:2404.14403, 2024.3 +[49] Ruicheng Wang, Jianfeng Xiang, Jiaolong Yang, and Xin Tong. Diffusion models are geometry critics: Single image 3d editing using pre-trained diffusion priors. arXiv preprint arXiv:2403.11503, 2024. 3 +[50] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024. + +[51] Rishubh Parihar, Vaibhav Agrawal, Sachidanand VS, and Venkatesh Babu Radhakrishnan. Compass control: Multi object orientation control for text-to-image generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2791-2801, 2025. 3 +[52] Jiraphon Yenphraphai, Xichen Pan, Sainan Liu, Daniele Panozzo, and Saining Xie. Image sculpting: Precise object editing with 3d geometry control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4241-4251, 2024. 3 +[53] Qihang Zhang, Yinghao Xu, Chaoyang Wang, Hsin-Ying Lee, Gordon Wetzstein, Bolei Zhou, and Ceyuan Yang. 3ditscene: Editing any scene via language-guided disentangled gaussian splatting. arXiv preprint arXiv:2405.18424, 2024. 3 +[54] Nupur Kumari, Grace Su, Richard Zhang, Taesung Park, Eli Shechtman, and Jun-Yan Zhu. Customizing text-to-image diffusion with camera viewpoint control. arXiv preprint arXiv:2404.12333, 2024. 3 +[55] Or Patashnik, Rinon Gal, Daniel Cohen-Or, Jun-Yan Zhu, and Fernando de la Torre. Consolidating attention features for multi-view image editing. In ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia, 2024. 3 +[56] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 5, 6 +[57] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 5, 6 +[58] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. +[59] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In European Conference on Computer Vision, pages 150–168. Springer, 2024. +[60] Nataniel Ruiz, Yuanzhen Li, Neal Wadhwa, Yael Pritch, Michael Rubinstein, David E Jacobs, and Shlomi Fruchter. Magic insert: Style-aware drag-and-drop. arXiv preprint arXiv:2407.02489, 2024. 3 +[61] Daniel Winter, Matan Cohen, Shlomi Fruchter, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. arXiv preprint arXiv:2403.18818, 2024. 3 +[62] Mohamad Shahbazi, Liesbeth Claessens, Michael Niemeyer, Edo Collins, Alessio Tonioni, Luc Van Gool, and Federico Tombari. Inserf: Text-driven generative object insertion in neural 3d scenes. arXiv preprint arXiv:2401.05335, 2024. 3 +[63] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Edit + +ing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740-19750, 2023. 3 +[64] Yunhao Ge, Hong-Xing Yu, Cheng Zhao, Yuliang Guo, Xinyu Huang, Liu Ren, Laurent Itti, and Jiajun Wu. 3d copy-paste: Physically plausible object insertion for monocular 3d detection. Advances in Neural Information Processing Systems, 36, 2024. 3 +[65] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3, 4 +[66] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3 +[67] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In The Twelfth International Conference on Learning Representations. 3 +[68] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8227, 2024. 3 +[69] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023. 4, 5 +[70] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 5, 6 +[71] Abdou. vit-swin-base-224-gpt2-image-captioning. In https://huggingface.co/Abdou/vit-swin-base-224-gpt2-image-captioning, 2022. 5, 6 +[72] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 6 +[73] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 6 +[74] Mikołaj Binkowski, Dougal J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations, 2018. 6 +[75] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, + +Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[76] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 6 +[77] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6 +[78] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 6 \ No newline at end of file diff --git a/zeroshotdepthawareimageeditingwithdiffusionmodels/images.zip b/zeroshotdepthawareimageeditingwithdiffusionmodels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2c1e842d142725d92b310975390ba2f3b336e2de --- /dev/null +++ b/zeroshotdepthawareimageeditingwithdiffusionmodels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64f457188f63bab1a48d6e5137f905a86f74d65df559a7e0534be25dc1c1ccd1 +size 762194 diff --git a/zeroshotdepthawareimageeditingwithdiffusionmodels/layout.json b/zeroshotdepthawareimageeditingwithdiffusionmodels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e335e73793e597ef80f106e4543aac4bd747d7da --- /dev/null +++ b/zeroshotdepthawareimageeditingwithdiffusionmodels/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c44bce27ade5a70264dbf1f9a0ba9f5aa29ba15fa2506892f6a18454ae006902 +size 624258 diff --git a/zeroshotinexactcadmodelalignmentfromasingleimage/9abc6024-2f21-4050-a942-e3a3a7717cec_content_list.json b/zeroshotinexactcadmodelalignmentfromasingleimage/9abc6024-2f21-4050-a942-e3a3a7717cec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d3cb2ce6d68039e3c503d9f8e5cc72eb666c2d6b --- /dev/null +++ b/zeroshotinexactcadmodelalignmentfromasingleimage/9abc6024-2f21-4050-a942-e3a3a7717cec_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:557f51e8e550a106c9e36703ea5e26f10024c47464a345629c2dbdd13f9c3146 +size 87617 diff --git a/zeroshotinexactcadmodelalignmentfromasingleimage/9abc6024-2f21-4050-a942-e3a3a7717cec_model.json b/zeroshotinexactcadmodelalignmentfromasingleimage/9abc6024-2f21-4050-a942-e3a3a7717cec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ab1893d436916f4f8f8714d4a933a3cdec62d3f4 --- /dev/null +++ b/zeroshotinexactcadmodelalignmentfromasingleimage/9abc6024-2f21-4050-a942-e3a3a7717cec_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3eec080c340f49177c327e6cd334b7735020114170664640c39c0f7a0e1f9abf +size 106946 diff --git a/zeroshotinexactcadmodelalignmentfromasingleimage/9abc6024-2f21-4050-a942-e3a3a7717cec_origin.pdf b/zeroshotinexactcadmodelalignmentfromasingleimage/9abc6024-2f21-4050-a942-e3a3a7717cec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fa04a654529deb592481c8ef924a288c9813c7c2 --- /dev/null +++ b/zeroshotinexactcadmodelalignmentfromasingleimage/9abc6024-2f21-4050-a942-e3a3a7717cec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34a4a931467e1912a8c4e2ed00a11a4efa1775f03ebf378d00e8db2f79d44e77 +size 7463219 diff --git a/zeroshotinexactcadmodelalignmentfromasingleimage/full.md b/zeroshotinexactcadmodelalignmentfromasingleimage/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2d449cfd331b12d3d3eacdc01f685ab146b216ec --- /dev/null +++ b/zeroshotinexactcadmodelalignmentfromasingleimage/full.md @@ -0,0 +1,302 @@ +# Zero-shot Inexact CAD Model Alignment from a Single Image + +Pattaramanee Arsomngern $^{1}$ Sasikarn Khwanmuang $^{1}$ Matthias Nießner $^{2}$ Supasorn Suwajanakorn $^{1}$ $^{1}$ VISTEC, Thailand $^{2}$ Technical University of Munich, Germany + +https://zerocad9d.github.io/ + +![](images/f140c6eb11413ee801fbe011170b50e0ca19389d5d086fa8a1b62099a4822d2f.jpg) +Figure 1. We estimate the 9-DoF pose of an inexact CAD model to align it with a target object's pose in a 2D image without training on scene-level pose annotations. Our approach can generalize to unseen classes in real images even though it was trained on just 9 classes. + +# Abstract + +One practical approach to infer 3D scene structure from a single image is to retrieve a closely matching 3D model from a database and align it with the object in the image. Existing methods rely on supervised training with images and pose annotations, which limits them to a narrow set of object categories. To address this, we propose a weakly supervised 9-DoF alignment method for inexact 3D models that requires no scene-level pose annotations and generalizes to unseen categories. Our approach derives a novel feature space based on foundation features that ensure multi-view consistency and overcome symmetry ambiguities inherent in foundation features using a self-supervised triplet loss. Additionally, we introduce a texture-invariant pose refinement technique that performs dense alignment in normalized object coordinates, estimated through the enhanced feature space. We conduct extensive evaluations on the real-world ScanNet25k dataset, where our method outperforms SOTA weakly supervised baselines by $+4.3\%$ mean alignment accuracy and is the only weakly supervised approach to surpass the supervised ROCA by $+2.7\%$ . To assess generalization, we introduce SUN2CAD, a real-world test set with 20 novel object categories, where our method achieves SOTA results without prior training on them. + +# 1. Introduction + +Recovering 3D scene structure from a single image is highly ill-posed, not only due to depth prediction ambiguities but + +also because large portions of objects are occluded, making full object reconstruction difficult. One way to address this is by aligning existing 3D models to objects in the input image [16, 18, 23, 24]. This approach leverages artist-crafted 3D models, which offer detailed and complete geometry, even for occluded regions, and is particularly well-suited for applications like VR and gaming, where visual realism takes priority over exact geometric fidelity. + +Unlike 6-DoF CAD alignment tasks [25, 37, 49], where the 3D model exactly matches the object, this task involves aligning an inexact 3D model, retrieved from a database, that may differ in shape or texture or lack texture entirely, to its object's pose in an image. In this paper, we propose an approach that eliminates the need for pose annotations to solve this task, enabling generalization to novel objects in unseen categories in a zero-shot manner. + +Existing methods [18, 24, 26] tackle this problem by relying on extensive 3D supervision from annotated tuples of RGB images, depth maps, CAD models, and 9-DoF poses. However, by training on annotated poses from a limited set of categories, they struggle to generalize to unseen categories that differ significantly from the training set [2, 11]. + +To address annotation scarcity, several studies utilize synthetic data for training [16, 49]. DiffCAD [16] relies on the dataset 3D-FRONT [15], which provides 9-DoF pose annotations for CAD models in synthetic indoor scenes. However, its pose estimator is category-specific and still requires CAD models from that category to synthesize training data, making it difficult to scale and infeasible for unseen categories. FoundationPose [49] constructs its own + +synthetic dataset, which covers over 1,000 object categories with diverse textures through augmentation. However, it is designed for 6-DoF tasks with matching textures and models, and its performance drops on inexact matches. Other 6-DoF studies [5, 37] leverage semantic features from 2D foundation models like DINOv2 [36] to establish zero-shot 2D-3D correspondences for pose estimation. As our study shows, DINOv2-based techniques also struggle in our 9-DoF task with inexact matches or unseen categories. + +This performance drop partly stems from key limitations of foundation features inherited by these methods. First, foundation features often appear similar in terms of vector distance for symmetrical parts, such as the left and right legs of a chair, as reported by many studies [34, 53]. While this property benefits semantic understanding tasks, where both legs can be grouped semantically, our task critically relies on precise differentiation between such parts. Second, these features are sensitive to texture variations [5, 52], making matching less consistent, especially for textureless models common in online collections. + +To this end, we introduce a technique to enhance foundation features and integrate them into a novel 3D alignment pipeline with a coarse-to-fine estimation scheme: The first step estimates a coarse 9-DoF pose by utilizing a new geometry-aware feature space, derived from DINOv2 features, that is more robust to object part symmetries. The second step refines the pose through dense alignment optimization in a texture-invariant space called the Normalized Object Coordinate (NOC) [48], for which we also propose a new NOC estimator that generalizes better. Unlike prior work, which requires real or synthetic training scenes with pose-annotated objects, our pipeline uses only easily accessible front-aligned CAD models for supervision. + +In coarse alignment, object pixels and 3D model parts are encoded into a shared feature space, where correspondences are found via nearest neighbors and used to estimate pose with least squares. The key challenge is designing an effective feature space and encoder. Our solution trains a small feature adapter network that converts foundation features, computed from an image or a 3D model rendering, into custom features. This network enforces multiview consistency, ensuring features for the same part remain similar across views, while distinguishing features for symmetrical parts that are not well separated in the foundation feature space. Leveraging direct access to CAD models [7], we formulate these objectives into a self-supervised triplet loss [43]. This new feature space improves geometric awareness while allowing useful semantics in the foundation features to be retained. + +In fine alignment, we use dense image-based alignment to optimize the 3D pose by matching the 3D model's rendering to the input image. However, instead of comparing in RGB space, which is impractical due to mismatched tex + +ture, we convert both the input and the rendering into NOC maps [4, 48] for comparison. These maps assign pixels from the same object part to a shared normalized 3D coordinate, allowing direct matching. To predict NOC maps, we leverage our feature space and perform nearest neighbor matching, as in coarse alignment. Since nearest neighbor matching is invariant to global scaling and shifting in the feature space, our NOC maps can offer improved robustness to domain gaps and have been found to generalize better to real-world images, even outperforming direct NOC regressors trained on the same synthetic renderings. + +We evaluate our method on ScanNet [11] and outperform weakly supervised 9-DoF SOTA, DiffCAD [16], by $+4.2\%$ in mean alignment accuracy, the supervised baseline, ROCA [18], by $+2.7\%$ , and an adapted Foundation-Pose [49] in the inexact 6-DoF setting by $+6.9\%$ . Our features improve DINoV2 by $+4.4\%$ , while NOC optimization refines coarse poses by $+4.1\%$ . We also introduce SUN2CAD, an inexact 9-DoF test set with 20 unseen categories, where we surpass the supervised SOTA, SPARC [26], and weakly supervised baselines, achieving state-of-the-art generalization with a large margin $+12.7\%$ . + +To summarize, our contributions are: + +- A zero-shot single-view 3D alignment approach that handles inexact 3D model matches with state-of-the-art generalization to unseen categories. +- A technique to enhance foundation features with object geometry and part symmetry awareness, enabling better 2D-3D part matching. +- A technique for predicting NOC maps that enables texture-invariant, smooth dense alignment optimization for fine-grained pose estimation. +- A new 9-DoF alignment benchmark, SUN2CAD, featuring 20 new categories not seen in any existing benchmark. + +# 2. Related Work + +Single-view CAD Model Alignment. Pose estimator frameworks often integrate multiple techniques in a coarse-to-fine manner, starting with a rough pose estimate and refining it for accuracy [28]. Template matching [1, 25, 35] finds an initial pose by comparing the image with multiview rendering templates. Alternatively, an initial pose can be derived through transformation estimation [47] using sparse [21] or dense [9, 44, 48] 2D-3D correspondences. Pose regression can be used for both initial prediction [18] and refinement [27], but these methods require extensive pose annotations and are limited to known instances or categories. Recent studies [5, 37] address this by exploring training-free pose estimators that leverage features from foundation models [36]. However, these pipelines require exact-match CAD models for texture-dependent refinement, limiting their use with retrieved or user-selected inexact-match CAD models. In contrast, model-free meth + +ods [46, 49] employ neural implicit representations to generate both 3D objects and poses but require multiple reference views and often produce lower-quality 3D models. + +To achieve visual realism and support inexact models, supervised end-to-end CAD model retrieval and alignment methods, such as Mask2CAD and Patch2CAD, are introduced [23, 24]. These methods extract features from input images and CAD models by training a feature decoder for shared image-CAD space and pose regression. Instead of using only feature vectors, ROCA [18] predicts dense 2D-3D correspondences to improve retrieval and pose estimation, while SPARC [26] refines pose predictions of retrieved CAD models via iterative feature learning. However, these studies are still limited by the availability of object categories and pose annotations. To address limitations in annotations, DiffCAD [16] explores training a retrieve-and-align framework on photorealistic scene renderings [15] using a probabilistic diffusion model. However, the domain gap hinders accuracy on real images, making it necessary to generate multiple outputs to reliably achieve accurate poses. Plus, the availability of photorealistic scene renderings remains limited to a few categories. In contrast, our method supports handling inexact 3D models and requires only synthetic renderings without pose annotations to train, enabling zero-shot pose estimation for unseen categories. + +Vision foundation models. Recent advances in computer vision have introduced foundation models trained on large-scale datasets, enabling them to generalize to new tasks without needing task-specific data. These models serve as either feature encoders or task-specific tools, such as for object detection [22, 30], depth estimation [50, 51], or image generation [38, 42]. Feature encoders, such as DINO [6, 36], are trained with self-supervised methods [8] on unlabeled data or on text-image pairs [39], providing rich zero-shot features or serving as reliable pre-trained weights. While DINO and SD [42] have been shown in multiple studies to be effective for zero-shot feature matching, they often lack geometric awareness [53] and fail to represent textureless objects. To address this, post-processing algorithms and feature adapters have been proposed to improve accuracy through labels or self-supervision based on 2D spatial location [34]. In our work, we adapt DINOv2 [36] for zero-shot 2D-3D matching and introduce a feature adapter trained on 3D renderings to enhance geometric awareness. + +# 3. Proposed Method + +Given an input image containing a target object, specified by a bounding box, and its closely matching CAD model, our goal is to predict the model's 9-DoF pose in the camera coordinate system, parameterized by a 6D rigid transformation $\mathbf{T} \in \mathrm{SE}(3)$ and an anisotropic scaling vector $\mathbf{s} \in \mathbb{R}^3$ . The matching model can be provided by the user or obtained from a retrieval system. In our work, we use ROCA's + +retrieval system [18]. We assume known camera intrinsics. + +This setup challenges traditional CAD alignment techniques [18, 23, 24, 26] as well as recent methods based on foundation features [5, 37], since the retrieved model may differ in shape and texture from the object in the input image. To solve this, we propose a novel coarse-to-fine pose estimation method, outlined in Figure 2. We first predict an initial pose by finding 2D-3D correspondence using a novel geometry-aware feature space (Section 3.1). Then, we refine this pose through dense alignment (Section 3.2). + +# 3.1. Coarse Alignment + +To perform coarse alignment, we encode the target object image into a 2D feature map and the CAD model into a 3D feature voxel grid within a shared feature space, then use nearest-neighbors matching to establish 2D-3D correspondences for pose estimation. This section first explains how we derive the shared feature space and train its encoder, then how to encode images and 3D models, and finally, how to establish correspondences and solve for pose. + +# 3.1.1. Geometry-aware feature space + +A recent trend in zero-shot correspondence matching repurposes foundation models like DINOv2 [36] as image encoders that output 2D feature maps. This approach has been applied to related tasks, such as 6-DoF pose estimation with an exact model shape and texture [5, 37]. However, these features are found to be sensitive to texture variations [52], which is problematic in our setting where the CAD model may lack matching textures or even be textureless. Furthermore, these features do not prioritize 3D part differentiation or accurately capture geometry and spatial locations, making it difficult to distinguish symmetrical parts, such as chair legs or car wheels. Nonetheless, we hypothesize that these features already contain latent information for predicting part locations, which could be leveraged if restructured. + +To this end, we propose training a light-weight MLP feature adapter, $E_{\theta}$ , to transform DINO's feature map into a geometry-aware feature map. This adapter is trained on synthetic renderings of CAD models from ShapeNet [7] with augmentations. In this training set, each CAD model is first aligned to a canonical pose, scaled to fit within a unit cube, and rendered from multiple views using an orbiting camera. For augmentations, each rendering is transformed with a conditional diffusion model [55] to generate variations with added textures and backgrounds, which are added to the training set (Appendix 7). The training is done in an encoder-decoder scheme, with $E_{\theta}$ as the encoder and another MLP, $D_{\phi}$ , as the decoder, using two objectives. + +NOC prediction loss. Our first objective encourages our feature to be predictive of 3D locations by solving a coordinate prediction task. Specifically, the decoder's output should predict the Normalized Object Coordinates (NOC) + +![](images/d12274f2649a5aac60e86b40f1d26b99a71278be64b114bbaf131a211de95701.jpg) +Figure 2. Method Overview. From an image-3D model pair, we first construct 2D and 3D feature grids from DINOv2 and our geometry-aware adapter, trained with 3D self-supervision, then use nearest-neighbor matching to establish correspondences for initial pose solving. Finally, we refine the pose through dense alignment between the predicted NOC map and the 3D model's rendered NOC map. + +map [48] for each CAD rendering, where each pixel in this map encodes the 3D coordinate of the corresponding point on the model. Let $\mathbf{R}_i$ be a CAD rendering and $\mathbf{N}_i$ its corresponding NOC map, downsampled to the spatial dimension $h\times w$ of the decoder's output. We minimize: + +$$ +\mathcal {L} _ {\mathrm {N O C}} = \frac {1}{n \cdot h \cdot w} \sum_ {i = 1} ^ {n} \left\| D _ {\phi} \left(E _ {\theta} (\operatorname {D I N O} \left(\mathbf {R} _ {i}\right))\right) - \mathbf {N} _ {i} \right\| _ {2} ^ {2}, \tag {1} +$$ + +where DINO outputs a 2D feature map from its penultimate layer [52], and the MLP architectures of $E_{\theta}$ and $D_{\phi}$ process each pixel in the feature map independently. We also experimented with architectures that incorporate spatial context, like CNNs and Transformers, but their inductive biases often caused overfitting on limited data—e.g., over-relying on memorized overall shapes rather than utilizing individual DINO features at each pixel to predict NOCs. + +Geometry-consistent triplet loss. Another objective ensures that features for the same part appear consistent across viewing angles, while features for parts that are distant in 3D space are distinct. We achieve both tasks with a triplet loss. For each anchor point on a training CAD model: + +Its positive set gathers model points within a small Euclidean distance of $\tau_{\mathrm{dist}}^{+}$ from the anchor. + +Its negative set gathers model points that are at least $\tau_{\mathrm{dist}}^{-}$ away from the anchor and whose feature vectors have a cosine similarity above $\tau_{\mathrm{feat}}^{-}$ with the anchor's feature. + +We sample various anchors from various models, along with their positive and negative samples, forming (anchor a, positive $\mathbf{p}$ , negative $\mathbf{n}$ ) into a set $\mathcal{T}$ , then compute: + +$$ +\mathcal {L} _ {\text {t r i p l e t}} = \frac {1}{| \mathcal {T} |} \sum_ {(\mathbf {a}, \mathbf {p}, \mathbf {n}) \in \mathcal {T}} [ d (\mathbf {a}, \mathbf {n}) - d (\mathbf {a}, \mathbf {p}) + \alpha ] _ {+}, \tag {2} +$$ + +where $d(\mathbf{x},\mathbf{y})$ computes the cosine similarity between the encoded features of model points $\mathbf{x}$ and $\mathbf{y}$ , which may come from different rendering views of the same model, and $\alpha$ is a margin threshold used in standard triplet training [43]. + +Training. We train both $E_{\theta}$ and $D_{\phi}$ , which is later discarded, using the combined loss: $\mathcal{L}_{\mathrm{adapter}} = (1 - \beta)\mathcal{L}_{\mathrm{NOC}} + \beta \mathcal{L}_{\mathrm{triplet}}$ , where $\beta$ balances the two objectives. These objectives enable $E_{\theta}$ to become more geometry-aware, as illustrated in Fig. 3, mitigating the 3D perception limitations of prior methods without relying on annotated real datasets. + +# 3.1.2. Encoding images & 3D models into shared space + +While the feature from $E_{\theta}$ enhances geometry-awareness, we found it beneficial to concatenate it with the original DINOv2 features for improved matching. In particular, the final feature $E_{f}(\mathbf{I}) = (1 - \omega)\cdot \hat{\mathrm{DINO}} (\mathbf{I})\oplus \omega \cdot \hat{E}_{\theta}(\mathrm{DINO}(\mathbf{I}))$ , where $\hat{\cdot}$ denotes normalization of the output to unit length. + +Given $E_{f}$ , encoding the input image $\mathbf{I}$ into a 2D feature map is simply $E_{f}(\mathbf{I})$ . For the retrieved CAD model, we apply the same function to its multi-view renderings, yielding $E_{f}(\mathbf{R}_{i})$ for each $\mathbf{R}_{i}$ . Using each $\mathbf{R}_{i}$ 's z-buffer and known extrinsics, these features are back-projected into the same 3D world space and stored in a lower-resolution voxel grid (with many voxels empty due to background areas or self-occlusion). Finally, we average across the grids from all $\mathbf{R}_{i}$ to produce a single feature voxel grid. We further apply simple smoothing to the voxel grid, detailed in Appendix 6. + +# 3.1.3. Feature matching and pose solving. + +For matching, since $E_{f}(\mathbf{I})$ computes features for the entire input image, including unwanted background elements, we first use SAM [22] to segment the object from the given bounding box, then consider only features within the object mask. For each of these features, we find the closest feature in the voxel grid in terms of cosine similarity, resulting in a set of correspondences between 2D feature coordinates and 3D voxel coordinates for initial pose estimation. + +Given such correspondences, it is possible to solve for pose using standard techniques such as PnP algorithms [17]. However, this approach can be quite sensitive to scene scale ambiguity. Instead, we leverage a recent metric depth estimator [50] to predict the depth map of the input image, + +![](images/d0bb59021b4be024c8dcb07f5c3185c39299ac8639f4a1632529278f1e472670.jpg) +Figure 3. Visualization of our learned space. DINOv2 and our geometry-aware features are dimensionally reduced with PCA and color-coded by NOC (see Appendix 6.7). Nearby object parts are more clearly separated in our features compared to DINOv2. + +which lifts the 2D coordinates of each feature at $(u, v)$ in the feature map to 3D $(u, v, d)$ . We then convert these to standard $(x, y, z)$ via back-projection and estimate pose from 3D-3D correspondences using a RANSAC-based pose estimator [14, 47]. The final coarse pose is estimated using all inlier correspondences, as is standard practice. + +In our experiments, we compare our method to an alternative that directly predicts NOC for 2D-3D correspondence. Specifically, each feature in the 2D feature map can use $D_{\phi}(E_{\theta}(\mathbf{I}))$ , which already encodes normalized 3D coordinates for correspondence. We found that our nearest neighbor matching is more robust to distribution shifts, which are particularly significant when training on synthetic data and testing on real images, compared to directly using the NOC output from a neural network, which can be less predictable with out-of-distribution images. + +# 3.2. Dense Image-based Alignment + +This step aims to refine the coarse pose using dense image-based alignment, which traditionally optimizes the pose parameters so that the rendering matches the input. To increase invariance to object textures, recent methods like [5, 37] perform the comparison in feature space instead, using DINO's features. While replacing DINO's features with our own in this refinement step readily improves performance (see Section 4.4), we instead propose performing the comparison in NOC space, along with depth and mask spaces, for two reasons. First, each optimization step requires rendering the 3D model in the target space, and rendering it as a NOC map requires only a rasterizer, as opposed to a costly network inference, and consistently yields smooth images. These smooth images help provide stable gradients in dense alignment, where every pixel is consid- + +ered. Second, a NOC map of the input image is inferred only once using nearest neighbor matching in our feature space, providing efficiency, robustness to global shifting and scaling in the feature space, and texture invariance. + +To convert the input image $\mathbf{I}$ to a NOC map, denoted by $\mathbf{N}^{\mathbf{I}}$ , we use the same process as before: first computing a feature map $E_{f}(\mathbf{I})$ and link each feature to its closest 3D voxel, whose location already represents normalized object coordinates (NOC). The $h\times w$ NOC map is upsampled to the $H\times W$ input size via bilinear interpolation. At optimization step $t$ , we render the posed 3D model into a NOC map, denoted by $\mathbf{N}^t$ , using a differentiable renderer [40]. The total loss consists of $\lambda_{\mathrm{NOC - A}}\mathcal{L}_{\mathrm{NOC - A}}(t) + \lambda_{\mathrm{m}}\mathcal{L}_{\mathrm{mask}}(t) + \lambda_{\mathrm{d}}\mathcal{L}_{\mathrm{depth}}(t)$ , where $\lambda_{(\cdot)}$ are balancing weights. + +NOC alignment loss ensures that the predicted NOC matches the rendered NOC from the posed 3D model: + +$$ +\mathcal {L} _ {\mathrm {N O C - A}} (t) = \frac {1}{m} \left\| \mathbf {M} \odot \left(\mathbf {N} ^ {\mathbf {I}} - \mathbf {N} ^ {t}\right) \right\| _ {1}, \tag {3} +$$ + +where $\odot$ is element-wise multiplication, $\mathbf{M}$ is a binary mask indicating overlapping pixels between $\mathbf{N}^{\mathbf{I}}$ and $\mathbf{N}^{t}$ , and $m$ is the count of these pixels. + +Silhouette loss ensures that the silhouettes of the 3D model and the input image match: + +$$ +\mathcal {L} _ {\text {m a s k}} (t) = \frac {1}{H W} \left\| \mathbf {S} ^ {\mathbf {I}} - \mathbf {S} ^ {t} \right\| _ {1}, \tag {4} +$$ + +where $\mathbf{S}^{\mathbf{I}}$ denotes the object mask computed from SAM (Section 3.1.3), and $\mathbf{S}^t$ is a differentiable soft mask of the 3D model rendered using SoftRasterizer [29, 40]. + +Depth loss ensures consistency between the depth rendered from the model and the predicted metric depth from the input image. Unlike NOC and silhouette losses, which focus on 2D projection and are invariant to depth scaling (e.g., enlarging and moving objects away from the camera), this loss is crucial for predicting overall scale and translation, especially along the camera's forward axis. + +$$ +\mathcal {L} _ {\mathrm {d e p t h}} (t) = \frac {1}{m} \left\| \mathbf {M} \odot (\mathbf {D} ^ {\mathbf {I}} - \mathbf {D} ^ {t}) \right\| _ {1}, \tag {5} +$$ + +where $\mathbf{D}^{\mathrm{I}}$ is the predicted metric depth map from the input image, and $\mathbf{D}^t$ is the rendered depth map. We minimize the total loss via gradient descent by backpropagating through differentiable renderings to obtain the final pose estimate. + +Our use of the three spaces for pose estimation shares some similarities with prior work [4]. However, they apply these losses for hypothesis sampling, whereas we use them for dense, differentiable pose optimization. + +# 4. Experiments + +We compare our method against two supervised baselines, SPARC [26] and ROCA [18], and two weakly supervised baselines, DiffCAD [16] and FoundationPose [49], on + +
Dataset / MetricPoseSupMethodBathtubBedBinBkshlfCabinetChairDisplaySofaTableAvg Cat.↑Avg Inst.↑
ScanNet25kNMS alignment accuracy [33]9D✓ROCA [18]22.510.029.314.215.841.030.415.914.621.527.4
✓SPARC [26]26.725.726.717.523.852.622.532.717.727.333.9
★FoundationPose [49] (for 9D)20.022.927.60.93.141.823.615.017.519.225.7
★Ours16.718.622.812.79.249.324.138.116.523.130.1
6D★FoundationPose [49]22.521.437.56.15.844.530.429.227.124.931.1
★Ours (for 6D)20.825.727.619.822.756.151.845.120.132.238.0
ScanNet25k(DiffCAD's split)Single-viewaccuracy. [16]9D★DiffCAD (GT) [16]-27.1-24.433.065.9-46.318.335.841.9
★DiffCAD (Mean) [16]-5.2-1.15.125.2-6.51.97.511.7
★DiffCAD (Err) [16]-7.7-7.610.431.1-15.34.312.716.7
★Ours-2.8-8.212.041.3-21.49.815.920.9
Ablation StudyCoarseFine
ScanNet25kNMS alignment accuracy [33]9D-DINOv2-14.25.76.08.56.934.39.924.87.213.118.8
-DINOv2FM14.25.79.19.07.333.99.423.07.413.218.8
-DINOv2Ours16.710.013.87.19.042.613.631.910.517.324.2
★NOC-S-5.01.41.34.23.534.017.812.40.58.915.9
★NOC-SOurs5.87.11.75.76.538.229.827.42.513.919.9
★Ours (ω=1)-15.84.36.010.86.238.517.323.07.614.421.0
★Ours (ω=1)Ours16.711.421.611.86.241.517.323.910.317.924.3
★Ours-16.710.010.815.17.346.816.231.010.718.326.0
★OursFM17.511.410.315.17.746.115.730.311.018.326.1
★OursOurs16.718.622.812.79.249.324.138.116.523.130.1
+ +Table 1. Comparison on ScanNet25k [2] & Ablation study. We compare across 3 groups: 9-DoF, 6-DoF, and DiffCAD's split (Section 4.1 and 4.2), where our method outperforms all non-supervised baselines in mean accuracy. Supervised, weakly supervised, and unsupervised baselines are marked with $\checkmark$ , $\star$ , and $-$ , respectively. We ablate coarse and fine alignment alternatives, detailed in Section 4.4. + +
PoseSupMethodbasketbicycleblenderbroomclockcofflmkrorchfireextkeybrdladderlampmugpianoprinterremoteshoephoneovenvasebottleCat.↑Inst.↑
#7#14#7#2#13#19#18#15#66#4#132#59#18#92#8#3#18#92#8#3#47#14#9#3#20#550
9D✓SPARC [26]14.37.10.050.00.00.00.00.00.00.00.00.00.03.00.027.814.10.00.00.00.00.00.00.00.00.00.00.022.20.00.06.94.9
+ +Table 2. Comparison on unseen categories in SUN2CAD. We report single-view accuracy [16], and achieve the highest mean accuracies. + +inexact-CAD pose estimation. SPARC is the current SOTA in 9-DoF pose refinement; however; it requires category-specific median scaling values, which are unavailable in other competitors' settings. Thus, we also include ROCA, the second-best supervised 9-DoF baseline, which predicts pose from scratch, similar to our setting. Among weakly supervised baselines, DiffCAD is the SOTA in 9-DoF alignment but is limited to training categories. In contrast, FoundationPose represents the best open-source pose estimator that supports unseen categories and textures but estimates only 6-DoF poses without 3-DoF scaling. + +Due to differences in setup and requirements, not every baseline can be evaluated in all scenarios: We compare against SPARC, ROCA, and FoundationPose on ScanNet25K [11] (Section 4.1); DiffCAD on its own test subset of ScanNet25K (Section 4.2); SPARC and FoundationPose on our new SUN2CAD dataset with unseen categories. Finally, we conduct ablation studies on our coarse-to-fine pipeline and hyperparameters (Sections 4.4 and 4.5). See Appendices 10-13 for additional results and failure cases. + +Implementation details. Our feature adapter $E_{\theta}$ uses a 2-layer MLP, and the decoder $D_{\phi}$ uses a 1-layer MLP. Both are trained from scratch using the AdamW optimizer [31]. For training data, following Scan2CAD [2], we use 9 cate + +gories in ShapeNet [7] dataset to render and augment templates, resulting in 300k images. We use DINOv2- ViT-L [36] to create input feature maps with the size of 1024 per patch. The dense alignment is implemented using PyTorch3D [40] differentiable rendering and optimized with Adam. We use DepthAnything [50] as a metric depth estimator, which was fine-tuned on the training data of each benchmark dataset. All experiments were conducted on a single V100 GPU. See Appendix 6 for additional details. + +# 4.1. Comparison on ScanNet25k Dataset + +Following previous studies [18, 23, 26], we evaluate performance on ScanNet25k [11] using pose annotations from Scan2CAD [2] and the standard accuracy metric [18, 26, 33]. The dataset provides scene videos, where each object appears in multiple frames, totaling $20\mathrm{K}$ training and $5\mathrm{K}$ test images. The accuracy metric first applies nonmaximum suppression (NMS) [33] to select the pose prediction with the best ROCA's retrieval confidence score across frames, then computes the percentage of predictions with translation, rotation, and scaling errors of $\leq 20\mathrm{cm}, \leq 20^{\circ}$ , and $\leq 20\%$ relative to the ground truth. Despite being designed for video-based alignment [33], this NMS-based metric is adopted by single-view methods [18, 26]. + +9-DoF ROCA [18] and SPARC [26]: As seen in Table 1, our method is the only weakly supervised method to surpass ROCA [18] in the average category-wise and instance-wise NMS scores by $\{+2.7\%, +1.6\}$ . SPARC [26] achieves higher average scores but lower scores on Display and Sofa. Note that SPARC requires per-category median scales for pose initialization, which alone yield $65.2\%$ scaling accuracy without further processing, and cannot handle unseen categories (Table 2). We observe that our underperforming categories (e.g., table, bathtub, bed) often face large occlusions and cropping, affecting scale and rotation estimates, whereas supervised methods can be less impacted by learning from similar distributions. (Appendix 10). + +6-DoF FoundationPose [49]: For a fair comparison, we evaluate under two setups, shown in Table 1: (1) An inexact 6-DoF estimation task, which is closest to the FoundationPose's original setup. In this setup, we adapt our method by fixing the scales to the ground truth values used by FoundationPose. (2) An inexact 9-DoF estimation task, where we adapt FoundationPose and provide it with the scales predicted by ROCA (76.8% accuracy in ScanNet25k [11]). We provide these setups for reference only, as their method is not designed for 9-DoF or inexact match tasks. + +The adapted 6-DoF version of our method outperforms FoundationPose in 6 out of 9 classes ( $+7.3\%, +6.9\%$ ) mean accuracies), and our method also outperform its adapted 9-DoF version in 5 out of 9 classes. We observe that their method struggles with rotation in large objects, such as bookshelves and cabinets, which differ from the smaller objects FoundationPose was trained on. + +# 4.2. Comparison on ScanNet25k—DiffCAD's split + +We compare our method to the SOTA 9-DoF weakly supervised method, DiffCAD [16] in Table 1. DiffCAD was trained on six categories from 3D-FRONT [15] and tested on their own subset of ScanNet25K with the same categories in their paper. DiffCAD's metric does not apply NMS as used with ScanNet25K and involves generating multiple hypotheses and uses the one closest to the ground truth to report scores (DiffCAD(GT) in Table 1; reproduced from official code for reference only). However, our problem setup does not assume access to the ground-truth pose for such hypothesis selection, thus we instead compute two other metrics. (1) DiffCAD (Err), which selects the best hypothesis based on projection errors from its 2D-3D correspondences and the solved pose. (2) DiffCAD (Mean), which computes the mean error across all hypotheses. For a fair comparison, our method uses a non-fine-tuned depth estimator since DiffCAD was not trained on ScanNet's depth maps, and we use DiffCAD's own test set for evaluation. + +Our method surpasses both DiffCAD (Err) and DiffCAD (Mean) in 5 of 6 classes, achieving $\{+3.3\%, +4.3\%$ and $\{+8.4\%, +9.2\%$ mean accuracies. + +![](images/e0c3b604791dc04b656d772aeb8078d8e19e1b6308abe91e7462832ad1299265.jpg) + +![](images/625759280e2a05d72b4b6617326a1d61d420c7ac0b2a540ee818fde3da771d96.jpg) +Figure 4. Qualitative results in ScanNet25k dataset. + +# 4.3. Comparison on SUN2CAD Dataset + +To assess zero-shot capability on unseen categories, we introduce SUN2CAD, a new inexact 9-DoF test set that extends beyond the nine categories in ScanNet25k [11]. SUN2CAD comprises 550 samples across 20 categories, featuring diverse shapes and sizes, including bicycles, pianos, fire extinguishers, and lamps. Annotations are derived from 3D bounding boxes in SUN RGB-D [45] and further refined through manual adjustments to improve accuracy. + +Table 2 and Fig. 5 compare our method with competitors. Note that we can not evaluate against DiffCAD [16] because its category-specific pose estimators are incompatible with unseen categories. While SPARC [26] performs well in seen categories, it struggles to generalize to SUN2CAD (see Appendix 10 for further analysis). Compared to DINOv2, known for its zero-shot semantic understanding [36], our method achieves superior results, suggesting zero-shot capabilities of geometry-aware features. We outperform SPARC in 18/20 and DINOv2 in 14/20 categories, with mean accuracy gains of $\{+17.6\%, +12.7\%$ and $\{+12.9\%, +10.3\%\}$ , respectively. In the inexact 6-DoF setting, where ground-truth scales are provided to all competitors, we surpass FoundationPose [49], which is specialized in novel but exact object pose estimation, in 13/20 categories and $\{+14.2\%, +10.3\%$ gains. + +# 4.4. Ablation Study on Coarse-to-fine Pipeline + +We evaluate our pipeline against combinations of coarse and fine alignment alternatives in Table 1. + +Coarse alignment. We test three alternatives: (1) Using + +DINOv2 [36] features for nearest neighbor matching instead of our fused geometry-aware features (Section 3.1.3). (2) Using only geometry-aware features without DINOv2 $(\omega = 1)$ . (3) Predicting NOC maps directly from DINOv2 features using a neural network trained on the same synthetic data as our method (denoted by NOC-S in Table 1). See Appendix 12.3 for additional details. + +Fine alignment. We test two alternatives: (1) No refinement. (2) Performing dense alignment using features used in the coarse step, as proposed by FoundPose [37] (denoted as FM in the "Fine" column). Note that this is a reimplementation, as FoundPose does not provide source code. + +The results using only coarse alignment show that our fused features (Ours, -) improve accuracy from using pure DINOv2 features (DINOv2, -) by $\{+5.2\%, +7.3\}$ . Pure adapted features (Ours ( $\omega = 1$ ), -) also outperform DINOv2 but underperform our fused version. When all methods use fine alignment, (Ours, Ours) clearly outperforms (DINOv2, Ours) and surpasses (NOC-S, Ours) in 8 out of 9 classes and in mean accuracy, highlighting the effectiveness of our geometry-aware features. For fine alignment, our NOC dense optimization improves accuracy compared to both using no refinement $\{+4.8\%, +4.1\}$ and using dense alignment based on features ( $\cdot$ , FM) by $\{+4.8\%, +4.0\}$ . + +![](images/b797efe470d9a208fffb813621ff3460ad5563ed88027893a60b0a7b0aa70318.jpg) +Figure 5. Qualitative results in SUN2CAD dataset. + +# 4.5. Ablation Study on Hyperparameters + +Feature adapter loss study. The impact of $\beta$ for balancing feature adapter loss in Eq 1 and Eq 2 is shown in Fig. 6, + +
Dense alignment losses9 CategoriesAccuracy per parameter ↑
+L_NOC-A+L_Mask+L_DepthCat. ↑Inst. ↑Tr ↑Sc ↑Ro ↑
---19.3123.9344.8944.0548.18
--✓20.4724.8745.1846.9249.72
-✓-15.0118.9236.9548.3336.62
✓--18.4523.7641.4648.6348.90
-✓✓22.2726.5745.8649.5650.44
✓✓-16.8721.7739.6949.0241.56
✓-✓22.5826.9545.8651.8852.40
✓✓✓23.6028.3646.6752.7352.49
+ +Table 3. Ablation study in dense image-based alignment loss + +where a weight of 0.1 yields the best NOC error. + +Feature fusion. We test $\omega$ for fusing DINOv2 with our geometry-aware features and find that $\omega = 0.5$ minimizes NOC error (Fig. 6), balancing both feature types. + +![](images/20e4e0c08a5fa294ac5465ec43566fe492eb2f762a11b2caba59806288f0745e.jpg) +Figure 6. Ablation studies on feature adapter losses (upper) and feature fusion weights (below). + +Dense alignment loss study. Table 3 presents the impact of each loss term. $\mathcal{L}_{\mathrm{NOC - A}}$ (Eq 3) improves scaling and rotation accuracy but degrades translation, while $\mathcal{L}_{\mathrm{depth}}$ (Eq 5) enhances translation and rotation. Combining both losses improves all three parameters, and incorporating silhouette information from $\mathcal{L}_{\mathrm{mask}}$ yields the best results. + +# 5. Conclusion + +We propose a zero-shot 9-DoF pose estimation method to align an inexact 3D model with its object in a single image. Our method enhances a foundation feature space to be geometry-aware, mitigating ambiguities from symmetrical parts, and leverages it to enable initial pose estimation via simple nearest-neighbor matching. This space also facilitates pose refinement via dense alignment in a normalized coordinate space, which enhances robustness to texture mismatches and geometric differences between the input image and the 3D model. Our method is fully self-supervised, requiring only a lightweight adapter trained on a small number of ScanNet25K categories, with no scene-level pose annotations. On nine seen categories from ScanNet25K, it trails the state-of-the-art supervised SPARC $(-3.8\%)$ but outperforms weakly supervised competitors and the supervised ROCA $(+4.2\%, +2.7\%)$ . However, on the newly introduced SUN2CAD dataset with 20 unseen categories, it achieves state-of-the-art generalization with a large margin $(+12.7\%)$ , demonstrating its potential for scalable, category-agnostic alignment. + +# References + +[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 463-469. IEEE, 2024. 2 +[2] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 1, 6, 4, 8 +[3] Alexey Bochkovskiy, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. In The Thirteenth International Conference on Learning Representations, 2025. 2 +[4] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 536-551. Springer, 2014. 2, 5 +[5] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. ECCV, 2024. 2, 3, 5 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630-9640, 2021. 3 +[7] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, L. Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. ArXiv, abs/1512.03012, 2015. 2, 3, 6, 1, 4 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 3 +[9] Jaime Corsetti, Davide Boscaini, Changjae Oh, Andrea Cavallaro, and Fabio Poiesi. Open-vocabulary object 6d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18071-18080, 2024. 2 +[10] Camille Couprie, Clément Farabet, Laurent Najman, and Yann Lecun. Indoor semantic segmentation using depth information. In First International Conference on Learning Representations (ICLR 2013), pages 1-8, 2013. 2 +[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2432-2443, 2017. 1, 2, 6, 7, 3, 4, 5 + +[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 4 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021, 2021. 1, 7, 8 +[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5 +[15] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 1, 3, 7 +[16] Daoyi Gao, David Rozenberszki, Stefan Leutenegger, and Angela Dai. Diffcad: Weakly-supervised probabilistic cad model retrieval and alignment from an rgb image. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 1, 2, 3, 5, 6, 7, 4, 8, 11, 12, 13, 14 +[17] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930–943, 2003. 4 +[18] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4022-4031, 2022. 1, 2, 3, 5, 6, 7, 4, 12, 13, 14 +[19] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 4 +[20] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[21] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10095-10105, 2024. 2 +[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. 2023 IEEE/CVF In + +ternational Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 3, 4 +[23] Weicheng Kuo, Anelia Angelova, Tsung-Yi Lin, and Angela Dai. Mask2cad: 3d shape prediction by learning to segment and retrieve. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 260-277. Springer, 2020. 1, 3, 6 +[24] Weicheng Kuo, Anelia Angelova, Tsung-Yi Lin, and Angela Dai. Patch2cad: Patchwise embedding learning for in-the-wild shape retrieval from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12589-12599, 2021. 1, 3 +[25] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapore: 6d pose estimation of novel objects via render and compare. In Proceedings of The 6th Conference on Robot Learning, pages 715-725. PMLR, 2023. 1, 2 +[26] F. Langer, G. Bae, I. Budvytis, and R. Cipolla. Sparc: Sparse render-and-compare for cad model alignment in a single rgb image. In Proc. British Machine Vision Conference, London, 2022. 1, 2, 3, 5, 6, 7, 4, 12, 13, 14, 15, 16, 17 +[27] Yuelong Li, Yafei Mao, Raja Bala, and Sunil Hadap. Mrcnet: 6-dof pose estimation with multiscale residual correlation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 10476-10486. IEEE, 2024. 2 +[28] Jian Liu, Wei Sun, Hui Yang, Zhiwen Zeng, Chongpei Liu, Jin Zheng, Xingyu Liu, Hossein Rahmani, Nicu Sebe, and Ajmal Mian. Deep learning-based object pose estimation: A comprehensive survey. arXiv preprint arXiv:2405.07801, 2024. 2 +[29] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. The IEEE International Conference on Computer Vision (ICCV), 2019. 5 +[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chun yue Li, Jianwei Yang, Hang Su, Jun-Juan Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ArXiv, abs/2303.05499, 2023. 3 +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 6, 1, 9 +[32] Wufei Ma, Qihao Liu, Jiahao Wang, Angtian Wang, Xiaoding Yuan, Yi Zhang, Zihao Xiao, Guofeng Zhang, Beijia Lu, Ruxiao Duan, et al. Generating images with 3d annotations using diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2 +[33] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Vid2cad: Cad model alignment using multiview constraints from videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:1320-1327, 2020. 6, 8 +[34] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided + +spherical maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521-19530, 2024. 2, 3 +[35] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9903-9913, 2024. 2 +[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2, 3, 6, 7, 8, 4 +[37] Evin Pinar Örnek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodač. Found- pose: Unseen object pose estimation with foundation features. European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 5, 8 +[38] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *SDXL: Improving latent diffusion models for high-resolution image synthesis*. In *The Twelfth International Conference on Learning Representations*, 2024. 3 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 3 +[40] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 5, 6, 3 +[41] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 4 +[42] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 3, 2, 7, 8 +[43] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2, 4 +[44] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE transactions on pattern analysis and machine intelligence, 44(11):7417-7435, 2021. 2 + +[45] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 7, 1, 3 +[46] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825-6834, 2022. 3 +[47] Shinji Umeyama. An eigendecomposition approach to weighted graph matching problems. IEEE transactions on pattern analysis and machine intelligence, 10(5):695-703, 1988. 2, 5 +[48] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 2, 4 +[49] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868-17879, 2024. 1, 2, 3, 5, 6, 7, 4, 8, 12, 13, 14, 15, 16, 17 +[50] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371-10381, 2024. 3, 4, 6, 1, 2 +[51] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. NeurIPS, 2024. 3 +[52] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements DINO for zero-shot semantic correspondence. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 3, 4 +[53] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3076-3085, 2023. 2, 3 +[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[55] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3, 2 \ No newline at end of file diff --git a/zeroshotinexactcadmodelalignmentfromasingleimage/images.zip b/zeroshotinexactcadmodelalignmentfromasingleimage/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4d7095da152490c2028896180be1f1822d0d8f70 --- /dev/null +++ b/zeroshotinexactcadmodelalignmentfromasingleimage/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f674da9ae967b5d4d23938e0517ede019bd1c6a1f13f83e72adc5ff4c644fc5 +size 710601 diff --git a/zeroshotinexactcadmodelalignmentfromasingleimage/layout.json b/zeroshotinexactcadmodelalignmentfromasingleimage/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a7d4821520ac2445d82876818838d188e526289b --- /dev/null +++ b/zeroshotinexactcadmodelalignmentfromasingleimage/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0193f2e91e86682e4e34cef79c0ca5af06e855b47f544c5ad4bcd6031b1c144 +size 397877 diff --git a/zeroshotvisionencodergraftingviallmsurrogates/7a116134-5667-436c-b482-aee8bdfd9f65_content_list.json b/zeroshotvisionencodergraftingviallmsurrogates/7a116134-5667-436c-b482-aee8bdfd9f65_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..72a9c684b269f873f325bc4182e68c96498a3c8c --- /dev/null +++ b/zeroshotvisionencodergraftingviallmsurrogates/7a116134-5667-436c-b482-aee8bdfd9f65_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b795a6eecee91f2510528414f4ad95c7479cf3ecc0a08240bfd02d0cd2198b42 +size 96459 diff --git a/zeroshotvisionencodergraftingviallmsurrogates/7a116134-5667-436c-b482-aee8bdfd9f65_model.json b/zeroshotvisionencodergraftingviallmsurrogates/7a116134-5667-436c-b482-aee8bdfd9f65_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e7272323f3a8a1bfe8b3da8c9aa9040999425f3 --- /dev/null +++ b/zeroshotvisionencodergraftingviallmsurrogates/7a116134-5667-436c-b482-aee8bdfd9f65_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33ba7fa0b410ffbfe925099b95aa73abc4e5b104fd23ea6cb73dde2fd775ef51 +size 118686 diff --git a/zeroshotvisionencodergraftingviallmsurrogates/7a116134-5667-436c-b482-aee8bdfd9f65_origin.pdf b/zeroshotvisionencodergraftingviallmsurrogates/7a116134-5667-436c-b482-aee8bdfd9f65_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..79881bb89daa1be2cf33730d26134837a3efacd5 --- /dev/null +++ b/zeroshotvisionencodergraftingviallmsurrogates/7a116134-5667-436c-b482-aee8bdfd9f65_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d5b13058b9f0f34e396309f9da9cfafe44fa3685c5a6dbfa03f98c22dd14ed1 +size 9089911 diff --git a/zeroshotvisionencodergraftingviallmsurrogates/full.md b/zeroshotvisionencodergraftingviallmsurrogates/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7f437765f81bcd9c058f52b726396db1b93e9cb --- /dev/null +++ b/zeroshotvisionencodergraftingviallmsurrogates/full.md @@ -0,0 +1,422 @@ +# Zero-Shot Vision Encoder Grafting via LLM Surrogates + +Kaiyu Yue Vasu Singla Menglin Jia† John Kirchenbauer + +Rifaa Qadri Zikui Cai Abhinav Bhatele Furong Huang Tom Goldstein + +University of Maryland †Meta + +https://github.com/facebookresearch/zero + +![](images/8dc714c90c15bf40fcf42fd3ac41296e9a93d801d15b60fbe8e18fe2f4f60f41.jpg) +Figure 1. Zero-shot vision encoder grafting via a small language surrogate (srgt) model to trigger the target LLM to perform visual understanding task without any additional training. + +# Abstract + +Vision language models (VLMs) typically pair a modestly sized vision encoder with a large language model (LLM), e.g., Llama-70B, making the decoder the primary computational burden during training. To reduce costs, a potential promising strategy is to first train the vision encoder using a small language model before transferring it to the large one. We construct small "surrogate models" that share the same embedding space and representation language as the large target LLM by directly inheriting its shallow layers. Vision encoders trained on the surrogate can then be directly transferred to the larger model, a process we call zero-shot grafting $^1$ - when plugged directly into the full-size target LLM, the grafted pair surpasses the encoder-surrogate pair and, on some benchmarks, even performs on par with full decoder training with the target LLM. Furthermore, our surrogate training approach reduces overall VLM training costs by $\sim 45\%$ when using Llama-70B as the decoder. + +Most modern auto-regressive VLMs are built by extracting visual features from images using an encoder like CLIP [37] or SigLIP [47, 52], and placing these features into + +![](images/684afe00e6aa5ff35f975a42c6a4c08df647093600fa37cc1ba046c240f7e842.jpg) +Figure 2. Reducing full decoder training cost with our surrogate-trained encoder for Llama-70B in VLMs. Hollow $\circ$ indicates the average score of the surrogate-trained encoder on the left. + +![](images/d55866e3f1d981894283d1ec5d529914a453ade8b2b63fc26e81ed26bb97cc0e.jpg) + +the context window of an LLM. The image features must be aligned with the representation space of the LLM, and this is achieved by training the entire pipeline end-to-end. The cost of such training is often severely dominated by the language model. For example, plugging CLIP (approx 400M parameters) into Llama-70B [9] results in a pipeline where vision encoder training occupies almost none of the required memory and computation. + +In this paper, we explore methods of performing encoder alignment using relatively small lightweight language models, and transferring the resulting features to a large language model. We train small surrogate language models with the same representation space as a larger target LLM. After training the vision encoder on this small surrogate model, we can then transfer it to the larger model, either directly (grafting) or with fine-tuning. + +A major focus of our work is on understanding how to construct small surrogate models that accurately mock larger target LLMs. Our method of creating such small models stems from analyzing the internal prediction dynamics of LLMs, particularly how predictions evolve across layers. This analysis reveals two distinct phases in the prediction trajectory, separated by a clear transition point. We construct our small models by preserving the layers that participate in the early feature extraction phase of inference, and + +condensing all other layers. Since the small model inherits its shallow parameters from the target LLM, it shares the same embedding space as the original larger model and can effectively stand in as its surrogate. Our surrogate model has two major advantages: + +Zero-shot grafting capability. Vision features trained on a smaller and less resource-intensive surrogate can be directly used by the larger target LLM without any fine-tuning, as depicted in Figure 1. This zero-shot grafting demonstrates these surrogate-trained encoders effectively trigger visual understanding in target LLMs. + +Fast-converging VLM training. The encoders trained on surrogate models can be further fine-tuned with the full-size target LLM. Since they are already aligned with the LLM's embedding space, they achieve high performance with comparatively little full-scale training. Our experiments show a $\sim 45\%$ cost reduction for full decoder training with Llama70B, as shown in Figure 2, highlighting the efficiency of our surrogate-trained encoders. + +Table of Main Contents + +- Section 1: We detail the method of constructing our surrogate model, providing analysis that demonstrates how we discovered, developed, and validated our approach through experimental ablations. +- Section 2: We show our surrogate models for giant LLMs like Llama-70B, producing encoders with a strong zero-shot grafting ability, which can also accelerate the full decoder training of giant language models for VLMs. + +# 1. Building Surrogate Models + +In this section, we present our approach for building small surrogate models for target LLMs. First, we analyze the LLM's hidden features to identify the critical transition point between shallow and deep information processing layers. Next, we observe that the second/deep phase of inference contributes very little to encoder transferability, and observe that image features transfer well between models when they share their early/shallow processing layers. Finally, we validate these findings and propose to construct surrogate models by preserving the early-phase layers while replacing late-phase layers with a translator. + +# 1.1. Analyzing the Prediction Trajectory + +For a target LLM and input array of $N$ text token IDs $\mathbf{t}\in \mathbb{Z}^N$ , we trace the evolution of features over a forward pass of the model. By propagating these tokens through all $L$ transformer layers, we obtain intermediate hidden states $\mathbf{X}^{\ell}\in \mathbb{R}^{N\times D}$ from each layer, where $\ell \in [0,L - 1]$ denotes the + +layer index and $D$ is the hidden dimension. The final hidden states $\mathbf{X}^{L - 1}$ are passed through a normalization layer and the final linear layer $\mathbf{W}\in \mathbb{R}^{V\times D}$ to produce the logits, where $V$ is the vocabulary size. The probability distribution for the predicted next token can be computed for all positions: + +$$ +\mathbf {P} = \operatorname {s o f t m a x} \left(\operatorname {n o r m} \left(\mathbf {X} ^ {L - 1}\right) \mathbf {W} ^ {\top}\right) \in \mathbb {R} ^ {N \times V}. \tag {1} +$$ + +The probability for the next output token at each individual position is + +$$ +\mathbf {p} = \mathbf {P} [ : - 1, \mathbf {t} [ 1: ] ] \in \mathbb {R} ^ {N - 1}, \tag {2} +$$ + +where $\mathbf{P}[: -1, \mathbf{t}[1:]]$ shifts $\mathbf{t}$ one position forward and indexes by $\mathbf{P}$ up to the second-to-last position, aligning each token's probability with its following token in the sequence. + +For each layer's hidden states $\mathbf{X}^{\ell}$ , we compute the intermediate probability distribution $\mathbf{q}^{\ell}$ following the same procedure: + +$$ +\mathbf {q} ^ {\ell} = \operatorname {s o f t m a x} \left(\operatorname {n o r m} \left(\mathbf {X} ^ {\ell}\right) \mathbf {W} ^ {\top}\right) [: - 1, \mathbf {t} [ 1: ] ]. \tag {3} +$$ + +To capture the trajectory of evolving predictions, we calculate the KL divergence between the normalized layer-wise distribution $\mathbf{q}^{\ell}$ and the final distribution $\mathbf{p}$ : + +$$ +D _ {K L} \left(\mathbf {q} ^ {\ell} \| \mathbf {p}\right) = \mathbf {1} ^ {\top} \left(\mathbf {q} ^ {\ell} \log \frac {\mathbf {q} ^ {\ell}}{\mathbf {p}}\right), \tag {4} +$$ + +where $\mathbf{1} \in \mathbb{R}^{N-1}$ is a vector of ones, $\log$ is applied elementwise. Eq. (4) quantifies the deviation of each layer's prediction from the final model output, offering insight into how much each layer's distribution shifts along the prediction trajectory. This measure enables a deeper understanding of each layer's role in shaping the model's eventual output distribution. + +In Figure 3, we plot Eq. (4) across different layers of the Llama-3B, 8B, and $70\mathrm{B}^3$ models by feeding $^{4}300$ random samples from GenQA [5]. To demonstrate the same curve pattern in a different model family, Gemma-2B is also included. Each model displays a distinct phase transition where the curves abruptly coalesce and then monotonically converge to the final distribution. For example, in Llama-8B, this point appears to occur around layer 17 whereas for Llama-70B it is closer to layer 40. We speculate that this point marks a transition in the type of position-wise information processing occurring in the model, where the internal states shift from early phase before the transition point to the late phase after it. The layers in the early phase + +![](images/dcec7941ef75217073b9163d0397fb938ba65ec2e318c0b0bf2d96c42580b651.jpg) +Figure 3. The trajectory of prediction across different layers of Llama-3B, 8B, and 70B, and Gemma-2B from a different model family. The arrow marks the transition point where the trajectories of 300 random samples converge. + +![](images/3f18d4c0f2ef2307de4bd9b228c3b3a07b629955f5bd8dc5c24976222fe7e553.jpg) + +![](images/ccad0660600c136c65f912401a2d44edfdb148f15d5772b36b0581c025ffd445.jpg) + +![](images/cb05fdce128df88230c65074b2701ce2cbc09efccc31efb201b0997053f86bc6.jpg) + +process information from individual token embeddings and combine simple representations together to form higher order concepts, then layers in the late phase converge towards a specific next-token prediction. + +![](images/0d417b5d7456f216cdb47f5806266f4569b586a9020659121bf51313965b69d4.jpg) +Figure 4. Replacing layers with a translator. Despite the relative size in the illustration, our translator is simply an identical transformer layer inherited from the target LLM. The translator bypasses many network layers, and is initialized from the shallowest original layer that it replaced. + +# 1.2. Studying the Transition Phases + +To test our hypothesis on the transition point, we experiment with Llama-3B5 by replacing consecutive layers of each phase with a single transformer layer called a translator (terminology adopted from [3]), as depicted in Figure 4. From Llama-3B's 28 layers, we preserve the first $(\ell = 0)$ and last $(\ell = 27)$ layers while replacing two groups of eleven layers each with a translator $\mathcal{T}$ : layers from $\ell = 1$ to 11 indicated as $\mathcal{T}(1,11)$ for the early phase before Llama-3B's transition point $(\ell = 16$ in Figure 3) and $\mathcal{T}(16,26)$ for the late phase after it. Each is a 2B small model. + +Next, we examine the two transition phases by evaluating vision encoders trained on models $\mathcal{T}(1,11)$ and $\mathcal{T}(16,26)$ . To understand their differences and how they affect encoder transferability to the target LLM, we construct two LLaVA-like VLMs using these small models as decoders. We employ a two-stage training approach to conduct the initial experiments: + +1) First, we simultaneously pre-train a vision adapter (a two-layer MLP) and the translator on 1M instructions6, combin- + +ing LLaVA-1.5-665K [27] vision-language instructions and random GenQA [5] 500K text instructions, for one epoch. 2) Then, we fine-tune the encoder (ViT-L/14@336px) and vision adapter with the frozen decoders on the LLaVA-1.5-665K instructions for one epoch. + +
small modelmmluhellaswagatcasyatcchallengewinograndepiqaboolqopenbookqa
Llama-3B60.773.071.152.770.677.178.939.2
T(1,11)26.642.550.327.753.566.657.532.4
T(16,26)58.957.254.838.564.367.678.232.6
T(16,26)*56.957.357.540.764.370.179.935.2
+ +Table 1. Accuracy $(\%)$ of small models for Llama-3B on text benchmarks. * is a control experiment added later in the study. + +Evaluating decoders. After fine-tuning translators in the first stage, we evaluate models $\mathcal{T}(1,11)$ and $\mathcal{T}(16,26)$ on text benchmarks $^7$ (Table 1). The first row is the baseline performance of Llama-3B. The second and third rows show the performance of the decoders with early- and late-phase layers replaced, respectively. A significant performance drop occurs when replacing early-phase layers, underscoring their critical role in understanding and generation. + +Evaluating encoders. During the second stage, encoders are fine-tuned with small models $\mathcal{T}(1,11)$ and $\mathcal{T}(16,26)$ . We also train an encoder with the full-size Llama-3B as our baseline, listed in the first row of Table 2. For each model, $\mathcal{T}(1,11)$ and $\mathcal{T}(16,26)$ , we report two results: a) performance with their respective encoders, and b) performance with these encoders zero-shot grafted to Llama-3B. For case b), since Llama-3B is never trained on vision-language instructions, it cannot consistently follow special instructions in benchmarks like MME [10] and POPE [26] that expect "yes" or "no" answers by prompting with "single word or phrase". For these benchmarks, we prompt the model with binary prompts, directing it to answer with "yes" or "no" to ensure measurable responses. + +![](images/6a5a00efcc18cae2e664ff23b5a1877629a1d540db1d4082e9c97aca7ca7d4ea.jpg) +Input image +Input prompt +Figure 5. Qualitative results on zero-shot grafting capability of encoders trained with small models for Llama-3B. For comparison, we also include responses from the encoder trained with Llama-3B and the fine-tuned Llama-3B. The encoder trained on $\mathcal{T}(16,26)$ achieves strong zero-shot transfer to Llama-3B. Response is sampled with greedy decoding. A $\rightarrow$ B denotes plugging A into B. + +Based on the visual elements captured in this image of $< |$ image tokens $|$ , describe the image briefly. + +# Encoder fine-tuned by small model $\mathcal{T}(1,11)$ + +![](images/4beca03f5341fa09c7574df9f3dbc7787d36cf2cf6c73dae562a9dd0037563a0.jpg) + +Llama 3B + +Zero-shot grafting: failure + +There is no image provided. The text appears to be a collection of phrases and sentences that seem to be related to a topic, possibly about a image or a visual element, but no actual image is included. + +# Encoder fine-tuned by small model $\mathcal{T}(16,26)$ + +![](images/49abbd53a45f13265c668763b5aac43286ae2fdaa6401bb52f3c8f440750c2e0.jpg) + +Llama 3B + +Zero-shot grafting: success + +The image shows two lambs standing side by side in a yard, with one lamb having a distinctive tuft of wool on its forehead. The lambs appear to be in a natural setting, possibly in a farm or a rural area. + +# Encoder fine-tuned by Llama 3B $\longrightarrow$ Llama 3B + +The image features two sheep standing next to each other, with one sheep looking directly at the camera. The other sheep is slightly turned away, but still visible in the scene. The sheep are standing on a dirt path, and there is a wooden fence in the background. + +# Pre-trained Encoder $\longrightarrow$ fine-tuned Llama 3B + +The image features two adorable baby llamas standing next to each other in a brick courtyard. They are facing the camera, with one of the llamas looking directly at the viewer. The scene appears to be set in a zoo, as there is a bench visible in the background. The two llamas are the main focus of the image, showcasing their unique and charming appearance. + +
encoder fine-tuned onMMEper +binaryPEbinarySEED-BenchMMvetLLaVA-WildMMbenperformance
Llama-3B102881.754.224.142.941.8
model T(1,11)59963.225.814.337.20.6
zero-shot grafting5402.725.36.926.39.2
model T(16,26)92370.453.220.642.745.4
zero-shot grafting102280.153.123.156.647.4
model T(16,26)*116284.459.825.048.150.3
zero-shot grafting71422.340.711.234.230.1
+ +Table 2. Accuracy (\%) of encoders fine-tuned by small models for Llama-3B on VLM benchmarks. * indicates a control experiment added later in the study. + +Table 2 clearly shows that the encoder trained with early-phase layers preserved model $\mathcal{T}(16,26)$ outperforms the one with early-phase layers discarded model $\mathcal{T}(1,11)$ . Remarkably, performance improves further when the encoder fine-tuned on $\mathcal{T}(16,26)$ is zero-shot grafted to Llama-3B, as shown in the third row block. This improvement highlights that the encoder trained with $\mathcal{T}(16,26)$ can produce image features that are interpretable by Llama-3B. + +In Figure 5, we present qualitative results showcasing the zero-shot grafting capability of the encoders trained via $\mathcal{T}(1,11)$ and $\mathcal{T}(16,26)$ . The responses enhance the above results that replacing the early-phase layers causes the encoder to fail in generating image features that are directly interpretable by the full-size Llama-3B. + +# Are early layers the most critical for encoder transfer? + +The shallow phase of inference plays a crucial role in transferring a pre-trained encoder to the full-size target LLM. To concretely verify this observation, we conduct a control + +run based on $\mathcal{T}(16,26)$ , in which we unfreeze every other layer before the translator and train them alongside it during the first stage. This control experiment is designed to disrupt the original early-phase parameters, allowing us to completely assess their impact on encoder transferability. We denote this modified model as $\mathcal{T}(16,26)^{*}$ . + +First, back in Table 1, the last row indicates that fine-tuning additional layers alongside the translator leads to better performance on text benchmarks. However, in Table 2, when evaluating the encoder trained on $\mathcal{T}(16,26)^*$ , we actually observe a huge loss of zero-shot grafting ability. This suggests that modifying early-phase parameters in $\mathcal{T}(16,26)$ enhance performance on both text and VLM benchmarks when evaluated through itself, but fails to preserve the encoder's zero-shot grafting capability as the embedding space of $\mathcal{T}(16,26)^*$ drifts away from the target model. + +How many early-phase layers should be preserved? If retaining the original early-phase parameters is necessary, the next question is how many layers to preserve for effective zero-shot grafting. In other words, we seek to confirm the transition point in Figure 3 as the optimal starting point for layer removal and translator insertion. + +To ensure generalization, we conduct this ablation with Llama-8B, where the transition point is around layer $\ell = 17$ . To validate the transition point, we create three small models by progressively reducing the replaced layers before the transition point: $\mathcal{T}(6,30)$ , $\mathcal{T}(12,30)$ , and $\mathcal{T}(17,30)$ . As shown in Table 3, the performance of these three models on text benchmarks corroborates our findings from Llama-3B, demonstrating that early layers are indeed important. Keeping more early layers leads to better performance, with the best achieved by the model $\mathcal{T}(17,30)$ . + +![](images/6f5d257dd6b60ccdc6a6d8aef2141f891f719612e586757455d985d5015f5950.jpg) +Input image +Input prompt + +Based on the visual elements captured in this image of $<\text{image\_tokens}>$ , where is the cat? And what is it doing? + +# Encoder fine-tuned by small model $\mathcal{T}(6,30)$ + +![](images/9fded7cbac7ee9e5666af26e4e34ded211383892ebcd270b524a300756fa5294.jpg) + +Llama 8B + +Zero-shot grafting: failure + +It seems like you're having a bit of fun with the text! To answer your question, the image you're referring to is not actually an image, but rather a text prompt. I'm here to help you with any questions or topics you'd like to discuss. If you'd like to start fresh, feel free to ask me anything, and I'll do my best to help! + +# Encoder fine-tuned by small model $\mathcal{T}(12,30)$ + +![](images/efbf113c029830125bed64fa067c28edf2afcc91e62abe1bb3db7a69f430d6a5.jpg) + +Llama 8B + +Zero-shot grafting: success + +The cat is sitting on a colorful, hand-painted canvas, which is likely a wall or a piece of art. The cat appears to be looking directly at the viewer with a curious expression, as if it is observing its surroundings. + +# Encoder fine-tuned by small model $\mathcal{T}(17,30)$ + +![](images/8c4caec2512286d61a697527033fcb4d26542a85510bb3919c19b00a9c1c10d0.jpg) +Figure 6. Qualitative results on zero-shot grafting capability of encoders trained with surrogate models for Llama-8B. For comparison, we also include responses from the encoder trained with Llama-8B and the fine-tuned Llama-8B. More early-phase layers preserved lead to stronger zero-shot grafting capability. Responses are sampled with greedy decoding. A $\rightarrow$ B denotes plugging A into B. + +Llama 8B + +Zero-shot grafting: success + +The cat is sitting on a colorful, striped cushion or couch, looking directly at the viewer with its eyes. It appears to be relaxed and calm, possibly enjoying the comfortable seating. + +# Encoder fine-tuned by Llama 8B $\longrightarrow$ Llama 8B + +The cat is sitting on a yellow couch, looking directly at the camera. + +# Pre-trained Encoder $\longrightarrow$ fine-tuned Llama 8B + +The cat is sitting on a couch, which is located in a room. + +
small modelmmluhellaswagarceasyarchallengewinograndepiqaboolqopenbookqa
Llama-8B68.480.579.861.877.381.585.444.8
T( 6,30)25.531.836.824.750.958.861.325.8
T(12,30)25.442.940.529.059.662.669.429.6
T(17,30)66.861.259.344.870.971.069.334.2
+ +Table 3. Accuracy (%) of small models for Llama-8B on text benchmarks. + +
encoder fine-tuned byMME binaryPEP binarySEED-BenchMM VetLLaVA-WildMMBenperformance
Llama-8B116584.757.523.247.644.9
model T(6,30)58373.325.88.922.48.6
zero-shot grafting76720.630.913.428.1-↓
model T(12,30)98377.926.913.629.10.43
zero-shot grafting102281.750.720.547.945.4↑
model T(17,30)104181.355.420.942.049.7
zero-shot grafting104483.456.125.256.853.5↑
+ +Table 4. Accuracy (%) of encoders fine-tuned by small models for Llama-8B on VLM benchmarks. + +We evaluate the zero-shot grafting capability of the encoders trained with these three models to Llama-8B in Table 4. Performance improves with more preserved early layers, showing particularly strong results when retaining all early-phase layers before the transition point in $\mathcal{T}(17,30)$ . + +Figure 6 depicts a qualitative example demonstrating the zero-shot grafting capability of three trained encoders. As expected, the encoder trained by the model $\mathcal{T}(6,30)$ fails to generate readable image features for Llama-8B as most of the early-phase layers are removed. The encoder trained + +by $\mathcal{T}(12,30)$ performs better, but its image features lack fine-grained detail (e.g., no couch in response), which explains the zero-shot performance gap in Table 4. The encoder trained by $\mathcal{T}(17,30)$ generates more detailed and accurate image features, achieving the best zero-shot grafting response among the three, which covers the cat's color, expression, eye direction, position, the presence of a couch, and even the atmosphere. + +# In summary, our entire analysis of the prediction trajectory reveals: + +- The early phase ${}^{a}$ plays a pivotal role in the encoder's transferability to the target LLM. +- Retaining the original parameters of the early phase is critical for maintaining the encoder's zero-shot grafting capability. +- The transition point in Figure 3 is a good starting point for late-phase removal and translator insertion. + +aWe believe the early phase of LLMs has potential beyond building surrogate models, enabling more creative applications. + +Based on these three key findings, now we define the model $\mathcal{T}(16,26)$ as our surrogate model for Llama-3B, the model $\mathcal{T}(17,30)$ as our surrogate for Llama-8B, by inheriting the early-phase layers and replacing the late-phase layers with a translator, which can be fine-tuned with a small set of text instructions, e.g., 500K for one epoch. + +# 2. Generalizing to Giant Models + +Having validated our approach at relatively small scales, we now expand our methodology to construct surrogate models for giant LLMs - Llama-70B. In this section, our ex + +![](images/ec189bcd3799531ed5bca6236528ee437b7be6abb61c81f2b8215a619cc2a9a6.jpg) +Input image +Input prompt + +Based on the visual elements captured in this image of $< |$ image tokens $|>$ ,the input question is displayed on the right side. + +# Encoder fine-tuned by surrogate $\mathcal{T}(40,78)$ + +# Question 1 (creativity) + +can you write two funny Haiku poems for this image? One is about the subject with fine-grained details, and the other is about how the subject vividly interacts with the background. + +# Haiku 1: Fine-grained details + +Fuzzy sheep on foam +Whimsical, playful, and sweet +Winter's warm delight + +# Haiku 2: Interaction with background + +Frothy white sheep +Dances on the coffee sea +Joy in every sip + +![](images/975a5fe827fbb53b0ff4364449903c3b1e058c3feca9eb7c4262369b629ea408.jpg) + +# Llama 70B + +# Question 2 (negation) + +is this a sad picture? And why? + +No, this is not a sad picture. The image depicts a cheerful and playful scene, featuring a white ceramic mug with a cartoon sheep design on top of a foamy coffee drink. The sheep's facial expression appears to be smiling, which adds to the overall lighthearted and uplifting atmosphere of the image. + +# Question 3 (reasoning) + +guess what I'm doing now? + +I'm enjoying a warm and delicious cup of coffee with a cute sheep-shaped marshmallow on top! + +periments demonstrate two key advantages of our surrogate approach: First, surrogates can bring a strong zero-shot grafting ability to encoders, enabling them to trigger target LLMs to perform visual understanding tasks without additional training. Second, training target LLM decoders on surrogate-trained encoders significantly reduces cost by providing a warm start for fine-tuning. + +# 2.1. A Surrogate for Llama-70B + +We analyze the prediction trajectory for Llama-70B in Figure 3 to identify the transition point that marks the end of token processing, which occurs around layer $\ell = 40$ . Then we keep the first $(\ell = 0)$ and last layer $(\ell = 79)$ , insert a translator at $\ell = 40$ , and remove the late phase from $\ell = 41$ to $\ell = 78$ , to build a 37B surrogate $\mathcal{T}(40,78)$ . Text benchmark results of this surrogate are shown in Table 5, and VLM benchmark results in Table 6. Table 6 shows the performance of the encoder trained using surrogate $\mathcal{T}(40,78)$ on VLM benchmarks, highlighting a significant improvement through zero-shot grafting. These experiments show that our approach can be scaled up to giant models, holding the same principles of early phase preservation. + +# 2.1.1. Results: Zero-shot Grafting + +In Table 7, with encoder-only training, our surrogate outperforms the full-size Llama-70B on most VLM benchmarks, except for VisWiz. This demonstrates the effectiveness of our surrogate models. The last row shows the performance of zero-shot grafting the surrogate-trained encoder into Llama-70B. Notably, the performance of zero-shot grafting surpasses the full-size Llama-70B decoder training on some benchmarks by a big margin, demonstrating that our surrogate-trained encoder effectively prompts LLaMA-70B to handle complex visual understanding tasks. + +![](images/a525e5b0c2e0cd7d6a51a68f0682f2f7591461b41aee9e25ebde20ac48699c7f.jpg) +Figure 7. Qualitative results on strong zero-shot grafting ability of surrogate-trained encoder for Llama-70B, which produces fine-grained image features to trigger Llama-70B to perform complex visual understanding tasks. Response is sampled with greedy decoding. +Input image + +# Input prompt + +Based on the visual elements captured in this image of $< |$ image tokens $|$ , what is the text in the image? And where is it? + +The text in the image reads "Shark Bar" and is located on the side of a plastic cup. + +Figure 8. Qualitative OCR results on strong zero-shot grafting ability of surrogate-trained encoder for Llama-70B. The input image size is $336^2$ . + +
small modelmmluhellaswagarceasyarchallengewinograndepiqaboolqopenbookka
Llama-70B82.686.983.471.285.483.789.147.6
surrogate-37B80.870.467.356.677.973.986.937.8
+ +Table 5. Accuracy(%) of surrogate model for Llama-70B on text benchmarks. + +
encoder fine-tuned onMMEperbinaryPOPEbinarySEED-BenchMMVetLLaVA-WildMMBenperformance
Llama-70B129483.459.827.045.658.8
surrogate-37B128784.864.229.654.259.5
zero-shot grafting131586.164.137.459.760.7↑
+ +Table 6. Accuracy (%) of encoder fine-tuned by surrogate for Llama-70B on VLM benchmarks. + +Figure 7 presents qualitative results showcasing the strong zero-shot grafting capability of our surrogate-trained encoder, including questions about creativity, negation, and reasoning. Additionally, Figure 8 demonstrates its effectiveness on OCR tasks, showing that our surrogate models are able to squeeze robust and detailed visual information into encoders. + +
training methodMME binaryMMEPOPE binaryPOPESEED-BenchMMLLaVAMMB enCV-BenchGQAVis-Wiz
cogperccogpercacc.f1acc.f1allimgvid-Vet-Wild2d3davg
Llama-70B decoder3271545345152484.983.184.882.963.668.943.735.567.571.861.873.367.562.453.0
Llama-70B encoder2851294288132183.482.682.781.259.865.438.627.045.658.862.259.460.854.447.4
surrogate-37B encoder3121329291125085.583.986.385.065.971.146.228.854.363.164.764.064.356.522.7
zero-shot grafting2951348303129886.886.187.086.465.470.745.332.868.965.663.267.265.251.940.0
+ +Table 7. Accuracy (%) for Llama-70B on VLM benchmarks. The bold numbers indicate the best performance between the full-size decoder training and our surrogate-trained encoder by zero-shot grafting. A special clarification for LLaVA-Wild is in Sec. A.13. + +
methodX%avg. scoreMMEbinaryMMEPOPEbinaryPOPESEED-BenchMMLLaVAMMB enCV-BenchGQAVis-Wiz
cogperccogpercacc.f1acc.f1allimgvid-Vet-Wild2d3davg
baseline100.51273011178333113974.573.174.373.946.049.234.117.730.445.652.461.757.145.443.9
200.51532551061277111671.461.171.661.352.255.539.520.641.358.356.666.861.747.538.5
300.62773141447316139985.584.885.584.859.864.840.930.857.966.860.671.165.957.155.1
600.64443531511358151584.883.184.482.462.467.941.532.464.570.961.572.266.861.348.1
1000.65383271545345152484.983.184.882.963.668.943.735.567.571.861.873.367.562.452.9
grafting-0.62592951348303129886.886.187.086.465.470.745.332.868.965.663.267.265.251.940.0
ours100.66123401404342143087.386.884.982.767.172.745.937.669.770.966.670.668.660.157.9
200.67013691435361148686.785.686.484.967.272.846.038.870.573.265.774.870.360.652.4
300.67043741449349149087.987.786.885.567.072.047.838.969.373.966.672.869.761.449.2
+ +Table 8. Convergence comparison with using X% of training data between baseline and our surrogate training approach for Llama-70B decoder training. The gray row indicates the training hours reported in the Table 9 with 20% of training data for ours. See Table A.1 for extra columns with additional benchmarks. + +While the surrogate-trained encoder enables zero-shot conversion of the giant LLM into a VLM, its performance still lags behind that of full-size decoder training. What benefits can we expect from this surrogate-trained encoder? Next, we demonstrate that it can accelerate training convergence and improves the performance of full-size decoder training. + +# 2.2. Reducing Full Decoder Training Cost + +In the previous sections, we conduct the experiments with a two-stage training strategy, where we simultaneously train the vision adapter in encoder and the translator in decoder during the first stage, and then fine-tune the encoder atop the surrogate in the second stage. Currently, we are interested in training the full-size decoder, which is the final third training stage. First, we introduce the training setup, and recipes are introduced in Sec. A.9. + +Models. As in previous sections, we use the CLIP-L/14 encoder with an input image size of $336^2$ . The vision adapter is a two-layer MLP, consisting of consecutive linear layers with a GELU activation in between. Notably, we maintain a fixed vision adapter size across all model scales, unlike prior works [6, 24] that scale it with model size. This design choice ensures that variations in adapter size do not introduce unknown effects on the encoder's zero-shot grafting capability, allowing for a controlled initial study. For state-of-the-art performance, however, the vision adapter can be scaled up with the model size. The decoders are our surrogate-37B, i.e., $\mathcal{T}(40,78)$ , and full-size Llama-70B. + +Data. In the third training stage, the training data is still the same as in the previous two stages – the LLaVA-1.5-665K [27] instructions (without text-only samples). This choice is based on the following considerations: + +1) The first training stage focuses on the adapter and translator. Commonly, vision adapters are trained on captions instead of instructions, but we found no significant difference in experimental outcomes. Thus, to simplify training, we merge the training of the vision adapter and translator into a single stage using vision-language and text-only instructions. When forwarding the text-only instructions, gradients backpropagated to the vision adapter are zero. +2) The second stage trains encoders with surrogates, aiming to efficiently compress data into encoders while preparing to transfer knowledge to the full-size decoder. To ensure consistency, we use the same training data for the second and third stages. It is recommended to use larger and more diverse datasets for those two stages. + +# 2.2.1. Results: Convergence and Training Cost + +In Table 8, we compare performance of the typical baseline method and our surrogate training approach across different percentages of training data used for training decoders. The baseline trains Llama-70B with the original CLIP encoder, while ours trains it with our surrogate-trained encoder (the third row in Table 7). First, the gray row represents the performance of zero-shot grafting the surrogate-trained encoder to Llama-70B, which nearly matches the baseline with $30\%$ of the data. Second, after training on just $10\%$ of data, our approach achieves nearly the same performance as the baseline with $100\%$ of the data, except for MME. For other benchmarks, our $10\%$ performance even outperforms the final baseline result. With continued training, performance remains unchanged, suggesting saturation after $20\%$ of the data. In Figure 2, we plot the normalized average score of each $\mathrm{X}\%$ data utilization for our method and baseline. + +We also visualize Table 9 in Figure 2 for a direct comparison of training hours for each training stage. First, our pretraining time is longer than the baseline because we train both the vision adapter and the translator with additional text instructions. However, the key advantage of our surrogate training approach is seen in the decoder training, which is the real bottleneck in common methods. With $20\%$ of the data used for our decoder training, we achieve a training cost reduction of $\sim 45\%$ . This reduction is the minimum, as our performance in the $10\%$ case already exceeds the final result of the baseline, as shown in Table 8. + +
method#gpupre.ft. enc.ft. dec.total hours
baseline1287.010.0027.8834.79
ours1289.364.255.5619.17
+ +Table 9. Training hours comparison between baseline and our surrogate approach for training VLMs with Llama-70B, including the time for pretraining (pre.), fine-tuning encoder (ft. enc.), and fine-tuning decoder (ft. dec.). Checkpoint loading and saving times are excluded. More details in Sec. A.8. + +# 3. Related Work Overview + +Understanding LLMs is a key topic in mechanistic interpretability [39]. [1] uses linear classifiers (probes) to understand the dynamics of intermediate layers in neural networks. For LLMs, [36] directly employs the output embedding matrix as a probe to classify layer-wise representations, illustrating how input tokens shift from current positions to next ones. Tuned Lens [3] extends this idea with a trainable probe for broader applicability to modern LLMs. [41] conceptualizes transformer layers as "painters" that iteratively refine representations and suggests that middle layers share the same representation. In contrast, we identify two distinct transition phases in LLMs. + +The shared representation in middle layers suggests redundancy. [41] further concludes that some middle layers can be removed without a significant performance drop. Prunning LLMs largely is based on such insight of redundancy. Notably, both [12] and [34] found that deep layers are not essential and can be removed. Interestingly, our surrogate models also replace deep layers in the late phase. However, our method differs in how we identify the transition point and in our objective. Unlike pruning, which aims to remove layers while preserving performance, our focus is on the efficiency of surrogate models for encoder transferability. While our surrogate models consistently underperform compared to their target LLMs, they serve a distinct purpose in producing efficient encoders for VLMs. + +Our surrogate-trained encoders can directly prompt target LLMs to generate the expected responses without any finetuning. This zero-shot grafting ability aligns with the concept of steering LLMs, a lightweight alternative to finetuning LLMs [14, 20]. Prior works show that language + +models can be guided to perform specific tasks without extensive fine-tuning. Similarly, in our case, image features from surrogate-trained encoders act as steering tokens, enabling target LLMs to interpret visual content and answer various complicated questions. + +This capability provides a warm start for further decoder fine-tuning, helping to mitigate the expensive training cost of VLMs [2, 6, 25, 46, 48, 49]. The costs surged as decoder sizes scale from relatively small models (3B, 8B) to much larger ones, such as 70B [24], 110B [22, 28]. Additionally, increasing the number of image tokens for high-resolution inputs further escalates the computational burden. LoRA [16] could be applied for training VLMs. While LoRA improves efficiency, it underperforms full fine-tuning, especially in giant LLMs, when applied with small rank (e.g., 8) and alpha (e.g., 32) to query and key decoder layers – a common practice in LLM training. Closing this gap needs applying LoRA to entire transformer layers with large rank and alpha (e.g., rank 128 with alpha 256 as in [24] for 13B decoder training). Then LoRA takes about the same time as full decoder fine-tuning. This limitation likely explains why current VLMs still rely on full decoder fine-tuning. Critically, contrasting to our surrogate training approach, LoRA does not accelerate convergence. See more in Sec. A.2. + +Additionally, the idea of using small models to train encoders before applying them to larger decoders has been depicted in [18]. However, this work is not directly related to ours, as it employs a progressive multi-stage training strategy to just scale up model size and refine image processing from coarse to fine. No further details are provided on this method, leaving it unclear how it reduces costs. In contrast, our approach provides a well-defined framework for constructing efficient surrogate models specifically tailored for any target LLM. Plus, we plug the surrogate-trained encoders directly into their target LLMs, converting them into VLMs without any fine-tuning to perform complex visual understanding tasks. Further, with our surrogate-trained encoders, the decoder needs only a few full-scale fine-tuning steps to achieve the desired performance. + +# 4. Conclusion + +In this work, we show that vision encoders trained with our surrogate models can accelerate VLM training. We also note that our surrogate models are not limited to vision encoders. The main limitation of our approach is the need for a well-designed surrogate, which ideally should be small. Although our layer-dropping strategy works in principle for any LLM, resulting models are still half the size of their target LLMs, for example, our surrogate-37B for Llama-70B. This underscores the practical value of surrogate models and highlights the need for ways to create them more efficiently and with better compression. + +# Acknowledgements + +# References + +[1] Guillaume Alain and Yoshua Bengio. Understanding Intermediate Layers Using Linear Classifier Probes. In ICLR, 2017. 8 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a Visual Language Model for Few-Shot Learning. In NeurIPS, 2022. 8 +[3] Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. Eliciting Latent Predictions From Transformers With the Tuned Lens. arXiv preprint arXiv:2303.08112, 2023. 3, 8 +[4] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. PIQA: Reasoning about Physical Commonsense in Natural Language. In AAAI, 2020. 11 +[5] Jiuhai Chen, Rifaa Qadri, Yuxin Wen, Neel Jain, John Kirchenbauer, Tianyi Zhou, and Tom Goldstein. GenQA: Generating Millions of Instructions from a Handful of Prompts. arXiv preprint arXiv:2406.10323, 2024. 2, 3, 12 +[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks. In CVPR, 2024. 7, 8, 13 +[7] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions. In ACL Anthology, 2019. 11 +[8] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge. arXiv preprint arXiv:1803.05457, 2018. 11 +[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 Herd of Models. arXiv preprint arXiv:2407.21783, 2024. 1, 12 +[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models. arXiv preprint arXiv:2306.13394, 2023. 3, 11 + +We sincerely appreciate Meta's support in providing GPUs for our experiments. The authors at University of Maryland were supported by DAPRA TIAMAT, the ONR MURI program, the National Science Foundation (IIS-2212182), and the NSF TRAILS Institute (2229885). Additional commercial support was provided by the Amazon Research Award program, Open Philanthropy, and Capital One Bank. +[11] Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A Framework for Few-Shot Language Model Evaluation, 2024. 11 +[12] Andrey Gromov, Kushal Tirumala, Hassan Shapourian, Paolo Glorioso, and Daniel A Roberts. The Unreasonable Ineffectiveness of the Deeper Layers. In ICLR, 2025. 8 +[13] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. VizWiz Grand Challenge: Answering Visual Questions from Blind People. In CVPR, 2018. 11 +[14] Chi Han, Jialiang Xu, Manling Li, Yi Fung, Chenkai Sun, Nan Jiang, Tarek Abdelzaher, and Heng Ji. Word Embeddings Are Steers for Language Models. In ACL Anthology, 2024. 8 +[15] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring Massive Multitask Language Understanding. In ICLR, 2021. 11 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-Rank Adaptation of Large Language Models. In ICLR, 2022. 8, 9 +[17] Drew A Hudson and Christopher D Manning. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. In CVPR, 2019. 11 +[18] InternVL2. InternVL2: Better than the Best—Expanding Performance Boundaries of Open-Source Multimodal Models with the Progressive Scaling Strategy. https://internvl.github.io/blog/2024-07-02-InternVL-2.0/, 2024.8 +[19] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A Diagram Is Worth a Dozen Images. In ECCV, 2016. 11 +[20] Md Kowsher, Nusrat Jahan Prottasha, and Prakash Bhat. Propulsion: Steering LLM with Tiny Fine-Tuning. In ACL Anthology, 2024. 8 +[21] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. SEED-Bench: Benchmarking Multimodal LLMs with Generative Comprehension. In CVPR, 2024. 11 +[22] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. LLaVA-Next: Stronger Llms Supercharge Multimodal Capabilities in the Wild. https://llava-v1.github.io/blog/2024-05-10-llava-next-stronger-llms/, 2024.8 +[23] Bo Li, Peiyuan Zhang, Kaichen Zhang, Fanyi Pu, Xinrun Du, Yuhao Dong, Haotian Liu, Yuanhan Zhang, Ge Zhang, Chunyuan Li, and Ziwei Liu. LMMs-Eval: Accelerating the Development of Large Multimodal Models. https://github.com/EvolvingLMMs-Lab/lmms-eval, 2024.11 + +[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. LLaVA-OneVision: Easy Visual Task Transfer. arXiv preprint arXiv:2408.03326, 2024. 7, 8 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. In ICML, 2023. 8 +[26] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating Object Hallucination in Large Vision-Language Models. In EMNLP, 2023. 3, 11 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning. In CVPR, 2024. 3, 7, 11, 13 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-Next: Improved Reasoning, Ocr, and World Knowledge. https://llava-v1.github.io/blog/2024-01-30-llavanext/, 2024.8 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In NeurIPS, 2024. 11, 13 +[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. MMBench: Is Your Multi-modal Model an All-around Player? In ECCV, 2025. 11 +[31] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A Benchmark for Question Answering About Charts With Visual and Logical Reasoning. In ACL Findings, 2022. 11 +[32] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A Dataset for VQA on Document Images. In WACV, 2021. 11 +[33] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. InfographicVQA. In WACV, 2022. 11 +[34] Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. ShortGPT: Layers in Large Language Models are More Redundant Than You Expect. arXiv preprint arXiv:2403.03853, 2024. 8 +[35] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering. In ACL Anthology, 2018. 11 +[36] nostalgicbraist. interpreting GPT: the logit lens. https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens, 2020.8 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1 +[38] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WinoGrande: An Adversarial Winograd Schema Challenge at Scale. In ACM Communications, 2021. 11 + +[39] Lee Sharkey, Bilal Chughtai, Joshua Batson, Jack Lindsey, Jeff Wu, Lucius Bushnaq, Nicholas Goldowsky-Dill, Stefan Heimersheim, Alejandro Ortega, Joseph Bloom, et al. Open Problems in Mechanistic Interpretability. arXiv preprint arXiv:2501.16496, 2025. 8 +[40] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards VQA Models That Can Read. In CVPR, 2019. 11 +[41] Qi Sun, Marc Pickett, Aakash Kumar Nain, and Llion Jones. Transformer Layers as Painters. arXiv preprint arXiv:2407.09298, 2024. 8 +[42] Gemma Team. Gemma-2. Kaggle, 2024. 12 +[43] Gemma Team. Gemma-3. Kaggle, 2025. 9 +[44] Qwen Team. Qwen2.5 Technical Report. arXiv preprint arXiv:2412.15115, 2025. 10 +[45] Qwen Team. Qwen3 Technical Report. arXiv preprint arXiv:2505.09388, 2025. 10 +[46] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs. In NeurIPS, 2025. 8, 11 +[47] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features. arXiv preprint arXiv:2502.14786, 2025.1 +[48] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 8 +[49] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. CogVLM: Visual Expert for Pretrained Language Models. In NeurIPS, 2024. 8 +[50] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities. In ICML, 2024. 11 +[51] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag: Can a Machine Really Finish Your Sentence? In ACL Anthology, 2019. 11 +[52] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid Loss for Language Image Pre-Training. In ICCV, 2023. 1 +[53] Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallel. arXiv preprint arXiv:2304.11277, 2023. 11 \ No newline at end of file diff --git a/zeroshotvisionencodergraftingviallmsurrogates/images.zip b/zeroshotvisionencodergraftingviallmsurrogates/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b37ccf3ebcc9246e3e22cab7f80e5fa0d608d1d6 --- /dev/null +++ b/zeroshotvisionencodergraftingviallmsurrogates/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69f29e13e86170d3f02fb10a00f722db685943fb9d099f29c5dc36a74d2a5462 +size 538230 diff --git a/zeroshotvisionencodergraftingviallmsurrogates/layout.json b/zeroshotvisionencodergraftingviallmsurrogates/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4bdfa3baec2e8408c8d560ce45c416c13dba2c67 --- /dev/null +++ b/zeroshotvisionencodergraftingviallmsurrogates/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f5448c34d8d2ac2b9b937dbf633e880ec515975d63628f187fe3ba424ff30bd +size 488905 diff --git a/zerostereozeroshotstereomatchingfromsingleimages/76bb9552-1902-4fcd-a11d-bcea4458ddbf_content_list.json b/zerostereozeroshotstereomatchingfromsingleimages/76bb9552-1902-4fcd-a11d-bcea4458ddbf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e6ed6923191d7b9b43507245a50cf6c0388c4439 --- /dev/null +++ b/zerostereozeroshotstereomatchingfromsingleimages/76bb9552-1902-4fcd-a11d-bcea4458ddbf_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9338119f2928dd6ccd17f761da159730a7e376e33e3bc20589ddf82e7256a756 +size 81964 diff --git a/zerostereozeroshotstereomatchingfromsingleimages/76bb9552-1902-4fcd-a11d-bcea4458ddbf_model.json b/zerostereozeroshotstereomatchingfromsingleimages/76bb9552-1902-4fcd-a11d-bcea4458ddbf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e980ee9cff494850172d535477759227af4be5d --- /dev/null +++ b/zerostereozeroshotstereomatchingfromsingleimages/76bb9552-1902-4fcd-a11d-bcea4458ddbf_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:204e39e03f2b4e63deb2f3f4480aafa940054efd47276d8182d8233602eabe95 +size 104549 diff --git a/zerostereozeroshotstereomatchingfromsingleimages/76bb9552-1902-4fcd-a11d-bcea4458ddbf_origin.pdf b/zerostereozeroshotstereomatchingfromsingleimages/76bb9552-1902-4fcd-a11d-bcea4458ddbf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c72977bba0b76a48bd97c993a078fdd507f756b7 --- /dev/null +++ b/zerostereozeroshotstereomatchingfromsingleimages/76bb9552-1902-4fcd-a11d-bcea4458ddbf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d15ab5b2264c5f9a3d2040cdd6ede9e90ad72d996a93eaa38c4ac18fa3a0add +size 7180274 diff --git a/zerostereozeroshotstereomatchingfromsingleimages/full.md b/zerostereozeroshotstereomatchingfromsingleimages/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b7ef640cc69455a002a866cf3a6eb7d142bc27c --- /dev/null +++ b/zerostereozeroshotstereomatchingfromsingleimages/full.md @@ -0,0 +1,338 @@ +# ZeroStereo: Zero-shot Stereo Matching from Single Images + +Xianqi Wang $^{1}$ , Hao Yang $^{1}$ , Gangwei Xu $^{1}$ , Junda Cheng $^{1}$ , Min Lin $^{1}$ , Yong Deng $^{2}$ , Jinliang Zang $^{2}$ , Yurui Chen $^{2}$ , Xin Yang $^{3,1\dagger}$ + +\(^{1}\)Huazhong University of Science and Technology \(^{2}\)Autel Robotics \(^{3}\)Optics Valley Laboratory \(\{xianqiw, haoyang2002, gwxu, jundacheng, minlin, xinyang2014\}@hust.edu.cn + +![](images/40b5088a46b3dfa070b04a98212fb811a1ff5052a43a6e8e57e6ef4c826dab1c.jpg) +Figure 1. Zero-shot generalization results by RAFT-Stereo [25] trained under our ZeroStereo pipeline. + +# Abstract + +State-of-the-art supervised stereo matching methods have achieved remarkable performance on various benchmarks. However, their generalization to real-world scenarios remains challenging due to the scarcity of annotated real-world stereo data. In this paper, we propose ZeroStereo, a novel stereo image generation pipeline for zero-shot stereo matching. Our approach synthesizes high-quality right images from arbitrary single images by leveraging pseudo disparities generated by a monocular depth estimation model. Unlike previous methods that address occluded regions by filling missing areas with neighboring pixels or random backgrounds, we fine-tune a diffusion inpainting model to recover missing details while preserving semantic structure. Additionally, we propose Training-Free Confidence Generation, which mitigates the impact of unreliable pseudo labels without additional training, and Adaptive Disparity Selection, which ensures a diverse and realistic disparity distribution while preventing excessive occlusion and foreground distortion. Experiments demonstrate that models trained with our pipeline achieve state-of-the-art zero-shot generalization across multiple datasets, with only a dataset volume comparable to Scene Flow. Code: https://github.com/Windsrain/ZeroStereo. + +# 1. Introduction + +Stereo matching is a fundamental task in computer vision that estimates depth information by identifying correspond- + +ing points between stereo image pairs. By computing the disparity between matched pixels, stereo matching enables 3D scene reconstruction, which is essential for applications such as autonomous driving and robotic perception. + +With the advancement of deep learning, stereo matching has shifted from traditional handcrafted feature-based approaches to data-driven methods [2, 7, 9, 25, 53, 57-59]. While deep learning-based models achieve impressive performance on standard benchmarks, they struggle to generalize to real-world scenarios due to the scarcity of annotated real-world stereo data [48]. Most models rely on synthetic datasets [29, 52] or limited real-world datasets [40, 41] which fail to cover the full diversity of real-world environments. Several approaches have been proposed to mitigate this challenge. + +One direction involves learning domain-invariant feature representations from synthetic data [3, 10, 26, 38, 64]. However, a domain gap persists due to fundamental differences between synthetic and real-world data distributions. Another approach leverages self-supervised learning [45, 46], using photometric loss [11] as a proxy supervision signal on unlabeled stereo images. However, this method struggles with occlusions, ghosting artifacts, and ambiguities in ill-posed regions, while large-scale collection of high-quality stereo image pairs remains non-trivial. + +In recent years, view synthesis techniques [31, 35] have emerged as a promising approach to self-supervised stereo matching. These methods generate pseudo stereo images and corresponding disparity labels from single images or + +NeRF-rendered scenes. Early strategies [28, 54] employ monocular depth estimation [11, 21, 37] to derive pseudo disparity labels, followed by forward warping to synthesize the right image. However, this approach struggles with occluded regions, where missing pixels are typically filled using neighboring pixels [28] or random backgrounds [54], resulting in structural inconsistencies. To address this, NeRF-Stereo [47] has been proposed to generate stereo images from NeRF-rendered scenes. It leverages an implicit 3D representation, enabling it to synthesize occluded regions during rendering, rather than relying on post-processing heuristics. Additionally, it introduces Ambient Occlusion [33] as a confidence measure to enhance the reliability of pseudo disparity. However, NeRF-Stereo requires multiview inputs for scene reconstruction, limiting its flexibility compared to single-image-based methods. Moreover, NeRF's reconstruction quality for distant objects is often suboptimal, leading to degraded stereo generation in large-scale outdoor environments [8]. + +To overcome these challenges, we propose ZeroStereo, a novel stereo image generation pipeline for zero-shot stereo matching. Inspired by Marigold [17], we hypothesize that modern diffusion models, pre-trained on large-scale image datasets, can be effectively adapted for stereo matching. However, directly applying existing diffusion inpainting models is insufficient for stereo generation, as standard inpainting tasks do not account for the complex and structured occlusion patterns in stereo pairs. To address this, we fine-tune a diffusion inpainting model specifically for stereo image synthesis, ensuring it can handle the diverse and irregular inpainting masks encountered in occluded regions. This enables our method to recover missing background details more accurately, significantly preserving semantic consistency compared to previous heuristic filling approaches. In addition to high-quality image synthesis, training stability is another key factor in stereo matching. To mitigate the impact of unreliable pseudo disparities, we introduce Training-Free Confidence Generation, which derives confidence directly from a monocular depth estimation model. Furthermore, we propose Adaptive Disparity Selection, which dynamically adjusts the disparity distribution to prevent excessive occlusions and foreground distortions. By ensuring a wider yet realistic disparity range, this component enhances the model's ability to generalize across diverse scenarios. + +By integrating these components, ZeroStereo enables efficient and high-quality stereo image generation, leading to state-of-the-art zero-shot stereo matching. Remarkably, our method achieves this performance with a dataset volume comparable to Scene Flow [29], demonstrating its ability to generate highly effective training data without requiring large-scale real-world stereo pairs. + +Our main contributions can be summarized as follows: + +- We propose a novel stereo image generation pipeline ZeroStereo for zero-shot stereo matching, including a finetuned diffusion inpainting model adapting for complex inpainting masks in stereo matching. +- We propose Training-Free Confidence Generation and Adaptive Disparity Selection to improve stereo training stability and enhance disparity diversity. +- We demonstrate that models trained with our pipeline achieve state-of-the-art zero-shot generalization performance using only a synthesized dataset volume comparable to Scene Flow. + +# 2. Related Work + +Deep Stereo Matching. The advancement of deep learning has significantly improved stereo matching. Early methods [2, 6, 42], such as DispNet [29] and GC-Net [18], employed CNNs to construct cost volumes over a predefined disparity range. More recently, iterative refinement-based methods [25, 57, 58, 60], inspired by RAFT [44], have been introduced to iteratively update disparity predictions, improving accuracy and robustness. Additionally, transformer-based models [13, 22] leverage self-attention mechanisms to capture long-range feature dependencies, enabling more effective cost volume aggregation. + +Zero-shot Generalization in Stereo Matching. Despite these advancements, deep stereo models often struggle with generalization to real-world scenarios. DSMNet [64] addresses domain shifts by introducing domain normalization layers and non-local graph-based filters to enhance feature robustness. GraftNet [26] improves generalization by incorporating pre-trained features from large-scale datasets, while ITSA [10] mitigates shortcut learning using an information-theoretic approach. Inspired by masked representation learning, Rao et al. [38] propose a masking-based strategy to enhance stereo feature learning. Another line of research focuses on self-supervised learning using unlabeled images. Luo et al. [28] pioneers single-view stereo training on the KITTI dataset, while MfS-Stereo [54] generates stereo pairs from monocular images to enable training without ground-truth disparities. NeRF-Stereo [47] introduces NeRF to generate stereo images from 3D scene reconstructions. + +Diffusion Models for Image Synthesis. Denoising Diffusion Probabilistic Models (DDPMs) [15] have demonstrated success in image synthesis by progressively refining images through a denoising process. Latent Diffusion Models (LDMs) [39] further improve efficiency by performing diffusion steps in a lower-dimensional latent space. ControlNet [66] extends these models by introducing spatial conditioning mechanisms for better control over generated content. RePaint [27] proposes an inpainting method based on pre-trained DDPMs, showcasing the effectiveness of diffusion models in restoring missing visual details. + +![](images/cd474383a63cb18c0e8d220121e95f0ca871e7e1289ec189bca6af6b31c691a9.jpg) +Figure 2. Overview of ZeroStereo. Given a left image, a monocular depth estimation model infers the normalized inverse depth. Our Training-Free Confidence Generation (Sec. 3.3) and Adaptive Disparity Selection (Sec. 3.4) modules extract the confidence and pseudo disparity. Forward warping is then applied to generate a warped image and corresponding masks, which are processed by a diffusion inpainting model (Sec. 3.2) to synthesize the right image. The final stereo images and pseudo labels are used for stereo training (Sec. 3.5). + +# 3. Method + +In this section, we present the overview of our ZeroStereo pipeline (Fig. 2) and details of our proposed modules. + +# 3.1. Overview + +Given a single image as the left image $\mathbf{I}_l$ , we first obtain a normalized inverse depth map $\mathbf{D}$ using a monocular depth estimation model (we use Depth Anything V2 [63], referred to as DAv2). This depth map is then converted into a pseudo disparity map $\mathbf{d}$ via our Adaptive Disparity Selection (ADS) module. Using the forward warping technique from [54], we generate a warped image $\tilde{\mathbf{I}}_r$ , a non-occlusion mask $\mathbf{M}_{noc}$ (pixels only visible in $\mathbf{I}_l$ ), and an inpainting mask $\mathbf{M}_{inp}$ (pixels invisible in $\mathbf{I}_l$ ). $\tilde{\mathbf{I}}_r$ and $\mathbf{M}_{inp}$ are then processed by a fine-tuned diffusion inpainting model to synthesize a high-quality right image $\mathbf{I}_r$ . To improve training stability, we introduce the Training-Free Confidence Generation (TCG) module, which computes confidence $\mathbf{C}$ . Finally, the synthesized stereo image pairs and associated pseudo labels are used to train stereo matching models. + +# 3.2. Image Inpainting + +We fine-tune a diffusion inpainting model based on Stable Diffusion V2 Inpainting (SDv2I) [39]. Although the pretrained inpainting model can be directly applied, it is not specifically designed for stereo image synthesis. There exist differences between standard image inpainting and image inpainting in stereo matching. + +First, there is no explicit textual guidance for inpainting. As a text-to-image model, SDv2I is trained on both text-conditioned and unconditioned data. However, no reliable textual prompt effectively directs the model to inpaint occluded regions in stereo matching. Second, unlike standard image inpainting, which typically restores or replaces specific objects or regions, occlusion masks in stereo matching exhibit diverse and irregular shapes. As a result, directly applying a pre-trained model yields suboptimal performance, necessitating fine-tuning to achieve effective results. + +Fine-tuning Protocol. For fine-tuning, we utilize synthetic stereo datasets like Scene Flow [29] which provide dense disparity maps as ground truth. Similar to Marigold [17], synthetic data is essential because it offers dense and complete ground truth, enabling per-pixel warping. Moreover, synthetic images are free from real-world noise, ensuring cleaner training data. Given a warped image $\tilde{\mathbf{I}}_r$ , an inpainting mask $\mathbf{M}_{inp}$ , and a right image $\mathbf{I}_R$ , we employ a frozen Variational Auto-Encoder (VAE) [19] to encode $\tilde{\mathbf{I}}_r$ and $\mathbf{I}_R$ into the latent space. The inpainting mask $\mathbf{M}_{inp}$ is resized to match the latent space resolution. We then sample Gaussian noise $\epsilon$ and add it to the latent right image. Finally, these latent features and the resized inpainting mask are concatenated as input to the U-Net, which predicts noise $\tilde{\epsilon}$ . The network is optimized using an L2 loss function: + +$$ +\mathcal {L} _ {u} = \left| \left| \tilde {\epsilon} - \epsilon \right| \right| _ {2} ^ {2} \tag {1} +$$ + +![](images/c90e9a3b186406ec47aba107efd53f7f9d62bd090f94cbe1f3bfac43be6b0877.jpg) +Figure 3. Overview of our diffusion inpainting protocol. During training, we freeze VAE and only fine-tune U-Net. + +Inference Protocol. Given a warped image $\tilde{\mathbf{I}}_r$ and an inpainting mask $\mathbf{M}_{inp}$ , we first encode them into the latent space. Next, we sample standard Gaussian noise to initialize the latent right image and concatenate it with the encoded inputs for the U-Net. During inference, we iteratively denoise the latent representation over $\mathbf{T}$ steps. Finally, the VAE decoder reconstructs the denoised image $\mathbf{I}_d$ , yielding the final inpainted image $\mathbf{I}_r$ : + +$$ +\mathbf {I} _ {r} = \mathbf {M} _ {i n p} \odot \mathbf {I} _ {d} + (1 - \mathbf {M} _ {i n p}) \odot \tilde {\mathbf {I}} _ {r} \tag {2} +$$ + +# 3.3. Training-Free Confidence Generation + +Assessing confidence in depth predictions remains challenging for previous monocular depth estimation models, which require additional training or auxiliary modules. Some post-processing methods rely on gradients [16] or probabilistic distributions [56] to estimate uncertainty. + +Modern monocular depth estimation models [23, 62, 63] tend to predict relative depth, often represented as inverse depth in disparity space. This representation captures the relative distances between pixels, independent of camera parameters. Therefore, when an image is flipped horizontally, the predicted relative depth between pixels is expected to remain unchanged. + +Given a left image $\mathbf{I}_l$ , we apply a horizontal flip operation $\mathbf{H}(\mathbf{x})$ to obtain the flipped image $\mathbf{I}_l'$ . Both images are then processed separately by the monocular depth estimation model to generate their respective relative depth maps. Since we use DAv2 [63], which does not impose constraints on its predicted depth range, we normalize these outputs into the normalized inverse depth maps $\mathbf{D}$ and $\mathbf{D}'$ . The confidence $\mathbf{C}$ of $\mathbf{D}$ is then measured: + +$$ +\begin{array}{l} \mathbf {u} = 1 - \left| \mathbf {D} - \mathbf {H} ^ {- 1} \left(\mathbf {D} ^ {\prime}\right) \right| \\ \mathbf {C} = \frac {\mathbf {u} - \operatorname* {m i n} (\mathbf {u})}{\operatorname* {m a x} (\mathbf {u}) - \operatorname* {m i n} (\mathbf {u})} \tag {3} \\ \end{array} +$$ + +As shown in Fig. 4, low-confidence regions typically appear along edges, textureless areas, and thin objects, which are also ambiguous in stereo matching. These unreliable labels are suppressed to mitigate their impact on learning. + +![](images/ce433479b598b138ddad17e55abdc55dc4c821f17375c263141e71e58dc0081b.jpg) +Figure 4. Visualization of confidence map. + +![](images/8e13590c146958591aa3f5a790bf20a28a45f33b8cb06a557365dcbd4b0d1eda.jpg) + +![](images/ce75c63b9bbc72999dbf4396de5442dd2a65ca7bfa92aa0532232ac46b770179.jpg) + +# 3.4. Adaptive Disparity Selection + +The previous method, MfS-Stereo [54], generates disparity maps by uniformly sampling the maximum disparity value from the range $[d_{min}, d_{max}]$ . However, this fixed-range approach introduces several limitations. + +First, when the image resolution is low, the disparity-to-width ratio becomes relatively large, potentially causing foreground distortion during forward warping or failure in the diffusion inpainting model due to excessive occlusion. Second, when the image resolution is high, the disparity-to-width ratio becomes relatively small, reducing the perceptible differences between the left and right images. + +Therefore, we compute the disparity map $\mathbf{d}$ by multiplying $\mathbf{D}$ with the scaling factor $\mathbf{s} \cdot \mathbf{w}$ , where $\mathbf{w}$ represents the image width and $\mathbf{s}$ is sampled from the distribution: + +$$ +\mathbf {s} \in \left\{ \begin{array}{l l} (\mathbf {c} - 2 \mathbf {r}, \mathbf {c} - \mathbf {r}), & \mathbf {p} = \mathbf {p} _ {s} \\ (\mathbf {c} - \mathbf {r}, \mathbf {c} + \mathbf {r}), & \mathbf {p} = \mathbf {p} _ {c} \\ (\mathbf {c} + \mathbf {r}, \mathbf {c} + 2 \mathbf {r}), & \mathbf {p} = \mathbf {p} _ {l} \end{array} \right. \tag {4} +$$ + +where $\mathbf{c}$ is the center, $\mathbf{r}$ is the radius, and $\mathbf{p}_i$ ( $i \in \{s, c, l\}$ ) is the probability. This sampling strategy ensures that most warped images maintain high-quality while occasionally introducing extreme disparity values to enhance the robustness of stereo training. Furthermore, since single-image datasets vary in resolution, this approach allows for the generation of large disparities, effectively covering a wide range of scenarios. + +![](images/7946b9c71feae24eaa788b6eca52eadc5bb384b4e877889e6b9ac8438ebeee85.jpg) +Warped Image + +![](images/d88f8122197af460cc50fe6ef5c4913076a49eff6a8b0da5aed86bb332df3873.jpg) +Inpainting Mask + +![](images/d78cb9dfbe59f367a593f243bddbba466cb042cca830f959468df6c77f96a43f.jpg) +Inpainted w/MfS + +![](images/bfec235965d6d98521596f1a4d045f543089f70e80db0033217228cfe893a0aa.jpg) +Inpainted w/ pre-trained SDv2I + +![](images/26bead5d900323d33a906d939d555312a69529fd02f2523281e0b17465c0118a.jpg) +Inpainted w/ fine-tuned SDv2I + +![](images/26147decd60037dc993c544d728c07c7dcc1de8581610a6bb6d10020d7733a09.jpg) +Figure 5. Visualization of different inpainting methods. +Figure 6. Visualization of StereoDiffusion [51] and our fine-tuned SDv2I. + +# 3.5. Stereo Training + +Given a stereo pair $(\mathbf{I}_l, \mathbf{I}_r)$ , a disparity map $\mathbf{d}$ , a confidence map $\mathbf{C}$ , a non-occlusion mask $\mathbf{M}_{noc}$ , an inpainting mask $\mathbf{M}_{inp}$ , and an estimated disparity map $\tilde{\mathbf{d}}$ , we train stereo matching models using our proposed ZeroStereo loss. + +Disparity Loss. We adopt the same L1 loss as used in previous supervised stereo matching methods: + +$$ +\mathcal {L} _ {d} = \left| \left| \tilde {\mathbf {d}} - \mathbf {d} \right| \right| _ {1} \tag {5} +$$ + +Non-occlusion Photometric Loss. We backward warp $\mathbf{I}_r$ using the estimated disparity $\tilde{\mathbf{d}}$ to obtain $\mathbf{I}_l^r$ . The ordinary photometric loss is then computed as: + +$$ +\mathcal {L} _ {p} = \beta \cdot \frac {1 - S S I M (\mathbf {I} _ {l} , \mathbf {I} _ {l} ^ {r})}{2} + (1 - \beta) \cdot | | \mathbf {I} _ {l} - \mathbf {I} _ {l} ^ {r} | | _ {1} \tag {6} +$$ + +$\mathbf{I}_l^r$ includes pixels inpainted by the diffusion inpainting model. To exclude these pixels, we backward warp $1 - \mathbf{M}_{inp}$ to obtain $\tilde{\mathbf{M}}_{inp}^r$ . Besides, $\mathbf{I}_l^r$ contains ghosting artifacts due to backward warping, which we filter using $\mathbf{M}_{noc}$ . Thus, our loss is computed as follows: + +$$ +\mathcal {L} _ {n p} = \mathbf {M} _ {n o c} \odot \tilde {\mathbf {M}} _ {i n p} ^ {r} \odot \mathcal {L} _ {p} \tag {7} +$$ + +ZeroStereo Loss. The above two terms are summed as: + +$$ +\mathcal {L} _ {\text {Z e r o}} = \mathbf {C} \odot \mathcal {L} _ {d} + \mu \cdot (1 - \mathbf {C}) \odot \mathcal {L} _ {n p} \tag {8} +$$ + +# 4. Experiments + +In this section, we present our implementation details, evaluation datasets, ablation study, and experimental results. + +# 4.1. Implementation Details + +All experiments are implemented with PyTorch on NVIDIA RTX 4090 GPUs. + +Diffusion Inpainting Model Fine-tuning. We utilize the Stable Diffusion V2 Inpainting (SDv2I) [39], disabling text conditioning and applying the DDPM noise scheduler [15] with 1,000 diffusion steps. We use a collection of synthetic stereo datasets, including Tartan Air [52], CREStereo Dataset [20], Scene Flow [29], VKITTI 2 [1], etc. Fine-tuning takes 50K steps with a batch size of 32 + +
BaselineADSInpaintingTCG\( {\mathcal{L}}_{Zero} \)KITTI-15 AllMidd-T (H) NocETH3D Noc
EPE>3pxEPE>2pxEPE>1px
✓1.524.892.718.410.252.38
✓1.244.842.287.460.282.27
✓1.444.852.347.590.231.92
✓1.094.782.137.270.272.16
✓✓1.064.742.266.680.232.05
✓✓✓1.054.712.187.110.232.01
✓✓✓✓1.044.732.097.070.221.90
+ +Table 1. Ablation study of proposed modules trained with IGEV-Stereo [58]. Baseline denotes that we train the model under the pipeline in [54]. ADS denotes our Adaptive Disparity Selection module and TCG denotes our Training-free Confidence Generation module. + +
DatasetDiWCOCODIODEADE20KMapillary
Mean18.8920.7033.8331.0681.34
Max75.0096.00153.45627.54751.54
+ +Table 2. Disparity statistics results based on ADS. + +
MethodKITTI-15 AllMidd-T (H) NocETH3D Noc
EPE>3pxEPE>2pxEPE>1px
Pre-trained1.184.772.277.210.251.90
Fine-tuned1.064.742.266.680.232.05
+ +Table 3. Ablation study of SDv2I. + +
MethodResolution (px)Memory (G)Time (s)
RePaint [27]256 × 2562.7156.5
StereoDiffusion [51]512 × 51214.631.2
Ours512 × 5125.81.9
+ +gradient accumulation for 4 steps). We use the AdamW optimizer and a one-cycle learning rate schedule with a learning rate of $2\mathrm{e} - 5$ . Besides, we apply a crop size of $512 \times 512$ and symmetric color augmentation. + +Stereo Image Generation. We use the DDIM scheduler [43] and perform 50 sampling steps. Following MfS-Stereo [54], we sample images from Depth in the Wild [4], COCO 2017 [24], DIODE [50], ADE20K [68], and Mapillary Vistas [34]. We randomly sample $35\mathrm{K}$ to create a dataset called MfS35K. For disparity generation, we set: $\mathbf{c} = 0.1$ , $\mathbf{r} = 0.05$ , $\mathbf{p}_c = 0.8$ , and $\mathbf{p}_s = \mathbf{p}_l = 0.1$ . + +Stereo Matching Model Training. We use RAFT-Stereo [25] and IGEV-Stereo [58]. Models are trained on MfS35K with a batch size of 8 and a crop size of $384 \times 512$ . We follow all the source codes' settings and train 200k steps from scratch. In addition to the data augmentation from RAFT-Stereo, we introduce Gaussian augmentation on the right image, as done in MfS-Stereo [54]. For ZeroStereo loss, we set $\beta = 0.85$ , $\mu = 0.1$ the same as NeRF-Stereo [47]. + +# 4.2. Evaluation Datasets + +KITTI 2015 [30] includes 200 training pairs with lidar ground truth for outdoor driving scenarios. ETH3D [41] contains 27 training pairs of grayscale images, covering out + +Table 4. Ablation study of different synthesis methods. + +
DatasetSizeKITTI-15 >3pxMidd-T >2pxETH3D >1px
Falling Things [49]62K5.145.8128.63
CREStereo [20]200K6.2610.912.56
Tartan Air [52]307K4.915.472.69
FoundationStereo [55]1106K4.695.112.52
MfS35K (Ours)35K4.534.452.13
+ +Table 5. Ablation study of datasets. + +![](images/ca93b1c1ce242320d0e1906e6dbe6e1054c5ea16db3b5d8afe04d71fa7447f1e.jpg) +Figure 7. Visualization of different disparity selection. + +![](images/a27a4d7370111bdfbac13330cdf8e125a6b61b87f8ff4fe2190cb2f1fdc78c27.jpg) + +door and indoor scenarios. For Middlebury [40], we select the Middlebury V3 Benchmark Training Set (Midd-T), which consists of 15 training pairs for high-resolution indoor scenarios. + +# 4.3. Ablation Study + +In this section, we evaluate models with different settings to verify the effectiveness of our proposed pipeline. + +Effectiveness of proposed modules. Tab. 1 shows the ablation study of our proposed modules. By adding ADS or TCG, we observe a notable reduction in EPE for both KITTI-15 and Midd-T. ADS improves the model's ability to handle large disparities. As shown in Tab. 2, the disparity range adjusts adaptively according to the dataset resolution. TCG suppresses unreliable labels, particularly in edges and textureless regions. However, adding Inpainting alone results in only a slight improvement. As shown in Fig. 7, the large disparity ratio causes separation and distortion in the foreground, which hinders the effectiveness of the diffusion inpainting model. When ADS is combined with Inpainting, we observe a significant performance improvement. + +
MethodKITTI-15 >3pxMidd-T F (>2px Noc)H (>2px Noc)ETH3D >1px
AllNocD <192D <AllD <192D <AllAllNoc
NS-IGEV-Stereo5.885.5815.8927.918.5510.674.023.58
Zero-IGEV-Stereo (Ours)4.734.5713.0326.784.737.072.641.90
NS-RAFT-Stereo‡ [47]5.415.2316.6112.046.406.452.952.55
NS-RAFT-Stereo5.655.4415.0513.419.099.443.302.79
Zero-RAFT-Stereo (Ours)4.534.339.518.364.214.452.752.13
+ +![](images/2552e83c101248efd4c157912e0b84e06eb0f22eeaf510cecb5f2117117582f4.jpg) +Figure 8. Visualization of Middlebury, including Midd-T, 2014 and 2021. + +Table 6. Comparison with NeRF-Stereo [47]. Models are trained with the same augmentation. Exception: $\ddagger$ official weights. + +
DatasetKITTI-15 AllMidd-T NocETH3D Noc
EPE>3pxEPE>2pxEPE>1px
Scene Flow [29]1.448.172.4915.810.293.79
NS65K [47]1.397.602.2411.990.324.58
MfS35K (Ours)1.377.412.3511.920.243.38
+ +Table 7. Analysis of errors on edge (RAFT-Stereo). + +When ADS, Inpainting, and TCG are all added, the performance consistently improves. Moreover, by introducing the final ZeroStereo Loss, the model can learn with $\mathcal{L}_{Zero}$ in low-confidence regions. This loss helps maintain the model's robustness, enabling it to achieve optimal performance across various datasets. A more detailed analysis can be found in our supplementary materials. + +Effectiveness of fine-tuned SDv2I. As shown in Fig. 5, pixels inpainted using our fine-tuned SDv2I preserve minimal noise and maintain the semantic structure closest to the background. As shown in Tab. 3, the fine-tuned SDv2I outperforms the pre-trained model. It suggests that fine-tuning enhances the inpainting model's ability to capture the semantic structure of the background more accurately. + +Comparison with other synthesis methods. StereoDiffusion [51] introduces a training-free method for generating stereo images using the pre-trained SDv2. However, the inherent inconsistency arises because the warping oper + +ation is performed in the latent space. As shown in Fig. 6, StereoDiffusion suffers from structural distortions, texture inconsistencies, and poor occlusion handling, leading to unrealistic right-image generation. As shown in Tab. 4, our fine-tuned SDv2I is significantly faster and more memory-efficient than StereoDiffusion [51], while still achieving high-resolution synthesis. A more detailed discussion is available in our supplementary materials. + +Comparison with other synthetic datasets. In recent years, many synthetic datasets with better diversity and rendering realism have been proposed. As shown in Tab. 5, our MfS35K outperforms others, despite being significantly smaller. It reveals that rather than the absolute size of datasets, the diversity of scenarios is more beneficial for zero-shot generalization. + +# 4.4. Comparison with NeRF-Stereo + +We compare our method with NeRF-Stereo [47], a leading method for generating stereo images. As shown in Tab. 6, all models are trained with the same data augmentation, except for NS-RAFT-Stereo $\ddagger$ , which uses official weights. Since the official IGEV-Stereo limits $\mathcal{L}_d$ supervision to a disparity of 192, we split the Midd-T table by disparity to provide a more detailed evaluation. + +
MethodKITTI-15 >3pxF (>2px)Midd-T H (>2px)Q (>2px)ETH3D >1px
AllNocAllNocAllNocAllNocAllNoc
Training SetScene Flow with GT
DSMNet [64]5.505.1941.9638.5418.7414.4913.759.444.033.62
CFNet [42]5.995.7935.2130.0521.9917.6914.2110.516.085.48
Graft-PSMNet [26]5.345.0039.9236.3017.6513.3613.929.2311.4310.70
ITSA-CFNet [10]4.734.6734.0130.1416.4812.3212.288.545.435.17
HVT-PSMNet [3]4.844.6340.7437.6015.6612.5510.127.006.075.65
RAFT-Stereo [25]5.475.2715.6311.9411.208.6610.257.442.602.29
IGEV-Stereo [58]6.035.7630.9428.9811.909.458.886.204.043.60
NMRF-Stereo [12]5.315.1437.6335.2513.3610.907.875.073.803.48
Mocha-Stereo [5]5.975.7330.2328.2610.189.457.964.874.023.47
DKT-RAFT [65]4.954.7416.059.2610.186.1010.396.642.772.53
Former-RAFT [67]5.184.93--13.2710.298.515.613.963.50
Training SetReal-world data without GT
Mfs-PSMNet [54]5.184.9126.4220.9117.5613.4512.079.098.177.44
NS-RAFT-Stereo [47]5.415.2316.3812.049.706.458.094.852.952.55
Zero-IGEV-Stereo (Ours)4.734.5729.4726.789.717.077.074.462.641.90
Zero-IGEV-Stereo* (Ours)4.894.7318.8314.878.455.546.994.382.852.00
Zero-RAFT-Stereo (Ours)4.534.3312.408.367.864.457.244.502.752.13
+ +Table 8. Zero-shot generalization benchmark. DKT-RAFT [65] is trained on SceneFlow [29] and fine-tuned on Booster [36]. Zero-IGEV-Stereo* denotes that we expand the $\mathcal{L}_d$ supervision same as RAFT-Stereo [25]. We highlight first, second, third bests. + +Re-training NS-RAFT-Stereo with our data augmentation shows no improvement, confirming that the gains are not solely due to augmentation. Zero-RAFT-Stereo outperforms NS-RAFT-Stereo $\ddagger$ by over $20\%$ , with only a $10\%$ drop in Midd-T (F) for $\mathbf{D} < 192$ , whereas NS-RAFT-Stereo $\ddagger$ declines more, likely due to its dataset's imperfect reconstruction of distant objects. Fig. 8 highlights NS-RAFT-Stereo's failures in textureless regions, while our Zero-RAFT-Stereo shows over $40\%$ improvement in handling such cases. Moreover, as shown in Tab. 7, the model trained on MfS35K surpasses both the synthetic SceneFlow [29] and the NeRF-based NS65K [47], achieving the lowest edge error and superior edge accuracy. + +# 4.5. Zero-shot Generalization Benchmark + +Following NeRF-Stereo [47], we construct a zero-shot generalization benchmark. All methods are evaluated across the entire disparity range. For Zero-IGEV-Stereo, we train two versions: one using the original code settings for disparity supervision, and the other expanding the supervised range, consistent with RAFT-Stereo [25]. + +As shown in Tab. 8, our models demonstrate state-of-the-art zero-shot generalization performance across multiple datasets, both under the SceneFlow with ground truth (GT) and Real-world data without GT. Notably, Zero-RAFT-Stereo achieves the best or near-best results, particularly excelling in handling complex, real-world scenes. + +Zero-IGEV-Stereo*, with an expanded supervised range of $\mathcal{L}_d$ , shows improved results on Middlebury's large-disparity scenarios, although this leads to a slight performance trade-off on other datasets. + +# 5. Conclusion + +We propose ZeroStereo, a novel stereo data generation pipeline for zero-shot stereo matching. The fine-tuned SDv2I adapts to complex inpainting masks and recovers background details. To handle unreliable pseudo labels, the TCG module leverages the spatial invariance of relative depth to compute confidence, helping to suppress uncertain labels. Besides, the ADS module generates a broader disparity distribution while avoiding foreground distortion. Finally, experiments demonstrate that our models achieve state-of-the-art zero-shot generalization performance. + +Limitations. The fine-tuned SDv2I still struggles in some complex scenarios, and there may be occasional color inconsistencies due to fine-tuning on synthetic datasets. Furthermore, forward warping performs poorly in ill-posed regions, such as transparent areas or net-like objects. + +Acknowledgement. This research is supported by the National Key R&D Program of China (2024YFE0217700), National Natural Science Foundation of China (62472184, 623B2036), the Fundamental Research Funds for the Central Universities, and the Innovation Project of Optics Valley Laboratory (Grant No. OVL2025YZ005). + +# References + +[1] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv preprint arXiv:2001.10773, 2020. 5 +[2] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 1, 2 +[3] Tianyu Chang, Xun Yang, Tianzhu Zhang, and Meng Wang. Domain generalized stereo matching via hierarchical visual transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9559-9568, 2023. 1, 8 +[4] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 6 +[5] Ziyang Chen, Wei Long, He Yao, Yongjun Zhang, Bingshu Wang, Yongbin Qin, and Jia Wu. Mocha-stereo: Motif channel attention network for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27768-27777, 2024. 8 +[6] Junda Cheng, Xin Yang, Yuechuan Pu, and Peng Guo. Region separable stereo matching. IEEE Transactions on Multimedia, 25:4880-4893, 2022. 2 +[7] Junda Cheng, Gangwei Xu, Peng Guo, and Xin Yang. Coatsnet: Fully exploiting convolution and attention for stereo matching by region separation. International Journal of Computer Vision, 132(1):56-73, 2024. 1 +[8] Junda Cheng, Wei Yin, Kaixuan Wang, Xiaozhi Chen, Shijie Wang, and Xin Yang. Adaptive fusion of single-view and multi-view depth for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10138-10147, 2024. 2 +[9] Junda Cheng, Longliang Liu, Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Yong Deng, Jinliang Zang, Yurui Chen, Zhipeng Cai, and Xin Yang. Monster: Marry monodepth to stereo unleashes power. arXiv preprint arXiv:2501.08643, 2025. 1 +[10] WeiQin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, Alireza Bab-Hadiashar, and David Suter. Itsa: An information-theoretic approach to automatic shortcut avoidance and domain generalization in stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13022-13032, 2022. 1, 2, 8 +[11] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 1, 2 +[12] Tongfan Guan, Chen Wang, and Yun-Hui Liu. Neural markov random field for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2024. 8 +[13] Weiyu Guo, Zhaoshuo Li, Yongkui Yang, Zheng Wang, Russell H Taylor, Mathias Unberath, Alan Yuille, and Yingwei Li. Context-enhanced stereo transformer. In European Con + +ference on Computer Vision, pages 263-279. Springer, 2022. 2 +[14] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view view synthesis in the wild with learned adaptive multiplane images. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 1, 2, 3 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 5 +[16] Julia Hornauer and Vasileios Belagiannis. Gradient-based uncertainty for monocular depth estimation. In European Conference on Computer Vision, pages 613-630. Springer, 2022. 4 +[17] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 2, 3 +[18] Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision, pages 66-75, 2017. 2 +[19] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[20] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Zwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 5, 6 +[21] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 2 +[22] Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6197-6206, 2021. 2 +[23] Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji-aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth for 4k resolution accurate metric depth estimation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 17070–17080, 2025. 4 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[25] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 1, 2, 6, 8 + +[26] Biyang Liu, Huimin Yu, and Guodong Qi. Grafnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13012-13021, 2022. 1, 2, 8 +[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 2, 6 +[28] Yue Luo, Jimmy Ren, Mude Lin, Jiahao Pang, Wenxiu Sun, Hongsheng Li, and Liang Lin. Single view stereo matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 155-163, 2018. 2 +[29] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 1, 2, 3, 5, 7, 8 +[30] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 6 +[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1 +[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023. 1 +[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2 +[34] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 6, 1 +[35] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1 +[36] Pierluigi Zama Ramirez, Fabio Tosi, Matteo Poggi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Open challenges in deep stereo: the booster dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21168-21178, 2022. 8 +[37] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 2 + +[38] Zhibo Rao, Bangshu Xiong, Mingyi He, Yuchao Dai, Renjie He, Zhelun Shen, and Xing Li. Masked representation learning for domain generalized stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5435-5444, 2023. 1, 2 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 1 +[40] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 31-42. Springer, 2014. 1, 6 +[41] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 1, 6 +[42] Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906-13915, 2021. 2, 8 +[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.6 +[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 2 +[45] Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learning to adapt for stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9661-9670, 2019. 1 +[46] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 195-204, 2019. 1 +[47] Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Matteo Poggi. Nerf-supervised deep stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 855-866, 2023. 2, 6, 7, 8 +[48] Fabio Tosi, Luca Bartolomei, and Matteo Poggi. A survey on deep stereo matching in the twenties. International Journal of Computer Vision, pages 1-32, 2025. 1 +[49] Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 2038-2041, 2018. 6 +[50] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Moham- + +madreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 6 +[51] Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7416-7425, 2024. 5, 6, 7, 1, 2 +[52] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 1, 5, 6 +[53] Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701-19710, 2024. 1 +[54] Jamie Watson, Oisin Mac Aodha, Daniyar Turmukhambetov, Gabriel J Brostow, and Michael Firman. Learning stereo from single images. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 722-740. Springer, 2020. 2, 3, 4, 6, 8, 1 +[55] Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. Foundationstereo: Zero-shot stereo matching. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5249-5260, 2025. 6 +[56] Mochu Xiang, Jing Zhang, Nick Barnes, and Yuchao Dai. Measuring and modeling uncertainty degree for monocular depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 4 +[57] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12981-12990, 2022. 1, 2 +[58] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. 2, 6, 8 +[59] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[60] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: Iterative multirange geometry encoding volumes for stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2 +[61] Guorun Yang, Xiao Song, Chaoqin Huang, Zhidong Deng, Jianping Shi, and Bolei Zhou. Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios. + +In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 899-908, 2019. 2 +[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 4 +[63] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2024. 3, 4, 1 +[64] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 420-439. Springer, 2020. 1, 2, 8 +[65] Jiawei Zhang, Jiahe Li, Lei Huang, Xiaohan Yu, Lin Gu, Jin Zheng, and Xiao Bai. Robust synthetic-to-real transfer for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20247-20257, 2024. 8 +[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[67] Yongjian Zhang, Longguang Wang, Kunhong Li, Yun Wang, and Yulan Guo. Learning representations from foundation models for domain generalized stereo matching. In European Conference on Computer Vision, pages 146-162. Springer, 2024. 8 +[68] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6 +[69] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 1 \ No newline at end of file diff --git a/zerostereozeroshotstereomatchingfromsingleimages/images.zip b/zerostereozeroshotstereomatchingfromsingleimages/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d56a0be6c6ac60fc0b793549f64614a2ed4c7dfe --- /dev/null +++ b/zerostereozeroshotstereomatchingfromsingleimages/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f52ab77b41abd0a04b97b7ae6ed5059b300699678a446f795f8237eafcb050b6 +size 1074182 diff --git a/zerostereozeroshotstereomatchingfromsingleimages/layout.json b/zerostereozeroshotstereomatchingfromsingleimages/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a6347f13cf36cd1f5bcd6c8df4a382367c800031 --- /dev/null +++ b/zerostereozeroshotstereomatchingfromsingleimages/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89ddbe87cf9484dff93ec38dc401fc424ced81609b3dcdb7d6557719029ab6e0 +size 418562 diff --git a/zerothorderfinetuningofllmsinrandomsubspaces/30dae0e3-0c17-444c-9824-db8146e69eb8_content_list.json b/zerothorderfinetuningofllmsinrandomsubspaces/30dae0e3-0c17-444c-9824-db8146e69eb8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3426a36356ade4939b8441910d01184355f70525 --- /dev/null +++ b/zerothorderfinetuningofllmsinrandomsubspaces/30dae0e3-0c17-444c-9824-db8146e69eb8_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4486198e516e4fc08e078dcf94afad62352afb2479cc82e1e492a6ffd82fb2c8 +size 101375 diff --git a/zerothorderfinetuningofllmsinrandomsubspaces/30dae0e3-0c17-444c-9824-db8146e69eb8_model.json b/zerothorderfinetuningofllmsinrandomsubspaces/30dae0e3-0c17-444c-9824-db8146e69eb8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1b85083ec1c14137903bc5ec91edab0ddece3705 --- /dev/null +++ b/zerothorderfinetuningofllmsinrandomsubspaces/30dae0e3-0c17-444c-9824-db8146e69eb8_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6a1e845a9bb14e4ee4670089846731bdb5f0a979eb37f584f324a6fcd274d6d +size 131142 diff --git a/zerothorderfinetuningofllmsinrandomsubspaces/30dae0e3-0c17-444c-9824-db8146e69eb8_origin.pdf b/zerothorderfinetuningofllmsinrandomsubspaces/30dae0e3-0c17-444c-9824-db8146e69eb8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ae6197f2171ba830f005662b78c375dbc635770c --- /dev/null +++ b/zerothorderfinetuningofllmsinrandomsubspaces/30dae0e3-0c17-444c-9824-db8146e69eb8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4f1b96deb4ed70be7b0f90ae1a0f4aa73cc4fc633b0564220ef9de7d1ff43bf +size 625527 diff --git a/zerothorderfinetuningofllmsinrandomsubspaces/full.md b/zerothorderfinetuningofllmsinrandomsubspaces/full.md new file mode 100644 index 0000000000000000000000000000000000000000..52b336cce6029bc45efb0fa750e2bba1b7d6c8da --- /dev/null +++ b/zerothorderfinetuningofllmsinrandomsubspaces/full.md @@ -0,0 +1,443 @@ +# Zeroth-Order Fine-Tuning of LLMs in Random Subspaces + +Ziming Yu1, Pan Zhou2, Sike Wang1, Jia Li1,4†, Mi Tian3, Hua Huang1,4 +1 Beijing Normal University 2 Singapore Management University 3 TAL Education Group +4 Engineering Research Center of Intelligent Technology and Educational Application (MOE) {zimingyu, sikewang}@mail.bnu.edu.cn, {jiali, huahuang}@bnu.edu.cn, panzhou@smu.edu.sg, tianmi@tal.com + +# Abstract + +Fine-tuning Large Language Models (LLMs) has proven effective for a variety of downstream tasks. However, as LLMs grow in size, the memory demands for backpropagation become increasingly prohibitive. Zeroth-order (ZO) optimization methods offer a memory-efficient alternative by using forward passes to estimate gradients, but the variance of gradient estimates typically scales linearly with the model's parameter dimension—a significant issue for LLMs. In this paper, we propose the random Subspace Zeroth-order (SubZero) optimization to address the challenges posed by LLMs' high dimensionality. We introduce a low-rank perturbation tailored for LLMs that significantly reduces memory consumption while improving performance. Additionally, we prove that our gradient estimation closely approximates the backpropagation gradient, exhibits lower variance than traditional ZO methods, and ensures convergence when combined with SGD. Experimental results show that SubZero enhances fine-tuning performance and achieves faster convergence compared to standard ZO approaches like MeZO across various language modeling tasks. Code is available at https://github.com/zimingyy/SubZero. + +# 1. Introduction + +Large Language Models (LLMs), such as the GPT and LLaMA series [55, 64], have recently demonstrated impressive capabilities in natural language processing tasks and beyond [1, 52]. These models utilize deep learning, particularly the transformer architecture [56], to learn complex patterns in language data. However, LLMs can struggle with specialized tasks that require domain-specific knowledge [49]. Fine-tuning presents an effective solution by slightly adjusting pre-trained LLMs with domain data, enabling them to adapt to specific tasks more effectively. + +For fine-tuning, first-order (FO) optimizers, such as SGD [3] or Adam [30], are commonly used to achieve + +promising performance on domain datasets. However, as LLMs grow in size, FO optimizers demand increasingly memory consumption due to the gradient computations required by backpropagation (BP) [67]. Additionally, they are unable to directly handle non-differentiable objectives. To enhance memory efficiency, MeZO [40] first introduces the zeroth-order (ZO) optimizer to LLM fine-tuning without BP. It only requires forward passes and calculates gradient estimates using finite differences of training loss values, enabling it to directly handle non-differentiable objectives. Nevertheless, the variance of ZO gradient estimates scales linearly with the perturbation dimension, which corresponds to the number of model parameters. This can become extremely large in LLMs, leading to significant performance degradation compared to FO optimizers [20, 28, 39]. + +Reducing the variance of ZO gradient estimates generally results in faster convergence [62]. There are two main attempts to addressing the high variance of ZO gradient estimates. The first approach involves increasing batch size alongside training steps, which reduces gradient noise and variance in ZO gradient estimates [20, 28]. However, this leads to significant runtime and memory costs due to the large batch size in the later training stages. The second approach focuses on perturbing fewer parameters by employing sparse parameter perturbations, such as random and sparse pruning masks [39] and block-coordinate perturbations [65], or by reducing the number of trainable parameters through techniques like parameter-efficient fine-tuning (PEFT) [40, 65] and tensorized adapters [60]. Recent theoretical advancements have proposed using random projections to lessen the dimensionality dependence in ZO optimizers [31, 43, 46] by applying low-dimensional perturbations in random subspaces. Nonetheless, a major drawback of this approach is the need to store a huge projection matrix that scales with model parameter dimensionality, making it impractical for fine-tuning large LLMs. + +Contributions. In this work, we propose the first random Subspace Zeroth-order (SubZero) optimization to tackle the challenges of high-dimensional LLM fine-tuning. We intro- + +![](images/bac984c17841d8b58bb06565dd0e648f886a26279906ed5c941498ee3a492673.jpg) +(a) Cosine Similarity + +![](images/a4338d382f94eace79469042ad48f347c9fdbfd8d112ddc5c039d82ed3c5c391.jpg) +(b) Relative Variance +Figure 1. Visualization of cosine similarity $\mathbb{E}\left[\cos \mathrm{i}\left(\boldsymbol {g},\hat{\boldsymbol{g}}\right)\right]$ , relative variance $\operatorname {Var}\left[\| \hat{\boldsymbol{g}}\| \right] / \| \boldsymbol {g}\| ^2$ , training loss, and peak total GPU memory cost with OPT-1.3B on SST-2 in the prompt tuning scheme. All three methods utilize a batch size of 16 and run for $20\mathrm{K}$ steps. Here, $\hat{\pmb{g}}$ represents the gradient estimated by MeZO or our SubZero, and $\pmb{g}$ denotes the expected gradient $\mathbb{E}[\hat{g} ]$ . Theorem 1 (b) ensures that SubZero maintains a small distance between $\pmb{g}$ and the BP gradient in a subspace. (a) and (b) demonstrate that SubZero's estimated gradient $\hat{\pmb{g}}$ has lower angle error and variance than MeZO. (c) and (d) indicate that SubZero enhances convergence speed with minimal extra memory usage. + +![](images/ba92ee15364c40804a135581df2d81dbdc460d8753d58baf85c719acdd53daba.jpg) +(c) Training Loss + +![](images/1600881ea2aec5ed141602ff6c1ff000e0f906bec7cda9cb64f2d8e5a413de5c.jpg) +(d) Memory Cost + +duce a low-rank perturbation to estimate the gradient, specifically designed for LLM architecture, leading to reduced memory consumption and enhanced training performance. Our main contributions are as follows. + +Firstly, we propose a layer-wise low-rank perturbation approach for gradient estimation, specifically designed for fine-tuning LLMs. In each layer, we generate a low-rank perturbation matrix by combining two column-orthogonal matrices with a Gaussian random matrix, which is then used for gradient estimation. Unlike traditional ZO methods like MeZO [40] which apply non-low-rank perturbations to the entire model, our approach significantly reduces the variance of gradient estimates and the angle error between the estimated gradient and its expectation, as respectively shown in Fig. 1 (a) and (b). SubZero also improves upon random subspace ZO methods like S-RGF [43] by using smaller and layer-specific low-rank perturbation matrices instead of a large and model-scale projection matrix, thus cutting memory and computational costs. Additionally, we introduce a lazy update strategy, generating perturbations periodically rather than iteratively, further reducing overhead. Besides, we successfully apply SubZero to four popular LLM fin-tuning schemes, highlighting the compatibility of SubZero. + +Secondly, we provide theoretical guarantees for SubZero. We first convert our gradient estimation into an equivalent formulation, highlighting the key differences between our approach and existing traditional ZO methods [40], as well as random subspace ZO methods [43]. Then, we prove that the gradient estimated by SubZero closely approximates the BP gradient, i.e., the ground-truth gradient, and enjoys significantly lower gradient variance than traditional ZO methods like MeZO. Furthermore, we establish the theoretical convergence of SubZero when combined with the SGD optimizer. + +Finally, experimental results demonstrate SubZero's superior performance and memory efficiency compared to other ZO approaches in both full-parameter tuning and parameter-efficient fine-tuning (PEFT) schemes, such as LoRA, prefix tuning, and prompt tuning. For instance, SubZero improves upon MeZO by $7.1\%$ on LLaMA-7B and by $3.2\%$ on OPT1.3B under full-parameter tuning and prompt tuning, while + +maintaining nearly identical memory costs to MeZO. + +# 2. Related Work + +Zeroth-Order Fine-Tuning. ZO optimizers utilize just two forward passes to estimate gradient without BP. Malladi et al. [40] first used ZO optimization to fine-tune LLMs, significantly lowering the GPU hours and memory usage to levels similar to inference, which offers a considerable advantage over FO optimizers. They demonstrated that LLM fine-tuning benefits from a well-structured loss landscape by introducing suitable task-specific prompt templates. Convergence theories for ZO optimization have been elaborated in both convex [19, 25, 41] and non-convex settings [26, 37]. However, these convergence rates typically increase linearly with the number of trainable parameters [19, 25, 26, 37, 41]. + +Recently, more work in ZO has focused on improving the convergence rates and reducing gradient estimation variance for LLM fine-tuning. Increasing batch size can diminish noise in ZO gradient estimation [20, 28]. Perturbing a subset of model parameters also lowers gradient variance. This approach induces sparse parameter perturbations through random and sparse pruning masks [39] or block-coordinate perturbations [65]. Additionally, some approaches tried to reduce trainable parameters through PEFT [40, 65] and tensorized adapters [60]. + +Random Subspace Optimization. To lessen dependence on dimensionality, some research utilizes random projections and low-dimensional perturbations in subspaces [31, 43, 46]. However, these methods are hindered by the need to store a large projection matrix that increases with dimensionality, making it impractical for fine-tuning LLMs. + +Memory-Efficient Fine-Tuning. Fine-tuning generally employs FO optimizers like SGD [3] or Adam [30]. Various approaches have been developed to reduce the memory cost of BP, such as LoRA [24, 58], gradient sparsification [54], low-rank gradient projection [67], and optimizer state quantization [15, 35]. Additional methods to conserve activation and weight memory during forward and backward passes include gradient checkpointing [8], FlashAttention [12], QLoRA [16], and LLM.int8() [14]. + +# 3. Preliminaries + +Here we introduce the most popular ZO optimization approach and existing random subspace optimization methods. Notations. Let non-bold letter like $a$ and $A$ denote a scalar, a boldfaced lower-case letter like $\boldsymbol{w}$ denote a column vector, and a boldfaced upper-case letter like $\boldsymbol{W}$ denote a matrix. $\mathcal{N}(\mathbf{0},\mathbf{I})$ is a multivariate normal distribution with a zero mean vector and an identity covariance matrix. $\operatorname {vec}(W)$ denotes the vectorization of matrix $W$ which reshapes $W$ into a column vector by stacking the columns of $W$ vertically. $\pmb{A}\otimes \pmb{B}$ is the Kronecker product of matrices $\pmb{A}$ and $\pmb{B}$ . $\mathbb{E}[\pmb {x}]$ is the expected value of a random variable $\pmb{x}$ . $\mathrm{Var}[\pmb {x}]$ is the variance of a random variable $\pmb{x}$ . The $\ell_2$ -norm of a vector $\pmb{x}$ is $\| x\| = \sqrt{\sum_{i = 1}^{n}x_i^2}$ . The spectral norm of a matrix $\pmb{A}$ is $\| A\|$ . The Frobenius norm of a matrix $\pmb{A}$ is $\| A\| _F = \sqrt{\langle A,A\rangle}$ . $C_L^{s,p}(\mathcal{S})$ denotes the class of $s$ -th smooth and $p$ -th $L$ -smooth functions over the set $\mathcal{S}$ . bdiag $(A_{1},A_{2},\dots ,A_{l})$ is a block diagonal matrix with diagonal blocks $A_{1},A_{2},\dots ,A_{l}$ . + +We are interested in fine-tuning large LLMs [17]. These models typically comprise multiple layers, with trainable parameter vectors represented as $\pmb{w} = \left[\pmb{w}_1^{\mathsf{T}},\pmb{w}_2^{\mathsf{T}},\dots,\pmb{w}_l^{\mathsf{T}}\right]^{\mathsf{T}}\in \mathbb{R}^{d}$ , where $\pmb{w}_i$ denotes the flattened parameter vector from the $i$ -th layer and $d$ is model parameter dimension. Then training these models involves optimizing the problem: + +$$ +\min _ {\boldsymbol {w}} \mathcal {L} (\boldsymbol {w}), \tag {1} +$$ + +where $\mathcal{L}(\cdot)$ denotes the loss function. + +Zeroth-Order Optimization. ZO optimization is BP-free and estimates gradients via random perturbations. A classical gradient estimator is the simultaneous perturbation stochastic approximation (SPSA) [53], which is defined as + +$$ +\widehat {\nabla} \mathcal {L} (\boldsymbol {w}; \mathcal {B}) = \frac {\mathcal {L} (\boldsymbol {w} + \varepsilon \boldsymbol {z} ; \mathcal {B}) - \mathcal {L} (\boldsymbol {w} - \varepsilon \boldsymbol {z} ; \mathcal {B})}{2 \varepsilon} \boldsymbol {z}, \tag {2} +$$ + +where $\mathcal{L}(\boldsymbol{w};\mathcal{B})$ is the loss on a minibatch $\mathcal{B}$ of size $B$ uniformly sampled from the training dataset $\mathcal{D}$ , $\boldsymbol{z} \in \mathbb{R}^d$ represents a random perturbation sampled from $\mathcal{N}(\mathbf{0},\mathbf{I}_d)$ , and $\varepsilon$ is the perturbation scale. + +The SPSA in Eqn. (2) is an unbiased gradient estimator of the desired gradient $\nabla \mathbb{E}_z[\mathcal{L}(\boldsymbol{w} + \varepsilon \boldsymbol{z})]$ [41]. It only requires two forward passes to estimate the gradient and eliminates BP computation, greatly reducing computation cost and GPU memory. With this estimated gradient, one can integrate with existing FO optimizers like SGD to develop corresponding ZO optimizers, e.g., ZO-SGD defined as: + +$$ +\boldsymbol {w} ^ {t + 1} = \boldsymbol {w} ^ {t} - \eta^ {t} \hat {\nabla} \mathcal {L} \left(\boldsymbol {w} ^ {t}; \mathcal {B} ^ {t}\right), \tag {3} +$$ + +where $\eta^t$ is the learning rate at step $t$ . In practice, MeZO [40] implements ZO-SGD via in-place operations and uses a single random seed to facilitate efficient perturbation regeneration, greatly reducing memory overhead. + +Random Subspace Optimization. Recent theoretical work [43, 46] has explored using low-dimensional perturbations in random subspaces to reduce gradient variances and hence enhance convergence rates. The key to random subspace methods is the generation of the perturbation vector $\tilde{z}$ within a subspace spanned by $P$ : + +$$ +\tilde {z} = P z, \tag {4} +$$ + +where $\pmb{P} \in \mathbb{R}^{d \times q}$ is a random projection matrix with entries drawn from $\mathcal{N}(0,1)$ , $\pmb{z} \in \mathbb{R}^q$ is a low-dimensional random perturbation vector sampled from $\mathcal{N}(\mathbf{0},\pmb{I}_q)$ , and $q < d$ is the dimension of the subspace. Thus, the gradient estimator in the subspace is given as follows: + +$$ +\widehat {\nabla} \mathcal {L} (\boldsymbol {w}, \boldsymbol {P}; \mathcal {B}) = \frac {\mathcal {L} (\boldsymbol {w} + \varepsilon \boldsymbol {P} \boldsymbol {z} ; \mathcal {B}) - \mathcal {L} (\boldsymbol {w} - \varepsilon \boldsymbol {P} \boldsymbol {z} ; \mathcal {B})}{2 \varepsilon} \boldsymbol {P} \boldsymbol {z}. \tag {5} +$$ + +LLMs have a large model size, and thus their training and fine-tuning parameters can be very high-dimensional. This results in an excessively large matrix $\mathbf{P}$ which is $q$ times larger than the model size $d$ in full-parameter tuning [2] and is also large in other fine-tuning schemes e.g., LoRA [24]. Consequently, this approach significantly increases memory requirements and computational complexity. Therefore, it is crucial to develop an efficient subspace construction strategy with minimal memory consumption for LLM fine-tuning. + +# 4. Methodology + +Here we first elaborate on our SubZero, a powerful ZO framework for LLM fine-tuning. Then we present how to integrate SubZero into four representative fine-tuning schemes. + +# 4.1. Random Subspace Optimization for LLM Fine-Tuning + +Our intuition is that exploring update directions in a low-dimensional subspace may result in a reduced variance of the estimated gradient [43, 46] compared to the estimation in the vanilla space as used in MeZO. Moreover, recent work indicates that BP gradients in LLM fine-tuning rapidly converge to a small subspace [23, 40, 66, 67]. Accordingly, we propose the random Subspace Zeroth-order (SubZero) optimization framework tailored for LLM fine-tuning. This framework reduces gradient estimation variance, and minimizes the memory overhead associated with gradient estimation, such as the memory overhead caused by the projection matrix $P$ in Eqn. (5) used in [43, 46]. + +Layer-wise Random Subspace Perturbation. LLMs primarily consist of dense layers that perform matrix multiplication. We denote the trainable parameters of the $i$ -th layer in matrix form as $\boldsymbol{W}_i \in \mathbb{R}^{m_i \times n_i}$ . Then we will explain how to design its low-rank perturbation $\tilde{\boldsymbol{Z}}_i \in \mathbb{R}^{m_i \times n_i}$ . + +We propose a low-rank perturbation strategy for model parameter matrix of each layer, contrasting with previous random subspace methods that focus on the entire + +model's parameters [43, 46]. At each step, we generate a low-dimensional random matrix $Z_{i} \in \mathbb{R}^{r \times r}$ , where $r \ll \min\{m_{i}, n_{i}\}$ , and perform QR decomposition on two Gaussian random matrices with entries sampled from $\mathcal{N}(0,1)$ to create projection matrices $U_{i} \in \mathbb{R}^{m_{i} \times r}$ and $V_{i} \in \mathbb{R}^{n_{i} \times r}$ (see Algorithm 1). Both $U_{i}$ and $V_{i}$ are column-orthogonal matrices. Our experiments in Table 5 indicate that directly using Gaussian random projection matrices yields worse performance than using our designed column-orthogonal matrices. Then we combine these three matrices to yield a low-rank perturbation as follows: + +$$ +\tilde {\boldsymbol {Z}} _ {i} = \boldsymbol {U} _ {i} \boldsymbol {Z} _ {i} \boldsymbol {V} _ {i} ^ {\top}, \tag {6} +$$ + +where $\tilde{Z}_i$ is the perturbation matrix in a subspace spanned by $U_i$ and $V_i$ , and $Z_i$ represents the low-dimensional random perturbation matrix with entries sampled from $\mathcal{N}(0,1)$ . The projection matrices for SubZero can only be derived using the QR decomposition of the weights, activations, ZO gradients, or random matrix, without increasing memory overhead. As shown in Table 1, our proposed random matrix achieves the best performance. While the weight and ZO gradient matrices are feasible, their performance drops significantly. The activation matrix is ineffective due to its batch size dependency, requiring more sophisticated handling. Detailed experimental setups are in Appendix 8.4. + +Let the model consist of $l$ layers, with the parameter matrix set defined as $\mathcal{W} = \{\pmb{W}_i\}_{i=1}^l$ and the perturbation matrix set as $\tilde{\mathcal{Z}} = \{\tilde{\pmb{Z}}_i\}_{i=1}^l$ . Similar to Eqns. (2) and (5), we compute the loss difference: + +$$ +\rho = \frac {\mathcal {L} (\mathcal {W} + \varepsilon \tilde {\mathcal {Z}} ; \mathcal {B}) - \mathcal {L} (\mathcal {W} - \varepsilon \tilde {\mathcal {Z}} ; \mathcal {B})}{2 \varepsilon}. \tag {7} +$$ + +Note that multiplying a set by a scalar means that the scalar is multiplied by each element in the set. The addition of two sets means that the corresponding elements are added. This is only for mathematical expression, and $\rho$ in Eqn. (7) can be calculated by two forward passes through all the layers in practice. Then we obtain the gradient estimate for the $i$ -th layer as + +$$ +\widehat {\nabla} \mathcal {L} \left(\boldsymbol {W} _ {i}; \mathcal {B}\right) = \rho \tilde {\boldsymbol {Z}} _ {i} = \rho \boldsymbol {U} _ {i} \boldsymbol {Z} _ {i} \boldsymbol {V} _ {i} ^ {\top}. \tag {8} +$$ + +In Sec. 5, we analyze the effectiveness of this new gradient estimation (8). Specifically, Theorem 1 proves the close distance between our gradient estimate (8) and the vanilla gradient computed by BP in FO methods, while Theorem 2 shows smaller variance and angle error of our gradient estimate in Eqn. (8) compared to the gradient estimate (2) in MeZO [40]. See more theoretical details in Sec. 5. + +Then, one can use estimated gradient in (8) to replace the gradient in any FO optimizer such as SGD: + +$$ +\boldsymbol {W} _ {i} ^ {t + 1} = \boldsymbol {W} _ {i} ^ {t} - \eta^ {t} \widehat {\nabla} \mathcal {L} \left(\boldsymbol {W} _ {i} ^ {t}; \boldsymbol {\mathcal {B}} ^ {t}\right) = \boldsymbol {W} _ {i} ^ {t} - \eta^ {t} \rho^ {t} \boldsymbol {U} _ {i} ^ {t} \boldsymbol {Z} _ {i} ^ {t} \boldsymbol {V} _ {i} ^ {t} ^ {\top}. \tag {9} +$$ + +Table 1. Projection matrix generation for SubZero in full-parameter tuning with OPT-1.3B and LLaMA2-7B on SST-2 and OPT-13B on RTE. + +
Matrix1.3B7B13B
Weight91.591.765.3
Activation51.552.953.1
ZO Gradient89.692.067.5
Random93.494.574.0
+ +Table 2. Memory cost in full-parameter tuning with RoBERTa-large on SST-2. + +
MethodMem.(GB)
SGD6.063
MeZO [40]2.683
S-RGF [43]23.845
SubZero2.690
+ +Here we choose SGD as the default optimizer of SubZero. Theorem 3 in Sec. 5 guarantees the convergence of SubZero with SGD as basic optimizer and gives its convergence rate. The choice of FO optimizers is orthogonal to ZO optimization. Also, some empirical work indicates that adaptive optimizers like Adam [30] do not necessarily enhance convergence of ZO approaches during LLM fine-tuning [22, 65]. So the combination of SubZero and Adam is included in Appendix 8.1 due to the limited space. We apply the primitive ZO approach. There are other ZO optimizers that utilize stochastic momentum [28] and second-order information [68] to facilitate faster convergence. While SubZero can be adapted to these ZO optimizers, we leave a comprehensive evaluation of these approaches for future work. + +We compare the memory overhead of SubZero with the existing random subspace method S-RGF [43] using identical experimental settings, including layer-wise perturbation and matching subspace dimension, with all methods utilizing the SGD optimizer. As shown in Table 2, S-RGF's memory usage is roughly four times greater than SGD and 8.8 times that of MeZO [40], while our SubZero's memory usage is comparable to MeZO. See more experimental comparisons on OPT-13B in Table 10 of Appendix 8.1. + +Lazy Low-rank Subspace Update. According to Eqn. (9), at the $t$ -th step, the gradient estimate of the parameter matrix in the $i$ -th layer, $\widehat{\nabla}\mathcal{L}(W_i^t;\mathcal{B}^t)$ , lies within a subspace defined by the projection matrices $U_{i}^{t}$ and $V_{i}^{t}$ . Specifically, $U_{i}^{t}$ spans the column subspace, while $V_{i}^{t}$ determines the row subspace, with both matrices generated iteratively, leading to extra computational overhead to LLM fine-tuning. + +However, for LLM fine-tuning, enhancing the computational efficiency and the accuracy of gradient subspace approximation is crucial. An excessively short update interval for $\pmb{U}_i$ and $\pmb{V}_i$ , such as generating them iteratively, can incur high computational costs and limit exploration of the gradient subspace they established. Conversely, a long interval may result in inaccuracies in subspace approximation and fail to capture the evolving nature of the gradient subspace. Accordingly, we propose a lazy subspace update strategy that periodically regenerates the projection matrices $\pmb{U}_i$ and $\pmb{V}_i$ . Specifically, these matrices are generated at the first step of every $F > 1$ training steps and remain unchanged for the subsequent $F - 1$ steps (see lines 4-7 in Algorithm 3). We utilize QR decomposition on two different random matrices for generating the column-orthogonal matrices $\pmb{U}_i$ and $\pmb{V}_i$ , + +as summarized in Algorithm 1. This lazy subspace update strategy is both efficient and effective in all our experiments. + +Algorithm 1 GenerateProjMatrix $(m,n,r)$ + +Input: size of parameter matrix $m \times n$ , rank $r$ . + +1: Generate random matrices $\mathbf{R}_1 \in \mathbb{R}^{m \times r}$ and $\mathbf{R}_2 \in \mathbb{R}^{n \times r}$ whose entries are sampled from $\mathcal{N}(0, 1)$ +2: $U, - \leftarrow$ QR_Decomposition $(R_{1})$ +3: $V, - \leftarrow$ QR_Decomposition $(R_{2})$ +4: return $U, V$ + +Algorithm 2 PerturbParams $(\mathcal{W},\mathcal{U},\mathcal{V},r,\varepsilon ,s)$ + +Input: model parameter set $\mathcal{W}$ , projection matrix sets $\mathcal{U}$ and $\mathcal{V}$ , rank $r$ , perturbation scale $\varepsilon$ , seed $s$ . + +1: Reset random number generator with seed $s$ +2: for $i = 1, 2, \dots, l$ do +3: Generate the perturbation matrix $Z_{i} \in \mathbb{R}^{r \times r}$ whose entries are sampled from $\mathcal{N}(0,1)$ +4: $\pmb{W}_i\gets \pmb{W}_i + \varepsilon \pmb {U}_i\pmb {Z}_i\pmb {V}_i^{\mathrm{T}}$ +5: return $\mathcal{W}$ + +Algorithm 3 SubZero + +Input: parameter matrix in the $i$ -th layer $\boldsymbol{W}_i \in \mathbb{R}^{m_i \times n_i}$ , $i = 1,2,\ldots,l$ , loss $\mathcal{L}$ , step budget $T$ , perturbation scale $\varepsilon$ , learning rate schedule $\{\eta^t\}$ , subspace change frequency $F$ , rank $r$ . + +1: for $t = 0, 1, \ldots, T - 1$ do +2: Sample a minbatch $\mathcal{B}^t\subset \mathcal{D}$ and a random seed $s^t$ +3: for $i = 1,2,\dots ,l$ do +4: if $t \mod F \equiv 0$ then +5: $\pmb{U}_i^t, \pmb{V}_i^t \gets \text{GenerateProjMatrix}(m_i, n_i, r)$ +6: else +7: $\pmb{U}_i^t\gets \pmb {U}_i^{t - 1},\pmb {V}_i^t\gets \pmb {V}_i^{t - 1}$ +8: // Note that $\mathcal{W}^t = \{\pmb{W}_i^t\}_{i=1}^l$ , $\mathcal{U}^t = \{\pmb{U}_i^t\}_{i=1}^l$ , $\mathcal{V}^t = \{\pmb{V}_i^t\}_{i=1}^l$ +9: $\mathcal{W}^t\gets$ PerturbParams $(\mathcal{W}^t,\mathcal{U}^t,\mathcal{V}^t,r,\varepsilon ,s^t)$ $\ell_{+}^{t}\gets \mathcal{L}(\mathcal{W}^{t};\mathcal{B}^{t})$ +10: $\mathcal{W}^t\gets$ PerturbParams $(\mathcal{W}^t,\mathcal{U}^t,\mathcal{V}^t,r, - 2\varepsilon ,s^t)$ $\ell_{-}^{t}\gets \mathcal{L}(\mathcal{W}^{t};\mathcal{B}^{t})$ +11: $\mathcal{W}^t\gets$ PerturbParams $(\mathcal{W}^t,\mathcal{U}^t,\mathcal{V}^t,r,\varepsilon ,s^t)$ +12: $\rho^t\gets \left(\ell_+^t -\ell_-^t\right) / (2\varepsilon)$ +13: Reset random number generator with seed $s^t$ +14: for $i = 1,2,\dots ,l$ do +15: Regenerate the perturbation matrix $\mathbf{Z}_i^t \in \mathbb{R}^{r \times r}$ whose entries are sampled from $\mathcal{N}(0,1)$ +16: $\pmb{W}_{i}^{t + 1}\gets \pmb{W}_{i}^{t} - \eta^{t}\rho^{t}\left(\pmb{U}_{i}^{t}\pmb{Z}_{i}^{t}\pmb{V}_{i}^{t}^{\top}\right)$ +17: return $\mathcal{W}^{t + 1}$ + +SubZero maintains just three small matrices per layer: a perturbation matrix $\mathbf{Z}_i \in \mathbb{R}^{r \times r}$ , and two column-orthogonal matrices $\mathbf{U}_i \in \mathbb{R}^{m_i \times r}$ and $\mathbf{V}_i \in \mathbb{R}^{n_i \times r}$ . This design enhances memory efficiency, as $r$ is generally much smaller than the size of the corresponding parameter matrix $\mathbf{W}_i \in \mathbb{R}^{m_i \times n_i}$ (i.e., $r \ll \min\{m_i, n_i\}$ ). Moreover, we employ + +in-place operations and per-layer parameter updates to estimate gradients and update parameters in parallel (see Appendix 8.2). Consequently, SubZero uses significantly less GPU memory than previous methods while achieving similar or better performance. For example, fine-tuning OPT-1.3B [64] on SST-2 [51] using SGD (without momentum) in full-parameter scheme as shown in Table 4, SubZero requires only 6.8GB GPU memory, compared to 11.5GB for SGD, yielding a $1.6 \times$ improvement in memory efficiency, similar as illustrated in Fig. 1 (d). + +Now we are ready to summarize the overall algorithm of SubZero in Algorithm 3. Each training step consists of three sequential phases. First, it obtains the projection matrices $\boldsymbol{U}_i^t$ and $\boldsymbol{V}_i^t$ using Algorithm 1 or directly adopts previous ones. Next, it computes the loss value difference $\rho$ with Eqn. (7) by applying Algorithm 2 to perturb all parameter matrices. Finally, SubZero updates all parameter matrices layer by layer, following Eqn. (9). + +# 4.2. Integration into Fine-Tuning Schemes + +We describe the integration of SubZero into full-parameter tuning [2] and three prominent PEFT schemes: LoRA [24], prefix tuning [36], and prompt tuning [33]. Typically, SubZero can be easily incorporated into these fine-tuning schemes. However, it encounters a challenge with extremely non-square parameter matrices, which have far more rows than columns or vice versa. This issue is particularly prevalent in LoRA, which employs two low-rank matrices $\pmb{A}_i \in \mathbb{R}^{m_i \times k}$ and $\pmb{B}_i \in \mathbb{R}^{k \times n_i}$ to approximate a full matrix $W_i' \in \mathbb{R}^{m_i \times n_i}$ , with $k \ll \min\{m_i, n_i\}$ , e.g., $k = 8$ while $\min\{m_i, n_i\} = 2048$ used in [65]. Consequently, it is impossible to find a smaller rank $r \ll k$ to compute the gradient estimates of $A_i$ and $B_i$ using Eqn. (6), imposing a challenge when applying SubZero to this scenario. + +To overcome this limitation, we propose a reshaping strategy that transforms the original non-square matrix into an approximate square matrix. For instance, we reshape $\mathbf{A}_i \in \mathbb{R}^{m_i \times k}$ into $\mathbf{A}_i' \in \mathbb{R}^{m_i' \times k'}$ such that $m_i k = m_i' k'$ and $m_i'$ is close to $k'$ . This reshaping allows us to apply Eqn. (6) to find a low-rank perturbation with rank $r$ significantly smaller than $\min\{m_i', k'\}$ , demonstrating the applicability of SubZero in the scenario. Table 8 in Sec. 6.4 shows the effectiveness of this reshaping strategy. + +# 5. Theoretical Analysis + +In this section, we theoretically analyze why SubZero can reduce the variance of gradient estimates and hence accelerate convergence. Before the analysis, we first define some necessary notations: + +$$ +\boldsymbol {P} = \operatorname {b d i a g} \left(\boldsymbol {V} _ {1} \otimes \boldsymbol {U} _ {1}, \dots , \boldsymbol {V} _ {l} \otimes \boldsymbol {U} _ {l}\right), \tag {10} +$$ + +$$ +\boldsymbol {z} = \left[ \operatorname {v e c} \left(\boldsymbol {Z} _ {1}\right) ^ {\mathsf {T}}, \dots , \operatorname {v e c} \left(\boldsymbol {Z} _ {l}\right) ^ {\mathsf {T}} \right] ^ {\mathsf {T}}, \tag {11} +$$ + +$$ +\tilde {\boldsymbol {z}} = \left[ \operatorname {v e c} \left(\tilde {\boldsymbol {Z}} _ {1}\right) ^ {\top}, \dots , \operatorname {v e c} \left(\tilde {\boldsymbol {Z}} _ {l}\right) ^ {\top} \right] ^ {\top}. \tag {12} +$$ + +Then we first state the main theoretical results on our gradient estimation in Eqn. (8). + +Theorem 1. For the gradient estimation in Eqn. (8), the following two properties hold. + +$\pmb{a}$ ) By using gradient estimation in (8), our estimated gradient $\hat{g}_{\varepsilon}(\pmb {x},\pmb {P},\pmb {z})$ is equivalent to + +$$ +\hat {g} _ {\varepsilon} (\boldsymbol {x}, \boldsymbol {P}, \boldsymbol {z}) = \frac {f (\boldsymbol {x} + \varepsilon \boldsymbol {P} \boldsymbol {z}) - f (\boldsymbol {x} - \varepsilon \boldsymbol {P} \boldsymbol {z})}{2 \varepsilon} \boldsymbol {P} \boldsymbol {z}, \tag {13} +$$ + +where $\pmb{z} \sim \mathcal{N}(\pmb{0},\pmb{I}_q)$ , $\varepsilon > 0$ , $\pmb{P} \in \mathbb{R}^{d \times q}$ satisfies $\pmb{P}^{\top}\pmb{P} = \pmb{I}_q$ with $d = \sum_{i=1}^{l} m_i n_i$ and $q = lr^2$ . + +b) Let $\mathbf{z} \sim \mathcal{N}(\mathbf{0},\mathbf{I}_q)$ , and $f \in C_{L_2}^{2,2}(\mathbb{R}^d)$ . Then we have + +$$ +\Phi (\boldsymbol {x}) = \| \mathbb {E} _ {\boldsymbol {z}} [ \hat {g} _ {\varepsilon} (\boldsymbol {x}, \boldsymbol {P}, \boldsymbol {z}) ] - \boldsymbol {P} \boldsymbol {P} ^ {\top} \nabla f (\boldsymbol {x}) \| _ {2} \leq \frac {\varepsilon^ {2}}{6} L _ {2} (q + 4) ^ {2}. +$$ + +See its proof in Appendix 8.6. Theorem 1 (a) provides the equivalent form (13) of our gradient estimation (8). By comparing this with the gradient estimation (5) in random subspace optimization [43, 46], we observe significant differences. First, our gradient estimation (13) accounts for the layer-wise structure of the network, requiring the projection matrix $\pmb{P}$ to be block-diagonal, whereas in random subspace optimization, $\pmb{P}$ is not. Additionally, our method introduces a layer-wise low-rank perturbation matrix, reflected by the block-diagonal structure of $\pmb{P}$ , with lazy updates to the column and row spaces defined by $U_{i}$ and $V_{i}$ . In contrast, random subspace optimization simply requires $\pmb{P}$ to be random. These distinctions highlight the key differences between our gradient estimation and existing methods in random subspace optimization. + +Theorem 1 (b) guarantees that the distance $\Phi (\pmb {x})$ between the expected gradient estimate and the BP gradient in the subspace spanned by $P$ is small. Moreover, by setting $\varepsilon = \frac{1}{q + 4}$ , the distance $\Phi (\pmb {x})$ is bounded by a constant $L_{2} / 6$ , independent of the parameter dimension $d$ . This implies that the error in our gradient estimation does not scale with the extremely high parameter dimensions of LLMs, providing accurate gradient estimation—crucial for optimizing LLMs. + +Next, we utilize a strictly convex quadratic loss to further analyze our gradient estimation in Eqn. (13). This choice is motivated by the fact that, after pretraining, the LLM parameters tend to converge toward a local minimum within a local basin, which can be well-approximated by a quadratic loss [42]. + +Theorem 2. Let $f(\pmb{x}) = \pmb{x}^{\top}\pmb{H}\pmb{x}$ and $z \sim \mathcal{N}(\mathbf{0},\pmb{I}_q)$ , where $\pmb{H} \in \mathbb{R}^{d\times d}$ is positive definite. We have + +$$ +\mathbb {E} _ {\boldsymbol {z}} \left[ \hat {g} _ {\varepsilon} (\boldsymbol {x}, \boldsymbol {P}, \boldsymbol {z}) \right] = \boldsymbol {P} \boldsymbol {P} ^ {\mathrm {T}} \nabla f (\boldsymbol {x}), \tag {14} +$$ + +$$ +\mathbb {E} _ {\boldsymbol {z}} \left[ \| \hat {g} _ {\varepsilon} (\boldsymbol {x}, \boldsymbol {P}, \boldsymbol {z}) \| ^ {2} \right] = (q + 2) \| \boldsymbol {P} ^ {\mathrm {T}} \nabla f (\boldsymbol {x}) \| ^ {2}, \tag {15} +$$ + +$$ +\mathbb {E} _ {\boldsymbol {z}} \left[ \frac {\langle \nabla f (\boldsymbol {x}) , \hat {g} _ {\varepsilon} (\boldsymbol {x} , \boldsymbol {P} , \boldsymbol {z}) \rangle^ {2}}{\| \boldsymbol {P} ^ {\top} \nabla f (\boldsymbol {x}) \| ^ {2} \| \hat {g} _ {\varepsilon} (\boldsymbol {x} , \boldsymbol {P} , \boldsymbol {z}) \| ^ {2}} \right] = \frac {1}{q}. \tag {16} +$$ + +See its proof in Appendix 8.6. Theorem 2 demonstrates several advantageous properties of our gradient estimation on the quadratic function. First, Eqn. (14) establishes the equivalence between the expected gradient estimation and the BP gradient within the subspace spanned by our projection matrix $\pmb{P}$ . Second, Eqn. (15) shows that, in this subspace, the variance of the gradient estimation scales linearly with the subspace dimension $q$ . In contrast, the variance of gradient estimation (2) in MeZO depends linearly on the model's parameter dimension $d$ , which is significantly larger than $q$ . Finally, Eqn. (16) reveals that the expected cosine similarity between our estimated gradient and the BP gradient within the subspace depends only on the subspace dimension $q \ll d$ , indicating that our gradient estimation provides a highly accurate parameter update direction. + +Building upon the above results, we can prove the convergence of our SubZero. + +Theorem 3. Let $f \in C_{L_1}^{1,1}(\mathbb{R}^d)$ be a non-convex function bounded below by $f^*$ . Suppose $\mathcal{E}_k = (z_0, z_1, \dots, z_k)$ with $z_k \sim \mathcal{N}(0, I_q)$ , and let $\mathcal{P}_j = (P_0, P_1, \dots, P_j)$ , where $P_j$ is defined in (10) with a fixed update frequency $F$ . Then, the sequence $\{\pmb{x}_k\}_{k > 0}$ generated by Algorithm 3 satisfies: + +$$ +\frac {1}{T} \sum_ {k = 0} ^ {T - 1} \mathbb {E} _ {\mathcal {E} _ {k}, \mathcal {P} _ {[ k / F ]}} \left[ \| \nabla f (\boldsymbol {x} _ {k}) \| ^ {2} \right] \leq \epsilon +$$ + +with $T = \mathcal{O}\left(\frac{d}{\epsilon}\right)$ if the perturbation scale satisfies $\varepsilon \leq \mathcal{O}\left(\frac{\epsilon^{1/2}}{q^{3/2}d^{1/2}L_1^{3/2}}\right)$ . Here $T = KF$ , where $K$ denotes the total number of subspace updates. + +See its proof in Appendix 8.6. Theorem 3 guarantees the convergence of our SubZero when the projection matrix $P$ is updated at a fixed frequency $F$ . + +# 6. Experiments + +In this section, we present comprehensive experiments to evaluate the effectiveness of SubZero. We conduct our experiments using medium-sized masked LLMs (RoBERTa-large [38]) and large-scale autoregressive LLMs (OPT-1.3B and 13B [64], LLaMA2-7B [55], and Mistral-7B [27]). Our exploration covers full-parameter tuning (FT) [2] and three PEFT schemes: LoRA [24], prefix tuning [36], and prompt tuning [33]. For comparison, we include leading ZO methods, such as MeZO [40], ZO-AdaMU [28], S-MeZO [39], HiZOO [68], and LOZO [9] alongside inference-only memory-efficient baselines like zero-shot, in-context learning (ICL) [7], and linear probing (LP) [32]. As the first and most popular ZO optimizer for LLM fine-tuning, MeZO is considered our primary competitor. We also use the FO optimizer SGD as a benchmark. Since appropriate prompts are critical for ZO optimization [40, 65], all experiments + +Table 3. Performance of fine-tuning OPT-13B on SuperGLUE with various experimental settings (with 1000 examples). AVG: average relative percentage difference with MeZO of all tasks. + +
Task type Taskclassification- multiple choice -generation -
SST-2RTECBBoolQWSCWICMultiRCCOPAReCoRDSQuADDROP
SGD(FT)94.982.385.778.465.365.874.290.082.488.035.5
Zero-shot58.859.646.459.038.555.046.980.081.246.214.6
ICL87.062.157.166.939.450.553.187.082.575.929.6
LP93.468.667.959.363.560.263.555.027.13.711.1
MeZO(FT)92.171.571.474.461.560.060.187.082.084.231.2
ZO-AdaMU(FT)92.172.967.973.061.560.763.089.083.082.432.0
S-MeZO(FT)92.376.975.076.561.158.263.387.071.277.931.9
HiZOO(FT)91.369.369.467.363.559.455.588.081.481.931.3
LOZO(FT)92.973.671.470.763.560.260.387.081.784.530.7
SubZero(FT)92.174.073.275.365.460.861.088.082.384.532.0
MeZO(LoRA)92.274.469.675.264.459.758.287.082.082.931.0
ZO-AdaMU(LoRA)88.072.071.672.660.156.458.988.083.276.832.4
S-MeZO(LoRA)90.862.275.072.951.955.856.486.069.976.431.7
HiZOO(LoRA)90.667.569.670.563.560.260.287.081.983.831.2
SubZero(LoRA)93.875.571.476.165.460.360.389.081.983.731.3
+ +![](images/7d372c90bf75073eec60d897c51ffc9e9f9530f3d3f9ea15db62fc814873df42.jpg) +(a) WIC: OPT-13B, FT + +![](images/2c4e1e5fc841e4062f274e1fc49cd247964009e85a8d461b607a0f0a3ac47657.jpg) +(b) SQuAD: OPT-13B, LoRA + +![](images/4d2ad1d9f409d4e56ebb7e40c8588b38c7722f32cf21f94b493531c0637d5076.jpg) +(c) CB: LLaMA2-7B, Prompt +Figure 2. Training loss curves of MeZO and SubZero. + +![](images/b434603097ece516f606b265ec524c3fb8d79184b569104cf5501e33a6939dff.jpg) +(d) MNLI: RoBERTa, FT, Non-diff + +incorporate prompt templates. Since a larger batch size reduces the variance of the estimated gradients, all compared methods use a batch size of 16 unless otherwise specified. All experimental settings are detailed in Appendix 8.2-8.5. + +# 6.1. Results under Different Experimental Settings + +Following the settings in MeZO [40], we evaluated Sub-Zero using OPT-13B on the SuperGLUE benchmark [57], which covers a diverse range of tasks, including classification, multiple-choice, and generation, as outlined in Table 3. The ZO methods were applied to both full-parameter tuning (FT) and LoRA fine-tuning schemes. The comparisons with vanilla LoRA and SGD with gradient accumulation are provided in Appendix 8.1. + +Table 3 presents the key findings, highlighting the best-performing ZO method in bold. The results show that ZO techniques significantly outperform baseline approaches like zero-shot, in-context learning, and linear probing, underscoring their ability to enhance a pre-trained model's performance on downstream tasks. From Table 3, one can also observer that MeZO, the first ZO optimizer for LLM fine-tuning, is highly competitive after carefully tuning its hyperparameters. Only ZO-AdaMU and LOZO, apart from SubZero, outperform MeZO in FT scheme. SubZero con + +sistently surpasses MeZO across all tasks and fine-tuning schemes. In FT scheme, S-MeZO showed competitive performance on several classification tasks. However, its performance on ReCoRD remained unsatisfactory even after hyperparameter tuning. Excluding ReCoRD, SubZero still outperforms S-MeZO with $2.05\%$ vs. $1.20\%$ in FT scheme and $1.74\%$ vs. $-4.90\%$ in LoRA scheme. + +We further extended our evaluation of SubZero using OPT-1.3B, LLaMA2-7B, and Mistral-7B in FT and three PEFT schemes: LoRA, prefix tuning, and prompt tuning. As shown in Table 4, SubZero outperformed MeZO across all models and fine-tuning schemes. Notably, while MeZO struggled in the prompt tuning scheme, SubZero excelled, achieving performance levels that closely matched those of the SGD optimizer. + +We also present several training loss curves in Fig. 2 (a)-(c), demonstrating that SubZero generally converges faster and achieves lower losses compared to MeZO. + +# 6.2. Results on Non-Differentiable Objectives + +Following MeZO [40], we apply SubZero to fine-tune RoBERTa-large and OPT-13B using full-parameter finetuning with two non-differentiable objectives—optimizing accuracy in classification tasks and F1 in the SQuAD task. + +Table 4. Performance of fine-tuning LLaMA2-7B and Mistral-7B on CB, and OPT-1.3B on SST-2. Table 5. Orthogonal projection matrix. + +
LLaMA2-7BMistral-7BOPT-1.3B
FTLoRAPrefixPromptFTLoRAPrefixPromptFTLoRAPrefixPrompt
SGD69.675.069.669.673.275.069.662.593.293.093.190.7
MeZO64.373.269.660.762.569.658.357.192.392.891.685.9
SubZero71.475.076.866.164.373.264.362.593.492.992.289.1
+ +
DatasetOrtho.Accuracy
RTEX67.5
✓74.0
WSCX59.6
✓65.1
+ +Table 6. Fine-tuning performance comparison between Sub-Zero and MeZO on RoBERTa-large and OPT-13B with non-differentiable objectives. + +
Model TaskRoBERTa-largeOPT-13B SQuAD
SST-2SST-5SNLIMNLI
Zero-shot79.035.550.248.846.2
Cross entropy (Adam)93.955.988.783.884.2
Cross entropy (MeZO)92.953.283.077.084.2
Cross entropy (SubZero)93.454.084.777.484.5
Accuracy/F1 (MeZO)92.446.581.973.980.2
Accuracy/F1 (SubZero)92.747.183.074.881.1
+ +As baselines, we also report results using the differentiable cross-entropy objective with Adam, MeZO, and SubZero. As shown in Table 6, SubZero consistently outperforms MeZO with the two non-differentiable objectives. The training loss curves in Fig. 2 (d) also demonstrate the advantage of SubZero over MeZO. + +# 6.3. Memory Usage and Wall-Clock Time Analysis + +We compare the memory consumption and wall-clock time of ZO methods (MeZO and SubZero), SGD, and inference-only approaches (zero-shot and in-context learning (ICL)) using OPT-13B (see Table 10 in Appendix 8.1). Since inference-only methods do not involve fine-tuning, they have zero wall-clock time and their memory usage reflects only the inference load. For fine-tuning, all methods were run for 20K steps. The ZO methods, including SubZero, achieved over a $1.8 \times$ reduction in memory usage compared to SGD. Notably, SubZero's memory footprint closely aligns with MeZO's, while offering improved performance. We use per-layer weight updates for MeZO and SubZero (see Appendix 8.2), resulting in nearly identical memory usage for FT and LoRA schemes when one decimal place is reserved. + +Although SubZero introduces additional computational overhead for generating projection matrices via QR decomposition, the observed extra time overhead remains below $9\%$ across all OPT model sizes, with a mean value of $4.79\%$ of the total wall-clock time (see Table 11 in Appendix 8.1). + +Note that due to differences in how steps are defined between ZO methods and SGD, direct wall-clock time comparisons between the two are not entirely meaningful. + +# 6.4. Ablation Study + +We conducted a thorough investigation of the effectiveness of our techniques. Table 5 shows that using a column-orthogonal projection matrix significantly outperforms a + +Gaussian random projection matrix, primarily due to the low-rank structure of the perturbation matrices (see experimental settings in Appendix 8.4). This low-rank perturbation is key to improving the quality of gradient estimation. + +Next, Table 7 explores the effects of subspace rank $r$ and update frequency $F$ in Algorithm 3. The results demonstrate that SubZero is robust to variations in the subspace rank. However, performance drops sharply when the update frequency is too low, as the optimization becomes constrained to a single subspace for too long, limiting its adaptability. + +Table 7. Subspace change frequency $F$ and rank $r$ . + +
F\r3264128
50072.670.072.2
100073.671.874.0
200072.273.372.2
2000070.471.168.6
+ +Table 8. Reshaping strategy for non-square matrices on SST-2 with OPT-1.3B in PEFT schemes. + +
MethodLoRAPrefixPrompt
MeZO92.891.685.9
SubZero(w/o)92.189.474.2
SubZero(w/)92.992.289.1
+ +Finally, Table 8 underscores the critical role of the reshaping strategy for handling highly non-square perturbation matrices, essential for ensuring effective perturbations in different layers of the model. Together, these results highlight the improvements brought by our design choices, particularly in terms of projection and reshaping strategies, and their impact on SubZero's robustness and performance. + +Due to limited space, the ablation studies of SubZero on random seed, batch size, and combination with Adam are given in Appendix 8.1. + +# 7. Conclusion + +We have demonstrated that SubZero effectively fine-tunes large LLMs across various tasks and schemes with a memory cost comparable to that of inference. Extra experiments indicate that SubZero can optimize non-differentiable objectives. Our theory explains how SubZero reduces the variance of gradient estimates and hence accelerates convergence. + +Limitation. In addition to the representative first-order and primitive zero-order optimizers, we have yet to investigate the combinations of SubZero with other first-order and zero-order optimizers to evaluate the implications on convergence speed. While SubZero is compatible with memory-efficient techniques like parameter quantization [35], we have not thoroughly explored the practical effects of these combinations. A theoretical analysis of the reshaping strategy is valuable but left for future work. + +# Acknowledgments + +This work is supported by the NSF of China (grant nos. 62131003 and 62102034), Beijing Municipal Science and Technology Project (Z241100001324011), the Ministry of Education, Singapore, under its AcRF Tier 2 Funding (Proposal ID: T2EP20224-0048). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of the Ministry of Education, Singapore. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv:2303.08774, 2023. 1 +[2] Armen Aghajanyan, Sonal Gupta, and Luke Zettlemoyer. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing, pages 7319-7328, 2021. 3, 5, 6 +[3] Shun-ichi Amari. Backpropagation and stochastic gradient descent method. Neurocomputing, 5(4-5):185-196, 1993. 1, 2 +[4] Roy Bar Haim, Ido Dagan, Bill Dolan, Lisa Ferro, Danilo Giampiccolo, Bernardo Magnini, and Idan Szpektor. The second PASCAL recognising textual entailment challenge. In Proceedings of the Second PASCAL Challenges Workshop on Recognising Textual Entailment, 2006. 16 +[5] Luisa Bentivogli, Peter Clark, Ido Dagan, and Danilo Giampaiccolo. The fifth PASCAL recognizing textual entailment challenge. In Proceedings of the Second Text Analysis Conference, 2009. 16 +[6] Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 632-642, 2015. 16 +[7] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, pages 1877-1901, 2020. 6 +[8] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv:1604.06174, 2016. 2 +[9] Yiming Chen, Yuan Zhang, Liyuan Cao, Kun Yuan, and Zaiwen Wen. Enhancing zeroth-order fine-tuning for language models with low-rank structures. In Proceedings of the International Conference on Learning Representations, 2025. 6 +[10] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. BoolQ: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the Conference of the North + +American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2019. 16 +[11] Ido Dagan, Oren Glickman, and Bernardo Magnini. The PASCAL recognising textual entailment challenge. In Proceedings of the International Conference on Machine Learning Challenges: Evaluating Predictive Uncertainty Visual Object Classification, and Recognizing Textual Entailment, 2005. 16 +[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 2 +[13] Marie-Catherine de Marneffe, Mandy Simons, and Judith Tonhauser. The commitmentbank: Investigating projection in naturally occurring discourse. In Proceedings of Sinn und Bedeutung 23, 2019. 16 +[14] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettle-moyer. GPT3.int8(): 8-bit matrix multiplication for transformers at scale. Advances in Neural Information Processing Systems, 35:30318-30332, 2022. 2 +[15] Tim Dettmers, Mike Lewis, Sam Shleifer, and Luke Zettlemoyer. 8-bit optimizers via block-wise quantization. In Proceedings of the International Conference on Learning Representations, 2022. 2 +[16] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient finetuning of quantized LLMs. Advances in Neural Information Processing Systems, 36, 2024. 2 +[17] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. 3 +[18] Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2368-2378, 2019. 16 +[19] John C Duchi, Michael I Jordan, Martin J Wainwright, and Andre Wibisono. Optimal rates for zero-order convex optimization: The power of two function evaluations. IEEE Transactions on Information Theory, 61(5):2788-2806, 2015. 2 +[20] Tanmay Gautam, Youngsuk Park, Hao Zhou, Parameswaran Raman, and Wooseok Ha. Variance-reduced zeroth-order methods for fine-tuning language models. In Proceedings of the International Conference on Machine Learning, 2024. 1, 2, 16 +[21] Danilo Giampiccolo, Bernardo Magnini, Ido Dagan, and Bill Dolan. The third PASCAL recognizing textual entailment challenge. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, 2007. 16 +[22] Wentao Guo, Jikai Long, Yimeng Zeng, Zirui Liu, Xinyu Yang, Yide Ran, Jacob R Gardner, Osbert Bastani, Christopher De Sa, Xiaodong Yu, Beidi Chen, and Zhaozhuo Xu. Zeroth-order fine-tuning of LLMs with extreme sparsity. arXiv:2406.02913, 2024. 4, 15 + +[23] Yongchang Hao, Yanshuai Cao, and Lili Mou. Flora: Lowrank adapters are secretly gradient compressors. In Proceedings of the International Conference on Machine Learning, 2024. 3 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations, 2022. 2, 3, 5, 6 +[25] Kevin G Jamieson, Robert Nowak, and Ben Recht. Query complexity of derivative-free optimization. Advances in Neural Information Processing Systems, 25, 2012. 2 +[26] Kaiyi Ji, Zhe Wang, Yi Zhou, and Yingbin Liang. Improved zeroth-order variance reduced algorithms and analysis for nonconvex optimization. In Proceedings of the International Conference on Machine Learning, pages 3100-3109. PMLR, 2019. 2 +[27] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7B. arXiv:2310.06825, 2023. 6 +[28] Shuoran Jiang, Qingcai Chen, Youcheng Pan, Yang Xiang, Yukang Lin, Xiangping Wu, Chuanyi Liu, and Xiaobao Song. ZO-AdaMU optimizer: Adapting perturbation by the momentum and uncertainty in zeroth-order optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18363–18371, 2024. 1, 2, 4, 6, 13, 14 +[29] Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth. Looking beyond the surface: A challenge set for reading comprehension over multiple sentences. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 252–262, 2018. 16 +[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations, 2015. 1, 2, 4 +[31] David Kozak, Stephen Becker, Alireza Doostan, and Luis Tenorio. A stochastic subspace approach to gradient-free optimization in high dimensions. Computational Optimization and Applications, 79(2):339-368, 2021. 1, 2 +[32] Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In Proceedings of the International Conference on Learning Representations, 2022. 6 +[33] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, 2021. 5, 6 +[34] Hector Levesque, Ernest Davis, and Leora Morgenstern. The winograd schema challenge. In Proceedings of the International Conference on the Principles of Knowledge Representation and Reasoning, 2012. 16 +[35] Bingrui Li, Jianfei Chen, and Jun Zhu. Memory efficient optimizers with 4-bit states. Advances in Neural Information Processing Systems, 36, 2024. 2, 8 +[36] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the + +Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing, pages 4582-4597, 2021. 5, 6 +[37] Sijia Liu, Bhavya Kailkhura, Pin-Yu Chen, Paishun Ting, Shiyu Chang, and Lisa Amini. Zeroth-order stochastic variance reduction for nonconvex optimization. Advances in Neural Information Processing Systems, 31, 2018. 2 +[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach. arXiv:1907.11692, 2019. 6 +[39] Yong Liu, Zirui Zhu, Chaoyu Gong, Minhao Cheng, Chojui Hsieh, and Yang You. Sparse MeZO: Less parameters for better performance in zeroth-order LLM fine-tuning. arXiv:2402.15751, 2024. 1, 2, 6, 13, 16, 17 +[40] Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Fine-tuning language models with just forward passes. Advances in Neural Information Processing Systems, 36:53038-53075, 2023. 1, 2, 3, 4, 6, 7, 15, 16, 17, 18 +[41] Yurii Nesterov and Vladimir Spokoiny. Random gradient-free minimization of convex functions. Foundations of Computational Mathematics, 17:527-566, 2017. 2, 3, 19, 20, 22, 23 +[42] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in Neural Information Processing Systems, 33:512-523, 2020. 6 +[43] Ryota Nozawa, Pierre-Louis Poirion, and Akiko Takeda. Zeroth-order random subspace algorithm for non-smooth convex optimization. Journal of Optimization Theory and Applications, 204(3):53, 2025. 1, 2, 3, 4, 6 +[44] Mohammad Taher Pilehvar and Jose Camacho-Collados. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1267-1273, 2019. 16 +[45] Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, 2016. 16 +[46] Lindon Roberts and Clément W Royer. Direct search based on probabilistic descent in reduced spaces. SIAM Journal on Optimization, 33(4):3057-3082, 2023. 1, 2, 3, 4, 6 +[47] Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S. Gordon. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In Proceedings of the AAAI Spring Symposium Series, 2011. 16 +[48] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99-106, 2021. 15 +[49] Junhong Shen, Neil Tenenholtz, James Brian Hall, David Alvarez-Melis, and Nicolo Fusi. Tag-LLM: Repurposing general-purpose LLMs for specialized domains. In Proceed + +ings of the International Conference on Machine Learning, pages 44759-44773, 2024. 1 +[50] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2013. 16 +[51] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, 2013. 5 +[52] Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Release strategies and the social impacts of language models. arXiv:1908.09203, 2019. 1 +[53] James C Spall. Multivariate stochastic approximation using a simultaneous perturbation gradient approximation. IEEE Transactions on Automatic Control, 37(3):332-341, 1992. 3 +[54] Xu Sun, Xuancheng Ren, Shuming Ma, and Houfeng Wang. meProp: Sparsified back propagation for accelerated deep learning with reduced overfitting. In Proceedings of the International Conference on Machine Learning, pages 3299-3308, 2017. 2 +[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 6 +[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1 +[57] Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. SuperGLUE: A stickier benchmark for general-purpose language understanding systems. arXiv: 1905.00537, 2019. 7, 16 +[58] Junjie Wang, Guangjing Yang, Wentao Chen, Huahui Yi, Xiaohu Wu, Zhouchen Lin, and Qicheng Lao. MLAE: Masked LoRA experts for visual parameter-efficient finetuning. arXiv:2405.18897, 2024. 2 +[59] Adina Williams, Nikita Nangia, and Samuel Bowman. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2018. 16 +[60] Yifan Yang, Kai Zhen, Ershad Banijamali, Athanasios Mouchtaris, and Zheng Zhang. AdaZeta: Adaptive zeroth-order tensor-train adaption for memory-efficient large language models fine-tuning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 977–995, 2024. 1, 2, 16, 17 + +[61] Hongwei Yong, Ying Sun, and Lei Zhang. A general regret bound of preconditioned gradient method for dnn training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7866-7875, 2023. 17 +[62] Pengyun Yue, Long Yang, Cong Fang, and Zhouchen Lin. Zeroth-order optimization with weak dimension dependency. In Proceedings of the Annual Conference on Learning Theory, pages 4429-4472. PMLR, 2023. 1 +[63] Sheng Zhang, Xiaodong Liu, Jingjing Liu, Jianfeng Gao, Kevin Duh, and Benjamin Van Durme. ReCoRD: Bridging the gap between human and machine commonsense reading comprehension. arXiv preprint 1810.12885, 2018. 16 +[64] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. OPT: Open pre-trained transformer language models. arXiv:2205.01068, 2022. 1, 5, 6 +[65] Yihua Zhang, Pingzhi Li, Junyuan Hong, Jiaxiang Li, Yimeng Zhang, Wenqing Zheng, Pin-Yu Chen, Jason D. Lee, Wotao Yin, Mingyi Hong, Zhangyang Wang, Sijia Liu, and Tianlong Chen. Revisiting zeroth-order optimization for memory-efficient LLM fine-tuning: A benchmark. In Proceedings of the International Conference on Machine Learning, pages 59173-59190, 2024. 1, 2, 4, 5, 6, 15, 17 +[66] Zhong Zhang, Bang Liu, and Junming Shao. Fine-tuning happens in tiny subspaces: Exploring intrinsic task-specific subspaces of pre-trained language models. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, pages 1701-1713, 2023. 3 +[67] Jiawei Zhao, Zhenyu Zhang, Beidi Chen, Zhangyang Wang, Anima Anandkumar, and Yuandong Tian. Galore: Memory-efficient LLM training by gradient low-rank projection. In Proceedings of the International Conference on Machine Learning, 2024. 1, 2, 3 +[68] Yanjun Zhao, Sizhe Dang, Haishan Ye, Guang Dai, Yi Qian, and Ivor W Tsang. Second-order fine-tuning without pain for LMMs: A Hessian informed zeroth-order optimizer. In Proceedings of the International Conference on Learning Representations, 2025. 4, 6, 13 \ No newline at end of file diff --git a/zerothorderfinetuningofllmsinrandomsubspaces/images.zip b/zerothorderfinetuningofllmsinrandomsubspaces/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9f4dee20e984f824493447c72ee084504d586b40 --- /dev/null +++ b/zerothorderfinetuningofllmsinrandomsubspaces/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a4e672922ef20092694ab850ca45810ec7b6103d80eb975105b7e01fb00c207 +size 475242 diff --git a/zerothorderfinetuningofllmsinrandomsubspaces/layout.json b/zerothorderfinetuningofllmsinrandomsubspaces/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..85139770adba80648b9b32ea1810a78e3a07b4cf --- /dev/null +++ b/zerothorderfinetuningofllmsinrandomsubspaces/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d2538a708057366efc2ee3721ef209fb8f820886803b7646224be35936527d9 +size 622970 diff --git a/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/cb6554ce-9075-4801-b1eb-ec7a6cc16e64_content_list.json b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/cb6554ce-9075-4801-b1eb-ec7a6cc16e64_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..301aa26507e27a2ad0baf76a5a2f9b5ca8a6b6cb --- /dev/null +++ b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/cb6554ce-9075-4801-b1eb-ec7a6cc16e64_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83c5c6dd63c9adb1b6f3a8d71be5bf09bc4b7d355b470a4540e2f42f8d259fbb +size 90330 diff --git a/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/cb6554ce-9075-4801-b1eb-ec7a6cc16e64_model.json b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/cb6554ce-9075-4801-b1eb-ec7a6cc16e64_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d6e994f5448d2a84f7b1fa5344cd8c955dd535fb --- /dev/null +++ b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/cb6554ce-9075-4801-b1eb-ec7a6cc16e64_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:475c5af9caafa4ebfd7ad0704964b8deee33d6bb44d24cfccc4c2a8658e86bac +size 115735 diff --git a/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/cb6554ce-9075-4801-b1eb-ec7a6cc16e64_origin.pdf b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/cb6554ce-9075-4801-b1eb-ec7a6cc16e64_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c1541f3117d8d7a16e0f3ae8bd44edc25d9b017d --- /dev/null +++ b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/cb6554ce-9075-4801-b1eb-ec7a6cc16e64_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:341ae159dfa2e72cb0a91245178af0664e7aa4dccfc7004d1847d23f37299f01 +size 1750413 diff --git a/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/full.md b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..41e40ec16a5fa9ad97858956ee44ebf16d7cbadc --- /dev/null +++ b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/full.md @@ -0,0 +1,410 @@ +# ZFusion: Efficient Deep Compositional Zero-shot Learning for Blind Image Super-Resolution with Generative Diffusion Prior + +Alireza Esmaeilzehi $^{1*}$ , Hossein Zaredar $^{1*}$ , Yapeng Tian $^{2}$ , Laleh Seyyed-Kalantari $^{1,3}$ $^{1}$ York University, Toronto, ON, Canada $^{2}$ The University of Texas at Dallas, Richardson, TX, US $^{3}$ Vector Institute, Toronto, ON, Canada + +{alirezae, hzaredar, lsk}@yorku.ca, {yapeng.tian}@utdallas.edu + +# Abstract + +Deep blind image super resolution (Blind SR) schemes strive to provide high performances under various image degradation processes. Despite the significant advancement in the area of Blind SR, the performances of these methods still may not be as high as one would desire in the case of real-world degradation operations. In this paper, we develop a novel diffusion-based Blind SR method, which, by leveraging compositional zero-shot learning, is able to provide superior performances for both synthetic and real-world unknown degradation processes. Specifically, we first extract both synthetic and real-world degradation embeddings from the input visual signal in a compositional zero-shot fashion. Next, we have efficiently embedded such degradation embeddings in the architecture of our diffusion-based scheme for guiding the diffusion feature generation process. The results of extensive experiments have demonstrated the effectiveness of the proposed Blind SR method over the state-of-the-art algorithms. Our code is available at https://github.com/HosseinZaredar/ZFusion. + +# 1. Introduction + +Image super-resolution (SR) is a fundamental challenge in low-level vision, aiming to reconstruct photo-realistic high-resolution (HR) images from low-resolution (LR) inputs. Deep SR networks [6-8, 14, 15, 20, 43, 63, 66] have achieved state-of-the-art (SOTA) performance by effectively learning feature mappings between LR and HR image spaces. However, many of these methods are trained with data generated from a single degradation process, such as bicubic downsampling, leading to significant performance degradation when encountering different real-world degradations during inference. + +To address this limitation, blind image super-resolution (Blind SR) has emerged [11, 25-27, 43, 60, 61]. These ap + +![](images/da0fe6cd58fd1d124b1e9bc5198cc2e6ef0d109c9b9632fc31b393b075d5a13f.jpg) +(a) + +![](images/f54d40f25d97cc459105296ebbc16dddb7814a4db4400e90107c6c7f4fc5b1ca.jpg) +(b) + +![](images/dc0b730d52d5da77522a8eeb0053cc14fad9d61bf4915948a3d7d1fcdbf3c5a9.jpg) +(c) +Figure 1. Compositional zero-shot learning for SR. (a) LR image. (b) Schematic of LR degradations as a composition of synthetic and real-world degradations. (c) SR image. (d) Ground truth (GT). + +![](images/c163e40485f9c012fde6ccffca5ac8e6c1c56118c58b9c70888123163127b6cb.jpg) +(d) + +proaches enhance generalization by training on diverse LR images generated from various degradation processes, including blurring, downsampling, and noise, enabling them to handle unknown degradation. Furthermore, advanced learning paradigms like unpaired learning [22, 28], contrastive learning [47, 50], and self-supervised learning [4, 65] have been employed to improve Blind SR performance. + +Very recently, generative diffusion priors have been exploited in Blind SR to restore images with realistic details [5, 21, 57]. Diffusion models excel at restoring high-frequency components often lost during degradation. Many diffusion-based Blind SR methods incorporate text prompts alongside LR images. These prompts, representing seman + +tic information or degradation characteristics, are embedded and integrated into the network architecture to guide feature generation. This approach leverages the power of text embeddings to encode both semantic content and degradation processes, improving the reconstruction of realistic textures and structures. + +However, challenges remain in text-prompted diffusion-based Blind SR [5, 49, 57]. Firstly, ensuring alignment between textual degradation embeddings and visual feature maps is critical. Existing methods often rely on pre-trained vision-language models, which may not be optimized for SR, potentially leading to misalignment. Secondly, reliance on synthetic degradation embeddings limits applicability to real-world scenarios, where complex and unknown degradations prevail. This necessitates learning algorithms that generalize to unseen degradation operators. + +In this paper, to address the above challenges, we propose a new diffusion-based Blind SR scheme utilizing text-prompt image processing to model a wide range of degradations. We introduce a compositional zero-shot learning (ZSL) algorithm to handle unknown degradations encountered during inference without relying on real-world LR data. Drawing inspiration from compositional ZSL [10, 23], we argue that realistic unknown degradation embeddings can be synthesized as compositions of degradation attributes learned during training. These degradation attributes could be generated from text prompts associated with various synthetic degradations and non-reference-based image quality assessment (IQA) metrics during training without using any real-world low-quality data. + +Our zero-shot degradation learning algorithm operates solely in the feature space, enhancing the network's generalization ability during inference. By integrating these learned degradation embeddings into a diffusion-based Blind SR architecture, both artifact suppression and diffusion denoising are guided by the estimated degradation characteristics, leading to superior SR performance. Fig. 1 illustrates an LR image from RealSR [2] dataset with real-world unknown degradation, alongside its feature representations obtained by the proposed Blind SR scheme. It is seen from this figure that, thanks to compositional ZSL, our Blind SR method effectively obtains the degradation embeddings of an LR image as a composition of learned embeddings during training. This is carried out without using any prior knowledge of the degradation process of LR image. Further, the SR image obtained by our method has superior visual quality. + +Our contributions can be summarized as follows: + +- We propose a novel diffusion-based Blind SR scheme, which can handle both synthetic and real-world degradations by leveraging text-prompt image processing. +- Our Blind SR method uses compositional ZSL to model real-world degradations at inference as the composition + +of attributes learned during the training. + +- In order to efficiently synthesize degradation attributes for our compositional ZSL algorithm, we employ the parameters of degradation operators at the training stage, as well as those of IQA metrics, to handle both synthetic and unknown real-world degradation processes. +- We finally incorporate the degradation embeddings obtained by our compositional ZSL algorithm into the artifact suppression and denoising parts of the diffusion-based Blind SR network for achieving a superior performance. + +# 2. Related Works + +# 2.1. Blind Image Super-Resolution + +As deep learning-based Blind SR methods are not constrained to handling a specific image degradation process, they can be used in various pragmatic situations [16, 19, 24, 58]. SRMD [60] is the premier blind image SR network that processes the input low-quality image, along with its degradation parameters, e.g., blurring kernel and noise information, to super-resolve the low-quality images. The scheme of DAN [11] employs an alternating optimization process between blurring kernel estimation and high-quality image reconstruction to provide superior restored images. + +Diffusion-based Blind SR schemes have yielded high-quality visual signals with realistic textures and structures. In addition, these methods leverage the powerful image generation capacity of the diffusion models to produce more realistic textures in the high-frequency areas [48, 52, 53, 64]. DiffBIR [21] is an effective diffusion model for the task of Blind SR, which performs the denoising diffusion process in the latent space. StableSR [40] is another SOTA Blind SR model that uses a feature recalibration mechanism in the diffusion denoiser architecture in order to provide improved results. In contrast to the existing diffusion-based Blind SR methods, which explicitly do not strive to generalize to the real-world unknown degradation processes, we leverage compositional ZSL in order to enhance SR generalization capability. + +# 2.2. Compositional Zero-shot Learning + +Compositional ZSL algorithms [9, 10, 17, 23, 29, 32, 55] strive to understand novel concepts at inference by combining the attributes learned during training. Huang et al. [10] proposed a multi-path ZSL-based algorithm that uses cross-modal traction for shifting the embeddings of various modalities towards each other. The method developed in [23] utilizes a dual transformer-based neural network in order to efficiently leverage both visual features and attribute embeddings for inferring from a visual signal. In [41], Wang et al. designed a novel attribute learning paradigm for zeros-shot learning. Specifically, this scheme + +uses an attribute meta learner and base learner for learning conditional attributes that are well aligned with the visual features. In this work, we extended compositional ZSL to address the blind SR task by decoupling LR degradation embedding as a composition of synthetic and real-world degradation attributes. + +# 3. Methodology + +# 3.1. Motivation + +Let $y \in \mathbb{R}^{H \times W \times 3}$ denote the LR image of spatial resolution $H \times W$ , and $x \in \mathbb{R}^{aH \times aW \times 3}$ be its corresponding HR ground truth version with the scaling factor $a$ . The relation between $y$ and $x$ can be modeled as: + +$$ +y = f (x) = D H x + n \tag {1} +$$ + +where $D$ , $H$ , and $n$ , are, respectively, downsampling, blurring, and noise operators, and $f(\cdot)$ denotes the image degradation process. The challenge is that each of the degrading operators $D$ , $H$ , and $n$ can be unknown in real-world scenarios when we want to restore $x$ given $y$ . + +Incorporating the image degradation characteristics into the training of deep Blind SR networks has been shown to provide superior results [11, 26, 51, 60]. In order to effectively integrate the image degradation embeddings into the deep Blind SR network, two crucial factors must be taken into account. Firstly, the LR images during inference are often degraded by unknown real-world degradation, which has not been encountered during training. Hence, the degradation learning process of Blind SR must be sufficiently generalizable. Secondly, the degradation embeddings of LR images can guide SR feature generation only if they are well aligned with the LR feature maps. + +To consider the above two critical points simultaneously, we use the concept of compositional ZSL alongside vision-language models. Zero-shot learning is a paradigm in which the model infers from the data that was not seen during training [33]. Compositional ZSL is a category of ZSL wherein the model understands the unseen inference sample from the composition of attributes that are learned during training [54]. In this case, the model is able to bridge between the seen and unseen data distributions. + +In Blind SR, the network must be able to handle various unknown real-world, as well as synthetic decimation processes. Since both of these degradations have correlations with the degradation operators used in the training phase, they could be yielded as the compositions of training degradation attributes. For producing these training degradation attributes in a way that they are aligned with the LR feature maps, we leverage the vision-language models. + +Specifically, we synthesize two types of degradation attributes, one corresponding to the synthetic degradation operators with specific parameters, and the other associated + +with unknown real-world processes. For synthetic degradation attributes, we utilize vision-language models with textual data of synthetic operators and their parameters. For real-world degradation attributes, the vision-language models with non-reference-based IQA metrics are employed. The proposed compositional ZSL algorithm then reduces the bias between learning from these two types of degradation attributes, which in turn makes the Blind SR network robust for both synthetic and real-world LR images. + +# 3.2. Proposed Scheme + +Compositional Zero-shot Learning for Image Degradation Learning. During the training process, most of the diffusion-based Blind SR schemes apply synthetic image degradation operations based on the pipeline of BSRGAN [61] for generating LR images. Specifically, the combination of blurring, downsampling, noise, and JPEG compression with parameters selected from certain ranges are used to form $f(\cdot)$ . However, during the inference stage, the operations of $f(\cdot)$ may be unknown and different than those used in training. + +Following the points in Sec. 3.1, we want to extract degradation attributes $\mathbf{f}_f$ associated with the image degradation process $f(\cdot)$ in the embedding space. Specifically, for performing efficient Blind SR, we aim at modeling the image degradation process of $y$ as a composition of embeddings that are optimized in training. Since, during training, the degradation process $f(\cdot)$ is synthetic and known (e.g. pipeline of [61]), the first set of degradation attributes are produced by feeding the textual data of synthetic known decimation process of LR image to the text encoder of a vision-language model. + +Using only the synthetic degradation attributes leads to a poor SR generalization, as at inference, LR image $y$ decimation is real-world. In this regard, we further extract the real-world degradation attributes for our algorithm. Specifically, we argue that the level of degradation of $y$ with real-world decimation can be assessed by non-reference-based IQA metrics. We next feed the textual data of such IQA values to the text encoder of a vision-language model to produce real-world degradation attributes. + +For any LR image $y$ with unknown $f(\cdot)$ , our compositional ZSL algorithm extracts its degradation embedding $\mathbf{f}_f$ as the composition of the above two types of degradation attributes optimized during training. Further, the proposed compositional ZSL algorithm strives to reduce the bias between the synthetic and real-word degradation attributes, thus improving the generalization capability of the Blind SR task. We now describe the details of extracting the above two types of degradation attributes. + +Synthetic Degradation Attribute Learning. Fig. 2 shows the overall flow of our compositional zero-shot learning algorithm for degradation learning (CZSD). It consists of two + +![](images/423f16d4f48cbe6abf5c554668fbf48c0789038523e7dd04c5b59c7cdc86789d.jpg) +Figure 2. Architecture flow of the proposed compositional zero-shot degradation learning (CZSD). The synthetic and real-world degradation attributes are initially obtained in the text embedding space by $T(\cdot)$ . Afterwards, the visually-aligned features $\mathbf{f}_1$ and $\mathbf{f}_3$ corresponding to such degradation embeddings are produced by the vision encoder $V(\cdot)$ (Best viewed in color version). + +modules: synthetic degradation attribute learning (SDAL) and real-world degradation attribute learning (RDAL). + +In the SDAL module, our objective is to extract the synthetic degradation attributes from $y$ . During the training process, the synthetic degradation process of $y$ , along with its parameters, is known. In this regard, we first form four textual sentences, i.e., image is degraded by blurring (e.g. isotropic kernel with kernel std of 1.5)/downsampling (e.g. bicubic with scaling factor 2)/noise (e.g. Gaussian noise with noise level 5)/JPEG compression (e.g. quality factor 60), each of which corresponds to a single synthetic operation in $f(\cdot)$ . We then feed these four textual sentences associated with the input LR image $y$ to the CLIP [34] text encoder $T(\cdot)$ : + +$$ +\mathbf {s} = T \left(\left[ s _ {1}, s _ {2}, s _ {3}, s _ {4} \right]\right) \tag {2} +$$ + +where $s_i$ 's $(i = 1,2,3,4)$ are the four textual sentences explained above. The degradation attribute $\mathbf{s}$ is of size $\mathbb{R}^{512}$ . + +It is seen from Fig. 2 that the degradation attribute $\mathbf{s}$ is the composition of synthetic degradation operators in the text embedding space. The use of text embedding space for image degradation learning is crucial, as it is able to map various degradation operations with different properties into a single space. For example, image blurring operation can be modeled by the convolution operator with limited kernel size, while noise map is added to the visual signal with the same spatial resolution. As another example, image downsampling is linear, while JPEG compression blocking is produced through a highly non-linear process. Hence, all these inconsistent image degradation operations must be mapped into a single embedding space. + +It has been mentioned in Sec. 3.1 that the degradation attributes must be aligned with the LR feature maps to improve the SR performance. However, the degradation em + +beddings $\mathbf{s}$ are obtained from the textual modality. Further, during the inference, we do not know the degradation process of $y$ , and therefore, $s_i$ 's cannot be formed. By leveraging the power of vision-language models, we further feed the input LR image $y$ to the CLIP vision encoder $V(.)$ , followed by an adapter $A_1(.)$ . The adapter $A_1(.)$ is formed from 3 fully connected (FC) layers interleaved by the ReLU activation function. Therefore, the synthetic degradation attributes that are aligned with visual modality are yielded as: + +$$ +\mathbf {f} _ {1} = A _ {1} (V (y)) \tag {3} +$$ + +During the training, we freeze the parameters of the CLIP text encoder $T(\cdot)$ , and update the parameters of the pre-trained CLIP vision encoder $V(\cdot)$ , as well as those of the adapter $A_{1}(\cdot)$ using: + +$$ +l _ {S y n} = \left\| \mathbf {s} - \mathbf {f} _ {1} \right\| _ {2} ^ {2} \tag {4} +$$ + +The degradation attributes $\mathbf{f}_1$ of the SDAL are able to estimate synthetic degradation features and, therefore, are crucial for training with synthetic degradation operations. However, in real-world cases, LR images can be decimated with unknown and complex operators associated with camera systems. To account for this important point, we now develop the RDAL module. + +Real-world Degradation Attribute Learning. Similar to the synthetic degradation attributes $\mathbf{f}_1$ , we need to produce real-world degradation attributes corresponding to $y$ . The challenge is that this is completely unsupervised since there does not exist any real-world degradation ground truth attribute for $y$ . In order to address this, we resort to the non-reference-based IQA metrics, which can evaluate the severity of the real-world degradation without using any ground truth HR image. + +Specifically, we form four textual sentences corresponding to the LR image $y$ with four non-reference- + +based IQA metrics (HyperIQA [36], BRISQUE [30], TOPIQ [3], and NIMA [37]). These four textual sentences are, respectively, image is degraded with unknown realistic degradation resulting in HyperIQA (e.g. 0.4)/BRISQUE (e.g. 72)/TOPIQ (e.g. 0.24)/NIMA (e.g. 5.7) image quality assessment value. Afterward, we feed these four textual sentences to the CLIP text encoder $T(\cdot)$ : + +$$ +\mathbf {r} = T ([ s _ {5}, s _ {6}, s _ {7}, s _ {8} ]) \tag {5} +$$ + +where $s_i$ 's $(i = 5,6,7,8)$ are the four textual sentences with IQA values described above. The real-world degradation attribute $\mathbf{r}$ is of size $\mathbb{R}^{512}$ . In Supplemental Material (SM), we will assure that degradation attributes $\mathbf{r}$ are of high quality. Please see SM. + +Despite the fact that the real-world degradation features $\mathbf{r}$ are able to model the unknown degradation processes in-wild, they are not suitable to be incorporated into the SR network. Specifically, the statistics of $\mathbf{r}$ are not aligned with those of the synthetic degradation attributes $\mathbf{s}$ , which leads to training instability. In order to alleviate this, we employ an adapter $A_{2}(.)$ , whose role is to align the feature statistics of the synthetic and real-world degradation attributes. The adapter $A_{2}(.)$ is formed from 3 FC layers interleaved by the ReLU activation function. Therefore, the final real-world degradation attributes are obtained as: + +$$ +\mathbf {f} _ {2} = A _ {2} (\mathbf {r}), \tag {6} +$$ + +Here, the parameters of the adapter $A_{2}(.)$ are updated with: + +$$ +l _ {\text {A l i g n}} = \left\| \mu_ {\mathbf {f} _ {2}} - \mu_ {\mathbf {s}} \right\| _ {2} ^ {2} + \left\| \sigma_ {\mathbf {f} _ {2}} - \sigma_ {\mathbf {s}} \right\| _ {2} ^ {2} \tag {7} +$$ + +where $\mu_{\mathbf{f}_2}$ and $\mu_{\mathbf{s}}$ are the mean of the synthetic and realworld degradation attributes in a batch, and $\sigma_{\mathbf{f}_2}$ and $\sigma_{\mathbf{s}}$ are their corresponding standard deviations. + +The $l_{Align}$ term aligns the statistics of the synthetic and real-world types of degradation attributes for a given LR image batch. This is crucial for obtaining reliable real-world degradation attribute $\mathbf{f}_2$ that is the composition of real-world degradation embeddings with different IQA values in the text embedding space. In fact, because of the $l_{Align}$ term, these real-world degradation embeddings can be viewed as corresponding to the synthetic embeddings within the realistic degradation space, maintaining one-to-one mapping that preserves the structural relationship. This would assist the Blind SR network to effectively generalize to the unseen degradation domain. + +The final issue in the design of RDAL is that the degradation attributes $\mathbf{f}_2$ are not obtained from the visual modality, hindering the SR performance. In this regard, on the top of the CLIP vision encoder $V(.)$ , we add another adapter operation $A_3(.)$ , which is formed from a cascade of 3 FC layers interleaved by the ReLU activation function. Hence, + +![](images/b14a0e22e7312f6a619e520ccd71faba6626e6949a643980212004e634c5d489.jpg) +Figure 3. Architecture of the proposed residual degradation-aware self-attention block (RDSB) using degradation embeddings for feature recalibration (Best viewed in color version). + +![](images/43087478e757b88fa6b28fe5d85b05cbd515671e6946680fb12d468eed27a2b0.jpg) +Figure 4. Architecture of the proposed diffusion-based Blind SR scheme using compositional zero-shot degradation learning (Best viewed in color version). + +the real-world degradation attributes that are aligned with the visual modality are: + +$$ +\mathbf {f} _ {3} = A _ {3} (V (y)) \tag {8} +$$ + +where the parameters of the adapter $A_{3}(.)$ are updated by the guidance of degradation attributes $\mathbf{f}_2$ as the supervisory signal: + +$$ +l _ {\text {R e a l}} = \left\| \mathbf {f} _ {2} - \mathbf {f} _ {3} \right\| _ {2} ^ {2} \tag {9} +$$ + +The overall loss function for training CZSD is summarized as: + +$$ +l _ {T o t a l} = l _ {S y n} + l _ {R e a l} + \alpha l _ {A l i g n} \tag {10} +$$ + +where $\alpha$ is the balancing hyper-parameter (we set it to 0.1). We use the same coefficients for $l_{Syn}$ and $l_{Real}$ to emphasize their equal importance. + +We have explained how to obtain the degradation attributes from a given LR image $y$ . It is seen that our compositional ZSL-based approach reduces the network's bias towards learning synthetic degradation operations, which is a key factor for the task of Blind SR. We now describe how to use this information in the training of our diffusion-based Blind SR network. + +LR Image Degradation Suppression. Following [21], we first reduce the artifacts of the LR image $y$ . Without loss of generality, we employ the SwinIR architecture [18], a self-attention-based neural network, and incorporate the degradation attributes into it. The input LR image $y$ is first passed through a convolutional layer (180 filters with kernel size 3) and its features are extracted. Next, these features are made + +
NetworkComponentsMetrics
RDALSDALAlign LossCLIP-IQA↑NIQE↓MUSIQ-AVA↑MANIQA↑CNN-IQA↑
\(Variant_1\)×××0.64964.95534.92900.51640.6496
\(Variant_2\)✓××0.66384.79054.97160.52850.6473
\(Variant_3\)×✓×0.66304.72644.99820.53080.6437
\(Variant_4\)✓✓×0.66094.97674.94840.52580.6425
Proposed✓✓✓0.68484.71525.11150.56870.6798
+ +Table 1. Impact of different components of the proposed ZFusion on SR performance. Red indicates the best. + +to undergo a sequence of 8 residual degradation-aware self-attention blocks (RDSB), whose architecture will be detailed in the next paragraph. The features obtained by the last RDSB are finally fed into a regular convolutional layer (3 filters with kernel size 3) in order to reconstruct the suppressed-artifact SR image $\tilde{x}$ . + +The architecture of the proposed RDSB is shown in Fig. 3. The input feature tensor $\mathbf{h}_1^i$ of the $i$ -th RDSB is first passed through 6 window-based self-attention operation (6 heads using 30 embeddings) followed by a convolutional layer (180 filters with kernel size 3) and the feature tensor $\mathbf{h}_2^i$ is yielded: + +$$ +\mathbf {h} _ {2} ^ {i} = \operatorname {C o n v} \left(W S A \left(\mathbf {h} _ {1} ^ {i}\right)\right) \tag {11} +$$ + +where $WSA(.)$ and $Conv(.)$ are the set of window-based self-attention and convolution operations explained above. + +Next, we recalibrate the feature tensor $\mathbf{h}_2^i$ using the degradation attributes. Specifically, as mentioned above, we obtained two sets of image degradation attributes, i.e., synthetic attributes $\mathbf{f}_1$ and real-world attributes $\mathbf{f}_3$ . In order to decouple the LR image as the composition of synthetic and real-world degradation attributes $\mathbf{f}_1$ and $\mathbf{f}_3$ , we form a fused degradation attribute $\mathbf{f}_f$ , which is obtained as: + +$$ +\mathbf {f} _ {f} = F C \left(\boldsymbol {\gamma} \mathbf {f} _ {1} + \boldsymbol {\delta} \mathbf {f} _ {3}\right) \tag {12} +$$ + +where $\gamma$ and $\delta$ are learnable parameters for fusing the two types of embeddings, and $FC(.)$ is a cascade of 3 FC layers with ReLU activations (it adjusts the fused degradation embeddings for each hierarchical level). This provides the Blind SR network with rich clues for feature generation. Next, the feature tensor $\mathbf{h}_2^i$ that is obtained from a set of self-attention operations and a convolutional layer, is multiplied by $\mathbf{f}_f$ and the final output feature tensor of the $i$ -th RDSB block $\mathbf{h}_3^i$ is generated: + +$$ +\mathbf {h} _ {3} ^ {i} = \mathbf {h} _ {2} ^ {i} \odot \mathbf {f} _ {f} + \mathbf {h} _ {2} ^ {i} \tag {13} +$$ + +where $\odot$ is the element-wise multiplication. + +We train the degradation suppression network with the $\ell_1$ norm loss function. Even though the SR image $\check{x}$ obtained by this network has a visual quality superior to its low-quality counterpart $y$ , it still suffers from the lack of + +sharp textures and photo-realistic structures [21, 43]. To alleviate this, we feed this image to a diffusion-based neural network for recovering high-frequency components. + +Generative Diffusion Prior. We now employ the StableDiffusion [35] as the generative diffusion prior for our Blind SR network and embed the degradation attributes as conditional signal into it. We utilize the StableDiffusion [35] as the backbone architecture for restoring the detailed textures and structures (please see Fig. 4). Specifically, we have used the same network architecture for the encoder, denoiser, and decoder blocks of our diffusion model as those used in [35]. The architecture of the controlnet is the same as that of the encoder part of the denoiser block. In order to integrate the degradation attributes $\mathbf{f}_f$ from LR image as a condition for the diffusion model, we develop a degradation-aware adapter (DA) module $A_D^t$ ( $t = 0, \dots, T$ ) at every $t$ -th diffusion time step. This DA module consists of 3 FC layers followed by the ReLU activation function. Following [21, 40], we freeze the parameters of the encoder, denoiser, and decoder blocks of the diffusion model and only update those of the controlnet and DA module with: + +$$ +l _ {D i f f} = \| \epsilon - \epsilon_ {\Theta} (\mathbf {x} _ {t}, t, \check {\mathbf {x}}, A _ {D} ^ {t} (\mathbf {f} _ {f})) \| _ {2} ^ {2} \tag {14} +$$ + +where $\mathbf{x}_t = \sqrt{\bar{\alpha}_t}\mathbf{x} + \sqrt{1 - \bar{\alpha}_t}\epsilon$ , $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , $\mathbf{x} = \mathcal{E}(x)$ and $\breve{\mathbf{x}} = \mathcal{E}(\breve{x})$ are the outputs of the encoder block for $x$ and $\breve{x}$ , respectively, and $\Theta$ is the set of the parameters of the diffusion model. It is seen from Eq. (14) that the proposed Blind SR scheme uses the degradation embeddings at each time step in order to guide the diffusion reverse process. This contributes in generating feature maps by the diffusion model that are well suited for the task of Blind SR. + +In view of the fact that the proposed scheme uses compositional zero-shot learning for diffusion-based Blind SR, we refer to it as ZFusion. + +# 4. Experimental Results + +# 4.1. Implementation Details + +We use the images of DIV2K [1], Flickr2K [38], and OutdoorSceneTraining [42] to train the proposed diffusion-based Blind SR scheme. To have a fair comparative study with other diffusion-based Blind SR schemes, we only employ synthetic degradation operations during the training + +
MetricReal-ESRGANResShiftPASDSeeSRSinSRStableSRDiffBIRZFusion (Ours)
PSNR↑13.8521.8821.3120.8521.4720.3221.0621.05
SSIM↑0.36840.59160.56910.54690.56400.52960.53700.5547
LPIPS↓0.52710.27890.31760.27870.28940.28030.29040.2694
CLIP-IQA↑0.58430.59680.63230.73890.65550.67540.76440.7882
NIQE↓3.91395.89894.11314.07055.07604.01234.47824.3005
MUSIQ-AVA↑5.06735.08905.47805.64535.22195.05325.68065.7624
MANIQA↑0.22030.19710.25540.33950.23870.21100.38000.6141
CNN-IQA↑0.61860.58970.62570.68760.64850.60550.69770.7044
+ +Table 2. Comparison between various SOTA Blind SR schemes on the images of DIV2K dataset with synthetic degradations. Red indicates the best, while Blue shows the second best. + +
MetricReal-ESRGANResShiftPASDSeeSRSinSRStableSRDiffBIRZFusion (Ours)
PSNR↑23.6622.9324.1722.5622.0220.3622.5521.05
SSIM↑0.73740.68320.72020.64820.64960.63090.62400.5788
LPIPS↓0.20300.22290.20170.23550.22700.22420.23680.2456
CLIP-IQA↑0.51650.58600.56800.64890.67520.63270.66670.6848
NIQE↓5.21066.91625.00004.89845.61925.43765.11714.7152
MUSIQ-AVA↑4.75654.78494.90505.08554.90564.67985.10225.1115
MANIQA↑0.25070.23050.26790.33710.30820.24040.36320.5687
CNN-IQA↑0.62560.59460.60000.66450.66970.59760.67930.6798
+ +Table 3. Comparison between performances of various Blind SR schemes on the images of RealSR dataset with real-world unknown degradations. Red indicates the best, while Blue shows the second best. + +process using the image degradation pipeline of BSRGAN [61]. However, thanks to efficiently leveraging the compositional ZSL technique, our ZFusion is able to generalize to the unknown complex decimations at inference. Further implementation details of ZFusion are given SM. + +In order to perform a thorough comparative study, we employ 8 IQA metrics, namely, PSNR, SSIM [45], LPIPS [62], CLIP-IQA [39], NIQE [31], MUSIC-AVA [13], MANIQA [56], and CNN-IQA [12]. It is to be noted that these IQA metrics are different than those employed in RDAL module of our ZFusion. Since it has been shown [21], [40] that non-reference-based IQA metrics are better aligned for assessing the generative restoration capability of diffusion-based Blind SR methods, we report the results of ablation studies in terms of non-reference-based IQA metrics. However, we employ all the 8 IQA metrics mentioned above for comparative experiments. + +# 4.2. Ablation Studies + +The principal contribution of the proposed diffusion-based Blind SR scheme is its improved generalization capability by employing compositional ZSL. Specifically, we have designed the CZSD module in order to improve the performance of the proposed method in the case of LR images with unknown and complex degradation processes. In order to investigate the effectiveness of our compositional ZSL-based approach on the super resolution performance, we remove the CZSD module from our ZFusion (Variant1). + +In this case, the diffusion-based SR network does not employ any degradation embeddings in its architecture. The results of this experiment in the case of images of RealSR [2] dataset are shown in the first row of Tab. 1. As seen from this table, removing the CZSD module from ZFusion leads to significantly deteriorating the SR performance. This is justifiable in view of the fact that the CZSD module results in improving the generalization capacity of the proposed scheme to the realistic in-wild situations, which is the main objective of the task of Blind SR. + +We now form variants of ZFusion that only use synthetic degradation attributes or realistic degradation attributes individually (Variant $_2$ and Variant $_3$ ). The results of this experiment on the LR images of RealSR dataset are given in the second and third rows of Tab. 1. As seen from these results, by fusing the two types of degradation attributes, the decimation process of the LR image is efficiently decoupled as the composition of those synthetic and real-world embeddings optimized in training. This is indeed effective for providing superior SR performances. + +As pointed out in Sec. 3, the alignment loss $(l_{\text{Align}})$ term guarantees that the CZSD module mitigates bias towards synthetic degradations used in training. In order to scrutinize the effect of the alignment loss term, we remove it from the CZSD module $(\text{Variant}_4)$ . The results of this experiment on the images of RealSR [2] dataset are shown in the fourth row of Tab. 1. It can be concluded that without the alignment loss term, the network performance drops + +![](images/6a9d4a945bf2eeb86a0dfb78aecd602af536666d8c17f64ca995012ed1ba2398.jpg) +RealSR Canon 025 + +![](images/ea2c7b461a72eee4b77c9def0e50ed0d86fe399ea529736112b243593d99b513.jpg) +LR + +![](images/d23bf64a9b28ccad6d7f1a82aabc990bf23d18c5f2c9a18a863bcb6f5119127e.jpg) +ResShift + +![](images/a74fd619828e258ada2526a03c861411cf123590121a3e31b06fb1eeb94fceac.jpg) +PASD + +![](images/74c47889989f6a1bb6afa9a7f33d20abe617ca6e5c4905a40d276eaec57bfdea.jpg) +SeeSR + +![](images/785fd1089335354fea555069c4cf0060fd5661585632f94077acec1268a67551.jpg) +SinSR + +![](images/33939a15b35d7e53bc270adc9d6c3d5f8f641a9e07bc05c0234cdb433ade5b38.jpg) +StableSR + +![](images/77dd84ba9878e5009213c795f1d0422eefc9ec6e7cf381d04be9123d3c2a8eb9.jpg) +DiffBIR + +![](images/f7e732cbc689733b2edd8391314f91c66d38903cec287663ae6238b5edeff175.jpg) +Ours + +![](images/531c4cc6f310d0f364aa44863c811b280707a77fa34eacf1d65a6ca0383c6399.jpg) +RealSR Nikon 021 + +![](images/4d9c28b7336079c7eca48908af93a5d423cc68a6c136b6eb9aa3f33f40a80ea7.jpg) +LR + +![](images/5e7646a4172b813a201bd86377c072044301ce50fed7fb6dbefd179bde8c438b.jpg) +ResShift + +![](images/0219ccc351c7a18e2b7a71866dca58fe31b95572bb7c2e20e406275f74ec37fd.jpg) +PASD + +![](images/a1dc44bd39621594ce61a595b5aa1152cd3851a7fbb1cd580e103b4b47336358.jpg) +SeeSR + +![](images/a009dfe0349b552fe730a18cbfcfa244fc13f656f016d91c4765f38654fd1a2e.jpg) +SinSR +Figure 5. Visual qualities of images from the RealSR dataset with realistic unknown degradations that are super resolved by various state-of-the-art image super resolution schemes. + +![](images/6f40f74c4f811808ff479cc0db86cb3c0dc03226cb09cb9405eed4f628994a42.jpg) +StableSR + +![](images/791c51bddc6916250e7bd4deac22dc9a7c22d1c313ed72862385b4c0846fe7b7.jpg) +DiffBIR + +![](images/971be1ce4a90d9279e42d4fb16d024781ae15d4cd0d8db04a3178190a0f2149b.jpg) +Ours + +remarkably. This is obvious since, in this case, the SR network does not generalize beyond the synthetic degradation operations employed during the training. The results of further ablation experiments are given in SM. + +# 4.3. Comparison with the State-of-the-art Schemes + +We now compare the performance of the proposed diffusion-based Blind SR scheme with those of 7 SOTA methods, namely, Real-ESRGAN [43], ResShift [59], PASD [57], SeeSR [49], SinSR [44], StableSR [40], and DiffBIR [21]. The results of various SOTA Blind SR schemes on 3000 sub-image pairs of size $512 \times 512 \times 3$ extracted from the DIV2K validation set [1] are given in Tab. 2. It has to be noted that these sub-images are degraded by the decimation process of BSRGAN with various randomly selected synthetic operations. It is seen from the results of this table that our ZFusion provides 5 superior IQA values out of a total 8 number of metrics. This shows that when the LR images are generated using the common synthetic degradation operators, namely blurring, downsampling, noise, and JPEG compression, our scheme is able to effectively suppress these artifacts, as well as reconstruct the high-frequency components. + +Since the main purpose of Blind SR is to improve the quality of the images degraded by real-world decimation processes, we now compare the performances of our ZFusion with those of the other SOTA schemes in the case of the images of RealSR [2]. These LR images are acquired with a camera realistic degradation pipeline. The results of various methods on the images of this dataset are given in Tab. 3. As seen from this table, our proposed ZFusion is able to outperform the other methods in 5 out of 8 cases + +of IQA values. DiffBIR, standing out as the second best performing method, yields 4 second-best metrics. This indicates that by employing the compositional ZSL-based algorithm in the design of ZFusion, it can better generalize to the real-world degradation operations, even though our method is only trained with the LR images obtained by synthetic operations. We also perform a comparative study in SM on the images of DRealSR [46] dataset, which are obtained by real-world degradations. Please see SM. + +In order to investigate the qualitative results generated by our ZFusion and the other SOTA schemes, we show the results of various methods in Fig. 5. As seen from this figure, generally the proposed scheme produces more authentic sharp textures and structures than the other methods do. For instance, it is seen that the hand's texture is sharper and more natural in the super-resolved image of our ZFusion. We provide several more visual quality comparisons in SM. All of them further confirm the effectiveness of ZFusion. + +The time complexity comparison between various Blind SR schemes is shown in SM. + +# 5. Conclusion + +In this paper, we have designed an effective diffusion-based Blind SR method, which employs compositional zero-shot learning for better generalizing to the unseen cases of LR images. The degradation attributes obtained by the compositional ZSL have been efficiently integrated into the architecture of our diffusion-based scheme for generating SR images with superior visual qualities. The results of extensive experiments have confirmed the effectiveness of our approach over the other SOTA methods. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 6, 8 +[2] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE International Conference on Computer Vision, 2019, 2, 7, 8 +[3] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 5 +[4] Honggang Chen, Xiaohai He, Hong Yang, Yuanyuan Wu, Linbo Qing, and Ray E Sheriff. Self-supervised cycle-consistent learning for scale-arbitrary real-world single image super-resolution. Expert Systems with Applications, 212: 118657, 2023. 1 +[5] Zheng Chen, Yulun Zhang, Jinjin Gu, Xin Yuan, Linghe Kong, Guihai Chen, and Xiaokang Yang. Image superresolution with text prompt diffusion. arXiv preprint arXiv:2311.14282, 2023. 1, 2 +[6] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 1 +[7] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. +[8] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European conference on computer vision, pages 222-241. Springer, 2024. 1 +[9] Shaozhe Hao, Kai Han, and Kwan-Yee K Wong. Learning attention as disentangler for compositional zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15315-15324, 2023. 2 +[10] Siteng Huang, Biao Gong, Yutong Feng, Min Zhang, Yiliang Lv, and Donglin Wang. Troika: Multi-path cross-modal traction for compositional zero-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24005-24014, 2024. 2 +[11] Yan Huang, Shang Li, Liang Wang, Tieniu Tan, et al. Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems, 33:5632-5643, 2020. 1, 2, 3 +[12] Le Kang, Peng Ye, Yi Li, and David Doermann. Convolutional neural networks for no-reference image quality assessment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1733-1740, 2014. 7 +[13] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. + +In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 7 +[14] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 1 +[15] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 1 +[16] Xiaoming Li, Chaofeng Chen, Xianhui Lin, Wangmeng Zuo, and Lei Zhang. From face to natural image: Learning real degradation for blind image super-resolution. In European Conference on Computer Vision, pages 376-392. Springer, 2022. 2 +[17] Xiangyu Li, Xu Yang, Kun Wei, Cheng Deng, and Muli Yang. Siamese contrastive embedding network for compositional zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9326-9335, 2022. 2 +[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 5 +[19] Jingyun Liang, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Mutual affine network for spatially variant kernel estimation in blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4096-4105, 2021. 2 +[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1 +[21] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffir: Toward blind image restoration with generative diffusion prior. In European Conference on Computer Vision, pages 430-448. Springer, 2024. 1, 2, 5, 6, 7, 8 +[22] Huan Liu, Mingwen Shao, Yuanjian Qiao, Yecong Wan, and Deyu Meng. Unpaired image super-resolution using a lightweight invertible neural network. Pattern Recognition, 144:109822, 2023. 1 +[23] Man Liu, Feng Li, Chunjie Zhang, Yunchao Wei, Huihui Bai, and Yao Zhao. Progressive semantic-visual mutual adaptation for generalized zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15337–15346, 2023. 2 +[24] Fangzhou Luo, Xiaolin Wu, and Yanhui Guo. And: Adversarial neural degradation for learning blind image superresolution. Advances in Neural Information Processing Systems, 36:21255-21267, 2023. 2 +[25] Zhengxiong Luo, Yan Huang, Shang Li, Liang Wang, and + +Tieniu Tan. End-to-end alternating optimization for blind super resolution. arXiv preprint arXiv:2105.06878, 2021. 1 +[26] Ziwei Luo, Haibin Huang, Lei Yu, Youwei Li, Haoqiang Fan, and Shuaicheng Liu. Deep constrained least squares for blind image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17642-17652, 2022. 3 +[27] Shunta Maeda. Unpaired image super-resolution using pseudo-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 291-300, 2020. 1 +[28] Shunta Maeda. Unpaired image super-resolution using pseudo-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 291-300, 2020. 1 +[29] Massimiliano Mancini, Muhammad Ferjad Naeem, Yongqin Xian, and Zeynep Akata. Open world compositional zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5222-5230, 2021. 2 +[30] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 5 +[31] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 7 +[32] Muhammad Ferjad Naeem, Yongqin Xian, Federico Tombari, and Zeynep Akata. Learning graph embeddings for compositional zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 953-962, 2021. 2 +[33] Farhad Pourpanah, Moloud Abdar, Yuxuan Luo, Xinlei Zhou, Ran Wang, Chee Peng Lim, Xi-Zhao Wang, and QM Jonathan Wu. A review of generalized zero-shot learning methods. IEEE transactions on pattern analysis and machine intelligence, 45(4):4051-4070, 2022. 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6 +[36] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667-3676, 2020. 5 +[37] Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE transactions on image processing, 27(8): 3998-4011, 2018. 5 +[38] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single + +image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 6 +[39] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, pages 2555-2563, 2023. 7 +[40] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929-5949, 2024. 2, 6, 7, 8 +[41] Qingsheng Wang, Lingqiao Liu, Chenchen Jing, Hao Chen, Guoqiang Liang, Peng Wang, and Chunhua Shen. Learning conditional attributes for compositional zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11197-11206, 2023. 2 +[42] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6 +[43] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 6, 8 +[44] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25796-25805, 2024. 8 +[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[46] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 8 +[47] Gang Wu, Junjun Jiang, and Xianming Liu. A practical contrastive learning framework for single-image superresolution. IEEE Transactions on Neural Networks and Learning Systems, 2023. 1 +[48] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. Advances in Neural Information Processing Systems, 37:92529-92553, 2024. 2 +[49] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 2, 8 +[50] Bin Xia, Yucheng Hang, Yapeng Tian, Wenming Yang, Qingmin Liao, and Jie Zhou. Efficient non-local contrastive + +attention for image super-resolution. In Proceedings of the AAAI conference on artificial intelligence, pages 2759-2767, 2022. 1 +[51] Bin Xia, Yulun Zhang, Yitong Wang, Yapeng Tian, Wenming Yang, Radu Timofte, and Luc Van Gool. Knowledge distillation based degradation estimation for blind super-resolution. arXiv preprint arXiv:2211.16928, 2022. 3 +[52] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 2 +[53] Rui Xie, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Jian Yang, and Ying Tai. Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 2 +[54] Guangyue Xu, Parisa Kordjamshidi, and Joyce Y Chai. Zero-shot compositional concept learning. arXiv preprint arXiv:2107.05176, 2021. 3 +[55] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. Advances in Neural Information Processing Systems, 33:21969-21980, 2020. 2 +[56] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1191-1200, 2022. 7 +[57] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, pages 74-91. Springer, 2024. 1, 2, 8 +[58] Zongsheng Yue, Qian Zhao, Jianwen Xie, Lei Zhang, Deyu Meng, and Kwan-Yee K Wong. Blind image super-resolution with elaborate degradation modeling on noise and kernel. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2128-2138, 2022. 2 +[59] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36:13294-13307, 2023. 8 +[60] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3262-3271, 2018. 1, 2, 3 +[61] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4791-4800, 2021. 1, 3, 7 +[62] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 + +[63] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. In European conference on computer vision, pages 649-667. Springer, 2022. 1 +[64] Yuzhe Zhang, Jiawei Zhang, Hao Li, Zhouxia Wang, Luwei Hou, Dongqing Zou, and Liheng Bian. Diffusion-based blind text image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25827-25836, 2024. 2 +[65] Zhilu Zhang, Ruohao Wang, Hongzhi Zhang, Yunjin Chen, and Wangmeng Zuo. Self-supervised learning for real-world super-resolution from dual zoomed observations. In European Conference on Computer Vision, pages 610-627. Springer, 2022. 1 +[66] Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, and Chen Change Loy. Cross-scale internal graph neural network for image super-resolution. Advances in neural information processing systems, 33:3499-3509, 2020. 1 \ No newline at end of file diff --git a/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/images.zip b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f792cfab29840bf4a3816f51101c012a1f6d4f92 --- /dev/null +++ b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a749372204ca12f1d7bed15252ba3177c90e7f60dde420db17f8a463e177e84b +size 465622 diff --git a/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/layout.json b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1153575d37252361d449bb1bf8e0f68c0b56aac4 --- /dev/null +++ b/zfusionefficientdeepcompositionalzeroshotlearningforblindimagesuperresolutionwithgenerativediffusionprior/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69491718144d3aa454c8d6b02bbde1b4f76dd7c2bf1083b8f4f4ad37be22f0cb +size 508905 diff --git a/zimzeroshotimagemattingforanything/ec21a9a6-f4ee-473f-b698-87f36a029477_content_list.json b/zimzeroshotimagemattingforanything/ec21a9a6-f4ee-473f-b698-87f36a029477_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..219b8aa4f619dcac001967f08fee83a39ec442fd --- /dev/null +++ b/zimzeroshotimagemattingforanything/ec21a9a6-f4ee-473f-b698-87f36a029477_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28866c2a4ee5052ae679551ca659d464ec751748e9ca4a5d6594d31010eeb374 +size 89971 diff --git a/zimzeroshotimagemattingforanything/ec21a9a6-f4ee-473f-b698-87f36a029477_model.json b/zimzeroshotimagemattingforanything/ec21a9a6-f4ee-473f-b698-87f36a029477_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fc6dc3767b3963be884d387f981278dc40aa59dd --- /dev/null +++ b/zimzeroshotimagemattingforanything/ec21a9a6-f4ee-473f-b698-87f36a029477_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bfa4c9ec12640d307a9012adc85327723bc0ea55ccfdad2f4c45d23fc4dfa06 +size 107044 diff --git a/zimzeroshotimagemattingforanything/ec21a9a6-f4ee-473f-b698-87f36a029477_origin.pdf b/zimzeroshotimagemattingforanything/ec21a9a6-f4ee-473f-b698-87f36a029477_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fdc84751995360d6780407ad92825bb9edde6068 --- /dev/null +++ b/zimzeroshotimagemattingforanything/ec21a9a6-f4ee-473f-b698-87f36a029477_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5127b1690f178103c81bab3d7d40bcef81cb9dc3d54b677e096234afc63a60c +size 3687660 diff --git a/zimzeroshotimagemattingforanything/full.md b/zimzeroshotimagemattingforanything/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8837e673194bf71af1e38c8e38f961a9ad42f084 --- /dev/null +++ b/zimzeroshotimagemattingforanything/full.md @@ -0,0 +1,388 @@ +# ZIM: Zero-Shot Image Matting for Anything + +Beomyoung Kim Se-Yun Lee + +Chanyong Shin +Sewhan Chun + +Joonhyun Jeong Dong-Hyun Hwang + +Hyungsik Jung +Joonsang Yu + +NAVER Cloud, ImageVision + +![](images/bca2a0802bf55979aae4eafdb736646976cb0131baeb81a5999ab491f93976db.jpg) +(a) +Image + +![](images/d1f5717d30d2815944404b50519d5a31c2caea771a0f9ff03a21506590126734.jpg) + +![](images/bec78fca0ec4c0d03bcc14ed5bfe55fbe75349625ad998a89684b491ecb5b9bd.jpg) +SAM + +![](images/122566fab7d6ca066e64a47a006766bfcccf1264e0def37d21bd4206ae7e79d1.jpg) + +![](images/d8c5675c8674fbb17582863b5e7a32c7a2bdd8e29847b2a8e5fe91b6329657cb.jpg) +HQ-SAM + +![](images/3d27751185c9f8ad4242aebc7bb2c7a531857bda708499a2f02608aa9b39d3b0.jpg) + +![](images/7adfd66d035a4b31556ac8c9f54a4e7aaa083a84523ea351c6ba9ae1b08b1018.jpg) +Matte-Any + +![](images/6fc94a0877d189a0ba44c1a253cbd0e0120ce3b5697b52d8aa4ad6769d70493b.jpg) + +![](images/bc181144df9227a816bb407ec937a1aa663e0cf173979da3aec29bdef462cd64.jpg) +Matting-Any + +![](images/185b220c78cfc3b268f1ec9fc9b919fb58d75df5b57f913351844a524f30af0b.jpg) + +![](images/4736664d05d4b392c9d4dcd3031205e8587d50721b98a311f239c298f346327a.jpg) +ZIM (ours) + +![](images/36d94e60c7c0840862b8a9ad6b8806191bab6f2803cf5a454d1655c419fed742.jpg) + +![](images/a04de1ff1c03cf70e9eeb11c33974ef14a0911fa8412fa589eaeba3ef5604532.jpg) +(b) + +![](images/494585604fa3e47b2bfe7a8b13db3ca341864839aaa8e9c49c0a4e142be4c7dc.jpg) + +![](images/647e186b23a91bf3bb8f501e6b85c3528b5512e3c12908d6cbe988aabb2ba3ff.jpg) + +![](images/97f4033e386bd469d788b2b5c23660e7d41700e0600af1528354b604704a117f.jpg) + +![](images/9381d1fc9a4207834549b830bbba8c5dc117715f64a1c24763ea8a8ba725d533.jpg) + +![](images/8710fbf67c6f87b047d3668d4f19ebd4230b19ec6e00560f41581f0c0a54625e.jpg) + +![](images/68113c6653d1bdc20e36c8db6413d3724c1268a1e6b1048f4f25191d6f38719d.jpg) + +![](images/422046ff47bb7bc95f9759c0f7d2b766806157c73059cf8527dfc4f7ff462b34.jpg) + +![](images/4b43ad841933ce3b1f7745e5cab24e3986facb0e000c28789828159da78188bf.jpg) + +![](images/97b886215e1f6b1c0fde79ddef85ad39c18445304557bc30d3ef6bc003c27bb7.jpg) + +![](images/bc5c174ac66a1171d7554c2d7ae556014d49372018ec42a84b870e0c400a9543.jpg) + +![](images/5c2121da667130276c44e40774a4b07d41764c2519bfc6619d7b480528fa079e.jpg) + +![](images/9266c6764e7b4baec6fe1c034bb6fbfcb201ca9f6ca8c0b5c2c396f81bfe51b6.jpg) +(c) + +![](images/0be2922e3b474361e0d108f5e5987da9b14c32eff5517f1ddad0c81c17ba1f41.jpg) + +![](images/e1a21c96937ed89b7167d9135d84289b4a86092f0330b09cdfec809f2a0e5f9c.jpg) + +![](images/ff666f502ca1c77eae492b060689df5db468c24ba04b4e783f5ebeae8c93f46f.jpg) +Figure 1. Qualitative comparison of ours with five existing zero-shot models (SAM [18], HQ-SAM [13], Matte-Any [50], and Matting-Any [25]). It showcases (a) box prompting results, (b) point prompting results, and (c) automatic mask generation results. + +![](images/6d3ea7de74129d09e82ae542a14c00581b6051c4c53766bb5436720508da70dd.jpg) + +![](images/17c1513d0f622f7d66f0cd7e5df262f76fa40b1636bf3916aea79d56f45acf08.jpg) + +![](images/1c6c666e4c78549deca1ee194e600f637060183137df36b4120b061fe217dcd5.jpg) + +![](images/a9e7569d41bc88f470e72bafddfe5ca33fff35b0b201295202b48313d521f93f.jpg) + +![](images/4806651806a289d442bad13483b409cc9660f650aea501384918c1e4c4836fea.jpg) + +![](images/175099d5d9086d5ac72b09a0bd7a41a8633960fb50128d520f2da5e8e2c13819.jpg) + +![](images/9785ac16519e6f6ec2c8cbd29e194351670a944397ad8e1f367ffe6866a9baa7.jpg) + +![](images/93038a79ecfa7bd27808fc86a895e4bae9668e91260da1ba097444bbfc6fbca4.jpg) + +# Abstract + +The recent segmentation foundation model, Segment Anything Model (SAM), exhibits strong zero-shot segmentation capabilities, but it falls short in generating fine-grained precise masks. To address this limitation, we propose a novel zero-shot image matting model, called ZIM, with two key contributions: First, we develop a label converter that transforms segmentation labels into detailed matte labels, constructing the new SA1B-Matte dataset without costly manual annotations. Training SAM with this dataset enables it to generate precise matte masks while maintaining its zero-shot capability. Second, we design the zero-shot matting model equipped with a hierarchical pixel decoder + +to enhance mask representation, along with a prompt-aware masked attention mechanism to improve performance by enabling the model to focus on regions specified by visual prompts. We evaluate ZIM using the newly introduced MicroMat-3K test set, which contains high-quality micro-level matte labels. Experimental results show that ZIM outperforms existing methods in fine-grained mask generation and zero-shot generalization. Furthermore, we demonstrate the versatility of ZIM in various downstream tasks requiring precise masks, such as image inpainting and 3D segmentation. Our contributions provide a robust foundation for advancing zero-shot matting and its downstream applications across a wide range of computer vision tasks. The code is available at https://naver-ai.github.io/ZIM. + +# 1. Introduction + +Image segmentation, which divides an image into distinct regions to facilitate subsequent analysis, is a fundamental task in computer vision. Recent breakthroughs in segmentation models have made significant strides in this area, particularly with the emergence of the segmentation foundation model, Segment Anything Model (SAM) [18]. SAM is trained on the SA1B dataset [18] containing 1 billion micro-level segmentation labels, where its extensiveness enables SAM to generalize effectively across a broad range of tasks. Its strong zero-shot capabilities, powered by visual prompts, have redefined the state of the art in zero-shot interactive segmentation and opened new avenues for tackling more complex tasks within the zero-shot paradigm. + +Despite these achievements, SAM often struggles to generate masks with fine-grained precision (see Figure 1). To address this limitation, recent studies [25, 50, 51] have extended SAM to the image matting task, which focuses on capturing highly detailed boundaries and intricate details such as individual hair strands. These approaches achieve enhanced mask precision by fine-tuning SAM on publicly available matting datasets [21, 36, 47]. However, this fine-tuning process can undermine the zero-shot potential of SAM, since most public matting datasets contain only macro-level labels (e.g., entire human portrait) rather than the more detailed micro-level labels (e.g., individual body parts), as illustrated in Figure 2. Fine-tuning with macro-level labels can cause SAM to overfit to this macro-level granularity, resulting in catastrophic forgetting of its ability to generalize at the micro-level granularity, as shown in Figure 1. Moreover, the scarcity of large-scale matting datasets with micro-level matte labels poses a significant obstacle in developing effective zero-shot matting solutions. + +In this paper, we introduce a pioneering Zero-shot Image Matting model, dubbed ZIM, that retains strong zero-shot capabilities while generating high-quality micro-level matting masks. A key challenge in this domain is the need for a matting dataset with extensive micro-level matte labels, which are costly and labor-intensive to annotate. To address this challenge, we propose a novel label conversion method that transforms any segmentation label into a detailed matte label. For more reliable label transformation, we design two effective strategies to reduce noise and yield high-fidelity matte labels (Section 3.1). Subsequently, we construct a new dataset, called SA1B-Matte, which contains an extensive set of micro-level matte labels generated by transforming segmentation labels from the SA1B dataset via the proposed converter (see Figure 2). By training SAM on the SA1B-Matte dataset, we introduce an effective foundational matting model with micro-level granularity while preserving the zero-shot ability of SAM (see Figure 1). + +To further ensure effective image matting, we enhance the major bottleneck in the network architecture of SAM + +that impedes capturing robust and detailed feature maps. Specifically, SAM employs a simple pixel decoder to generate mask feature maps with a stride of 4, which is susceptible to checkerboard artifacts and often falls short in capturing fine details. To mitigate this, we design a more elaborated pixel decoder, enabling more robust and richer mask representations (Section 3.2). Furthermore, we introduce a prompt-aware masked attention mechanism that leads to the improvement of interactive matting performance. + +To validate our zero-shot matting model, we present a new test set, called MicroMat-3K, consisting of 3,000 high-quality micro-level matte labels. Our experiments on this dataset demonstrate that while SAM exhibits strong zero-shot capabilities, it struggles to deliver precise mask outputs. In contrast, existing matting models show limited zero-shot performance. ZIM, however, not only maintains robust zero-shot functionality but also provides superior precision in mask generation. Additionally, we highlight the foundational applicability of ZIM in several downstream tasks requiring precise masks, such as image inpainting [54] and 3D segmentation [3]. We hope this work provides valuable insights to the research community, encouraging further development of zero-shot matting models. + +# 2. Related Work + +Image Segmentation. Image segmentation is a fundamental task in computer vision, enabling the division of an image into distinct regions. Recent advancements in segmentation models [6, 12, 19] have significantly improved the accuracy of segmentation tasks, including semantic, instance, and panoptic segmentation. The emergence of Segment Anything Model (SAM) [18] introduced a new paradigm in segmentation by leveraging visual prompts (e.g., points or boxes). SAM is designed as a foundational segmentation model capable of handling diverse tasks due to its robust zero-shot capabilities, showing remarkable versatility across a wide range of tasks and domains. However, despite its strengths, SAM struggles to produce high-precision masks. In this paper, we address this limitation by developing a novel zero-shot model that enhances mask precision while maintaining SAM's generalization capabilities. + +Image Matting. Image matting is a more complex task than image segmentation, as it focuses on estimating the soft transparency of object boundaries to capture fine details, which is critical in tasks like image compositing and background removal. Unlike segmentation, which assigns hard labels to each pixel, matting requires precise edge detection and soft labeling for smooth blending between objects and their background. Recent developments in zero-shot matting have aimed to build upon the foundational segmentation capabilities of SAM. Most approaches [25, 50, 51] fine-tune SAM on public matting datasets [21, 22, 36, 47]. + +![](images/ce7bfc31ac21d482d08bcf2fccff17c4908533a1db12ea1e887de489f2a60b5c.jpg) +Public Matting Datasets: Macro-Level Fine Labels + +![](images/a4b41f0d5a69e50cdfe89cbe4765df6c2a75d46b0744849cc1b19b8d61f7db39.jpg) +SA1B: Micro-Level Coarse Labels +Figure 2. Qualitative samples from each dataset: Public matting datasets [21, 22, 36, 47] with macro-level fine labels, the SA1B dataset [18] with micro-level coarse labels, and our proposed SA1B-Matte dataset incorporating the micro-level labels with fine details. + +![](images/037455585611bd2275133a3cb9c1a22d074d4cac0182cf45520c8192ec927ca9.jpg) +SA1B-Matte (Ours): Micro-Level Fine Labels + +![](images/ff1cb2508b72dd8bbba7fe85a560da9ee8040f59a9df98008f3c629b4eda5862.jpg) + +![](images/a254425772a0ed4a3db517fc163b3164f7befd7a9e39de93f89be1685daafb08.jpg) + +![](images/ee048359a27d59b98a8929089e9bcce2713dac1def331f6ffda05e9e922b902e.jpg) + +![](images/a4bb5a6120c28e8fab2f55f0c9f65e536e502f08255cffa5802e361de11a63b8.jpg) + +![](images/8a60ac83edd19d8fb1e7051f555963c1ac1b07eab8c4f535bc32c225cbc06752.jpg) + +However, these datasets predominantly contain macro-level labels, degrading SAM's ability to generalize on micro-level structures, such as individual body parts of a human. The reliance on these datasets can deteriorate the zero-shot generalization of the model. Furthermore, the lack of large-scale matting datasets with micro-level labels restricts progress in developing matting models with truly effective zero-shot ability. In this paper, we correspondingly construct a large-scale micro-level labeled matting dataset via our proposed label converter without laborious annotation procedures, enabling effective zero-shot matting modeling. + +# 3. Methodology + +Our contributions can be divided into two components: matting dataset construction (Section 3.1) and network architecture enhancements (Section 3.2). + +# 3.1. Constructing the Zero-Shot Matting Dataset + +Motivation. For effective zero-shot matting, a dataset with micro-level matte labels is essential. However, manually annotating matte labels at the micro-level requires extensive human labor and cost. To this end, we present an innovative Label Converter that transforms any segment label into a matte label, motivated by mask-guided matting works [35, 53]. We first collect public image matting datasets [20-22, 24, 44, 53] to train the converter. We derive coarse segmentation labels from matte labels by applying image processing techniques such as thresholding, resolution down-scaling, Gaussian blurring, dilation, erosion, and convex hull transformations. The converter takes an image and segmentation label as input source and is trained to produce a corresponding matte label, as illustrated in Figure 3a. + +Challenges. (1) Generalization to unseen patterns: Public matting datasets predominantly contain macro-level labels (e.g., entire portraits), as shown in Figure 2. Consequently, the converter trained on these datasets often struggles to generalize to unseen micro-level objects (e.g., individual + +body parts). This limitation leads to the generation of noisy matte labels when applied to micro-level segmentation (see the 4th column in Figure 6a). (2) Unnecessary fine-grained representation: Some objects, such as cars or boxes, commonly do not require fine-grained representation. However, since the converter is trained to always transform segmentation labels into fine-grained matte labels, it often generates unnecessary noise into the output matte, particularly for objects that do not benefit from fine-grained representation (see the 4th column in Figure 6b). + +Spatial Generalization Augmentation. To improve the converter's ability to generalize to diverse segmentation labels, we design Spatial Generalization Augmentation. This approach introduces variability into the training data by applying a random cut-out technique, as shown in Figure 3a. During training, both the segmentation label and the corresponding matte label are randomly cropped in the same regions. By exposing the converter to irregular and incomplete input patterns, this augmentation forces the converter to adapt to diverse spatial structures and unseen patterns, thus enhancing its generalization capability. This method ensures that the converter can better handle a variety of input segmentation labels, even those that deviate from training patterns (see the 3rd column in Figure 6a). + +Selective Transformation Learning. To prevent the unnecessary transformation of objects that do not require fine-grained details (e.g., cars or desks), we introduce Selective Transformation Learning. This technique enables the converter to selectively focus on objects requiring detailed matte conversion (e.g., hair, trees) while skipping finer transformations for coarse-grained objects. We incorporate these non-transformable samples into the training process by collecting coarse-grained object masks from public segmentation datasets [56] (see Figure 3b). During training, the ground-truth matte label for the non-transformable samples is set to identical to the original segmentation label, allowing the converter to learn that no transformation is required. This selective approach reduces noise in the output + +![](images/ad75904dfdfa709a2a663fd6d0df8a188f7eef8e45ef6e6293b9c9246fadf557.jpg) +(a) Spatial Generalization Augmentation + +![](images/e7360b1dda3e6f2a43ce03e1efbd773fbb215565611927bc3da8a16d9df75677.jpg) +(b) Selective Transformation Learning +Figure 3. Illustration of the key components of the Label Converter. (a) Overview of the training procedure of the converter using Spatial Generalization Augmentation (indicated by red dotted boxes). (b) Examples of transformable (fine) and non-transformable (coarse) samples used in Selective Transformation Learning for the converter. + +and ensures that fine-grained transformations are applied only when needed (see the 3rd column in Figure 6b). + +Training. We employ standard loss functions commonly used in matting tasks, namely using a linear combination of L1 and Gradient losses [14, 21, 22] to minimize pixel-wise differences between the ground-truth and predicted matte: + +$$ +L = L _ {l 1} + \lambda L _ {\text {g r a d}} \tag {1} +$$ + +$$ +L _ {l 1} = \left| M - M ^ {\prime} \right| \tag {2} +$$ + +$$ +L _ {g r a d} = \left| \nabla_ {x} (M) - \nabla_ {x} \left(M ^ {\prime}\right) \right| + \left| \nabla_ {y} (M) - \nabla_ {y} \left(M ^ {\prime}\right) \right| \tag {3} +$$ + +where $M$ and $M'$ represent the ground-truth and predicted matte label, respectively, and $\lambda$ is a loss weighting factor. In addition, $\nabla_x$ and $\nabla_y$ represent the gradients along the horizontal and vertical axes, respectively. Moreover, we set a probability parameter $p$ to control the random application of Spatial Generalization Augmentation during training. + +SA1B-Matte Dataset. After training the label converter, we transform segmentation labels in the SA1B dataset [18] to matte labels using the converter, constructing a new SA1B-Matte dataset. As shown in Figure 2, the coarse labels in the SA1B dataset are successfully transformed into high-quality precise matte labels. Compared to existing public matting datasets consisting of macro-level fine labels, the SA1B-Matte dataset is a large-scale image matting dataset with micro-level fine labels, providing an ideal foundation for developing zero-shot matting models. + +# 3.2. ZIM: Zero-Shot Image Matting Model + +Overview of ZIM. Our proposed model, ZIM, builds upon SAM [18] and consists of four components, as illustrated in Figure 4: (1) Image Encoder: extracts image features from the input image, producing an image embedding with a stride of 16. (2) Prompt Encoder: encodes point or box inputs into prompt embeddings concatenated with learnable token embeddings, serving a role similar to the [c1s] token in ViT [8]. (3) Transformer Decoder: takes the image and token embeddings to generate output token embeddings. It performs four operations: self-attention on the + +tokens, token-to-image cross-attention, an MLP layer, and image-to-token cross-attention that updates the image embedding. (4) Pixel Decoder: upsamples the output image embedding with a stride of 2. Lastly, the model produces matte masks by computing a dot product between the upsampled image embedding and output token embeddings. + +Motivation. While SAM has shown success in segmentation tasks, its pixel decoder, which comprises two straightforward transposed convolutional layers, is prone to generating checkerboard artifacts, especially when handling challenging visual prompts, such as multiple positive and negative points placed near object boundaries or box prompts with imprecise object region delineation, as shown in Figure 1. Furthermore, their upsampled embeddings with a stride of 4 are often insufficient for image matting, which benefits from finer mask feature representations. + +Hierarchical Pixel Decoder. To address these shortcomings, we introduce a hierarchical pixel decoder with a multilevel feature pyramid design, motivated by [49], as illustrated in Figure 4. The pixel decoder takes an input image and generates multi-resolution feature maps at strides 2, 4, and 8 using a series of simple convolutional layers. The image embedding is sequentially upsampled and concatenated with the corresponding feature maps at each resolution. The decoder is designed to be highly lightweight, namely adding only $10\mathrm{ms}$ of computational overhead compared to the original pixel decoder of SAM on a V100 GPU. Our hierarchical design serves two key purposes: First, it preserves high-level semantics while refining spatial details, reducing checkerboard artifacts and enhancing robustness to challenging prompts. Second, it generates high-resolution feature maps with a stride of 2, essential for capturing fine-grained structures in matting. + +Prompt-Aware Masked Attention. To further boost the interactive matting performance, we propose a Prompt-Aware Masked Attention mechanism, inspired by Mask2Former [6] (See Figure 4). This mechanism allows the model to dynamically focus on the relevant regions within the image based on visual prompts (e.g., points or + +![](images/e5fc19114fd0ba33f4fe95c553b734114b5867e0135f90727737be3df22f092e.jpg) +Figure 4. Overview of the ZIM architecture. Based on the SAM network architecture [18], we introduce two key improvements: (1) Hierarchical Pixel Decoder for more robust and higher-resolution mask feature map generation, and (2) Prompt-Aware Masked Attention mechanism to enhance interactive matting performance. + +boxes), enabling more attention to the areas of interest. + +For box prompts, we generate a binary attention mask $\mathcal{M}^b$ that indicates the specific bounding box region. The binary attention mask $\mathcal{M}^b\in \{0, - \infty \}$ is defined as: + +$$ +\mathcal {M} ^ {b} (x, y) = \left\{ \begin{array}{l l} 0 & \text {i f} (x, y) \in \text {b o x r e g i o n} \\ - \infty & \text {o t h e r w i s e} \end{array} \right. \tag {4} +$$ + +where $(x,y)$ represents the pixel coordinates. This forces the model to prioritize the region within the box prompt. + +For point prompts, we generate a soft attention mask using a 2D Gaussian map distribution with standard deviation $\sigma$ . The soft attention mask, $\mathcal{M}^p \in [0,1]$ , smoothly weighs the region around the point of interest, ensuring a graded focus that transitions smoothly to the surrounding regions. + +The attention mask is incorporated into the cross-attention blocks of the transformer decoder. Specifically, the attention mask modulates the attention map as follows: + +$$ +X _ {l} = \left\{ \begin{array}{l l} \operatorname {s o f t m a x} \left(\mathcal {M} ^ {b} + Q _ {l} K _ {l} ^ {\intercal}\right) V _ {l} + X _ {l - 1} & (\text {b o x p r o m p t}) \\ \operatorname {s o f t m a x} \left(\mathcal {M} ^ {p} \odot Q _ {l} K _ {l} ^ {\intercal}\right) V _ {l} + X _ {l - 1} & (\text {p o i n t p r o m p t}) \end{array} \right. \tag {5} +$$ + +where $\odot$ denotes element-wise multiplication, $X_{l}$ represents the query feature maps at the $l^{th}$ layer of the decoder, and $Q_{l}, K_{l},$ and $V_{l}$ implies the query, key, and value matrices, respectively, at the $l^{th}$ layer. This mechanism dynamically adjusts the model's attention according to the visual prompt, leading to performance improvement in prompt-driven interactive scenarios (see Table 3a). + +Training. We train ZIM using the SA1B-Matte dataset. From the ground-truth matte label, we extract a box prompt from the given min-max coordinates and randomly sample positive and negative point prompts following [41]. The model is optimized using the same matte loss functions defined in Eq. (1). + +# 4. MicroMat-3K: Zero-Shot Matting Test Set + +We introduce a new test set, named MicroMat-3K, to evaluate zero-shot interactive matting models. It consists of 3,000 high-resolution images paired with micro-level matte labels. It includes two types of matte labels: (1) Fine-grained labels (e.g., hair, tree branches) to primarily evaluate zero-shot matting performance, where capturing intricate details is critical. (2) coarse-grained labels (e.g., cars, desks) to allow comparison with zero-shot segmentation models, which is still essential in zero-shot matting tasks. Moreover, It provides pre-defined point prompt sets for positive and negative points and box prompt sets for evaluating interactive scenarios. More detailed information about the MicroMat-3K is described in the supplementary material. + +# 5. Experiments + +# 5.1. Experimental Setting. + +Training Dataset for Label Converter. To train the label converter, we collect six publicly available matting datasets (i.e., AIM-500 [21], AM-2K [22], P3M-10K [20], RWP-636 [53], HIM-2K [44], and RefMatte [24]), consisting of 20,591 natural images and 118,749 synthetic images in total. For non-transformable samples, we extract coarse object categories (e.g., car and desk) from the ADE20K dataset [56], sampling 187,063 masks from 17,768 images. + +Evaluation Metrics. We use widely adopted evaluation metrics for the image matting task, including Sum of Absolute Difference (SAD), Mean Squared Error (MSE), Gradient Error (Grad), and Connectivity Error (Conn). + +Implementation Details for Label Converter. The label converter model is based on MGMatting [53] with Hierabase-plus [40] backbone network. For training the con + +
MethodPromptMicroMat3K Fine-grainedMicroMat3K Coarse-grained
SAD↓MSE↓MAE↓Grad↓Conn↓SAD↓MSE↓MAE↓Grad↓Conn↓
SAM [18]point box68.07621.65123.30716.49667.73017.0935.5695.7564.80017.035
36.08611.05712.71414.86735.8343.5161.0441.2312.5513.450
HQ-SAM [13]point box110.68136.67438.33116.855110.42118.8426.4576.6454.59918.792
124.26242.45744.14413.673124.1138.4582.7332.9202.4728.400
Matte-Any [50]point box68.79720.84423.5648.11868.93919.7176.0536.6752.63319.506
34.6619.74612.1827.02134.8566.9501.9832.4452.1426.905
Matting-Any [25]point box275.39877.33597.14120.019270.722164.14536.18755.94323.244155.780
246.21468.37287.61719.185241.597109.63923.78038.66215.841102.439
ZIM (ours)point box31.2868.21310.7405.32431.0096.6451.7882.3201.4696.472
9.9611.8933.4264.8139.6551.8600.4480.6591.2811.807
+ +Table 1. Quantitative comparison of ZIM and six existing methods on the MicroMat-3K test set, evaluated separately on fine-grained and coarse-grained categories using point and box prompts across five evaluation metrics. + +![](images/36f744ff0ae11f75b8d5d1e1d279c5ddfdf953d254996ccd3debf759ebc2ab60.jpg) + +![](images/01587279e7012a10d97447baff55f46166ff1b3809ca7fa47089fa6e0e60844f.jpg) + +![](images/0becdb12e85174ecd35efa7bc4d6e017daeda345d01099f6ea66a973964b9f09.jpg) + +![](images/5941521baf9d56fe3181462ee8c64192a3caaa6c20fea7738cfa0f538d0dd8f6.jpg) +Figure 5. Downstream Transferability Evaluation by replacing SAM with ZIM as the segmentation foundation model in various downstream tasks: Matte-Any [50] and Matting-Any [25] are evaluated on MicroMat3K-Fine, AIM-500 [21], AM-2K [22], and P3M-500-NP [20] using the MSE metric. HQ-SAM [13] is assessed on DIS [37], COIFT [27], HRSOD [55], and ThinObject [27] using mIoU. Inpainting Anything [54] is tested on COCO [28] using CLIP distance. Following Medical Image Segmentation evaluation protocol [32] with five different prompting modes (mIoU). 3D segmentation follows the SA3D [3] framework and is evaluated on NVOS [39]. + +![](images/11e5da3893c66f3173009d32ebbbd433c4bf0cb75411f53c71cc4eebf9fdd764.jpg) + +![](images/5b4e1468e398b6d9f5a2b4bc5f1fdba7e0d8360d0b7231aeee2edf61fdfd80b5.jpg) + +verter, we set the input size to $1024 \times 1024$ , a batch size of 16, and a learning rate of 0.001 with cosine decay scheduling using the AdamW optimizer [30]. The training process runs for 500K iterations with the probability parameter $p$ of 0.5 and the loss weight $\lambda$ of 10. + +Implementation Details for ZIM. For the ZIM model, we use the same image encoder (i.e., ViT-B [8]) and prompt encoder as SAM. Leveraging the pre-trained weights from SAM, we fine-tune the ZIM model on $1\%$ of the SA1B-Matte dataset, which amounts to approximately 2.2M matte labels. We set the input size to $1024 \times 1024$ , batch size to 16, a learning rate to 0.00001 with cosine decay scheduling using the AdamW optimizer [30], and training iterations to 500K. The loss weight $\lambda$ is set to 10 and the $\sigma$ for the point-based attention mask is set to 21 by default. + +# 5.2. Experimental Results + +We evaluate ZIM against four related methods on MicroMat3K: SAM [18], HQ-SAM [13], Matte-Any [50], and Matting-Any [25]. All methods use the ViT-B [8] backbone network. Table 1 presents the evaluation scores across five metrics for both point and box prompts on fine-grained and coarse-grained masks. For coarse-grained results, SAM achieves reasonable zero-shot performance, while other methods (e.g., Matting-Any and HQ-SAM) struggle to generalize to unseen objects. This is likely due to their fine-tuning on macro-level labeled datasets, which degrades their zero-shot capabilities. The qualitative results in Figure 1 show that these methods often produce macro-level outputs, even provided with micro-level prompts. Moreover, + +
SGASTLFine-grainedCoarse-grained
SAD↓MSE↓Grad↓SAD↓MSE↓Grad↓
3.3240.2762.6640.7160.1170.684
✓2.4400.1222.1390.6970.0920.634
✓3.1530.2392.4570.6350.0890.653
✓✓1.9990.0801.7710.2810.0210.399
+ +Table 2. Quantitative analysis of key components of the label converter: Spatial Generalization Augmentation (SGA) and Selective Transformation Learning (STL). + +SAM tends to suffer from checkerboard artifacts when challenging prompts are introduced. In contrast, ZIM generates a more robust quality of masks, due to our hierarchical feature pyramid decoder. The fine-grained results in Table 1 highlight ZIM's superiority in producing high-quality matting outputs while maintaining strong zero-shot capabilities. + +# 5.3. Downstream Transferability + +We assess the transferability of ZIM compared to SAM across various downstream tasks. Specifically, we first integrate ZIM into existing matting methods, Matte-Any [50] and Matting-Any [25], and observe significantly improved box prompting MSE results across diverse matting benchmarks, including MicroMat3K-Fine, AIM-500 [21], AM2K [22], and P3M-500-NP [20]. Likewise, replacing SAM with ZIM in HQ-SAM [13] notably enhances performance in fine-grained segmentation datasets, including DIS [37], COIFT [27], HRSOD [55], and ThinObject [27]. + +Moreover, we extend our analysis to broader vision tasks, including image inpainting, medical image segmentation, and 3D segmentation. Specifically, we evaluate image inpainting using the Inpainting Anything framework [54] on the COCO dataset [28] via the CLIP Distance metric [9, 52]. For medical image segmentation, we utilize the zero-shot evaluation protocol from [32] on five diverse medical imaging datasets [1, 11, 33, 42], measured by mean IoU (mIoU). Additionally, we assess 3D segmentation performance within the SA3D framework [3] on the NVOS dataset [39]. These tasks inherently require highly precise segmentation masks for optimal outcomes. While replacing SAM with existing matting models (e.g., Matte-Any and Matting-Any) results in significant performance drops due to limited generalization capabilities, employing ZIM consistently boosts zero-shot performance across these diverse scenarios, highlighting its robust generalization and precise mask representation capabilities. Comprehensive qualitative and quantitative results for these tasks are provided in the supplementary materials. + +# 6. Ablation Study + +In this section, we analyze the impact of the key components of our method using the MicroMat-3K test set. + +![](images/e324934898141e6daa3121a9b89afc9e5bddca025ba6f400b653504c6b265668.jpg) +(a) + +![](images/a6b23682a429af2340ef72423351d49f6aeee621073138d5cc2a87d81c566d46.jpg) +(b) +Figure 6. Qualitative analysis of key components of the label converter: (a) without Spatial Generalization Augmentation (SGA) and (b) without Selective Transformation Learning (STL). + +Analysis of Label Converter. To analyze the effect of Spatial Generalization Augmentation (SGA) and Selective Transformation Learning (STL) strategies, we conduct an ablation study by removing each component individually. The SGA is designed to enhance the generalization ability of the converter by simulating diverse input patterns, particularly beneficial given that the converter is trained on macro-level labeled datasets. Without the SGA, the converter struggles to produce clear matte labels for unseen objects, as shown in Figure 6a. In addition, the STL is designed to help the converter avoid unnecessary label conversion for coarse objects. Without the STL, the converter attempts to transform every segmentation label into a matte label, resulting in noisy outputs for unseen coarse objects, as shown in Figure 6b. The quantitative results in Table 2 confirm that using both strategies yields the best label conversion performance on the MicroMat-3K test set. + +Analysis of ZIM Model. We conduct experiments to analyze the effect of the prompt-aware masked attention and hierarchical mask decoder. The prompt-aware masked attention is designed to direct the model's focus on the regions of interest to improve the promptable matting performance. Table 3a shows that leveraging the masked attention yields a substantial improvement to our ZIM model. In addition, the hierarchical mask decoder is designed to produce more robust and higher-resolution mask feature maps to alleviate checkerboard artifacts and capture finer representation, simultaneously. Its effectiveness is particularly evident in reducing the gradient error for fine-grained objects in Table 3a, since the enhanced pixel decoder generates more solid and detailed mask outputs. Notably, the decoder remains lightweight, adding only $10\mathrm{ms}$ of additional inference time. + +
AttnDecFine-grainedCoarse-grained
SAD↓MSE↓Grad↓SAD↓MSE↓Grad↓
13.6232.7186.5162.0710.4741.526
✓13.1982.5046.4452.0490.4711.486
✓11.0742.0945.4012.0690.4871.355
✓✓9.9611.8934.8131.8600.4481.281
+ +(a) + +
Attn MaskFine-grainedCoarse-grained
T2II2TSAD↓MSE↓Grad↓SAD↓MSE↓Grad↓
11.0742.0945.4012.0690.4871.355
✓9.9611.8934.8131.8600.4481.281
✓12.5262.6586.0322.3530.5541.481
✓✓10.4371.9975.0661.9990.4701.306
+ +(b) + +
ModelTrainsetFine-grainedCoarse-grained
SAD↓MSE↓Grad↓SAD↓MSE↓Grad↓
ZIMSA1B-Matte9.9611.8934.8131.8600.4481.281
Public-Matte120.57138.3326.7309.5062.7601.688
Matting-Any [25]SA1B-Matte41.24212.2677.7074.6261.2841.552
Public-Matte246.21468.37219.185109.63923.78015.841
+ +(c) + +Table 3. Analysis of ZIM using box prompt evaluations: (a) Effect of Attn (prompt-aware masked attention) and Dec (hierarchical pixel decoder). (b) Effect of the masked attention in T2I (token to image) and I2T (image to token) cross-attention layers. (c) Effect of trainset: our SA1B-Matte and public matting datasets. + +Moreover, we delve into the effect of prompt-aware masked attention in our transformer decoder, which comprises two kinds of cross-attention layers (see Figure 4): token-to-image (t2i) updating token embeddings (as queries) and image-to-token (i2t) updating the image embedding (as queries). As a result in Table 3b, applying masked attention to only the t2i layer leads to a meaningful improvement. This suggests that focusing attention on tokens based on visual prompts in the t2i layer enhances their ability to capture relevant features. In contrast, applying attention to specific regions within the image embedding in the i2t layer may disturb the capture of global features. + +Analysis of Training Dataset. Table 3c investigates the influence of the training dataset on ZIM's zero-shot matting performance. When ZIM is trained on publicly available matting datasets, which predominantly contain macro-level masks, the performance on the micro-level MicroMat3K dataset significantly degrades. This result underscores the necessity of training with micro-level mask annotations for effective generalization to multi-granularity details. Despite Matting-Any [25] also being trained on SA1B-Matte, ZIM still surpasses its performance by a considerable margin, highlighting the advancements of our network architectural improvements for the zero-shot interactive matting task. + +Discussion on Domain Shift. There is a distinct domain shift between traditional matting test sets and ZIM trained + +![](images/404b8d72ad8d50a66b1e39a4064352de5d28109e0b89f4ab5c9d9710e0d23c78.jpg) +Figure 7. Domain shift between traditional public matting testsets and ZIM trained with SA1B-Matte (object-level). + +![](images/eb258fc07822f91782a48abdc0e03ba7123af1e45f2b07305f53855f79e67a5b.jpg) +Figure 8. Quantitative comparison of ZIM (with varying numbers of point prompts) and existing matting methods (Matte-Any [50], Matting-Any [25], GFM [23], AIM [21], and SMat [51]) on the traditional matting test sets (AIM-500 [21] and P3M-500-NP [20]). The "auto" implies prompt-free mode. + +![](images/b2c311c8c184d99f1a6ca52c066ff65199d44ac90303914b8fe6322731d4ebb0.jpg) + +with SA1B-Matte. Traditional matting test sets [20, 21] typically regard entire salient objects as foreground, whereas ZIM mainly focuses on object- and part-level matting (Figure 7). This mismatch leads to substantial performance penalties for ZIM under box prompting on these test sets (Figure 7), which is inherent from SAM's prompt ambiguity issue. However, by leveraging dense multiple-point prompting, ZIM effectively mitigates this discrepancy, surpassing existing methods [21, 23, 25, 50, 51], even some of which are explicitly trained on datasets similar to the evaluation domain (Figure 8). This highlights the adaptability of our zero-shot interactive matting modeling. + +# 7. Conclusion, Limitation, and Future Work + +In this paper, we presented a pioneering zero-shot image matting model that advances the field by generating precise, fine-grained matte masks. We addressed the limitations of SAM, which struggles with high-detail segmentation tasks, by introducing a novel label conversion method and enhancing the network architecture with a hierarchical pixel decoder and prompt-aware masked attention mechanism. However, ZIM inherits inherent shortcomings from SAM, including ambiguous visual prompt handling and robustness in uncertain predictions. Future research should aim at resolving these distinct challenges through innovative approaches to prompt design and uncertainty-based modeling. We hope that the research community will continue to build on this work, exploring new applications in computer vision and enhancing its zero-shot matting performance. + +# References + +[1] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A. Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M. Summers, Bram van Ginneken, Michel Bilello, Patrick Bilic, Patrick F. Christ, Richard K. G. Do, Marc J. Gollub, Stephan H. Heckers, Henkjan Huisman, William R. Jarnagin, Maureen K. McHugo, Sandy Napel, Jennifer S. Golia Pernicka, Kawal Rhode, Catalina Tobon-Gomez, Eugene Vorontsov, James A. Meakin, Sebastien Ourselin, Manuel Wiesenfarth, Pablo Arbelaez, Byeonguk Bae, Sihong Chen, Laura Daza, Jianjiang Feng, Baochun He, Fabian Isensee, Yuanfeng Ji, Fucang Jia, Ildoo Kim, Klaus Maier-Hein, Dorit Merhof, Akshay Pai, Beomhee Park, Mathias Perslev, Ramin Rezaiifar, Oliver Rippel, Ignacio Sarasua, Wei Shen, Jaemin Son, Christian Wachinger, Liansheng Wang, Yan Wang, Yingda Xia, Daguang Xu, Zhanwei Xu, Yefeng Zheng, Amber L. Simpson, Lena Maier-Hein, and M. Jorge Cardoso. The medical segmentation decathlon. Nature Communications, 13(1), 2022. 7 +[2] Xue Bai and Guillermo Sapiro. A geodesic framework for fast interactive image and video segmentation and matting. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. +[3] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, et al. Segment anything in 3d with nerfs. Advances in Neural Information Processing Systems, 36:25971-25990, 2023. 2, 6, 7 +[4] Quan Chen, Tiezheng Ge, Yanyu Xu, Zhiqiang Zhang, Xinxin Yang, and Kun Gai. Semantic human matting. In Proceedings of the 26th ACM international conference on Multimedia, pages 618-626, 2018. +[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in neural information processing systems, 34:17864-17875, 2021. +[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 4 +[7] Henghui Ding, Hui Zhang, Chang Liu, and Xudong Jiang. Deep interactive image matting with feature propagation. IEEE Transactions on Image Processing, 31:2421-2432, 2022. +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6 +[9] Yigit Ekin, Ahmet Burak Yildirim, Erdem Eren Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models, 2024. 7 +[10] Leo Grady, Thomas Schiwietz, Shmuel Aharon, and Rüdiger + +Westermann. Random walks for interactive alpha-matting. In Proceedings of VIIP, pages 423-429. CiteSeer, 2005. +[11] Daniel Gut. X-ray images of the hip joints, 2021., Mendeley Data, V1, doi: 10.17632/zm6bxzhmfz.1.7 +[12] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 2 +[13] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 1, 6, 7 +[14] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1140-1147, 2022. 4 +[15] Beomyoung Kim, Joonhyun Jeong, Dongyoon Han, and Sung Ju Hwang. The devil is in the points: Weakly semi-supervised instance segmentation via point-guided mask representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11360-11370, 2023. +[16] Beomyoung Kim, Myeong Yeon Yi, Joonsang Yu, Young Joon Yoo, and Sung Ju Hwang. Towards label-efficient human matting: A simple baseline for weakly semi-supervised trimap-free human matting. arXiv preprint arXiv:2404.00921, 2024. +[17] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024. +[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3, 4, 5, 6 +[19] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 2 +[20] Jizhizi Li, Sihan Ma, Jing Zhang, and Dacheng Tao. Privacy-preserving portrait matting. In Proceedings of the 29th ACM international conference on multimedia, pages 3501-3509, 2021. 3, 5, 6, 7, 8 +[21] Jizhizi Li, Jing Zhang, and Dacheng Tao. Deep automatic natural image matting. arXiv preprint arXiv:2107.07235, 2021. 2, 3, 4, 5, 6, 7, 8 +[22] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 2, 3, 4, 5, 6, 7 + +[23] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 8 +[24] Jizhizi Li, Jing Zhang, and Dacheng Tao. Referring image matting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22448-22457, 2023. 3, 5 +[25] Jiachen Li, Jitesh Jain, and Humphrey Shi. Matting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775-1785, 2024. 1, 2, 6, 7, 8 +[26] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI conference on artificial intelligence, pages 11450-11457, 2020. +[27] Jun Hao Liew, Scott Cohen, Brian Price, Long Mai, and Ji-ashi Feng. Deep interactive thin object selection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 305–314, 2021. 6, 7 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7 +[29] Yuhao Liu, Jiake Xie, Xiao Shi, Yu Qiao, Yujie Huang, Yong Tang, and Xin Yang. Tripartite information mining and integration for image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7555-7564, 2021. +[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6 +[31] Hao Lu, Yutong Dai, Chunhua Shen, and Songcen Xu. Indices matter: Learning to index for deep image matting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3266-3275, 2019. +[32] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 6, 7 +[33] Anna Montoya, Hasnin, kaggle446, shirzad, Will Cukierski, and yffud. Ultrasound nerve segmentation. https://kaggle.com/competitions/ultrasound-nerve-segmentation, 2016. Kaggle. 7 +[34] GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. Matteformer: Transformer-based image matting via prior-tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11696-11706, 2022. +[35] Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Mask-guided matting in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1992–2001, 2023. 3 +[36] Yu Qiao, Yuhao Liu, Xin Yang, Dongsheng Zhou, Mingliang Xu, Qiang Zhang, and Xiaopeng Wei. Attention-guided hi + +erarchical structure aggregation for image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13676-13685, 2020. 2, 3 +[37] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In European Conference on Computer Vision, pages 38-56. Springer, 2022. 6, 7 +[38] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. +[39] Zhongzheng Ren, Aseem Agarwala†, Bryan Russell†, Alexander G. Schwing†, and Oliver Wang†. Neural volumetric object selection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. († alphabetic ordering). 6, 7 +[40] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 5 +[41] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 5 +[42] Yuxin Song, Jing Zheng, Long Lei, Zhipeng Ni, Baoliang Zhao, and Ying Hu. Ct2us: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data. Ultrasonics, 122:106706, 2022. 7 +[43] Yanan Sun, Chi-Keung Tang, and Yu-Wing Tai. Semantic image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11120–11129, 2021. +[44] Yanan Sun, Chi-Keung Tang, and Yu-Wing Tai. Human instance matting via mutual guidance and multi-instance refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2647–2656, 2022. 3, 5 +[45] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmentation. Advances in Neural information processing systems, 33:17721-17732, 2020. +[46] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Hanqing Zhao, Weiming Zhang, and Nenghai Yu. Improved image matting via real-time user clicks and uncertainty estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15374-15383, 2021. +[47] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 2, 3 +[48] Stephen DH Yang, Bin Wang, Weijia Li, YiQi Lin, and Conghui He. Unified interactive image matting. arXiv preprint arXiv:2205.08324, 2022. + +[49] Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang. Vitmatte: Boosting image matting with pretrained plain vision transformers. Information Fusion, 103: 102091, 2024. 4 +[50] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with segment anything model. Image and Vision Computing, 147: 105067, 2024. 1, 2, 6, 7, 8 +[51] Zixuan Ye, Wenze Liu, He Guo, Yujia Liang, Chaoyi Hong, Hao Lu, and Zhiguo Cao. Unifying automatic and interactive matting with pretrained vits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25585-25594, 2024. 2, 8 +[52] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models, 2023. 7 +[53] Qihang Yu, Jianming Zhang, He Zhang, Yilin Wang, Zhe Lin, Ning Xu, Yutong Bai, and Alan Yuille. Mask guided matting via progressive refinement network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1154-1163, 2021. 3, 5 +[54] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2, 6, 7 +[55] Yi Zeng, Pingping Zhang, Jianming Zhang, Zhe Lin, and Huchuan Lu. Towards high-resolution salient object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7234-7243, 2019. 6, 7 +[56] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 3, 5 \ No newline at end of file diff --git a/zimzeroshotimagemattingforanything/images.zip b/zimzeroshotimagemattingforanything/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f3b00411a662abb36aed2c2c12ece80b875f018d --- /dev/null +++ b/zimzeroshotimagemattingforanything/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6adc97fe56e33c36cdfe48ef837546d8c078f2d5c57de3c7e3d5e2d2684d6923 +size 764076 diff --git a/zimzeroshotimagemattingforanything/layout.json b/zimzeroshotimagemattingforanything/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38dd96f8d1e0da7ababdba86de192b6bf1fac78b --- /dev/null +++ b/zimzeroshotimagemattingforanything/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:def2a8a847b609ab79ee02149f8473d0c7edf9af7845a9d18b547063349b0c9a +size 414387 diff --git a/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/8cd7319e-aae8-4f84-8ca4-5b3ef6a42984_content_list.json b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/8cd7319e-aae8-4f84-8ca4-5b3ef6a42984_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8a3fcc4b5d405140f1aaf96b93c2ce44825b94bf --- /dev/null +++ b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/8cd7319e-aae8-4f84-8ca4-5b3ef6a42984_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5ed1cd2f4cb14031e2473362ef43d0ebfc09dd0f8743681f4cb1df3a1829662 +size 73222 diff --git a/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/8cd7319e-aae8-4f84-8ca4-5b3ef6a42984_model.json b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/8cd7319e-aae8-4f84-8ca4-5b3ef6a42984_model.json new file mode 100644 index 0000000000000000000000000000000000000000..01c81ff4e0d3860b6c574809e073facedacbf4b5 --- /dev/null +++ b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/8cd7319e-aae8-4f84-8ca4-5b3ef6a42984_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:889929518d0187ef4260895341c67ae3346c09adf1a372f877fbd5cda7261675 +size 90737 diff --git a/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/8cd7319e-aae8-4f84-8ca4-5b3ef6a42984_origin.pdf b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/8cd7319e-aae8-4f84-8ca4-5b3ef6a42984_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..083a38f0395ac62ad41fdf36afba7f760fa0ea66 --- /dev/null +++ b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/8cd7319e-aae8-4f84-8ca4-5b3ef6a42984_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5dddb8d755efaac81ca8a9fd420078fd86a8da6917a7478d0da5eb06374dfe0f +size 535020 diff --git a/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/full.md b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/full.md new file mode 100644 index 0000000000000000000000000000000000000000..04a3185900710777a1e1e29f29d371472f5c2431 --- /dev/null +++ b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/full.md @@ -0,0 +1,316 @@ +# ZipVL: Accelerating Vision-Language Models through Dynamic Token Sparsity + +Yefei He $^{1,2}$ * Feng Chen $^{3}$ Jing Liu $^{4}$ Wenqi Shao $^{2}$ + +Hong Zhou $^{1\dagger}$ Kaipeng Zhang $^{2\dagger}$ Bohan Zhuang $^{1}$ + +1Zhejiang University, China + +$^{2}$ Shanghai AI Laboratory, China + +3The University of Adelaide, Australia + +4ZIP Lab, Monash University, Australia + +# Abstract + +The efficiency of large vision-language models (LVLMs) is constrained by the computational bottleneck of the attention mechanism during the prefetch phase and the memory bottleneck of fetching the key-value (KV) cache in the decoding phase, particularly in scenarios involving high-resolution images or videos. Visual content often exhibits substantial redundancy, resulting in highly sparse attention maps within LVLMs. This sparsity can be leveraged to accelerate attention computation or compress the KV cache through various approaches. However, most studies focus on addressing only one of these bottlenecks and do not adequately support dynamic adjustment of sparsity concerning distinct layers or tasks. In this paper, we present ZipVL, an efficient inference framework designed for LVLMs through a dynamic ratio allocation strategy of important tokens. This ratio is adaptively determined based on the layer-specific distribution of attention scores, rather than fixed hyperparameters, thereby improving efficiency for less complex tasks while maintaining high performance for more challenging ones. Then we select important tokens based on their normalized attention scores and perform sparse attention mechanism solely on those important tokens, reducing the latency in the prefetch phase. Tokens deemed less important will be discarded to reduce KV cache size, alleviating the memory bottleneck in the decoding phase. Our experiments demonstrate that ZipVL can accelerate the prefetch phase by $2.3 \times$ and improve decoding throughput by $2.8 \times$ , with a minimal accuracy reduction of only $0.5\%$ on VQAv2 benchmark over LLaVA-Next-13B model, effectively enhancing the generation efficiency of LVLMs. + +# 1. Introduction + +With the recent advancement of large language models (LLMs) [1, 35, 36], many studies have extended their capabilities to comprehend and generate visual content. These models, commonly known as large vision-language models (LVLMs), have demonstrated remarkable performance in tasks such as image captioning and visual question answering [11, 12, 22, 24, 34]. Typically, to remain compatible with the next-token-prediction generation scheme of LLMs, images or videos are encoded into visual tokens through a pre-trained visual encoder, and concatenated with text tokens for input into the model. However, for high-resolution images or videos, the visual encoder generates excessive sequences of visual tokens, significantly limiting the generative efficiency of LVLMs. For example, a short video consisting of 128 frames is encoded into over 18,000 tokens by the LongVA model [43]. In such cases, the pre-fill phase suffers from the quadratic complexity of the attention mechanism, resulting in computational bottleneck and prolonged time-to-first-token (TTFT). In the decoding phase, each new token interacts with all preceding tokens, requiring to fetch the full key-value (KV) cache from memory and thereby inducing a memory bottleneck. Improving generative efficiency in both phases is essential for the practical deployment of LVLMs. + +To address computational complexity in the prefetch phase, sparse attention [18, 30, 45] has emerged as an effective strategy, particularly suitable for LVLMs where visual information exhibits considerable redundancy, resulting in higher sparsity in attention maps compared to LLMs [4, 37]. This sparsity can be implemented at various levels of granularity. Some studies pre-define several sparse patterns and assign them to the attention mask during inference [18, 45]. However, these predefined patterns are incompatible with efficient attention implementations such as FlashAttention [7] and require custom GPU kernels for each + +![](images/605a93872a2b19f5c46ae9bad4313e03686fe7df4e16de7f70dbefbba843c1ad.jpg) +(a) Layer 0, ChartQA + +![](images/db2044788d2b9af45e0232c373708576dade76125dd010e5686f8590ab5025d5.jpg) +(b) Layer 10, ChartQA + +![](images/2dfe695f725720bc313450f53777a52a6abab99b84a53742af9ae14bf63641e0.jpg) +(c) Layer 10, VQAv2 +Figure 1. The attention maps exhibit distinct sparse patterns across different layers (subfigures (a) and (b)) and vary significantly between tasks (subfigures (b) and (c)). Data was collected from the LLaVA-Next-7B model using input samples from the VQAv2 and ChartQA datasets. + +pattern. Alternatively, other approaches adopt token-level sparsity by identifying and discarding less important tokens [2, 4], allowing seamless integration with off-the-shelf efficient attention implementations. However, the optimal retention ratio of important tokens may vary across different layers or tasks due to distinct attention patterns, as illustrated in Figure 1. These methods rely on a fixed token retention ratio and do not dynamically adjust based on task difficulty, leading to suboptimal performance on complex tasks. + +To alleviate memory bottleneck in the decoding phase, various efforts have been made to reduce KV cache size, including token dropping [37], token merging [40], and quantization [15, 16]. However, these methods often rely on fixed compression ratios that are uniformly applied across all layers, failing to account for the distinct characteristics of attention maps in different layers. Moreover, despite the necessity of identifying important tokens for both sparse attention and KV cache compression, a unified inference optimization framework has yet to be developed. + +In this paper, we present ZipVL, an efficient inference framework tailored for LVLMs that jointly optimizes the prefill and decoding phases with a unified ratio of important tokens, as shown in Figure 2. To start with, we introduce a layer-wise adaptive ratio assignment scheme for important tokens. This ratio is adaptively determined based on the distribution of attention scores in each layer, rather than relying on predefined hyper-parameters [2, 4, 15, 44]. This adaptive approach allows the ratio to be adjusted according to task complexity, enhancing efficiency for simpler tasks while preserving performance for more complex ones. After determining the ratio, we then select important tokens with the highest normalized attention scores, follow + +ing prior work [15, 31]. To alleviate the computational bottleneck in the prefetch phase, sparse attention is performed at the token level by computing attention only for the selected important tokens. Notably, this approach seamlessly integrates with existing fast attention implementations without requiring custom GPU kernels. To tackle the memory bottleneck, the same set of important tokens is applied to compress the KV cache, evicting tokens deemed less important. Extensive experiments on multimodal benchmarks demonstrate that our method achieves nearly lossless performance while reducing prefetch phase latency by $2.3 \times$ and improving decoding throughput by $2.8 \times$ . + +In summary, our contributions are as follows: + +- We propose an adaptive layer-wise ratio assignment scheme for important tokens. The ratio is dynamically determined based on the distribution of attention scores and varies across different layers and tasks, enhancing performance and efficiency compared to a fixed ratio scheme. +- We introduce a unified approach to jointly optimize the prefetch and decoding stages through the adaptive allocation of important tokens. Tokens considered less important are excluded from attention computation to reduce computational complexity, and their KV cache will be discarded to alleviate the memory bottleneck in the decoding phase. +- By integrating these techniques, we present ZipVL, an efficient inference framework tailored for LVLMs. Comprehensive experiments across diverse benchmarks validate the efficacy of ZipVL, demonstrating that it achieves state-of-the-art performance in both accuracy and generation efficiency for LVLMs. + +![](images/233dac2c2ff0b5e6b8a6b1f3637578fcc164347b33ea4795d534701eccc9335a.jpg) +(1) Dynamic Ratio of Important Tokens + +![](images/ce3f89d69c4aea9871d1876cff5240c695110a67ff60186fb079d1771f43a1cb.jpg) +③ Token-level Sparse Attention (Prefix Phase) +(4) Token-level Sparse Attention (Decoding Phase) + +![](images/77c3cbc2d1d5fa25633271ed174cb54f64b8b2394e5c4d81c011aae5e8a67079.jpg) +Figure 2. Overview of the proposed ZipVL framework during the prefetch phase. Here, $\tau$ represents the threshold for retaining attention scores, $n$ and $p$ are the total number of tokens and the number of important tokens, respectively. After determining the ratio of important tokens and identifying them, we optimize both prefetch and decoding phases by exclusively computing attention for important tokens. The KV cache for tokens deemed less important will be discarded. + +# 2. Related Work + +# 2.1. Sparse Attention + +Attention scores have been widely observed to exhibit high sparsity in both LLMs and LVLMs [3, 37, 39, 42, 45]. This sparsity allows sparse attention to overcome the quadratic computational complexity of the standard attention mechanism by restricting each token to focus on only a subset of tokens within the input sequence [6, 18, 30, 32, 45]. Depending on the granularity of sparsity, sparse attention can be categorized into unstructured, semi-structured, and structured schemes. The unstructured scheme [14, 20] employs sparse attention masks without a fixed structure, making it hardware-unfriendly and challenging to achieve practical inference acceleration. The semi-structured sparse attention uses attention masks with predefined sparse patterns [18, 30, 45] or introduces N:M sparsity to attention weights [5]. However, it requires customized computational kernels for each sparse pattern or specific hardware to achieve acceleration. Structured sparse attention [2, 4] directly prunes tokens before the attention computation, enabling acceleration without the need for custom kernels. However, due to its coarse granularity, the pruning sparsity and the selection of tokens to prune significantly impact model performance. For instance, HiRED [2] selects patches with the highest responses based on the feature maps of the visual encoder without considering the input text prompt, leading to suboptimal performance. FastV [4] employs a fixed token sparsity pattern, discarding $50\%$ of + +visual tokens after the second layer of the model, which results in performance degradation in challenging tasks such as ChartQA [28]. In contrast, our approach achieves superior performance by dynamically adjusting the token sparsity ratio according to distinct models, layers, and task complexity. + +# 2.2. KV Cache Compression + +KV cache prevents re-computation in the decoding phase by storing the key and value states of previous tokens, but with a significant memory bottleneck in long-context scenarios. Previous efforts to compress the KV cache can be broadly categorized into three types: token dropping-based [10, 31, 44], token merging-based [25, 37, 38], and quantization-based approaches [15, 16, 19, 26, 41]. In this paper, we reduce the KV cache size by dropping tokens deemed less important in a layer-wise adaptive manner. Compared to approaches with a fixed dropping ratio [4, 44], our method boosts the overall compression ratio and accuracy. Additionally, it should be noted that our method can be combined with quantization-based approaches to achieve further reductions in KV cache size. + +# 3. Preliminary + +Attention block is the key module of Transformer-based LLMs. Each attention block contains three weight matrices $\mathbf{W}_{\mathrm{Q}}, \mathbf{W}_{\mathrm{K}}, \mathbf{W}_{\mathrm{V}} \in \mathbb{R}^{d \times d}$ , where $d$ is the dimension of the input data. Here, we use a single attention head and omit the + +output projection for clarity. In the prefetch phase, the input data $\mathbf{X} \in \mathbb{R}^{n \times d}$ with a sequence length of $n$ is first multiplied with three weight matrices to obtain the query, key and value states: + +$$ +\mathbf {Q} = \mathbf {X} \mathbf {W} _ {\mathrm {Q}}, \quad \mathbf {K} = \mathbf {X} \mathbf {W} _ {\mathrm {K}}, \quad \mathbf {V} = \mathbf {X} \mathbf {W} _ {\mathrm {V}}. \tag {1} +$$ + +Then the attention output is calculated as follows: + +$$ +\mathbf {A} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {T} + \mathbf {M}}{\sqrt {d}}\right), \mathbf {O} = \mathbf {A V}. \tag {2} +$$ + +Here, computing the product of $\mathbf{Q}\mathbf{K}^T$ has a quadratic complexity $O(n^{2})$ , which makes the prefetch phase compute-bound. $\mathbf{M} \in \mathbb{R}^{n \times n}$ is a lower triangular causal mask to ensure that each token can only attend to itself and previous tokens. Unstructured and semi-structured sparse attention introduces sparsity in the attention mask $\mathbf{M}$ with dynamic or fixed sparse pattern. With custom computing kernels, tokens in certain positions can be skipped when computing $\mathbf{Q}\mathbf{K}^T$ , thus accelerating the computation. On the other hand, structured sparse attention only computes attention scores for a subset of tokens $\mathbf{X}' \in \mathbb{R}^{n' \times d}$ , reducing computational complexity to $O(n'^2)$ and seamlessly integrating with existing fast attention implementations. + +For the decoding phase, the input data is the embedding of the current token $\mathbf{x} \in \mathbb{R}^{1 \times d}$ . To enable the interaction between the current token and all previous tokens, the KV cache of previous tokens needs to be fetched from memory, making the decoding phase memory-bound: + +$$ +\mathbf {q} = \mathbf {x} \mathbf {W} _ {\mathrm {Q}}, \quad \mathbf {k} = \mathbf {x} \mathbf {W} _ {\mathrm {K}}, \quad \mathbf {v} = \mathbf {x} \mathbf {W} _ {\mathrm {V}}. \tag {3} +$$ + +$$ +\mathbf {K} = \operatorname {C o n c a t} (\mathbf {K}, \mathbf {k}), \quad \mathbf {V} = \operatorname {C o n c a t} (\mathbf {V}, \mathbf {v}). \qquad (4) +$$ + +The attention outputs are then computed as follows with a computational complexity of $O(n)$ : + +$$ +\mathbf {a} = \operatorname {S o f t m a x} \left(\frac {\mathbf {q} \mathbf {K} ^ {T}}{\sqrt {d}}\right), \quad \mathbf {o} = \mathbf {a} \mathbf {V}. \tag {5} +$$ + +# 4. Method + +# 4.1. Layer-wise Adaptive Ratio Assignment for Important tokens + +Prior studies [2, 15, 26, 37, 44] typically adopt a fixed ratio of important tokens across all layers. However, as analyzed by the preceding study [4] and demonstrated in Figure 1(a) and (b), there are substantial variations in the attention map patterns across different layers. Moreover, Figure 1(b) and (c) illustrate that, even within the same layer, attention maps can differ depending on the task and input. In scenarios involving complex tasks, a limited, static ratio for important tokens can impair model performance. This raises the question: + +![](images/6fe07177c34bb439df63c8a134dec23b963971b8ba9f94ea07a209acc9df3991.jpg) +Figure 3. The ratio of important tokens distributed across layers. Data was collected from the LLaVA-Next-7B model using input samples from the VQAv2 and ChartQA datasets with a $\tau$ of 0.975. + +can the model dynamically determine the number of tokens required to solve a task? + +Intuitively, for simpler tasks, the model needs to concentrate on fewer tokens, leading to a more focused distribution of attention scores. Conversely, more demanding tasks require the model to engage with a broader array of tokens, resulting in a more uniform distribution of attention scores. Prior work [39] also highlights the criticality of preserving significant attention scores during inference within a constrained attention window. Building on these insights, we introduce a layer-wise adaptive scheme for assigning ratio of important tokens, ensuring the majority of significant attention scores are maintained within each layer. + +Consider an attention layer with $n$ input tokens, where the full attention score matrix is denoted as $\mathbf{A} \in \mathbb{R}^{n \times n}$ . The accumulated attention score for each token $j$ is calculated by summing the corresponding column: + +$$ +a _ {j} = \sum_ {c = 1} ^ {n} \mathbf {A} _ {c, j}. \tag {6} +$$ + +These accumulated attention scores are subsequently sorted in descending order, such that $a_{\mathrm{sorted}(j)}$ represents the $j$ -th highest attention score. The number of important tokens $p$ is determined by preserving the majority of attention scores with a minimal number of tokens, which can be expressed as: + +$$ +p = \min \left\{p \in \mathbb {Z} \mid \sum_ {j = 1} ^ {p} a _ {\text {s o r t e d} (j)} \geq \tau \times n \right\}. \tag {7} +$$ + +Here, $\tau$ is the threshold dictating the retention of attention scores and the sum of the attention scores in $\mathbf{A}$ is equal to $n$ due to the row-wise Softmax operation. As shown in Figure 2 and 3, our method can dynamically adjust the ratio + +of important tokens across distinct layers and tasks, thereby enhancing performance in complex tasks while improving efficiency in simpler tasks. Additional experimental results can be found in Section 5.2.1 and Figure 4. + +# 4.2. Inference Optimization with Unified Important Token Allocation + +After determining the number of important tokens $p$ for each layer, we partition all tokens into two sets: set $\mathbf{T}$ of important tokens with a size of $p$ and the set $\mathbf{U}$ for less important tokens with a size of $n - p$ . Following prior work [15, 31], we use normalized attention scores to assess token importance, calculated as follows: + +$$ +\tilde {a} _ {j} = \frac {\sum_ {c = 1} ^ {n} \mathbf {A} _ {c , j}}{\operatorname {n n z} (\mathbf {A} _ {: , j})}. \tag {8} +$$ + +Here, $\mathrm{nz}(\mathbf{A}_{:,j})$ denotes the number of non-zero elements in the $j$ -th column. Important tokens are then selected using the top- $k$ indexing method, while the remainder are considered less important: + +$$ +\mathbf {T} = \operatorname {t o p k} _ {\text {i n d e x}} (\tilde {a} _ {j}, p), \tag {9} +$$ + +$$ +\mathbf {U} = \{j \in \{1, 2, \dots , n \} \mid j \notin \mathbf {T} \}. \tag {10} +$$ + +The inference optimization is then performed based on the split of tokens. Specifically, to address the computational bottleneck in the prefetch phase, the attention mechanism is performed solely on these important tokens, thereby enhancing efficiency through token-level sparsity. Tokens excluded from this computation have their outputs padded to maintain the number of tokens consistent for subsequent layers. By leveraging token-level sparsity, our approach seamlessly integrates with off-the-shelf, fast attention implementations [7] to expedite the prefetch process. + +To mitigate the memory bottleneck, we retain the KV cache for important tokens while discarding the less important ones. During the decoding phase, the current token interacts solely with the cached important tokens, thereby reducing the memory overhead of fetching the KV cache for computation. + +Efficient approximation of full attention scores. Full attention scores are not accessible in fast attention implementation [7]. In such case, to circumvent the computation of full attention scores in Eqs. (6) and (8), we selectively compute and accumulate the attention scores for a subset of tokens, following previous literature [15, 18]. The obtained partial attention scores are then used as the approximation of the full attention scores. The size of this subset is small and fixed, ensuring that the computational burden for these tokens remains minimal in long-context scenarios. The accumulated and normalized attention scores for each token can then be approximated with partial attention scores. Details and ablation experiments on this approximation can be found in the supplementary material. + +Algorithm 1: The prefetch phase of ZipVL +procedure ZipVL Prefill: +Input: Input embedding $\mathbf{X}$ , threshold $\tau$ +Output: Attention output $\mathbf{O}$ , KV cache $(\mathbf{K},\mathbf{V})$ +Calculate query, key and value states $(\mathbf{Q},\mathbf{K},\mathbf{V})$ as per Eq. (1) +Select a subset of tokens $\mathbf{Q}'$ from query states and compute attention scores $\mathbf{A}' = \text{Softmax}(\mathbf{Q}'\mathbf{K}^T)$ +Determine the number of important tokens as per Eq. (7) +Calculate the normalized attention scores for each token as per Eq. (8) +Select a set of important tokens $\mathbf{T}$ as per Eq. (9) +// Token-level Sparse Attention with FlashAttention + $\mathbf{O} = \text{FlashAttention}(\mathbf{Q}[\mathbf{T}],\mathbf{K}[\mathbf{T}],\mathbf{V}[\mathbf{T}])$ +// KV Cache + $\mathbf{K} = \mathbf{K}[\mathbf{T}]$ $\mathbf{V} = \mathbf{V}[\mathbf{T}]$ +return $\mathbf{O}$ , $(\mathbf{K},\mathbf{V})$ + +Algorithm 2: The decoding phase of ZipVL +procedure ZipVL Decoding: +Input: Input embedding x, stored KV cache $(\mathbf{K}_{\mathrm{in}},\mathbf{V}_{\mathrm{in}})$ +Output: Attention output o, updated KV cache $(\mathbf{K}_{\mathrm{out}},\mathbf{V}_{\mathrm{out}})$ +Calculate query, key and value states (q,k,v) as per Eq. (1) +Fetch KV cache from memory: $\mathbf{K}_{\mathrm{out}} =$ Concat $(\mathbf{K}_{\mathrm{in}},\mathbf{k}),\quad \mathbf{V}_{\mathrm{out}} =$ Concat $(\mathbf{V}_{\mathrm{in}},\mathbf{v})$ +Compute attention output o = FlashAttention(q,Kout,Vout) +return o, $(\mathbf{K}_{\mathrm{out}},\mathbf{V}_{\mathrm{out}})$ + +Overall, the attention mechanisms for the prefetch and decoding phases are summarized in Algorithms 1 and 2, respectively. + +# 5. Experiments + +# 5.1. Implementation Details + +To assess the effectiveness of our proposed method, we conduct experiments on both image and video understanding tasks. For image understanding, we utilize widely adopted LVLMs: LLaVA [22] and LLaVA-Next [23]. These models are evaluated against five rigorous benchmarks: VQAv2 [13], TextVQA [33], GQA [17], MME [8], and ChartQA [28]. For video understanding, evaluations are + +Table 1. Performance comparisons of image LVLMs on various benchmarks. Here, "Ratio" denotes the proportion of tokens participating in attention computation. "†" denotes token-level sparsity is only employed in attention modules. + +
ModelMethodRatioVQAv2ChartQATextVQAGQAMME
LLaVA-v1.5-7BFull100%76.618.246.161.91507
FastV†53.1%75.817.745.560.21511
HiRED20%73.017.345.656.81368
HiRED40%75.517.645.659.51433
Ours (τ=0.96)44.1%76.118.045.061.31495
Ours (τ=0.975)52.8%76.517.945.761.71505
LLaVA-Next-7BFull100%80.354.864.864.11519
FastV†53.1%79.551.263.763.71490
HiRED20%77.542.061.461.41483
HiRED40%78.846.561.859.41474
Ours (τ=0.96)40.4%79.451.062.863.81473
Ours (τ=0.975)49.3%79.852.664.064.01497
LLaVA-Next-13BFull100%80.966.266.965.71570
FastV†53.1%76.851.659.762.91555
HiRED20%77.948.963.663.11545
HiRED40%79.353.765.264.11570
Ours (τ=0.96)30.6%79.757.863.664.31551
Ours (τ=0.975)36.8%80.458.564.964.81574
+ +![](images/d778d78f75b0b165d7ad0f71eb6dd70f528bedbb83342a9449ac723152b16bc6.jpg) +Figure 4. The ratio of important tokens across different methods on different tasks. The proposed ZipVL can adaptively determine this ratio based on the attention scores, assigning more ratio to important tokens on complex tasks. + +![](images/daa78681942abd248feaac68ebb1385b282fe3415277196374beeaa42ca6fb95.jpg) + +![](images/9e1668d44b331d0e3a9e027252f6a9189b3f58c75b3a9aacaf12046abbc8d4dc.jpg) + +conducted using the LongVA [43] model on the Video-MME [9] benchmark. To ensure reproducibility, all reported results are obtained using the Evaluation Suite of Large Multimodal Models [21]. + +# 5.2. Main Results + +# 5.2.1. Evaluation on Image Benchmarks + +We begin our evaluation on five image comprehension benchmarks and compare our results against well-established methods with token-level sparsity: FastV [4] and HiRED [2]. The results are presented in Table 1. Notably, HiRED determines the importance of patches through the feature map of the visual encoder, without considering the semantic information of the input prompt, resulting in a significant accuracy drop. In contrast, both FastV and our approach assess token importance via attention maps + +in the LVLMs. However, FastV employs a fixed token ratio and exhibits severe performance degradation on challenging tasks such as ChartQA [28]. By implementing layer-wise adaptive ratio assignment, our proposed ZipVL consistently surpasses FastV across all five benchmarks and three model architectures, while maintaining a smaller overall ratio of important tokens. As illustrated in Figure 4, our method dynamically adjusts the ratio across various tasks and models, slightly increasing the ratio of important tokens for difficult tasks to preserve performance while enhancing efficiency on simpler tasks. Moreover, the performance gap between our method and FastV becomes more pronounced over the LLaVA-Next-13B model. This discrepancy can be attributed to the varying attention maps across different models and that FastV's predefined hyperparameters are not universally applicable, whereas our dynamic approach + +demonstrates high robustness. + +# 5.2.2. Evaluation on Video Benchmarks + +We also assess the performance of our method on the Video-MME benchmark [9] over the LongVA model [43], which supports a maximum multimodal input length of 224K tokens. We compare our approach with semi-structured sparse attention methods such as MInference [18] and QK-sparse [30], as well as the structured sparse attention method FastV [4]. The results are summarized in Table 2. Among these sparse attention methods, FastV [4] consistently retains a fixed proportion of tokens while MInference [18] retains a fixed number of sparse blocks. Notably, our approach not only achieves the highest overall performance but also exhibits superior reductions in FLOPs within the attention module and KV cache size compared to other sparse attention methods. This demonstrates the effectiveness of employing dynamic token-level sparsity to accelerate the attention module in LVLMs. Furthermore, long videos inherently contain significant redundancy, and our method dynamically allocates the ratio of important tokens by analyzing the sparse attention maps, resulting in a higher FLOPs reduction ratio when processing 128-frame videos compared to 64-frame videos. + +Additional evaluation results on the VideoChatGPT [27] benchmark can be found in the supplementary material. + +# 5.3. Ablation Study + +# 5.3.1. Effect of the Layer-wise Adaptive Ratio + +In this subsection, we evaluate the efficacy of the proposed adaptive ratio assignment scheme by integrating it with sparse attention and KV cache compression, as detailed in Table 3. Initially, we implement a fixed sparse attention scheme on LongVA-7B model over Video-MME benchmark. In this scheme, the ratio for important tokens remains constant across all attention layers and is fixed. Although this approach shares the same overall important token ratio and FLOPs reduction ratio as our method, it suffers from significant performance degradation (51.1% vs. 52.6%) due to its failure to account for the varying attention maps across layers. In contrast, our method achieves nearly lossless performance (52.4% vs. 52.6%) while reducing the FLOPs of attention mechanism by 82.3%. + +To ablate the efficacy of our method for KV cache compression, we integrate the adaptive ratio assignment scheme with the mixed-precision KV cache quantization approach [15] and evaluate its performance over LLaMA3-8B [29] model on the GSM8k dataset. For both methods, important tokens are quantized to 4-bit, while other tokens are quantized to 2-bit. Notably, by adaptively determining the ratio of important tokens for each layer, our method achieves a significantly higher compression ratio $(6.18 \times$ vs. $4.69 \times)$ while maintaining superior accuracy $(54.06\%$ + +vs. $53.75\%$ . This demonstrates that our method also sets a new state-of-the-art for KV cache compression of LLMs. + +# 5.3.2. Overhead Analysis + +ZipVL introduces additional overhead to adaptively assign the ratio of important tokens and identify token importance. In this subsection, we provide a detailed latency breakdown analysis, as summarized in Table 4. The primary source of overhead arises from computing approximate attention scores, while other operations, such as sorting and top- $k$ selection, contribute minimal overhead. Nonetheless, the speed gains achieved through sparse attention compensate for these costs, leading to a reduction in the overall time-to-first-token (TTFT). Furthermore, as the input sequence length increases, the impact of overhead diminishes due to the use of a fixed number of probe tokens for approximating attention scores. + +Ablation experiments on the effect of $\tau$ across other models and benchmarks, as well as on probe token selection and token importance metrics, are provided in the supplementary material. + +# 5.4. Deployment Efficiency + +In this subsection, we present the prefetch phase latency and decoding throughput in Table 5 to illustrate the real efficiency improvements achieved by ZipVL. Specifically, FastV [4] requires computing full attention scores to assess the token importance, which significantly increases the memory overhead. Consequently, FastV can only inference with a sequence length shorter than $16\mathrm{K}$ and a small batch size. On the other hand, MInference exhibits significant additional overhead and is notably slower than FlashAttention [7] for sequence lengths below $32\mathrm{K}$ . In contrast, ZipVL achieves comparable latency to FlashAttention and FastV with short input sequences, while significantly reducing the prefetch phase latency as the sequence length exceeds $32\mathrm{K}$ . This can be attributed to the fact that the attention module's latency becomes the dominant factor in the total latency with long sequences. With an input sequence length of $128\mathrm{K}$ , ZipVL achieves a $2.3\times$ reduction in prefetch-phase latency. Moreover, MInference is not designed to reduce the KV cache size, while the proposed ZipVL jointly optimizes the attention computation and the KV cache through the dynamic token ratio. With the reduction in KV cache size, ZipVL can perform inference with a larger batch size. Consequently, ZipVL presents a $2.8\times$ higher decoding throughput compared to baseline method [7] with an input sequence length of $16\mathrm{K}$ . + +# 6. Conclusion and Future Work + +In this paper, we have proposed ZipVL, an efficient inference framework tailored for LVLMs. ZipVL jointly optimizes both the prefetch and decoding phases by assigning an + +Table 2. Performance comparisons of video LVLMs on Video-MME benchmark. Here, "Attn FLOPs Reduction" denotes the reduction in floating-point operations (FLOPs) of the attention mechanism. "†" denotes token-level sparsity is only employed in attention modules. + +
ModelFramesMethodAttn FLOPs Reduction (%)KV Cache Reduction (%)ShortMediumLongOverall
LongVA-7B64Full0061.450.945.052.4
QK-sparse47.0060.951.445.152.4
MInference54.2060.751.244.652.1
FastV†71.746.461.050.645.052.2
Ours(τ=0.975)77.052.061.051.445.052.4
128Full0061.150.446.252.6
QK-sparse46.9061.349.746.352.4
MInference77.1061.050.545.352.3
FastV†71.746.460.250.246.252.2
Ours(τ=0.975)82.357.961.150.546.152.6
+ +Table 3. The effect of the proposed adaptive ratio assignment scheme on sparse attention and KV cache compression. Here, "Ratio" denotes the proportion of important tokens. For Video-MME benchmark, the input videos consist of 128 frames. + +
Sparse Attention
MethodRatio (%)Attn FLOPs Reduction (%)Video-MME (%)
LongVA-7B100052.6
Fixed42.182.351.1
Ours42.182.352.4
KV Cache Compression
MethodRatio (%)Compression RatioGSM8k Acc. (%)
LLaMA3-8B1001×55.88
ZipCache [15]70.04.69×53.75
Ours28.66.18×54.06
+ +Table 4. The overhead of each operation in ZipVL over LLaVA-Next-7B model. Here, "TTFT" denotes time-to-first-token and is measured with a batch size of 1. + +
Input LengthMethodTTFT (s)
16KOriginal3.48
+ approximate attention4.38
+ sort & cumsum4.41
+ normalize & top-k4.43
+ sparse attention3.05
32KOriginal9.35
+ approximate attention10.95
+ sort & cumsum10.96
+ normalize & top-k11.02
+ sparse attention6.40
+ +adaptive ratio of important tokens. This ratio is dynamically adjusted based on the distribution of attention scores across each layer, ensuring that the majority of attention scores are preserved. After identifying important tokens through normalized attention scores, less significant tokens are excluded from attention computation to alleviate the computational bottleneck. Additionally, their KV cache is evicted, mitigating the memory bottleneck in the decoding phase. Extensive experiments have demonstrated that ZipVL sig + +Table 5. Comparisons of prefetch phase latency and decoding throughput across different sequence lengths. Data is collected from LLaVA-Next-13B model on an Nvidia A100 GPU. Here, "TTFT" denotes time-to-first-token and is measured with a batch size of 1. "OOM" indicates out-of-memory during the decoding process. Throughput is measured using the maximum batch size that can be supported by a single GPU. + +
Input LengthMethodTTFT↓(s)Throughput↑(token/s)
8KFlashAttention1.3327.79
MInference3.3027.79
FastV1.2329.99
Ours1.3175.54
16KFlashAttention2.6414.10
MInference4.5714.10
FastVOOMOOM
Ours2.6140.87
32KFlashAttention5.90OOM
MInference6.51OOM
FastVOOMOOM
Ours5.0415.31
64KFlashAttention15.61OOM
MInference10.92OOM
FastVOOMOOM
Ours10.5110.29
128KFlashAttention48.01OOM
MInference20.08OOM
FastVOOMOOM
Ours20.57OOM
+ +nificantly enhances the generation efficiency of LVLMs, achieving up to a $2.3 \times$ reduction in prefill phase latency and a $2.8 \times$ higher decoding throughput. However, a limitation of our approach is its focus on sparse attention only, while the multi-layer perceptron (MLP) modules in both phases remain dense. Future efforts may explore extending sparse computations to MLP modules to further reduce computational complexity. + +Acknowledgement This work was supported by the National Key Research and Development Program of China (2022YFC3602601) and the National Key Research and Development Program of China (2022ZD0160102). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments. arXiv preprint arXiv:2408.10945, 2024. 2, 3, 4, 6 +[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 3 +[4] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 1, 2, 3, 4, 6, 7 +[5] Zhaodong Chen, Zheng Qu, Yuying Quan, Liu Liu, Yufei Ding, and Yuan Xie. Dynamic n: M fine-grained structured sparse attention mechanism. In PPoPP, pages 369-379, 2023. 3 +[6] Steve Dai, Hasan Genc, Rangharajan Venkatesan, and Brucek Khailany. Efficient transformer inference with statically structured sparse attention. In 2023 60th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2023. 3 +[7] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In NeurIPS, 2022. 1, 5, 7 +[8] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5 +[9] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 6, 7 +[10] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive kv cache compression for llms. In ICLR, 2024. 3 +[11] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. In ICLR, 2024. 1 +[12] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehen + +sion and generation. arXiv preprint arXiv:2404.14396, 2024. 1 +[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 5 +[14] Team PyTorch: Horace He, Driss Guessous, Yanbo Liang, and Joy Dong. Flexattention: The flexibility of pytorch with the performance of flashattention. https://pytorch.org/blog/flexattention/, 2024.3 +[15] Yefei He, Luoming Zhang, Weijia Wu, Jing Liu, Hong Zhou, and Bohan Zhuang. Zipcache: Accurate and efficient kv cache quantization with salient token identification. In NeurIPS, 2024. 2, 3, 4, 5, 7, 8 +[16] Coleman Hooper, Sehoon Kim, Hiva Mohammadzadeh, Michael W Mahoney, Yakun Sophia Shao, Kurt Keutzer, and Amir Gholami. Kvquant: Towards 10 million context length llm inference with kv cache quantization. arXiv preprint arXiv:2401.18079, 2024. 2, 3 +[17] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 5 +[18] Huiqiang Jiang, YUCHENG LI, Chengruidong Zhang, Qianhui Wu, Xufang Luo, Surin Ahn, Zhenhua Han, Amir H Abdi, Dongsheng Li, Chin-Yew Lin, et al. Minference: Accelerating pre-filling for long-context llms via dynamic sparse attention. In Workshop on Efficient Systems for Foundation Models II@ ICML, 2024. 1, 3, 5, 7 +[19] Hao Kang, Qingru Zhang, Souvik Kundu, Geonhwa Jeong, Zaoxing Liu, Tushar Krishna, and Tuo Zhao. Gear: An efficient kv cache compression recipe for near-lossless generative inference of llm. arXiv preprint arXiv:2403.05527, 2024. 3 +[20] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.3 +[21] Bo Li, Peiyuan Zhang, Kaichen Zhang, Fanyi Pu, Xinrun Du, Yuhao Dong, Haotian Liu, Yuanhan Zhang, Ge Zhang, Chunyuan Li, and Ziwei Liu. Lmms-eval: Accelerating the development of large multimodal models, 2024. 6 +[22] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 5 +[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.5 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1 +[25] Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, and Jiwen Lu. Efficient inference of vision instruction-following models with elastic cache. arXiv preprint arXiv:2407.18121, 2024. 3 + +[26] Zirui Liu, Jiayi Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, and Xia Hu. Kivi: A tuning-free asymmetric 2bit quantization for kv cache. arXiv preprint arXiv:2402.02750, 2024. 3, 4 +[27] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 7 +[28] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3, 5, 6 +[29] Meta. Introducing meta llama 3: The most capable openly available llm to date. https://ai.meta.com/blog/meta-llama-3/, 2024.7 +[30] Matteo Pagliardini, Daniele Paliotta, Martin Jaggi, and François Fleuret. Fast attention over long sequences with dynamic sparse flash attention. In NeurIPS, 2023. 1, 3, 7 +[31] Siyu Ren and Kenny Q Zhu. On the efficacy of eviction policy for key-value constrained generative language model inference. arXiv preprint arXiv:2402.06262, 2024. 2, 3, 5 +[32] Luka Ribar, Ivan Chelombiev, Luke Hudlass-Galley, Charlie Blake, Carlo Luschi, and Douglas Orr. Sparq attention: Bandwidth-efficient lIm inference. In ICML, 2024. 3 +[33] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, 2019. 5 +[34] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1 +[35] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1 +[36] Raja Vakekanand and Kira Sam. Llama 3.1: An in-depth analysis of the next-generation large language model, 2024. 1 +[37] Zhongwei Wan, Ziang Wu, Che Liu, Jina Huang, Zhihong Zhu, Peng Jin, Longyue Wang, and Li Yuan. Look-m: Look-once optimization in kv cache for efficient multimodal long-context inference. arXiv preprint arXiv:2406.18139, 2024. 1, 2, 3, 4 +[38] Zheng Wang, Boxiao Jin, Zhongzhi Yu, and Minjia Zhang. Model tells you where to merge: Adaptive kv cache merging for llms on long-context tasks. arXiv preprint arXiv:2407.08454, 2024. 3 +[39] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. In ICLR, 2024. 3, 4 +[40] Dongjie Yang, XiaoDong Han, Yan Gao, Yao Hu, Shilin Zhang, and Hai Zhao. Pyramid infer: Pyramid kv cache compression for high-throughput lvm inference. arXiv preprint arXiv:2405.12532, 2024. 2 +[41] June Yong Yang, Byeongwook Kim, Jeongin Bae, Beomseok Kwon, Gunho Park, Eunho Yang, Se Jung Kwon, and + +Dongsoo Lee. No token left behind: Reliable kv cache compression via importance-aware mixed precision quantization. arXiv preprint arXiv:2402.18096, 2024.3 +[42] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences. In NeurIPS, 2020. 3 +[43] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 1, 6, 7 +[44] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. In NeurIPS, 2023. 2, 3, 4 +[45] Qianchao Zhu, Jiangfei Duan, Chang Chen, Siran Liu, Xiuhong Li, Guanyu Feng, Xin Lv, Huanqi Cao, Xiao Chuanfu, Xingcheng Zhang, et al. Near-lossless acceleration of long context llm inference with adaptive structured sparse attention. arXiv preprint arXiv:2406.15486, 2024. 1, 3 \ No newline at end of file diff --git a/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/images.zip b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f1e53d911d5cab5e36e199f9fba334fe7bbb41aa --- /dev/null +++ b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae318e348c1046eee8c603d797fd108e3c06d3feb694784efb827b028cb1632e +size 484581 diff --git a/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/layout.json b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0a13f1b1b5084392552a1e9af685e964ac947d76 --- /dev/null +++ b/zipvlacceleratingvisionlanguagemodelsthroughdynamictokensparsity/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be1923d7b2903766b4515f967204d6882124d7a90fbba8518998440b07082786 +size 360352 diff --git a/ziumzeroshotintentawareadversarialattackonunlearnedmodels/2ecc5de9-caed-4861-8871-7280fe728f53_content_list.json b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/2ecc5de9-caed-4861-8871-7280fe728f53_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6411b7eb6eb409144727502c4ce08b5ff82a605 --- /dev/null +++ b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/2ecc5de9-caed-4861-8871-7280fe728f53_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0786061fff2e875096c1cf78df8bd0ef0f28dc5eb03bb89be31188c48df4191 +size 72563 diff --git a/ziumzeroshotintentawareadversarialattackonunlearnedmodels/2ecc5de9-caed-4861-8871-7280fe728f53_model.json b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/2ecc5de9-caed-4861-8871-7280fe728f53_model.json new file mode 100644 index 0000000000000000000000000000000000000000..19d17eb5d3edb8e9188dca9f16009d9ae4fbc478 --- /dev/null +++ b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/2ecc5de9-caed-4861-8871-7280fe728f53_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21abe4bb1a0d2b206f955f51807c97bf431f1dd5bd631a9e58aa8f22cc75e3d5 +size 89316 diff --git a/ziumzeroshotintentawareadversarialattackonunlearnedmodels/2ecc5de9-caed-4861-8871-7280fe728f53_origin.pdf b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/2ecc5de9-caed-4861-8871-7280fe728f53_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b04a3b0f60b236bef4c2713acc7346b5eec4d7d --- /dev/null +++ b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/2ecc5de9-caed-4861-8871-7280fe728f53_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ce9b88bd423f2003b8b3e78066e30fe15206a54de15c232da847d0982e6db31 +size 2191833 diff --git a/ziumzeroshotintentawareadversarialattackonunlearnedmodels/full.md b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ec8dfbb03303118b7b400cfccd3d431d0624cc7 --- /dev/null +++ b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/full.md @@ -0,0 +1,280 @@ +# ZIUM: Zero-Shot Intent-Aware Adversarial Attack on Unlearned Models + +Hyun Jun Yook $^{1}$ Ga San Jhun $^{1}$ Jae Hyun Cho $^{1}$ Min Jeon $^{1}$ Donghyun Kim $^{2}$ Tae Hyung Kim $^{3}$ Youn Kyu Lee $^{1,\dagger}$ Chung-Ang University ${}^{2}$ Korea University ${}^{3}$ Hongik University + +{hyunjun6, bonuspoint, wogus2031, mulsoap0504, younkyul}@cau.ac.kr d_kim@korea.ac.kr taehyung@hongik.ac.kr + +# Abstract + +Machine unlearning (MU) removes specific data points or concepts from deep learning models to enhance privacy and prevent sensitive content generation. Adversarial prompts can exploit unlearned models to generate content containing removed concepts, posing a significant security risk. However, existing adversarial attack methods still face challenges in generating content that aligns with an attacker's intent while incurring high computational costs to identify successful prompts. To address these challenges, we propose ZIUM, a Zero-shot Intent-aware adversarial attack on Unlearned Models, which enables the flexible customization of target attack images to reflect an attacker's intent. Additionally, ZIUM supports zero-shot adversarial attacks without requiring further optimization for previously attacked unlearned concepts. The evaluation across various MU scenarios demonstrated ZIUM's effectiveness in successfully customizing content based on user-intent prompts while achieving a superior attack success rate compared to existing methods. Moreover, its zero-shot adversarial attack significantly reduces the attack time for previously attacked unlearned concepts. + +# 1. Introduction + +Machine Unlearning (MU) selectively removes specific data points or features from a trained deep learning model through reweighting or pruning [20, 41]. It helps protect privacy and prevent the generation of sensitive content [22, 24, 37]. Recently, MU has been used to prevent the generation of inappropriate images by eliminating concepts such as nudity and violence—commonly associated with NSFW (i.e., Not Safe For Work) content—from pre-trained text-to-image generation models [11, 17, 25, 29, 45]. However, even after MU is applied, these models (i.e., unlearned models) can still generate images of the removed concepts (i.e., unlearned concepts) when given adversarial prompts, posing a significant security risk [29, 31, 38]. + +![](images/c4c88a17c88f041e3123e3e2bae15fb2d94d39545d9bb18d1344d73e6ff64336.jpg) +Figure 1. Examples of generated images by ZIUM: 1st adversarial attack utilizing the user-intent prompt and 2nd adversarial attack without additional optimization for the same unlearned concept. + +Several adversarial attack methods targeting unlearned models have been proposed to exploit this. Specifically, approaches [5, 12, 23, 30, 47] have been developed to identify optimal adversarial prompts based on a target attack image containing an unlearned concept. These methods generate an image that incorporates the unlearned concept while closely resembling the target attack image by using the optimal adversarial prompt as input to the unlearned model. However, these methods heavily rely on the target attack image, making it challenging to generate an image that reflects both unlearned concepts and the attacker's intent (e.g., preferences and background). This is crucial for generating an image that not only includes the unlearned concepts but also aligns with the attacker's intentions in different forms. Recent approaches [12, 26, 30, 39, 44] have attempted to align with the attacker's intent using only a target prompt. However, without a target attack image, fully representing a single image in text form is challenging, and the attacker's prompt may lack sufficient semantic detail [17]. Therefore, + +a new attack method is required to exploit unlearned models, generating customized images in various forms while accurately reflecting the attacker's intent. Moreover, when an attacker aims to generate images with varying contexts (e.g., adding a new concept like "Japanese comics" as shown in Fig.1(b)), existing approaches require repeatedly identifying optimal adversarial prompts for each context, making the attack process costly [23, 37, 41]. Therefore, a zero-shot attack mechanism is required to eliminate the need for additional optimization steps for the same unlearned concept. + +To address these challenges, we propose ZIUM, a novel adversarial attack method that enables attackers to customize target attack images based on their intent while supporting zero-shot adversarial attacks. Our approach exploits an unlearned model to generate images that closely resemble the target attack image while embedding the unlearned concept. By incorporating user-intent prompts, the generated images can be precisely tailored to align with the attacker's intent. Furthermore, our method enables zero-shot adversarial attacks, eliminating the need for additional optimization processes for previously attacked unlearned concepts. Fig. 1 illustrates an adversarial attack using ZIUM. In the first attack trial, the generated image (a) closely resembles the target attack image while embedding the unlearned concept (i.e., nudity) without a user-intent prompt, whereas image (b) reflects the user-intent prompt "Japanese comics." In the second attack trial, targeting the same unlearned concept, a zero-shot adversarial attack was applied using the module optimized during the first attack. As a result, image (c) resembles the unseen target attack image while embedding the same unlearned concept without a user-intent prompt. Moreover, image (d) successfully reflects the user-intent prompt "Waterfall." + +ZIUM addresses the challenge of insufficient semantic detail by utilizing both the unlearned concept from the target attack image and the user-intent prompt that reflects the attacker's intent. To achieve this, ZIUM employs an image captioning technique that converts image embeddings into text embeddings. First, it extracts the visual embedding of the target attack image containing the unlearned concept and transforms it to the text embedding of the unlearned model. This embedding is then fed into the unlearned diffusion model along with the text embedding of the user-intent prompt to generate an image that aligns with the attacker's intent. Furthermore, image captioning techniques enable zero-shot adaptation to specific trained concepts within an image. Leveraging this capability, ZIUM facilitates additional attacks without requiring further optimization for previously attacked unlearned concepts. + +We evaluated ZIUM on representative unlearned models, including ESD [11], FMN [45], SLD [36], and AdvUnlearn [46]. Our method outperforms existing adversarial + +attack methods across various unlearned concept scenarios (e.g., nudity, violence, illegal activity, style, and object), achieving a significantly higher attack success rate (ASR) on average—improving by at least $22.6\%$ p and up to $62.0\%$ p. Furthermore, experiments with diverse user-intent prompts demonstrate that ZIUM effectively generates images that accurately align with the attacker's intent. Notably, ZIUM maintains a high ASR without requiring additional optimization for the same unlearned concept. + +# Our contributions are summarized as follows: + +1. Proposal of a novel adversarial attack method for unlearned models that effectively reflects various attacker intents while achieving a higher attack success rate. +2. Design of a zero-shot adversarial attack mechanism that enables targeting the same unlearned concept without requiring additional optimization. +3. A comprehensive evaluation demonstrating the effectiveness of ZIUM across different unlearned models and unlearned concept scenarios. + +This paper is structured as follows: Section 2 reviews related works, Section 3 introduces ZIUM, Section 4 presents the experimental results, and Section 5 concludes the paper. + +# 2. Related Works + +# 2.1. Machine Unlearning + +Generative models can produce large amounts of potentially inappropriate content (e.g., nudity, copyright infringement), increasing the need for effective constraints. To address this issue, Machine Unlearning (MU) has been actively studied. MU is a data removal mechanism designed to eliminate the influence of undesirable data points without requiring costly retraining while preserving model performance for inputs unrelated to the removed data [2]. Gandikota et al. [11] proposed Erased Stable Diffusion (ESD), which removes specific concepts without requiring additional training data by fine-tuning the weights of the stable diffusion model. Fan et al. [10] introduced Saliency Unlearning (SalUn), which modifies only a subset of the model's weights rather than the entire network. Jia et al. [15] incorporated model sparsity which involves reducing unimportant data or parameters through targeted operations. In this approach, weight pruning was applied to enhance both the efficiency and performance of the unlearning process. Zhang et al. [45] introduced Forget-Me-Not (FMN), which minimizes the attention map between text and images to facilitate unlearning in text-to-image models [33]. Schramowski et al. [36] proposed Safe Latent Diffusion (SLD), which extends the existing Classifier-Free Guidance [14] to align text prompt inputs while preventing the generation of images containing unlearned concepts. Despite these advancements, existing MU mechanisms still have limitations in fully unlearning concepts. In particular, unlearned models can still generate + +![](images/cf552aaadb42345371434f71998e485f4c71be36fb9208b723b3da474bedd042.jpg) +Initial Attack Phase + +![](images/0b1f933f4a72177aa499e474a8d3a209892c91aeefd844fca4da28b699ad8de2.jpg) +Zero-shot Attack Phase +Figure 2. An overview of the ZIUM's initial attack phase and zero-shot attack phase. + +images containing concepts that were supposed to be removed. These limitations have become even more evident through adversarial attacks on unlearned models [5, 47]. + +# 2.2. Adversarial Attacks on Machine Unlearning + +Several methods leveraging adversarial attacks on MU have been proposed [6, 12, 27, 28, 43]. Tsai et al. [39] proposed Ring-A-Bell, a method that identifies optimal adversarial prompts in a black-box setting. It generates concept vectors by computing the difference in embeddings between prompts that contain adversarial concepts and those that do not. Ma et al. [26] introduced Jailbreaking Prompt Attack (JPA), an adversarial attack method that bypasses MU mechanisms. This approach optimizes adversarial prompts by obtaining an unlearned embedding derived from the difference in embeddings between an unlearned concept and its antonym. Zhang et al. [47] proposed UnlearnDiffAtk, which identifies optimal adversarial prompts to evaluate the robustness of MU mechanisms in diffusion models. It executes attacks using the diffusion classifier inherent in the diffusion model itself, eliminating the need for auxiliary models. Chin et al. [5] proposed Prompt4Debugging (P4D), which evaluates the robustness of MU mechanisms. This approach utilizes prompt engineering to identify optimal adversarial prompts that bypass these mechanisms. Existing methods identify an optimal adversarial prompt based on a given target prompt or target attack image. However, incorporating the attacker's intent into the target prompt or obtaining target attack images that capture both the at + +tacker's intent and the unlearned concept remains a challenge. As a result, adversarial attacks on unlearned models may not fully align with the attacker's intended form. + +# 3. Method + +In this study, ZIUM performs the attack by identifying the optimal adversarial condition that enables the generation of an image incorporating both the unlearned concept and the attacker's intent, using the target attack image and user-intent prompt. + +As shown in Fig. 2, ZIUM comprises an initial attack phase and a zero-shot attack phase, which utilizes target attack images and user-intent prompts to perform adversarial attacks. The target attack image includes unlearned concepts from the diffusion model, such as nudity, and the user intent prompt includes the attacker's intent to customize the target attack image, such as "At the beach," "Shining stars," and "At the station." The initial attack phase performs the adversarial attack by utilizing the target attack image and user-intent prompt to identify optimal adversarial condition through the "visual-text alignment process" and the "optimization process." Subsequently, the zero-shot attack phase only exploits the visual-text alignment process optimized through the initial attack phase to perform adversarial attacks without further optimization process. The details will be described in the following. + +# 3.1. Initial Attack Phase + +# 3.1.1. Visual-text Alignment Process + +The visual-text alignment process in the initial attack phase transforms the visual embedding of the target attack image into text embedding to utilize the unlearned concept of the target attack image as a condition for the unlearned diffusion model. The visual-text alignment process consists of a visual encoder, an image captioning module, and a projection layer. The visual encoder extracts key features of the target attack image that embed the unlearned concept, representing them as a fixed $k$ -dimensional visual embedding. The extracted $k$ -dimensional embedding is fed into the image captioning module. The image captioning module based on the cross-attention mechanism was pre-trained with Image-Text Contrastive Learning (ITC) loss, ImageText Matching (ITM) loss, and Image-grounded Text Generation (ITG) loss [4, 9, 19, 21]. ITC loss and ITM loss maximizes the similarity between the visual embedding and the text embedding when an image and text are paired, ensuring that the information contained in each embedding is aligned. ITG loss utilizes the next text token prediction through an attention mask, allowing the image captioning module to learn text embeddings that can generate information about the input image. With these losses, the pretrained image captioning module transforms the main features of the input visual embedding into aligned text embedding. Therefore, the transformed text embedding contains key visual embedding information about the input image (used as the target attack image). Subsequently, the projection layer projects the dimensionality of the transformed text embedding into $L$ -dimensions to match the input size of the CLIP text encoder used in the unlearned diffusion model [33]. The $L$ -dimensional text embedding from this process is concatenated with the text embedding of the user intent prompt to serve as a condition for the unlearned diffusion model. + +# 3.1.2. Optimization Process + +The optimization process of the initial attack phase performs the attack by identifying an optimal adversarial condition that enables image generation containing the unlearned concept. Specifically, the optimization process is based on the diffusion classifier mechanism [3, 18, 47]. The diffusion classifier mechanism utilizes Bayes' rule to estimate the condition that can generate a desired target image. By Bayes' rule, the probability that an image $x$ is generated given a certain condition $c_{i}$ is expressed as follows, + +$$ +p _ {\theta} \left(c _ {i} \mid x\right) = \frac {p \left(c _ {i}\right) p _ {\theta} \left(x \mid c _ {i}\right)}{\sum_ {j} p \left(c _ {j}\right) p _ {\theta} \left(x \mid c _ {j}\right)} \tag {1} +$$ + +In Eq. 1, $\theta$ denotes the parameters of the diffusion model, $x$ denotes the image we want to generate via diffusion, and $c$ denotes the condition we want to estimate. Thus, $p_{\theta}(x|c_i)$ is + +the probability of an image being generated by a condition, and, $p(c)$ is the prior probability distribution of the condition. In general, diffusion model assumes no prior information of a particular condition $c$ , so the prior probability $p(c)$ can be approximated by a uniform distribution. Applying this, Eq. 1 simplifies as follows, + +$$ +p _ {\theta} \left(c _ {i} \mid x\right) = \frac {p _ {\theta} \left(x \mid c _ {i}\right)}{\sum_ {j} p _ {\theta} \left(x \mid c _ {j}\right)} \tag {2} +$$ + +In a diffusion model, $p_{\theta}(x|c_i)$ is proportional to the accuracy of the denoising process at timestep $t$ . Based on this, it is expressed as follows, + +$$ +p _ {\theta} \left(c _ {i} | x\right) \propto \exp \left(- E _ {t, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} \left(x _ {t} \mid c _ {i}\right) \| _ {2} ^ {2} \right]\right) \tag {3} +$$ + +In Eq. 3, $x_{t}$ is the sum of $x$ and the noise $\epsilon_{t}$ , which represents the noisy image generated at a specific timestep $t$ , and $\epsilon_{\theta}(x_{t}|c_{i})$ is the noise predicted by diffusion given $x_{t}$ and condition $c_{i}$ . Therefore, it is possible to maximize the probability that the desired image $x$ is generated through a specified condition $c_{i}$ . As a result, the final optimization process for the diffusion classifier mechanism can be conducted as follows, + +$$ +\underset {c _ {i}} {\text {m i n i m i z e}} E _ {t, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} \left(x _ {t} \mid c _ {i}\right) \| _ {2} ^ {2} \right] \tag {4} +$$ + +Based on Eq. 4, ZIUM's optimization process utilizes both the target attack image and user-intent prompt to estimate the optimal adversarial condition $c_{i}$ . First, the visual encoder $\mathcal{E}(\cdot)$ of the visual-text alignment process which extracts the visual embedding $e_{i}$ from the target attack image $x_{i}$ is expressed as follows, + +$$ +e _ {i} = \mathcal {E} \left(x _ {i}\right) \tag {5} +$$ + +Let $f_{\theta'}(\cdot)$ be the network consisting of an image captioning module and a projection layer that converts the extracted visual embedding $e_i$ into a text embedding. Then, the condition $c_i$ can be expressed as follows, + +$$ +c _ {i} = f _ {\theta^ {\prime}} \left(e _ {i}\right) \tag {6} +$$ + +In Eq. 6, $\theta^{\prime}$ refers to the parameters of the image captioning module and the projection layer. Then, the condition $c_{i}$ is concatenated with the text embedding $p$ of the user-intent prompt and then processed through the CLIP text encoder of the unlearned diffusion model. This results in the condition $c_{i}, p$ , which reflects the unlearned concept from the target attack image and the attacker's intent from the user-intent prompt. Therefore, Eq. 4 is expressed as follows, + +$$ +\underset {\theta^ {\prime}} {\text {m i n i m i z e}} E _ {t, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} \left(x _ {t} \mid f _ {\theta^ {\prime}} \left(e _ {i}\right), p\right) \| _ {2} ^ {2} \right] \tag {7} +$$ + +In Eq. 7, note that in the process of transforming the extracted visual embedding into a text embedding, the parameters of image captioning module and projection layer $\theta^{\prime}$ are only updated while excluding those of the visual encoder $\mathcal{E}(\cdot)$ to identify the optimal adversarial condition. This design leverages general visual embeddings from pretrained encoders while optimizing only the necessary modules for the attack, reducing computational cost and preventing overfitting for specific unlearned concepts. + +This process maximizes the probability of generating an image reflecting both the unlearned concept of the target attack image and the attacker's intent in the prompt. + +# 3.2. Zero-shot Attack Phase + +In the zero-shot attack phase, ZIUM utilizes the optimized image captioning module and projection layer through the initial attack phase to perform adversarial attacks without further optimization process. Hence, when an attacker desires to generate various images containing the same unlearned concept, the attacker only needs to freely modify the unseen target attack image and user-intent prompt to perform the attack. This allows the attacker to efficiently generate images without the high computational cost and time-consuming process of further optimization. + +Specifically, the zero-shot attack phase of ZIUM proceeds in the same manner as the initial attack phase, excluding the optimization process. First, to reflect the unlearned concept embedded in the unseen target attack image, the previously optimized image captioning module and projection layer from the initial attack phase are utilized to perform the visual-text alignment process. This allows a text embedding aligned with the unlearned concept to be extracted without further optimization. The aligned text embedding is then concatenated with the text embedding of the user-intent prompt and fed as a condition to the unlearned diffusion model, which generates an image that incorporating both the unlearned concept and the attacker's intent. Consequently, as shown in Fig. 2, the attacker can efficiently perform adversarial attacks using unseen target attack images that contain the unlearned concept along with various user-intent prompts (e.g., "Shining stars" and "At the station"). + +# 4. Experiments + +To evaluate the effectiveness of ZIUM, we formulated the following research questions: + +- RQ#1: How well does ZIUM achieve superior attack performance compared to existing methods? +- RQ#2: How well does ZIUM reflect diverse attacker intents? +- RQ#3: How well does ZIUM's zero-shot adversarial attack maintain high attack performance without additional optimization? + +# 4.1. Experimental Settings + +Implementation Details. The visual-text alignment process of ZIUM is designed based on BLIP2 [21], a representative image captioning model. The visual encoder utilizes CLIP's ViT-L/14, and the image captioning module adopts a Q-former structure [8, 32]. The projection layer utilizes a fully connected layer. All experiments were conducted using an NVIDIA RTX A100 (80G) with the following hyperparameters: $k = 768$ , $L = 768$ , optimizer=AdamW, learning rate=1e-4, weight decay=1e-2, and iterations=100. + +Prompt Datasets. To evaluate ZIUM across various unlearned concept scenarios (i.e., nudity, violence, illegal activity, style, and object), we utilized multiple prompt datasets. For the nudity, we adopted the NSFW dataset, using 142 prompts. For the violence and illegal activity, we adopted the I2P dataset (violence: 756 prompts, illegal activity: 727 prompts) [36]. Among these, we selected 334 prompts for violence and 248 prompts for illegal activity, where the proportion of inappropriate images classified by the Q16 classifier was greater than $50\%$ [26, 35, 39]. For the style, we selected Van Gogh's artistic style as the target, using 50 prompts, following the experimental setup in the existing study [47]. For the object, we selected two different objects (i.e., church and parachute) as targets, using 50 prompts each, also following the experimental setup in the existing study [47]. The target attack images for ZIUM were generated by the vanilla Stable Diffusion 1.4v model, using each unlearned concept scenario's prompt dataset. + +Unlearned Models. We selected four representative unlearned diffusion models—ESD [11], FMN [45], SLD [36], and AdvUnlearn [46]. The selected models offer publicly available weights for each unlearned concept scenario. Notably, for the style and object, only the ESD, FMN, and AdvUnlearn models were used, as SLD was not originally designed for these concepts [47]. We used the official implementations provided by the authors. + +Existing Attack Methods. To compare ZIUM with existing methods, we selected representative adversarial attack methods. For methods using target attack images, we selected UnlearnDiffAtk [47], a state-of-the-art adversarial attack method. For methods using target prompts, we selected P4D [5] and Ring-A-Bell [39], which are white-box and black-box attack methods, respectively [16]. We used the official implementations provided by the authors. + +Evaluation Metric. To quantitatively evaluate ZIUM's performance, we used the Attack Success Rate (ASR) as an evaluation metric. ASR measures the proportion of successful attacks across the dataset. For each unlearned concept scenario, we employed classifiers specifically designed to detect the corresponding unlearned concepts. For the nudity, we utilized NudeNet [1]. A generated image was classified as containing nudity if at least one of the predefined labels (FEMALE_BREAST_EXPOSED, + +
MethodsNudityViolenceIllegal ActivityVan GoghChurchParachuteAvg.
ESDFMNSLDAUESDFMNSLDESDFMNSLDESDFMNAUESDFMNAUESDFMNAU
No attack21.1%88.0%33.1%21.1%45.8%70.6%47.9%56.4%57.6%43.9%2.0%10.0%2.0%14.0%52.0%6.0%4.0%46.0%14.0%33.4%
UnlearnDiffAtk80.2%98.5%37.3%21.1%96.4%98.8%94.6%97.1%97.9%97.9%36.0%54.0%0.0%66.0%96.0%8.0%48.0%100.0%12.0%65.2%
P4D29.5%64.0%40.8%5.6%33.5%41.9%27.2%42.7%59.6%19.7%18.0%4.0%2.0%20.0%20.0%4.0%22.0%34.0%2.0%25.8%
Ring-A-Bell49.2%95.7%1.4%2.8%44.9%67.9%60.4%42.3%48.7%32.6%0.0%2.0%0.0%2.0%54.0%0.0%6.0%64.0%0.0%30.2%
ZIUM97.1%98.5%98.5%91.5%99.4%99.1%98.2%98.3%98.7%97.1%86.0%68.0%88.0%62.0%92.0%70.0%76.0%90.0%60.0%87.8%
+ +Table 1. ASR for ESD, FMN, SLD, and AdvUnlearn (AU) of ZIUM and existing adversarial attack methods under unlearned concept scenarios (nudity, violence, illegal activity, Van Gogh, church, and parachute). + +![](images/482506fef8d727974f30ef6b010d8310242950a9abc02eed5780ced301fa015c.jpg) +Figure 3. Examples of generated images for ESD by ZIUM and existing adversarial attack methods under NSFW unlearned concept scenarios (nudity, violence, and illegal activity). + +FEMALE_GENITALIA_EXPOSED, MALE_BREAST_EXP- OSED, MALE_GENITALIA_EXPOSED, BUTTOCKS_EXP- OSED, ANUS_EXPOSED, FEET_EXPOSED, BELLY_EXP- OSED, and ARMPITS_EXPOSED) was detected with a confidence score exceeding the threshold of 0.45. For the violence and illegal activity, we utilized the Q16 classifier [35]. For the style, we utilized ViT-L/14 [8, 47], which was pretrained on ImageNet [7] and fine-tuned on the WikiArt dataset [34]. For the object, we utilized ResNet-50 [13] pretrained on ImageNet. + +# 4.2. Experimental Results + +# 4.2.1. (RQ#1) Comparison of ZIUM's attack performance with existing adversarial methods + +To evaluate RQ#1, we compared the ASR of ZIUM and various existing adversarial attack methods under unlearned + +concept scenarios (nudity, violence, illegal activity, style, and object). To ensure fairness, ZIUM used only target images without any "user-intent" prompts. Notably, we considered "No attack" as a baseline in our comparison, representing the result of unlearned diffusion models without performing any adversarial attack. + +Table 1 presents the ASR of "No attack" and each adversarial attack method (UnlearnDiffAtk, P4D, and Ring-A-Bell, and ZIUM), targeting unlearned diffusion models (ESD, FMN, SLD, and AdvUnlearn) across the nudity, violence, illegal activity, Van Gogh, church, and parachute. + +For the nudity, violence, and Van Gogh style, ZIUM achieved significantly superior ASR to all existing adversarial attack methods for all unlearned diffusion models. In contrast, for the illegal activity, church, and parachute categories, ZIUM did not achieve a superior ASR compared + +![](images/702678c398aa8d4e32cac0679264be5824af0a25782f81207db38b142e69f407.jpg) +Figure 4. Examples of generated images by ZIUM: 1st attack utilizing various user-intent prompts and 2nd attack utilizing ZIUM's zero-shot attack phase. Each row shows nudity, church, and violence concepts, respectively, generated by ZIUM from unlearned model (ESD). + +to existing adversarial attack methods but instead produced comparable results, across all unlearned diffusion models. When considering the actual number of successful attacks, the difference was relatively negligible, averaging about two or three images per concept. + +Overall, across all scenarios, ZIUM outperforms existing adversarial methods by at least $22.6\%$ p, with a maximum of $62.0\%$ p on average. Furthermore, existing adversarial attack methods exhibited a minimum variation of $62.0\%$ p, depending on the unlearned scenario and model. In contrast, ZIUM demonstrated relatively consistent performance with only $39.4\%$ p variation. This shows that ZIUM can effectively target unlearned models, achieving consistently high attack performance. Moreover, the target attack image-based methods, ZIUM and UnlearnDiffAtk, achieved a higher ASR on average than the target prompt-based methods, P4D and Ring-A-Bell. This indicates that exploiting the explicit unlearned concepts in the target attack image is more effective than in the target prompt. + +Fig. 3 illustrates examples of images generated by a vanilla Stable Diffusion without MU applied, and images generated by each of the adversarial attack methods against the ESD model under three unlearned concept scenarios (nudity, violence, and illegal activity). Note that, more examples of generated images for ZIUM and existing adversarial attack methods can be found in Appendix A1. + +For the nudity, the images generated by UnlearnDiffAtk + +and P4D both included a female figure, resembling the image generated by a vanilla Stable Diffusion. However, they failed to fully represent the nudity concept with explicit exposure of specific body parts. Ring-A-Bell, in particular, failed to depict the human figure at all. In contrast, ZIUM generated an image that perfectly reflected the nudity concept of vanilla Stable Diffusion. + +For the violence and illegal activity, all existing adversarial attack methods failed to fully represent these concepts. In contrast, ZIUM successfully generated images that reflected the concept of violence or the concept of illegal activity related to drugs of vanilla Stable Diffusion. + +# 4.2.2. (RQ#2) Evaluation of ZIUM's customization effectiveness using user-intent prompts + +To evaluate RQ#2, we analyzed ZIUM's customized attack images based on user-intent prompts. The first attack trial in Fig. 4, utilizing ZIUM's initial attack phase, presents the generated images without a user-intent prompt (Prompt: None) and the images reflecting the attacker's intent through three different unlearned concepts (nudity, church, and violence). Note that, the first attack trial follows ZIUM's initial attack phase. More examples of ZIUM's customized attack images can be found in Appendix A2. + +Fig. 4(a) shows that the background and action (Prompt: "Swimming underwater," "Whipping," and "At the toilet") change according to the user-intent prompt, while maintaining the characteristic that the target attack image is of a + +
MethodsNudityVan GoghParachuteAttack Time (mins)
ESD
UnlearnDiffAtk80.2%36.0%48.0%24.4
P4D29.5%18.0%22.0%29.9
Ring-A-Bell49.2%0.0%6.0%9.1
ZIUM (Initial)97.1%86.0%76.0%9.0
ZIUM (Zero-shot)84.5%50.0%48.0%0.2
+ +Table 2. ASR and average Attack Time for ESD of ZIUM and existing adversarial attack methods under various unlearned concept scenarios (nudity, Van Gogh, and parachute). + +woman. Fig. 4(b) shows that the color of the building of the church in the target attack image is maintained, but at the same time, related objects (Prompt: “Rainbow,” “Palm tree,” and “Bike”) are generated together according to the user-intent prompt. Fig. 4(c) shows that the violence concept in the target attack image is maintained, but at the same time, the style of the object (Prompt: “Dreadlocks” and “Bleeding”) and the background (Prompt: “Construction site”) change according to the user-intent prompt. + +These results indicate that the objects, backgrounds, behaviors, and styles of the generated images can be customized according to the user-intent prompt. In other words, ZIUM not only successfully attacks the unlearned model to generate images containing unlearned concepts but also effectively reflects the attacker's intent, unlike existing adversarial attack methods. + +# 4.2.3. (RQ#3) Comparison of ZIUM's zero-shot attack performance with existing adversarial methods + +To evaluate RQ#3, we compared the ASR and elapsed attack time of ZIUM and existing adversarial attack methods under unlearned concept scenarios (nudity, Van Gogh, and parachute). Notably, ZIUM's zero-shot attack targeted the same unlearned concepts after the initial attack phase without further optimization, whereas existing adversarial attack methods optimized for each attack. + +Table 2 presents the ASR and average attack time of UnlearnDiffAtk, P4D, Ring-A-Bell, ZIUM (Initial), and ZIUM (Zero-shot) targeting ESD across the nudity, Van Gogh, and parachute. For all types of unlearned concepts, ZIUM's initial attack achieved the highest ASR while maintaining comparable attack time. + +Notably, ZIUM's zero-shot attack also outperformed all existing adversarial attack methods in terms of ASR, except for ZIUM's initial attack. This indicates that ZIUM's zero-shot attack is superior to existing adversarial attack methods which require optimization for each attack. Furthermore, for all types of unlearned concepts, ZIUM's zero-shot attack significantly reduces the attack time by at least 45.5 times and up to 149.5 times on average, compared to existing adversarial attack methods. + +In addition, as shown in Fig. 4(d)-(f), ZIUM's zero-shot attack successfully generates customized images in the second attack trial. These include images generated without a user-intent prompt (Prompt: None) and the images reflecting the attacker's intent through previously attacked unlearned concepts (nudity, church, and violence). + +For the nudity, Fig. 4(d) shows that the shape of the undressed man and woman in the unseen target attack image changes according to the user-intent prompt, transforming into a black-and-white drawing style. For the church, Fig. 4(e) shows that while the color of the church building is maintained from the unseen target attack image, related objects and backgrounds (Prompt: "Forest") are also generated based on the user-intent prompt. For the violence, Fig. 4(f) shows that even without a user-intent prompt, the concept of violence is reflected while maintaining the appearance of the figure in the unseen target attack image. + +Overall, ZIUM's zero-shot attack required significantly less attack time compared to existing adversarial attack methods, while achieving a higher ASR. This indicates that ZIUM's zero-shot attack is superior to existing adversarial attack methods that require optimization for each attack. Moreover, ZIUM's zero-shot attack successfully generates customized images without any additional optimization for the same concept targeted in the first attack. + +# 5. Conclusion + +In this paper, we proposed ZIUM, a novel zero-shot adversarial attack method for unlearned diffusion models, enabling customization to reflect various attacker intents. ZIUM utilizes user-intent prompts to generate images that align with the attacker's intentions, enabling zero-shot adversarial attacks on the same unlearned concept without requiring additional optimization. + +Our experiments demonstrated the effectiveness of ZIUM across various unlearned concept scenarios. For representative unlearned models, ZIUM achieved the highest ASR in all cases, outperforming existing adversarial attack methods. Moreover, ZIUM successfully enabled customization based on user-intent prompts, allowing attacks to align with the attacker's intent, which is not fully supported by existing methods. Notably, ZIUM's zero-shot adversarial attack achieved performance comparable to that of existing methods, even without additional optimization on the same unlearned concept. + +As future work, we plan to develop a prompt engineering mechanism that automates the transformation of given conditions into text prompts [40]. Moreover, we plan to develop a model-agnostic mechanism by applying transferable adversarial attack methods [42]. + +# Acknowledgments + +This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2025-00555277), and by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grants funded by MSIT (2021-000766, 2024-RS-2024-00436857, RS-2019-II190079). This research was also supported by the Ministry of Culture, Sports and Tourism and the Korea Creative Content Agency in 2024 (RS-2024-00345025). + +# References + +[1] P Bedapudi. Nudenet: Neural nets for nudity classification, detection and selective censoring. 2019. 5 +[2] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In IEEE Symposium on Security and Privacy, pages 463-480. IEEE, 2015. 2 +[3] Huanran Chen, Yinpeng Dong, Zhengyi Wang, Xiao Yang, Chengqi Duan, Hang Su, and Jun Zhu. Robust classification via a single diffusion model. arXiv preprint arXiv:2305.15241, 2023. 4 +[4] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 4 +[5] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, Pin-Yu Chen, and Wei-Chen Chiu. Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 1, 3, 5 +[6] Pucheng Dang, Xing Hu, Dong Li, Rui Zhang, Qi Guo, and Kaidi Xu. Diffzoo: A purely query-based black-box attack for red-teaming text-to-image generative model via zeroth order optimization. arXiv preprint arXiv:2408.11071, 2024. 3 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 6 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5, 6 +[9] Maksim Dzabraev, Alexander Kunitsyn, and Andrei Ivaniuta. Vlrm: Vision-language models act as reward models for image captioning. arXiv preprint arXiv:2404.01911, 2024. 4 +[10] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classi + +fication and generation. arXiv preprint arXiv:2310.12508, 2023. 2 + +[11] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 5 + +[12] Sensen Gao, Xiaojun Jia, Yihao Huang, Ranjie Duan, Jindong Gu, Yang Liu, and Qing Guo. Rt-attack: Jailbreaking text-to-image models via random token. arXiv preprint arXiv:2408.13896, 2024. 1, 3 + +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 + +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 +[15] Jinghan Jia, Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, Pranay Sharma, and Sijia Liu. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36:51584-51605, 2023. 2 +[16] Changhoon Kim, Kyle Min, and Yezhou Yang. Race: Robust adversarial concept erasure for secure text-to-image diffusion model. In European Conference on Computer Vision, pages 461-478. Springer, 2024. 5 +[17] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1 +[18] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2206-2217, 2023. 4 +[19] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36:30146-30166, 2023. 4 +[20] Guihong Li, Hsiang Hsu, Chun-Fu Chen, and Radu Marculescu. Machine unlearning for image-to-image generative models. arXiv preprint arXiv:2402.00351, 2024. 1 +[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, pages 19730-19742. PMLR, 2023. 4, 5 +[22] Na Li, Chunyi Zhou, Yansong Gao, Hui Chen, Zhi Zhang, Boyu Kuang, and Anmin Fu. Machine unlearning: Taxonomy, metrics, applications, challenges, and prospects. IEEE Transactions on Neural Networks and Learning Systems, 2025. 1 +[23] Han Liu, Yuhao Wu, Shixuan Zhai, Bo Yuan, and Ning Zhang. Riatig: Reliable and imperceptible adversarial text-to-image generation with natural prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20585-20594, 2023. 1, 2 + +[24] Ziyao Liu, Huanyi Ye, Chen Chen, Yongsen Zheng, and Kwok-Yan Lam. Threats, attacks, and defenses in machine unlearning: A survey. IEEE Open Journal of the Computer Society, 2025. 1 +[25] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 1 +[26] Jiachen Ma, Anda Cao, Zhiqing Xiao, Yijiang Li, Jie Zhang, Chao Ye, and Junbo Zhao. Jailbreaking prompt attack: A controllable adversarial attack against diffusion models. arXiv preprint arXiv:2404.02928, 2024. 1, 3, 5 +[27] Yizhuo Ma, Shanmin Pang, Qi Guo, Tianyu Wei, and Qing Guo. Coljailbreak: Collaborative generation and editing for jailbreaking text-to-image deep generation. Advances in Neural Information Processing Systems, 37:60335-60358, 2025. 3 +[28] Neil G Marchant, Benjamin IP Rubinstein, and Scott Alfeld. Hard to forget: Poisoning attacks on certified machine unlearning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7691-7700, 2022. 3 +[29] Yong-Hyun Park, Sangdoo Yun, Jin-Hwa Kim, Junho Kim, Geonhui Jang, Yonghyun Jeong, Junghyo Jo, and Gayoung Lee. Direct unlearning optimization for robust and safe text-to-image models. arXiv preprint arXiv:2407.21035, 2024. 1 +[30] Duo Peng, Qiuhong Ke, Mark He Huang, Ping Hu, and Jun Liu. Unified prompt attack against text-to-image generation models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1 +[31] Minh Pham, Kelly O. Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. In International Conference on Learning Representations, 2024. 1 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4 +[34] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 6 +[35] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In Proceedings of the ACM Conference on Fairness, Accountability, and Transparency, pages 1350-1361, 2022. 5, 6 +[36] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 2, 5 +[37] Thanveer Shaik, Xiaohui Tao, Haoran Xie, Lin Li, Xiaofeng Zhu, and Qing Li. Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy.(2023). arXiv preprint arXiv:2305.06360, 2023. 1, 2 +[38] Vu Tuan Truong, Luan Ba Dang, and Long Bao Le. Attacks and defenses for generative diffusion models: A comprehensive survey. ACM Computing Surveys, 2024. 1 +[39] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023. 1, 3, 5 +[40] Jiaqi Wang, Zhengliang Liu, Lin Zhao, Zihao Wu, Chong Ma, Sigang Yu, Haixing Dai, Qiushi Yang, Yiheng Liu, Songyao Zhang, et al. Review of large vision models and visual prompt engineering. Meta-Radiology, 1(3):100047, 2023. 8 +[41] Jie Xu, Zihan Wu, Cong Wang, and Xiaohua Jia. Machine unlearning: Solutions and challenges. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 1, 2 +[42] Hunmin Yang, Jongoh Jeong, and Kuk-Jin Yoon. Prompt-driven contrastive learning for transferable adversarial attacks. In European Conference on Computer Vision, pages 36-53. Springer, 2024. 8 +[43] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Tsung-Yi Ho, Nan Xu, and Qiang Xu. Mma-diffusion: Multimodal attack on diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7737-7746, 2024. 3 +[44] Yuchen Yang, Bo Hui, Haolin Yuan, Neil Gong, and Yinzhi Cao. Sneakyprompt: Jailbreaking text-to-image generative models. In IEEE Symposium on Security and Privacy, pages 897–912. IEEE, 2024. 1 +[45] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1755-1764, 2024. 1, 2, 5 +[46] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. Advances in neural information processing systems, 37:36748-36776, 2024. 2,5 +[47] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. In European Conference on Computer Vision, pages 385-403. Springer, 2024. 1, 3, 4, 5, 6 \ No newline at end of file diff --git a/ziumzeroshotintentawareadversarialattackonunlearnedmodels/images.zip b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eaff70eecb53cf62d8a2f15437e178925500b255 --- /dev/null +++ b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a1d007443bd7c1e4d03c46f1206fc556902407bbaa07d17f1a90b7152f52072 +size 529515 diff --git a/ziumzeroshotintentawareadversarialattackonunlearnedmodels/layout.json b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f85f3137cb01ec8bd4c27f5e4e60043b80d7aa12 --- /dev/null +++ b/ziumzeroshotintentawareadversarialattackonunlearnedmodels/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74605a863d3dfc5e99931feae6d7c41a2668521b87122bdec874b269be7f38c3 +size 315332