Daily Papers

Vision-language models (VLMs) have improved significantly in multi-modal tasks, but their more complex architecture makes their safety alignment more challenging than the alignment of large language models (LLMs). In this paper, we reveal an unfair distribution of safety across the layers of VLM's vision encoder, with earlier and middle layers being disproportionately vulnerable to malicious inputs compared to the more robust final layers. This 'cross-layer' vulnerability stems from the model's inability to generalize its safety training from the default architectural settings used during training to unseen or out-of-distribution scenarios, leaving certain layers exposed. We conduct a comprehensive analysis by projecting activations from various intermediate layers and demonstrate that these layers are more likely to generate harmful outputs when exposed to malicious inputs. Our experiments with LLaVA-1.5 and Llama 3.2 show discrepancies in attack success rates and toxicity scores across layers, indicating that current safety alignment strategies focused on a single default layer are insufficient.

Pre-trained vision encoders contain layer-wise visual representations that differ in spatial granularity, semantic abstraction, and sensitivity to local details. However, most Multimodal Large Language Models (MLLMs) rely on only the final or penultimate vision encoder representations or fixed aggregation rules, making visual abstraction largely query-agnostic and limiting access to fine-grained cues such as small objects, spatial details, text, and subtle visual attributes. In this work, we propose Mixture of Layers (MoL), an instruction-conditioned layer routing approach at the visual patch level that dynamically aggregates query-relevant latent representations from intermediate vision encoder layers. Given a text query, MoL predicts routing probabilities over vision encoder layers and performs a top-k sparse aggregation over selected hidden states at either the image level, patch level, or through a hybrid routing mechanism. In doing so, MoL enables query-adaptive access to layer-specific visual features for fine-grained visual reasoning. Our experiments across 7 fine-grained visual reasoning tasks demonstrate substantial performance improvements, especially across fine-grained visual grounding and understanding tasks such as +18.9% improvement on V* in overall accuracy, +4.5% on HRBench4K, and +16.3% on CharXiv compared to the baseline MLLMs, without resorting to multi-resolution inputs, simple interleaving of multiple vision encoders, or increasing the number of patch tokens. We study vision encoders' receptive field scales across different layers and their sampling behaviors to provide an in-depth analysis of why layer-wise sampling is helpful, demonstrating that conditional visual representations are a key step towards better visual perception and reasoning in MLLMs. Our project page is available at https://wjdghks950.github.io/mol.github.io/.

Vision-language models (VLMs) achieve strong performance on high-level multimodal tasks, yet numerosity perception, a cognitive ability that emerges in human infants before language acquisition, remains poorly understood in current models, as existing counting benchmarks entangle numerosity with correlated visual factors. We introduce a cognitively inspired diagnostic benchmark, NumerosityVLM, comprising 10,800 synthetic images across six controlled conditions. The benchmark orthogonally manipulates object size, spatial arrangement, and numerosity, while progressively ablating texture, shape, and color. Evaluating seven VLMs in a zero-shot setting, multi-factor analysis reveals that model architecture explains the largest proportion of performance variance (partial ω^{2}=0.325), far exceeding visual conditions. Layer-wise probing further shows that linearly separable numerosity signals consistently emerge at early stages of the vision encoder, while performance differences across evaluated models are primarily associated with the language model component. Code and data are publicly available at https://github.com/fuy3/NumerosityVLM-Benchmark, and https://huggingface.co/datasets/fuy3/NumerosityVLM.

Recent advances in Multi-modal Large Language Models (MLLMs) have shown significant progress in open-world Visual Question Answering (VQA). However, integrating visual information increases the number of processed tokens, leading to higher GPU memory usage and computational overhead. Images often contain more redundant information than text, and not all visual details are pertinent to specific questions. To address these challenges, we propose QG-VTC, a novel question-guided visual token compression method for MLLM-based VQA tasks. QG-VTC employs a pretrained text encoder and a learnable feed-forward layer to embed user questions into the vision encoder's feature space then computes correlation scores between the question embeddings and visual tokens. By selecting the most relevant tokens and softly compressing others, QG-VTC ensures fine-tuned relevance to user needs. Additionally, a progressive strategy applies this compression across different vision encoder layers, gradually reducing token numbers. This approach maximizes retention of question-relevant information while discarding irrelevant details. Experimental results show that our method achieves performance on par with uncompressed models using just 1/8 of the visual tokens. The code and model will be publicly available on GitHub.

Recent Multimodal Large Language Models (MLLMs) excel on benchmark vision-language tasks, yet little is known about how input visual quality shapes their responses. Does higher perceptual quality of images already translate to better MLLM understanding? We conduct the first systematic study spanning leading MLLMs and a suite of vision-language benchmarks, applying controlled degradations and stylistic shifts to each image. Surprisingly, we uncover a visual-quality paradox: model, task, and even individual-instance performance can improve when images deviate from human-perceived fidelity. Off-the-shelf restoration pipelines fail to reconcile these idiosyncratic preferences. To close the gap, we introduce Visual-Quality Test-Time Tuning (VQ-TTT)-a lightweight adaptation module that: (1) inserts a learnable, low-rank kernel before the frozen vision encoder to modulate frequency content; and (2) fine-tunes only shallow vision-encoder layers via LoRA. VQ-TTT dynamically adjusts each input image in a single forward pass, aligning it with task-specific model preferences. Across the evaluated MLLMs and all datasets, VQ-TTT lifts significant average accuracy, with no external models, cached features, or extra training data. These findings redefine ``better'' visual inputs for MLLMs and highlight the need for adaptive, rather than universally ``clean'', imagery, in the new era of AI being the main data customer.

Recently, Multimodal Large Language Models (MLLMs) have demonstrated impressive performance on instruction-following tasks by integrating pretrained visual encoders with large language models (LLMs). However, existing approaches often struggle to ground fine-grained visual concepts in complex scenes. In this paper, we propose MoDA (Modulation Adapter), a lightweight yet effective module designed to refine pre-aligned visual features through instruction-guided modulation. Our approach follows the standard LLaVA training protocol, consisting of a two-stage process: (1) aligning image features to the LLMs input space via a frozen vision encoder and adapter layers, and (2) refining those features using the MoDA adapter during the instructional tuning stage. MoDA employs a Transformer-based cross-attention mechanism to generate a modulation mask over the aligned visual tokens, thereby emphasizing semantically relevant embedding dimensions based on the language instruction. The modulated features are then passed to the LLM for autoregressive language generation. Our experimental evaluation shows that MoDA improves visual grounding and generates more contextually appropriate responses, demonstrating its effectiveness as a general-purpose enhancement for image-based MLLMs.

Vision-language models face a fundamental scaling bottleneck: the number of visual tokens grows with both temporal duration and spatial resolution, making long-video understanding expensive for the vision encoder and the language model. Existing methods often compress visual tokens after dense encoding, creating a mismatch between the representation used during training and the compact interface required at deployment. We present CoVisco, a codec-native vision encoder with native token compression for unified image-video understanding. By combining codec-native input support with segmented attention, CoVisco can encode long visual inputs in a single forward pass without forming dense patch-to-patch interactions across all frames. Each temporal segment is equipped with learnable abstract tokens that learn a compact segment-level representation, while fine-grained patch tokens remain available throughout the encoder. Alternating intra-segment and abstract-communication layers preserve video-level context through the abstract-token channel. A lightweight selector further exposes either abstract tokens alone or abstract tokens augmented with a runtime-selected subset of patch tokens, yielding a compact visual interface that reduces the visual context and prefill burden of downstream MLLMs while retaining fine-grained evidence when needed. Pretrained with contrastive objectives on 565M image--text pairs and 6.4M videos, CoVisco shows competitive performance on video-oriented embedding and multimodal understanding benchmarks. In the evaluated four-segment, 64-frame setting, abstract-only inference uses only 400 visual tokens while achieving video-understanding performance close to, and on some benchmarks exceeding, OneVision-Encoder. Selected patch tokens further improve fine-grained video reasoning. Project URL: https://github.com/ernie-research/CoVisco.git

This paper reveals that large language models (LLMs), despite being trained solely on textual data, are surprisingly strong encoders for purely visual tasks in the absence of language. Even more intriguingly, this can be achieved by a simple yet previously overlooked strategy -- employing a frozen transformer block from pre-trained LLMs as a constituent encoder layer to directly process visual tokens. Our work pushes the boundaries of leveraging LLMs for computer vision tasks, significantly departing from conventional practices that typically necessitate a multi-modal vision-language setup with associated language prompts, inputs, or outputs. We demonstrate that our approach consistently enhances performance across a diverse range of tasks, encompassing pure 2D and 3D visual recognition tasks (e.g., image and point cloud classification), temporal modeling tasks (e.g., action recognition), non-semantic tasks (e.g., motion forecasting), and multi-modal tasks (e.g., 2D/3D visual question answering and image-text retrieval). Such improvements are a general phenomenon, applicable to various types of LLMs (e.g., LLaMA and OPT) and different LLM transformer blocks. We additionally propose the information filtering hypothesis to explain the effectiveness of pre-trained LLMs in visual encoding -- the pre-trained LLM transformer blocks discern informative visual tokens and further amplify their effect. This hypothesis is empirically supported by the observation that the feature activation, after training with LLM transformer blocks, exhibits a stronger focus on relevant regions. We hope that our work inspires new perspectives on utilizing LLMs and deepening our understanding of their underlying mechanisms. Code is available at https://github.com/ziqipang/LM4VisualEncoding.

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 -- 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 ~45% when using Llama-70B as the decoder.

High-resolution Vision-Language Models (VLMs) have been widely used in multimodal tasks to enhance accuracy by preserving detailed image information. However, these models often generate excessive visual tokens due to encoding multiple partitions of the input image. Processing these excessive visual tokens is computationally challenging, especially in resource-constrained environments with commodity GPUs. To support high-resolution images while meeting resource constraints, we propose High-Resolution Early Dropping (HiRED), a token-dropping scheme that operates within a fixed token budget before the Large Language Model (LLM) stage. HiRED can be integrated with existing high-resolution VLMs in a plug-and-play manner, as it requires no additional training while still maintaining superior accuracy. We strategically use the vision encoder's attention in the initial layers to assess the visual content of each image partition and allocate the token budget accordingly. Then, using the attention in the final layer, we select the most important visual tokens from each partition within the allocated budget, dropping the rest. Empirically, when applied to LLaVA-Next-7B on NVIDIA TESLA P40 GPU, HiRED with a 20% token budget increases token generation throughput by 4.7, reduces first-token generation latency by 15 seconds, and saves 2.3 GB of GPU memory for a single inference.

Recent advancements in multi-modal models have significantly improved vision-language alignment in radiology. However, existing approaches struggle to effectively utilize complex radiology reports for learning, rely on low-resolution images, and offer limited interpretability in attention mechanisms. To address these challenges, we introduce RadZero, a novel similarity-based cross-attention framework for vision-language alignment in radiology with zero-shot multi-task capability. RadZero leverages large language models to extract minimal semantic sentences from radiology reports and employs a multi-positive contrastive learning strategy to effectively capture relationships between images and multiple relevant textual descriptions. It also utilizes a pre-trained vision encoder with additional trainable Transformer layers, allowing efficient high-resolution image processing. By computing similarity between text embeddings and local image patch features, RadZero enables zero-shot inference with similarity probability for classification and pixel-level cross-modal similarity maps for grounding and segmentation. Experimental results on public chest radiograph benchmarks show that RadZero outperforms state-of-the-art methods in zero-shot classification, grounding, and segmentation. Furthermore, cross-modal similarity map analysis highlights its potential for improving explainability in vision-language alignment. Additionally, qualitative evaluation demonstrates RadZero's capability for open-vocabulary semantic segmentation, further validating its effectiveness in medical imaging.

Unified multimodal models typically rely on pretrained vision encoders and use separate visual representations for understanding and generation, creating misalignment between the two tasks and preventing fully end-to-end optimization from raw pixels. We introduce Tuna-2, a native unified multimodal model that performs visual understanding and generation directly based on pixel embeddings. Tuna-2 drastically simplifies the model architecture by employing simple patch embedding layers to encode visual input, completely discarding the modular vision encoder designs such as the VAE or the representation encoder. Experiments show that Tuna-2 achieves state-of-the-art performance in multimodal benchmarks, demonstrating that unified pixel-space modelling can fully compete with latent-space approaches for high-quality image generation. Moreover, while the encoder-based variant converges faster in early pretraining, Tuna-2's encoder-free design achieves stronger multimodal understanding at scale, particularly on tasks requiring fine-grained visual perception. These results show that pretrained vision encoders are not necessary for multimodal modelling, and end-to-end pixel-space learning offers a scalable path toward stronger visual representations for both generation and perception.

Ternary Vision Transformers offer substantial model compression, however state-of-the-art methods only ternarize the encoder layers, leaving patch embeddings, LayerNorm parameters, and classifier heads in full precision. In compact models targeting resource-constrained processors, such as microcontrollers, these remaining full-precision components determine the total memory footprint, severely limiting deployment efficiency and on-device feasibility. In this work, we introduce a fully ternarized Vision Transformer in which all weight matrices and normalization parameters are ternarized (FTerViT). To this end, we introduce two novel operators : TernaryBitConv2d with per-channel scaling for patch embedding and TernaryLayerNorm. FTerViT is trained using knowledge distillation, followed by a lightweight quantization-aware recovery phase. Our ternary W2A8 DeiT-III-S at 384times384 resolution achieves 82.43\% ImageNet-1K top-1 at 6.09\,MB ({sim}15times compression, -2.42\,pp vs.\ FP32), outperforming prior ternary ViTs methods up to 8 pp. Finally, we demonstrate the first implementation of ternary vision transformers on a dual cores XTensa LX7 microcontroller inside the ESP32-S3 system-on-chip. By deploying FTerViT-Small (based on DeiT-III-Small at 224times224 resolution, 5.81\,MB), we achieve 79.64\% ImageNet-1K top-1 accuracy.

Large-scale pre-trained Vision-Language Models (VLMs) have significantly advanced transfer learning across diverse tasks. However, adapting these models with limited few-shot data often leads to overfitting, undermining their ability to generalize to new tasks. To address this, we propose Multi-Modal Representation Learning (MMRL), which introduces a shared, learnable, modality-agnostic representation space. MMRL generates space tokens projected into both text and image encoders as representation tokens, enabling more effective cross-modal interactions. Unlike prior methods that mainly optimize class token features, MMRL inserts representation tokens into higher encoder layers--where task-specific features are more prominent--while preserving general knowledge in the lower layers. During training, both class and representation features are jointly optimized: a trainable projection layer is applied to representation tokens for task adaptation, while the projection layer for class token remains frozen to retain pre-trained knowledge. To further promote generalization, we introduce a regularization term aligning class and text features with the frozen VLM's zero-shot features. At inference, a decoupling strategy uses both class and representation features for base tasks, but only class features for novel tasks due to their stronger generalization. Building upon this, we propose MMRL++, a parameter-efficient and interaction-aware extension that significantly reduces trainable parameters and enhances intra-modal interactions--particularly across the layers of representation tokens--allowing gradient sharing and instance-specific information to propagate more effectively through the network. Extensive experiments on 15 datasets demonstrate that MMRL and MMRL++ consistently outperform state-of-the-art methods, achieving a strong balance between task-specific adaptation and generalization.

In this work, we explore neat yet effective Transformer-based frameworks for visual grounding. The previous methods generally address the core problem of visual grounding, i.e., multi-modal fusion and reasoning, with manually-designed mechanisms. Such heuristic designs are not only complicated but also make models easily overfit specific data distributions. To avoid this, we first propose TransVG, which establishes multi-modal correspondences by Transformers and localizes referred regions by directly regressing box coordinates. We empirically show that complicated fusion modules can be replaced by a simple stack of Transformer encoder layers with higher performance. However, the core fusion Transformer in TransVG is stand-alone against uni-modal encoders, and thus should be trained from scratch on limited visual grounding data, which makes it hard to be optimized and leads to sub-optimal performance. To this end, we further introduce TransVG++ to make two-fold improvements. For one thing, we upgrade our framework to a purely Transformer-based one by leveraging Vision Transformer (ViT) for vision feature encoding. For another, we devise Language Conditioned Vision Transformer that removes external fusion modules and reuses the uni-modal ViT for vision-language fusion at the intermediate layers. We conduct extensive experiments on five prevalent datasets, and report a series of state-of-the-art records.

The representation of visual inputs of large vision-language models (LVLMs) usually involves substantially more tokens than that of textual inputs, leading to significant computational overhead. Several recent studies strive to mitigate this issue by either conducting token compression to prune redundant visual tokens or guiding them to bypass certain computational stages. While most existing work exploits token importance as the redundancy indicator, our study reveals that two largely neglected factors, namely, the diversity of retained visual tokens and their task relevance, often offer more robust criteria in token pruning. To this end, we design ToDRE, a two-stage and training-free token compression framework that achieves superior performance by pruning Tokens based on token Diversity and token-task RElevance. Instead of pruning redundant tokens, ToDRE introduces a greedy k-center algorithm to select and retain a small subset of diverse visual tokens after the vision encoder. Additionally, ToDRE addresses the "information migration" by further eliminating task-irrelevant visual tokens within the decoder of large language model (LLM). Extensive experiments show that ToDRE effectively reduces 90% of visual tokens after vision encoder and adaptively prunes all visual tokens within certain LLM's decoder layers, leading to a 2.6x speed-up in total inference time while maintaining 95.1% of model performance and excellent compatibility with efficient attention operators.

Vision-language models (VLMs) excel at extracting and reasoning about information from images. Yet, their capacity to leverage internal knowledge about specific entities remains underexplored. This work investigates the disparity in model performance when answering factual questions about an entity described in text versus depicted in an image. Our results reveal a significant accuracy drop - reaching 18% for some models - when the entity is presented visually instead of textually. To study this gap we present PopVQA, a dataset which allows separating entity recognition and question answering, and use it to benchmark several models. We hypothesize that this decline arises from limitations in how information flows from image tokens to query tokens. Thus, we use mechanistic interpretability tools to reveal that, although image tokens are preprocessed by the vision encoder, meaningful information flow from these tokens occurs only in the much deeper layers. Furthermore, critical image processing happens in the language model's middle layers, allowing few layers for consecutive reasoning, highlighting a potential inefficiency in how the model utilizes its layers for reasoning. These insights shed light on the internal mechanics of VLMs and offer pathways for enhancing their reasoning capabilities. PopVQA can be found at https://huggingface.co/datasets/idoco/PopVQA.

Recently, large language and vision models (LLVMs) have received significant attention and development efforts due to their remarkable generalization performance across a wide range of tasks requiring perception and cognitive abilities. A key factor behind their success is their simple architecture, which consists of a vision encoder, a projector, and a large language model (LLM). Despite their achievements in advanced reasoning tasks, their performance on fundamental perception-related tasks (e.g., MMVP) remains surprisingly low. This discrepancy raises the question of how LLVMs truly perceive images and exploit the advantages of the vision encoder. To address this, we systematically investigate this question regarding several aspects: permutation invariance, robustness, math reasoning, alignment preserving and importance, by evaluating the most common LLVM's families (i.e., LLaVA) across 10 evaluation benchmarks. Our extensive experiments reveal several intriguing properties of current LLVMs: (1) they internally process the image in a global manner, even when the order of visual patch sequences is randomly permuted; (2) they are sometimes able to solve math problems without fully perceiving detailed numerical information; (3) the cross-modal alignment is overfitted to complex reasoning tasks, thereby, causing them to lose some of the original perceptual capabilities of their vision encoder; (4) the representation space in the lower layers (<25%) plays a crucial role in determining performance and enhancing visual understanding. Lastly, based on the above observations, we suggest potential future directions for building better LLVMs and constructing more challenging evaluation benchmarks.

Attention sinks are defined as tokens that attract disproportionate attention. While these have been studied in single modality transformers, their cross-modal impact in Large Vision-Language Models (LVLM) remains largely unexplored: are they redundant artifacts or essential global priors? This paper first categorizes visual sinks into two distinct categories: ViT-emerged sinks (V-sinks), which propagate from the vision encoder, and LLM-emerged sinks (L-sinks), which arise within deep LLM layers. Based on the new definition, our analysis reveals a fundamental performance trade-off: while sinks effectively encode global scene-level priors, their dominance can suppress the fine-grained visual evidence required for local perception. Furthermore, we identify specific functional layers where modulating these sinks most significantly impacts downstream performance. To leverage these insights, we propose Layer-wise Sink Gating (LSG), a lightweight, plug-and-play module that dynamically scales the attention contributions of V-sink and the rest visual tokens. LSG is trained via standard next-token prediction, requiring no task-specific supervision while keeping the LVLM backbone frozen. In most layers, LSG yields improvements on representative multimodal benchmarks, effectively balancing global reasoning and precise local evidence.

Recently, the advances in vision-language models, including contrastive pretraining and instruction tuning, have greatly pushed the frontier of multimodal AI. However, owing to the large-scale and hence expensive pretraining, the efficiency concern has discouraged researchers from attempting to pretrain a vision language model from scratch. In this work, we propose Dynamic patch Reduction via Interpretable Pooling (DRIP), which adapts to the input images and dynamically merges tokens in the deeper layers of a visual encoder. Our results on both ImageNet training from scratch and CLIP contrastive pretraining demonstrate a significant GFLOP reduction while maintaining comparable classification/zero-shot performance. To further validate our proposed method, we conduct continual pretraining on a large biology dataset, extending its impact into scientific domains.

Vision transformer based models bring significant improvements for image segmentation tasks. Although these architectures offer powerful capabilities irrespective of specific segmentation tasks, their use of computational resources can be taxing on deployed devices. One way to overcome this challenge is by adapting the computation level to the specific needs of the input image rather than the current one-size-fits-all approach. To this end, we introduce ECO-M2F or EffiCient TransfOrmer Encoders for Mask2Former-style models. Noting that the encoder module of M2F-style models incur high resource-intensive computations, ECO-M2F provides a strategy to self-select the number of hidden layers in the encoder, conditioned on the input image. To enable this self-selection ability for providing a balance between performance and computational efficiency, we present a three step recipe. The first step is to train the parent architecture to enable early exiting from the encoder. The second step is to create an derived dataset of the ideal number of encoder layers required for each training example. The third step is to use the aforementioned derived dataset to train a gating network that predicts the number of encoder layers to be used, conditioned on the input image. Additionally, to change the computational-accuracy tradeoff, only steps two and three need to be repeated which significantly reduces retraining time. Experiments on the public datasets show that the proposed approach reduces expected encoder computational cost while maintaining performance, adapts to various user compute resources, is flexible in architecture configurations, and can be extended beyond the segmentation task to object detection.

Vision-Language Models (VLMs) have shown remarkable performance in integrating visual and textual information for tasks such as image captioning and visual question answering. However, these models struggle with compositional generalization and object binding, which limit their ability to handle novel combinations of objects and their attributes. Our work explores the root causes of these failures using mechanistic interpretability techniques. We show evidence that individual neurons in the MLP layers of CLIP's vision encoder represent multiple features, and this "superposition" directly hinders its compositional feature representation which consequently affects compositional reasoning and object binding capabilities. We hope this study will serve as an initial step toward uncovering the mechanistic roots of compositional failures in VLMs. The code and supporting results can be found https://github.com/Mystic-Slice/Do-VLMs-Have-Bad-Eyes .

To effectively reduce the visual tokens in Visual Large Language Models (VLLMs), we propose a novel approach called Window Token Concatenation (WiCo). Specifically, we employ a sliding window to concatenate spatially adjacent visual tokens. However, directly concatenating these tokens may group diverse tokens into one, and thus obscure some fine details. To address this challenge, we propose fine-tuning the last few layers of the vision encoder to adaptively adjust the visual tokens, encouraging that those within the same window exhibit similar features. To further enhance the performance on fine-grained visual understanding tasks, we introduce WiCo+, which decomposes the visual tokens in later layers of the LLM. Such a design enjoys the merits of the large perception field of the LLM for fine-grained visual understanding while keeping a small number of visual tokens for efficient inference. We perform extensive experiments on both coarse- and fine-grained visual understanding tasks based on LLaVA-1.5 and Shikra, showing better performance compared with existing token reduction projectors. The code is available: https://github.com/JackYFL/WiCo.

Vision-language models, such as contrastive language-image pre-training (CLIP)-based approaches, have reached state-of-the-art (SOTA) results in medical artificial intelligence. However, recent work reveals that CLIP-based models remain vulnerable to shortcuts. We investigate how real-world shortcuts manifest across different layers of the medical CLIP-based model, MedCLIP, and its vision encoder, a frozen ResNet-50. We attach 17 linear classification probes to the intermediate layers of the ResNet-50 and train them on three different dataset configurations and targets: NIH-CXR14 (pneumothorax) and PadChest (cardiomegaly and pneumothorax). This setup allows us to observe model behaviour during evaluation using subgroup-based calibration and layer-wise confidence curves. We find that the final linear probes achieve a high AUROC but poor calibration in the models. The layer-wise confidence analyses suggest that shortcuts emerge at different depths. Patterns consistent with localised shortcuts, such as drains, appear at later layers, while patterns consistent with diffuse shortcuts, such as scanner-specific noise patterns, emerge earlier, aligning with previous work. Finally, we conduct a manual analysis of the images, which reveals data quality issues in both NIH-CXR14 and PadChest. Our findings underscore that even SOTA models remain vulnerable to shortcuts, and the need for high-quality and well-annotated datasets to draw solid conclusions. Code can be found on our GitHub: https://github.com/nikodice4/MedCLIP_shortcuts.

Existing visual token compression methods for Multimodal Large Language Models (MLLMs) predominantly operate as post-encoder modules, limiting their potential for efficiency gains. To address this limitation, we propose LaCo (Layer-wise Visual Token Compression), a novel framework that enables effective token compression within the intermediate layers of the vision encoder. LaCo introduces two core components: 1) a layer-wise pixel-shuffle mechanism that systematically merges adjacent tokens through space-to-channel transformations, and 2) a residual learning architecture with non-parametric shortcuts that preserves critical visual information during compression. Extensive experiments indicate that our LaCo outperforms all existing methods when compressing tokens in the intermediate layers of the vision encoder, demonstrating superior effectiveness. In addition, compared to external compression, our method improves training efficiency beyond 20% and inference throughput over 15% while maintaining strong performance.

Referring image segmentation is a fundamental vision-language task that aims to segment out an object referred to by a natural language expression from an image. One of the key challenges behind this task is leveraging the referring expression for highlighting relevant positions in the image. A paradigm for tackling this problem is to leverage a powerful vision-language ("cross-modal") decoder to fuse features independently extracted from a vision encoder and a language encoder. Recent methods have made remarkable advancements in this paradigm by exploiting Transformers as cross-modal decoders, concurrent to the Transformer's overwhelming success in many other vision-language tasks. Adopting a different approach in this work, we show that significantly better cross-modal alignments can be achieved through the early fusion of linguistic and visual features in intermediate layers of a vision Transformer encoder network. By conducting cross-modal feature fusion in the visual feature encoding stage, we can leverage the well-proven correlation modeling power of a Transformer encoder for excavating helpful multi-modal context. This way, accurate segmentation results are readily harvested with a light-weight mask predictor. Without bells and whistles, our method surpasses the previous state-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref by large margins.

Existing encoder-free vision-language models (VLMs) are rapidly narrowing the performance gap with their encoder-based counterparts, highlighting the promising potential for unified multimodal systems with structural simplicity and efficient deployment. We systematically clarify the performance gap between VLMs using pre-trained vision encoders, discrete tokenizers, and minimalist visual layers from scratch, deeply excavating the under-examined characteristics of encoder-free VLMs. We develop efficient strategies for encoder-free VLMs that rival mainstream encoder-based ones. After an in-depth investigation, we launch EVEv2.0, a new and improved family of encoder-free VLMs. We show that: (i) Properly decomposing and hierarchically associating vision and language within a unified model reduces interference between modalities. (ii) A well-designed training strategy enables effective optimization for encoder-free VLMs. Through extensive evaluation, our EVEv2.0 represents a thorough study for developing a decoder-only architecture across modalities, demonstrating superior data efficiency and strong vision-reasoning capability. Code is publicly available at: https://github.com/baaivision/EVE.

Large vision-language models (VLMs) are increasingly deployed for optical character recognition (OCR) in healthcare settings, raising critical concerns about protected health information (PHI) exposure during document processing. This work presents the first systematic evaluation of inference-time vision token masking as a privacy-preserving mechanism for medical document OCR using DeepSeek-OCR. We introduce seven masking strategies (V3-V9) targeting different architectural layers (SAM encoder blocks, compression layers, dual vision encoders, projector fusion) and evaluate PHI reduction across HIPAA-defined categories using 100 synthetic medical billing statements (drawn from a corpus of 38,517 annotated documents) with perfect ground-truth annotations. All masking strategies converge to 42.9% PHI reduction, successfully suppressing long-form spatially-distributed identifiers (patient names, dates of birth, physical addresses at 100% effectiveness) while failing to prevent short structured identifiers (medical record numbers, social security numbers, email addresses, account numbers at 0% effectiveness). Ablation studies varying mask expansion radius (r=1,2,3) demonstrate that increased spatial coverage does not improve reduction beyond this ceiling, indicating that language model contextual inference - not insufficient visual masking - drives structured identifier leakage. A simulated hybrid architecture combining vision masking with NLP post-processing achieves 88.6% total PHI reduction (assuming 80% NLP accuracy on remaining identifiers). This negative result establishes boundaries for vision-only privacy interventions in VLMs, provides guidance distinguishing PHI types amenable to vision-level versus language-level redaction, and redirects future research toward decoder-level fine-tuning and hybrid defense-in-depth architectures for HIPAA-compliant medical document processing.

Large-scale Transformer models bring significant improvements for various downstream vision language tasks with a unified architecture. The performance improvements come with increasing model size, resulting in slow inference speed and increased cost for severing. While some certain predictions benefit from the full complexity of the large-scale model, not all of inputs need the same amount of computation to conduct, potentially leading to computation resource waste. To handle this challenge, early exiting is proposed to adaptively allocate computational power in term of input complexity to improve inference efficiency. The existing early exiting strategies usually adopt output confidence based on intermediate layers as a proxy of input complexity to incur the decision of skipping following layers. However, such strategies cannot apply to encoder in the widely-used unified architecture with both encoder and decoder due to difficulty of output confidence estimation in the encoder. It is suboptimal in term of saving computation power to ignore the early exiting in encoder component. To handle this challenge, we propose a novel early exiting strategy for unified visual language models, which allows dynamically skip the layers in encoder and decoder simultaneously in term of input layer-wise similarities with multiple times of early exiting, namely MuE. By decomposing the image and text modalities in the encoder, MuE is flexible and can skip different layers in term of modalities, advancing the inference efficiency while minimizing performance drop. Experiments on the SNLI-VE and MS COCO datasets show that the proposed approach MuE can reduce expected inference time by up to 50\% and 40\% while maintaining 99\% and 96\% performance respectively.

Transformers combined with convolutional encoders have been recently used for hand gesture recognition (HGR) using micro-Doppler signatures. We propose a vision-transformer-based architecture for HGR with multi-antenna continuous-wave Doppler radar receivers. The proposed architecture consists of three modules: a convolutional encoderdecoder, an attention module with three transformer layers, and a multi-layer perceptron. The novel convolutional decoder helps to feed patches with larger sizes to the attention module for improved feature extraction. Experimental results obtained with a dataset corresponding to a two-antenna continuous-wave Doppler radar receiver operating at 24 GHz (published by Skaria et al.) confirm that the proposed architecture achieves an accuracy of 98.3% which substantially surpasses the state-of-the-art on the used dataset.

Medical image segmentation remains a formidable challenge due to the label scarcity. Pre-training Vision Transformer (ViT) through masked image modeling (MIM) on large-scale unlabeled medical datasets presents a promising solution, providing both computational efficiency and model generalization for various downstream tasks. However, current ViT-based MIM pre-training frameworks predominantly emphasize local aggregation representations in output layers and fail to exploit the rich representations across different ViT layers that better capture fine-grained semantic information needed for more precise medical downstream tasks. To fill the above gap, we hereby present Hierarchical Encoder-driven MAE (Hi-End-MAE), a simple yet effective ViT-based pre-training solution, which centers on two key innovations: (1) Encoder-driven reconstruction, which encourages the encoder to learn more informative features to guide the reconstruction of masked patches; and (2) Hierarchical dense decoding, which implements a hierarchical decoding structure to capture rich representations across different layers. We pre-train Hi-End-MAE on a large-scale dataset of 10K CT scans and evaluated its performance across seven public medical image segmentation benchmarks. Extensive experiments demonstrate that Hi-End-MAE achieves superior transfer learning capabilities across various downstream tasks, revealing the potential of ViT in medical imaging applications. The code is available at: https://github.com/FengheTan9/Hi-End-MAE

Vision-Language (VL) models with the Two-Tower architecture have dominated visual-language representation learning in recent years. Current VL models either use lightweight uni-modal encoders and learn to extract, align and fuse both modalities simultaneously in a deep cross-modal encoder, or feed the last-layer uni-modal representations from the deep pre-trained uni-modal encoders into the top cross-modal encoder. Both approaches potentially restrict vision-language representation learning and limit model performance. In this paper, we propose BridgeTower, which introduces multiple bridge layers that build a connection between the top layers of uni-modal encoders and each layer of the cross-modal encoder. This enables effective bottom-up cross-modal alignment and fusion between visual and textual representations of different semantic levels of pre-trained uni-modal encoders in the cross-modal encoder. Pre-trained with only 4M images, BridgeTower achieves state-of-the-art performance on various downstream vision-language tasks. In particular, on the VQAv2 test-std set, BridgeTower achieves an accuracy of 78.73%, outperforming the previous state-of-the-art model METER by 1.09% with the same pre-training data and almost negligible additional parameters and computational costs. Notably, when further scaling the model, BridgeTower achieves an accuracy of 81.15%, surpassing models that are pre-trained on orders-of-magnitude larger datasets. Code and checkpoints are available at https://github.com/microsoft/BridgeTower.

The Vision Transformer architecture has shown to be competitive in the computer vision (CV) space where it has dethroned convolution-based networks in several benchmarks. Nevertheless, convolutional neural networks (CNN) remain the preferential architecture for the representation module in reinforcement learning. In this work, we study pretraining a Vision Transformer using several state-of-the-art self-supervised methods and assess the quality of the learned representations. To show the importance of the temporal dimension in this context we propose an extension of VICReg to better capture temporal relations between observations by adding a temporal order verification task. Our results show that all methods are effective in learning useful representations and avoiding representational collapse for observations from Atari Learning Environment (ALE) which leads to improvements in data efficiency when we evaluated in reinforcement learning (RL). Moreover, the encoder pretrained with the temporal order verification task shows the best results across all experiments, with richer representations, more focused attention maps and sparser representation vectors throughout the layers of the encoder, which shows the importance of exploring such similarity dimension. With this work, we hope to provide some insights into the representations learned by ViT during a self-supervised pretraining with observations from RL environments and which properties arise in the representations that lead to the best-performing agents. The source code will be available at: https://github.com/mgoulao/TOV-VICReg

In this paper, we present a novel methodology we call MDS-ViTNet (Multi Decoder Saliency by Vision Transformer Network) for enhancing visual saliency prediction or eye-tracking. This approach holds significant potential for diverse fields, including marketing, medicine, robotics, and retail. We propose a network architecture that leverages the Vision Transformer, moving beyond the conventional ImageNet backbone. The framework adopts an encoder-decoder structure, with the encoder utilizing a Swin transformer to efficiently embed most important features. This process involves a Transfer Learning method, wherein layers from the Vision Transformer are converted by the Encoder Transformer and seamlessly integrated into a CNN Decoder. This methodology ensures minimal information loss from the original input image. The decoder employs a multi-decoding technique, utilizing dual decoders to generate two distinct attention maps. These maps are subsequently combined into a singular output via an additional CNN model. Our trained model MDS-ViTNet achieves state-of-the-art results across several benchmarks. Committed to fostering further collaboration, we intend to make our code, models, and datasets accessible to the public.

Leveraging the powerful capabilities of diffusion models has yielded quite effective results in medical image segmentation tasks. However, existing methods typically transfer the original training process directly without specific adjustments for segmentation tasks. Furthermore, the commonly used pre-trained diffusion models still have deficiencies in feature extraction. Based on these considerations, we propose LEAF, a medical image segmentation model grounded in latent diffusion models. During the fine-tuning process, we replace the original noise prediction pattern with a direct prediction of the segmentation map, thereby reducing the variance of segmentation results. We also employ a feature distillation method to align the hidden states of the convolutional layers with the features from a transformer-based vision encoder. Experimental results demonstrate that our method enhances the performance of the original diffusion model across multiple segmentation datasets for different disease types. Notably, our approach does not alter the model architecture, nor does it increase the number of parameters or computation during the inference phase, making it highly efficient.

A targeted adversarial perturbation can drive a vision-language model's (VLM's) teacher-forced training loss for a fixed target caption to near zero, yet the same model, allowed to generate freely, produces the original, correct description with no trace of the target. We call this dissociation the train/inference gap, and give it a precise mechanistic account on Qwen2.5-VL-7B-Instruct using a controlled two-stage PGD attack on 200 held-out COCO images. First, we show that image-level pixel statistics, including a correctly re-implemented, texture-based attackability measure from the CNN robustness literature, have essentially no predictive power over which images are corrupted (best predictor r=-0.050, p=0.484; ridge regression R^2=0.069). Second, using the logit lens, we localise the gap to a single autoregressive step: the rank of the target token, conditioned on the correct first token already being generated, is fixed at exactly 3,488 out of 152,064 vocabulary entries for every image and every condition, with zero variance. Third, tracking target-token rank across all 28 LLM decoder layers reveals that the visual encoder corrupts every image's representation by a comparable margin regardless of eventual outcome, but the language model decoder then differentially arbitrates: amplifying the corrupted signal for susceptible images and actively suppressing it, past its clean-image baseline, for resistant ones (p<0.001, rank-biserial r=0.579). A linear probe on the merger hidden state separates these two outcomes with AUC=0.858, though we flag a circularity concern in this estimate. Together these results argue that adversarial robustness in autoregressive VLMs is substantially a property of the language decoder's prior, not the visual encoder, with direct implications for where faithfulness evaluations and defenses for deployed VLM systems should be targeted.

Most modern multimodal large language models (MLLMs) build on a pretrained visual encoder that provides a strong visual prior. Encoder-free MLLMs instead learn visual representations directly from raw pixels, offering a simple and unified architecture, but their scaling behavior has not been systematically characterized. To fill this gap, we compare scaling laws for encoder-free and encoder-based MLLMs and report three main findings: (1) Removing the visual encoder shifts the compute-optimal allocation for the multimodal objective toward larger models, while leaving that for text nearly unchanged. (2) The two architectures exhibit nearly overlapping loss--compute frontiers on the text objective, but diverge on the multimodal objective: encoder-free models underperform at small scales yet are predicted to catch up at around 10^{22} FLOPs, well within practical pretraining budgets. (3) Without a visual encoder, the language model learns to take over its role via vision-specific adaptation: bidirectional interactions among visual tokens become increasingly beneficial as training compute grows, visual processing shifts toward earlier layers, and expert routing for visual tokens becomes more concentrated. Overall, our results indicate that the advantage of the visual prior provided by a pretrained encoder diminishes with scale, positioning encoder-free architectures as a promising direction for multimodal pretraining.

In the realm of medical image segmentation, both CNN-based and Transformer-based models have been extensively explored. However, CNNs exhibit limitations in long-range modeling capabilities, whereas Transformers are hampered by their quadratic computational complexity. Recently, State Space Models (SSMs), exemplified by Mamba, have emerged as a promising approach. They not only excel in modeling long-range interactions but also maintain a linear computational complexity. In this paper, leveraging state space models, we propose a U-shape architecture model for medical image segmentation, named Vision Mamba UNet (VM-UNet). Specifically, the Visual State Space (VSS) block is introduced as the foundation block to capture extensive contextual information, and an asymmetrical encoder-decoder structure is constructed with fewer convolution layers to save calculation cost. We conduct comprehensive experiments on the ISIC17, ISIC18, and Synapse datasets, and the results indicate that VM-UNet performs competitively in medical image segmentation tasks. To our best knowledge, this is the first medical image segmentation model constructed based on the pure SSM-based model. We aim to establish a baseline and provide valuable insights for the future development of more efficient and effective SSM-based segmentation systems. Our code is available at https://github.com/JCruan519/VM-UNet.

Recent Large Vision-Language Models (LVLMs) have advanced multi-modal understanding by incorporating finer-grained visual perception and encoding. However, such methods incur significant computational costs due to longer visual token sequences, posing challenges for real-time deployment. To mitigate this, prior studies have explored pruning unimportant visual tokens either at the output layer of the visual encoder or at the early layers of the language model. In this work, we revisit these design choices and reassess their effectiveness through comprehensive empirical studies of how visual tokens are processed throughout the visual encoding and language decoding stages. Guided by these insights, we propose VScan, a two-stage visual token reduction framework that addresses token redundancy by: (1) integrating complementary global and local scans with token merging during visual encoding, and (2) introducing pruning at intermediate layers of the language model. Extensive experimental results across four LVLMs validate the effectiveness of VScan in accelerating inference and demonstrate its superior performance over current state-of-the-arts on sixteen benchmarks. Notably, when applied to LLaVA-NeXT-7B, VScan achieves a 2.91times speedup in prefilling and a 10times reduction in FLOPs, while retaining 95.4% of the original performance.

Pretrained Vision Transformers (ViTs) such as DINOv2 and MAE provide generic image features that can be applied to a variety of downstream tasks such as retrieval, classification, and segmentation. However, such representations tend to focus on the most salient visual cues in the image, with no way to direct them toward less prominent concepts of interest. In contrast, Multimodal LLMs can be guided with textual prompts, but the resulting representations tend to be language-centric and lose their effectiveness for generic visual tasks. To address this, we introduce Steerable Visual Representations, a new class of visual representations, whose global and local features can be steered with natural language. While most vision-language models (e.g., CLIP) fuse text with visual features after encoding (late fusion), we inject text directly into the layers of the visual encoder (early fusion) via lightweight cross-attention. We introduce benchmarks for measuring representational steerability, and demonstrate that our steerable visual features can focus on any desired objects in an image while preserving the underlying representation quality. Our method also matches or outperforms dedicated approaches on anomaly detection and personalized object discrimination, exhibiting zero-shot generalization to out-of-distribution tasks.

Cross-lingual, cross-task transfer is challenged by task-specific data scarcity, which becomes more severe as language support grows and is further amplified in vision-language models (VLMs). We investigate multilingual generalization in encoder-decoder transformer VLMs to enable zero-shot image captioning in languages encountered only in the translation task. In this setting, the encoder must learn to generate generalizable, task-aware latent vision representations to instruct the decoder via inserted cross-attention layers. To analyze scaling behavior, we train Florence-2 based and Gemma-2 based models (0.4B to 11.2B parameters) on a synthetic dataset using varying compute budgets. While all languages in the dataset have image-aligned translations, only a subset of them include image captions. Notably, we show that captioning can emerge using a language prefix, even when this language only appears in the translation task. We find that indirect learning of unseen task-language pairs adheres to scaling laws that are governed by the multilinguality of the model, model size, and seen training samples. Finally, we demonstrate that the scaling laws extend to downstream tasks, achieving competitive performance through fine-tuning in multimodal machine translation (Multi30K, CoMMuTE), lexical disambiguation (CoMMuTE), and image captioning (Multi30K, XM3600, COCO Karpathy).

Quantum Machine Learning (QML) integrates quantum computational principles into learning algorithms, offering improved representational capacity and computational efficiency. Nevertheless, the security and robustness of QML systems remain underexplored, especially under adversarial conditions. In this paper, we present a systematization of adversarial robustness in QML, integrating conceptual organization with empirical evaluation across three threat models-black-box, gray-box, and white-box. We implement representative attacks in each category, including label-flipping for black-box, QUID encoder-level data poisoning for gray-box, and FGSM and PGD for white-box, using Quantum Neural Networks (QNNs) trained on two datasets from distinct domains: MNIST from computer vision and AZ-Class from Android malware, across multiple circuit depths (2, 5, 10, and 50 layers) and two encoding schemes (angle and amplitude). Our evaluation shows that amplitude encoding yields the highest clean accuracy (93% on MNIST and 67% on AZ-Class) in deep, noiseless circuits; however, it degrades sharply under adversarial perturbations and depolarization noise (p=0.01), dropping accuracy below 5%. In contrast, angle encoding, while offering lower representational capacity, remains more stable in shallow, noisy regimes, revealing a trade-off between capacity and robustness. Moreover, the QUID attack attains higher attack success rates, though quantum noise channels disrupt the Hilbert-space correlations it exploits, weakening its impact in image domains. This suggests that noise can act as a natural defense mechanism in Noisy Intermediate-Scale Quantum (NISQ) systems. Overall, our findings guide the development of secure and resilient QML architectures for practical deployment. These insights underscore the importance of designing threat-aware models that remain reliable under real-world noise in NISQ settings.

In recent years, notable advancements have been made in the domain of visual document understanding, with the prevailing architecture comprising a cascade of vision and language models. The text component can either be extracted explicitly with the use of external OCR models in OCR-based approaches, or alternatively, the vision model can be endowed with reading capabilities in OCR-free approaches. Typically, the queries to the model are input exclusively to the language component, necessitating the visual features to encompass the entire document. In this paper, we present VisFocus, an OCR-free method designed to better exploit the vision encoder's capacity by coupling it directly with the language prompt. To do so, we replace the down-sampling layers with layers that receive the input prompt and allow highlighting relevant parts of the document, while disregarding others. We pair the architecture enhancements with a novel pre-training task, using language masking on a snippet of the document text fed to the visual encoder in place of the prompt, to empower the model with focusing capabilities. Consequently, VisFocus learns to allocate its attention to text patches pertinent to the provided prompt. Our experiments demonstrate that this prompt-guided visual encoding approach significantly improves performance, achieving state-of-the-art results on various benchmarks.

Vision-language models commonly project all tokens produced by a pretrained vision encoder into a large language model. However, final-layer features can discard text, local attributes, and spatial relationships, while high-resolution inputs substantially increase context length and inference latency. We introduce \method, a Multi-scale Adaptive Vision Encoder. \method uses position-dependent gates to fuse shallow, intermediate, and deep features from a vision Transformer, preserving global semantics while enhancing edges, text, and local structure. It then performs question-conditioned token routing according to question relevance, local information content, global semantics, and spatial coverage, with a token budget that adapts to image complexity. To mitigate compression loss, we further introduce full-to-compressed representation distillation and a spatial diversity regularizer. In an illustrative simulation under a unified 7B language-model framework, \method reduces the average number of SigLIP-SO400M visual tokens from 729 to 146 (approximately 80.0\%) and improves the mean score on VQAv2, GQA, TextVQA, ScienceQA-IMG, and MMBench by 2.2 percentage points, while reducing single-image time to first token from 228\,ms to 129\,ms. We provide the full model design and evaluation protocol. All reported numbers currently serve only as placeholders for paper organization and experimental design; formal claims require real training runs, independent replications, and official benchmark evaluation.

Despite significant advancements in Multimodal Large Language Models (MLLMs) for understanding complex human intentions through cross-modal interactions, capturing intricate image details remains challenging. Previous methods integrating multiple vision encoders to enhance visual detail introduce redundancy and computational overhead. We observe that most MLLMs utilize only the last-layer feature map of the vision encoder for visual representation, neglecting the rich fine-grained information in shallow feature maps. To address this issue, we propose \modelname, a simple yet effective multi-layer feature fuser that efficiently integrates deep and shallow features from Vision Transformers (ViTs). Specifically, it leverages semantically aligned deep features as queries to dynamically extract missing details from shallow features, thus preserving semantic alignment while enriching the representation with fine-grained information. Applied to the LLaVA-1.5 model, \modelname~achieves significant improvements in visual representation and benchmark performance, providing a more flexible and lightweight solution compared to multi-encoder ensemble methods. The code and model have been released at https://github.com/yuecao0119/MMFuser.

The field of computer vision has experienced significant advancements through scalable vision encoders and multimodal pre-training frameworks. However, existing approaches often treat vision encoders and large language models (LLMs) as independent modules, limiting the integration of hierarchical visual features. In this work, we propose HIVE (Hierarchical Pre-Training of Vision Encoders), a novel framework that enhances vision-language alignment by introducing hierarchical cross-attention between the vision encoder and LLM. Unlike conventional methods that flatten image embeddings, HIVE enables structured feature fusion across multiple layers, improving gradient flow and representation learning. To optimize this interaction, we introduce a three-stage training strategy that progressively aligns the vision encoder with the LLM, ensuring stable optimization and effective multimodal fusion. Empirical evaluations demonstrate that HIVE achieves superior performance not only in image classification but also on various vision-language tasks, outperforming self-attention-based methods in benchmarks such as MME, GQA, OK-VQA, and ScienceQA. Our results highlight the benefits of hierarchical feature integration, paving the way for more efficient and expressive vision-language models.

We present Florence-VL, a new family of multimodal large language models (MLLMs) with enriched visual representations produced by Florence-2, a generative vision foundation model. Unlike the widely used CLIP-style vision transformer trained by contrastive learning, Florence-2 can capture different levels and aspects of visual features, which are more versatile to be adapted to diverse downstream tasks. We propose a novel feature-fusion architecture and an innovative training recipe that effectively integrates Florence-2's visual features into pretrained LLMs, such as Phi 3.5 and LLama 3. In particular, we propose "depth-breath fusion (DBFusion)" to fuse the visual features extracted from different depths and under multiple prompts. Our model training is composed of end-to-end pretraining of the whole model followed by finetuning of the projection layer and the LLM, on a carefully designed recipe of diverse open-source datasets that include high-quality image captions and instruction-tuning pairs. Our quantitative analysis and visualization of Florence-VL's visual features show its advantages over popular vision encoders on vision-language alignment, where the enriched depth and breath play important roles. Florence-VL achieves significant improvements over existing state-of-the-art MLLMs across various multi-modal and vision-centric benchmarks covering general VQA, perception, hallucination, OCR, Chart, knowledge-intensive understanding, etc. To facilitate future research, our models and the complete training recipe are open-sourced. https://github.com/JiuhaiChen/Florence-VL

We introduce Perception Encoder (PE), a state-of-the-art encoder for image and video understanding trained via simple vision-language learning. Traditionally, vision encoders have relied on a variety of pretraining objectives, each tailored to specific downstream tasks such as classification, captioning, or localization. Surprisingly, after scaling our carefully tuned image pretraining recipe and refining with our robust video data engine, we find that contrastive vision-language training alone can produce strong, general embeddings for all of these downstream tasks. There is only one caveat: these embeddings are hidden within the intermediate layers of the network. To draw them out, we introduce two alignment methods, language alignment for multimodal language modeling, and spatial alignment for dense prediction. Together with the core contrastive checkpoint, our PE family of models achieves state-of-the-art performance on a wide variety of tasks, including zero-shot image and video classification and retrieval; document, image, and video Q&A; and spatial tasks such as detection, depth estimation, and tracking. To foster further research, we are releasing our models, code, and a novel dataset of synthetically and human-annotated videos.

Vision encoders are increasingly used in modern applications, from vision-only models to multimodal systems such as vision-language models. Despite their remarkable success, it remains unclear how these architectures represent features internally. Here, we propose a novel approach for interpreting vision features via image reconstruction. We compare two related model families, SigLIP and SigLIP2, which differ only in their training objective, and show that encoders pre-trained on image-based tasks retain significantly more image information than those trained on non-image tasks such as contrastive learning. We further apply our method to a range of vision encoders, ranking them by the informativeness of their feature representations. Finally, we demonstrate that manipulating the feature space yields predictable changes in reconstructed images, revealing that orthogonal rotations (rather than spatial transformations) control color encoding. Our approach can be applied to any vision encoder, shedding light on the inner structure of its feature space. The code and model weights to reproduce the experiments are available in GitHub.

Multi-modal Large Language Models (MLLMs) have made significant strides in expanding the capabilities of Large Language Models (LLMs) through the incorporation of visual perception interfaces. Despite the emergence of exciting applications and the availability of diverse instruction tuning data, existing approaches often rely on CLIP or its variants as the visual branch, and merely extract features from the deep layers. However, these methods lack a comprehensive analysis of the visual encoders in MLLMs. In this paper, we conduct an extensive investigation into the effectiveness of different vision encoders within MLLMs. Our findings reveal that the shallow layer features of CLIP offer particular advantages for fine-grained tasks such as grounding and region understanding. Surprisingly, the vision-only model DINO, which is not pretrained with text-image alignment, demonstrates promising performance as a visual branch within MLLMs. By simply equipping it with an MLP layer for alignment, DINO surpasses CLIP in fine-grained related perception tasks. Building upon these observations, we propose a simple yet effective feature merging strategy, named COMM, that integrates CLIP and DINO with Multi-level features Merging, to enhance the visual capabilities of MLLMs. We evaluate COMM through comprehensive experiments on a wide range of benchmarks, including image captioning, visual question answering, visual grounding, and object hallucination. Experimental results demonstrate the superior performance of COMM compared to existing methods, showcasing its enhanced visual capabilities within MLLMs. Code will be made available at https://github.com/YuchenLiu98/COMM.

Multimodal large language models (MLLMs) typically rely on a single late-layer feature from a frozen vision encoder, leaving the encoder's rich hierarchy of visual cues under-utilized. MLLMs still suffer from visually ungrounded hallucinations, often relying on language priors rather than image evidence. While many prior mitigation strategies operate on the text side, they leave the visual representation unchanged and do not exploit the rich hierarchy of features encoded across vision layers. Existing multi-layer fusion methods partially address this limitation but remain static, applying the same layer mixture regardless of the query. In this work, we introduce TGIF (Text-Guided Inter-layer Fusion), a lightweight module that treats encoder layers as depth-wise "experts" and predicts a prompt-dependent fusion of visual features. TGIF follows the principle of direct external fusion, requires no vision-encoder updates, and adds minimal overhead. Integrated into LLaVA-1.5-7B, TGIF provides consistent improvements across hallucination, OCR, and VQA benchmarks, while preserving or improving performance on ScienceQA, GQA, and MMBench. These results suggest that query-conditioned, hierarchy-aware fusion is an effective way to strengthen visual grounding and reduce hallucination in modern MLLMs.

Vision-Language Navigation (VLN) requires an embodied agent to interpret a natural-language instruction and predict actions from temporally ordered visual observations. Adapting a multimodal large language model to VLN requires visual-language alignment, compact temporal inputs, action-space grounding, and stable training on the target hardware. This technical report presents PGN (Pangu Navigator), an offline VLN action-prediction system built on OpenPangu-7B. Training proceeds in two stages. First, PGMM aligns a frozen EVA-ViT-G/14 vision encoder with the frozen language backbone by training a Q-Former and a two-layer MLP projector. Second, PGN adapts the aligned model to expert navigation trajectories using five-observation windows, epoch-dependent temporal sampling, and a reasoning-then-action output format; this stage freezes the aligned visual pathway and updates three structural-token embeddings and LoRA adapters. The implementation combines mixed-precision computation, selective FP32 computation, and DeepSpeed ZeRO-2 on eight Ascend 910B NPUs. Under teacher-forced, open-loop evaluation on 500 held-out expert trajectories, V9 reports a 62.29% Normalized Action Match (NAM) and a 100.00% Non-empty Rate (NER). These metrics quantify offline expert-action alignment rather than closed-loop navigation success; evaluating error accumulation, path efficiency, and goal completion remains future work.

The feature maps of vision encoders are fundamental to myriad modern AI tasks, ranging from core perception algorithms (e.g. semantic segmentation, object detection, depth perception, etc.) to modern multimodal understanding in vision-language models (VLMs). Currently, in computer vision, the frontier of general purpose vision backbones is Vision Transformers (ViT), typically trained using contrastive loss (e.g. CLIP). A key problem with most off-the-shelf ViTs, particularly CLIP, is that these models are inflexibly low resolution. Most run at 224 times 224px, while the "high-resolution" versions are around 378-448px, but still inflexible. We introduce a novel method to coherently and cheaply upsample the feature maps of low-resolution vision encoders while picking up on fine-grained details that would otherwise be lost due to resolution. We demonstrate the effectiveness of this approach on core perception tasks as well as within agglomerative model training using RADIO as a way of providing richer targets for distillation. Code available at https://github.com/NVlabs/FeatSharp .

Typographic attacks exploit multi-modal systems by injecting text into images, leading to targeted misclassifications, malicious content generation and even Vision-Language Model jailbreaks. In this work, we analyze how CLIP vision encoders behave under typographic attacks, locating specialized attention heads in the latter half of the model's layers that causally extract and transmit typographic information to the cls token. Building on these insights, we introduce Dyslexify - a method to defend CLIP models against typographic attacks by selectively ablating a typographic circuit, consisting of attention heads. Without requiring finetuning, dyslexify improves performance by up to 22.06% on a typographic variant of ImageNet-100, while reducing standard ImageNet-100 accuracy by less than 1%, and demonstrate its utility in a medical foundation model for skin lesion diagnosis. Notably, our training-free approach remains competitive with current state-of-the-art typographic defenses that rely on finetuning. To this end, we release a family of dyslexic CLIP models which are significantly more robust against typographic attacks. These models serve as suitable drop-in replacements for a broad range of safety-critical applications, where the risks of text-based manipulation outweigh the utility of text recognition.

Recent advances in video captioning are driven by large-scale pretrained models, which follow the standard "pre-training followed by fine-tuning" paradigm, where the full model is fine-tuned for downstream tasks. Although effective, this approach becomes computationally prohibitive as the model size increases. The Parameter-Efficient Fine-Tuning (PEFT) approach offers a promising alternative, but primarily focuses on the language components of Multimodal Large Language Models (MLLMs). Despite recent progress, PEFT remains underexplored in multimodal tasks and lacks sufficient understanding of visual information during fine-tuning the model. To bridge this gap, we propose Query-Adapter (Q-Adapter), a lightweight visual adapter module designed to enhance MLLMs by enabling efficient fine-tuning for the video captioning task. Q-Adapter introduces learnable query tokens and a gating layer into Vision Encoder, enabling effective extraction of sparse, caption-relevant features without relying on external textual supervision. We evaluate Q-Adapter on two well-known video captioning datasets, MSR-VTT and MSVD, where it achieves state-of-the-art performance among the methods that take the PEFT approach across BLEU@4, METEOR, ROUGE-L, and CIDEr metrics. Q-Adapter also achieves competitive performance compared to methods that take the full fine-tuning approach while requiring only 1.4% of the parameters. We further analyze the impact of key hyperparameters and design choices on fine-tuning effectiveness, providing insights into optimization strategies for adapter-based learning. These results highlight the strong potential of Q-Adapter in balancing caption quality and parameter efficiency, demonstrating its scalability for video-language modeling.

In human conversations, individuals can indicate relevant regions within a scene while addressing others. In turn, the other person can then respond by referring to specific regions if necessary. This natural referential ability in dialogue remains absent in current Multimodal Large Language Models (MLLMs). To fill this gap, this paper proposes an MLLM called Shikra, which can handle spatial coordinate inputs and outputs in natural language. Its architecture consists of a vision encoder, an alignment layer, and a LLM. It is designed to be straightforward and simple, without the need for extra vocabularies, position encoder, pre-/post-detection modules, or external plug-in models. All inputs and outputs are in natural language form. Referential dialogue is a superset of various vision-language (VL) tasks. Shikra can naturally handle location-related tasks like REC and PointQA, as well as conventional VL tasks such as Image Captioning and VQA. Experimental results showcase Shikra's promising performance. Furthermore, it enables numerous exciting applications, like providing mentioned objects' coordinates in chains of thoughts and comparing user-pointed regions similarities. Our code, model and dataset are accessed at https://github.com/shikras/shikra.

Concept Bottleneck Models (CBMs) ground predictions in human-understandable concepts but face fundamental limitations: the absence of a metric to pre-evaluate concept relevance, the "linearity problem" causing recent CBMs to bypass the concept bottleneck entirely, an accuracy gap compared to opaque models, and finally the lack of systematic study on the impact of different visual backbones and VLMs. We introduce CBM-Suite, a methodological framework to systematically addresses these challenges. First, we propose an entropy-based metric to quantify the intrinsic suitability of a concept set for a given dataset. Second, we resolve the linearity problem by inserting a non-linear layer between concept activations and the classifier, which ensures that model accuracy faithfully reflects concept relevance. Third, we narrow the accuracy gap by leveraging a distillation loss guided by a linear teacher probe. Finally, we provide comprehensive analyses on how different vision encoders, vision-language models, and concept sets interact to influence accuracy and interpretability in CBMs. Extensive evaluations show that CBM-Suite yields more accurate models and provides insights for improving concept-based interpretability.

Vision-language models (VLMs) have achieved remarkable multimodal understanding and reasoning capabilities, yet remain computationally expensive due to dense visual tokenization. Existing efficiency approaches either merge redundant visual tokens or drop them progressively in language backbone, often trading accuracy for speed. In this work, we propose DUET-VLM, a versatile plug-and-play dual compression framework that consists of (a) vision-only redundancy aware compression of vision encoder's output into information-preserving tokens, followed by (b) layer-wise, salient text-guided dropping of visual tokens within the language backbone to progressively prune less informative tokens. This coordinated token management enables aggressive compression while retaining critical semantics. On LLaVA-1.5-7B, our approach maintains over 99% of baseline accuracy with 67% fewer tokens, and still retains >97% even at 89% reduction. With this dual-stage compression during training, it achieves 99.7% accuracy at 67% and 97.6% at 89%, surpassing prior SoTA visual token reduction methods across multiple benchmarks. When integrated into Video-LLaVA-7B, it even surpasses the baseline -- achieving >100% accuracy with a substantial 53.1% token reduction and retaining 97.6% accuracy under an extreme 93.4% setting. These results highlight end-to-end training with DUET-VLM, enabling robust adaptation to reduced visual (image/video) input without sacrificing accuracy, producing compact yet semantically rich representations within the same computational budget. Our code is available at https://github.com/AMD-AGI/DUET-VLM.

Vision-language models encode continuous geometry that their text pathway fails to express: a 6,000-parameter linear probe extracts hand joint angles at 6.1 degrees MAE from frozen features, while the best text output achieves only 20.0 degrees -- a 3.3x bottleneck. LoRA fine-tuning (r=16, 2,000 images) narrows this gap to 6.5 degrees, providing evidence for a pathway-training deficit rather than a representational one. Training objective determines accuracy more than architecture: five encoders spanning self-supervised, contrastive, and hybrid paradigms converge to statistically equivalent accuracy (R^2 approximately 0.55, TOST-equivalent at delta=0.03) despite sharing as little as CKA=0.41 representational similarity -- functional convergence without representational convergence. Autoregressive generation damages geometric fidelity, but the damage originates in the generation process, not in language alignment: Qwen2.5-VL's LLM layers actually improve probe accuracy over its raw vision encoder. Layer-wise analysis reveals a universal mid-network accuracy peak across all architectures, with attention heads in layers 18-22 carrying disproportionate geometric signal. These findings enable a single frozen backbone to function as a multi-task geometric sensor through lightweight probes, without fine-tuning or text generation.

The rapid adoption of HDR-capable devices has created a pressing need to convert the 8-bit Standard Dynamic Range (SDR) content into perceptually and physically accurate 10-bit High Dynamic Range (HDR). Existing inverse tone-mapping (ITM) methods often rely on fixed tone-mapping operators that struggle to generalize to real-world degradations, stylistic variations, and camera pipelines, frequently producing clipped highlights, desaturated colors, or unstable tone reproduction. We introduce LumaFlux, a first physically and perceptually guided diffusion transformer (DiT) for SDR-to-HDR reconstruction by adapting a large pretrained DiT. Our LumaFlux introduces (1) a Physically-Guided Adaptation (PGA) module that injects luminance, spatial descriptors, and frequency cues into attention through low-rank residuals; (2) a Perceptual Cross-Modulation (PCM) layer that stabilizes chroma and texture via FiLM conditioning from vision encoder features; and (3) an HDR Residual Coupler that fuses physical and perceptual signals under a timestep- and layer-adaptive modulation schedule. Finally, a lightweight Rational-Quadratic Spline decoder reconstructs smooth, interpretable tone fields for highlight and exposure expansion, enhancing the output of the VAE decoder to generate HDR. To enable robust HDR learning, we curate the first large-scale SDR-HDR training corpus. For fair and reproducible comparison, we further establish a new evaluation benchmark, comprising HDR references and corresponding expert-graded SDR versions. Across benchmarks, LumaFlux outperforms state-of-the-art baselines, achieving superior luminance reconstruction and perceptual color fidelity with minimal additional parameters.

Recent vision-language models (VLMs) typically rely on a single vision encoder trained with contrastive image-text objectives, such as CLIP-style pretraining. While contrastive encoders are effective for cross-modal alignment and retrieval, self-supervised visual encoders often capture richer dense semantics and exhibit stronger robustness on recognition and understanding tasks. In this work, we investigate how to scale the fusion of these complementary visual representations for vision-language modeling. We propose CoME-VL: Complementary Multi-Encoder Vision-Language, a modular fusion framework that integrates a contrastively trained vision encoder with a self-supervised DINO encoder. Our approach performs representation-level fusion by (i) entropy-guided multi-layer aggregation with orthogonality-constrained projections to reduce redundancy, and (ii) RoPE-enhanced cross-attention to align heterogeneous token grids and produce compact fused visual tokens. The fused tokens can be injected into a decoder-only LLM with minimal changes to standard VLM pipelines. Extensive experiments across diverse vision-language benchmarks demonstrate that CoME-VL consistently outperforms single-encoder baselines. In particular, we observe an average improvement of 4.9% on visual understanding tasks and 5.4% on grounding tasks. Our method achieves state-of-the-art performance on RefCOCO for detection while improving over the baseline by a large margin. Finally, we conduct ablation studies on layer merging, non-redundant feature mixing, and fusion capacity to evaluate how complementary contrastive and self-supervised signals affect VLM performance.

Estimating accurate depth from a single image is challenging because it is an ill-posed problem as infinitely many 3D scenes can be projected to the same 2D scene. However, recent works based on deep convolutional neural networks show great progress with plausible results. The convolutional neural networks are generally composed of two parts: an encoder for dense feature extraction and a decoder for predicting the desired depth. In the encoder-decoder schemes, repeated strided convolution and spatial pooling layers lower the spatial resolution of transitional outputs, and several techniques such as skip connections or multi-layer deconvolutional networks are adopted to recover the original resolution for effective dense prediction. In this paper, for more effective guidance of densely encoded features to the desired depth prediction, we propose a network architecture that utilizes novel local planar guidance layers located at multiple stages in the decoding phase. We show that the proposed method outperforms the state-of-the-art works with significant margin evaluating on challenging benchmarks. We also provide results from an ablation study to validate the effectiveness of the proposed method.

Hypothesis. Artificial general intelligence is, at its core, a compression problem. Effective compression demands resonance: deep learning scales best when its architecture aligns with the fundamental structure of the data. These are the fundamental principles. Yet, modern vision architectures have strayed from these truths: visual signals are highly redundant, while discriminative information, the surprise, is sparse. Current models process dense pixel grids uniformly, wasting vast compute on static background rather than focusing on the predictive residuals that define motion and meaning. We argue that to solve visual understanding, we must align our architectures with the information-theoretic principles of video, i.e., Codecs. Method. OneVision-Encoder encodes video by compressing predictive visual structure into semantic meaning. By adopting Codec Patchification, OV-Encoder abandons uniform computation to focus exclusively on the 3.1%-25% of regions rich in signal entropy. To unify spatial and temporal reasoning under irregular token layouts, OneVision-Encoder employs a shared 3D RoPE and is trained with a large-scale cluster discrimination objective over more than one million semantic concepts, jointly capturing object permanence and motion dynamics. Evidence. The results validate our core hypothesis: efficiency and accuracy are not a trade-off; they are positively correlated. When integrated into LLM, it consistently outperforms strong vision backbones such as Qwen3-ViT and SigLIP2 across 16 image, video, and document understanding benchmarks, despite using substantially fewer visual tokens and pretraining data. Notably, on video understanding tasks, OV-Encoder achieves an average improvement of 4.1% over Qwen3-ViT. Codec-aligned, patch-level sparsity is a foundational principle, enabling OV-Encoder as a scalable engine for next-generation visual generalists.

Convolutional neural networks typically encode an input image into a series of intermediate features with decreasing resolutions. While this structure is suited to classification tasks, it does not perform well for tasks requiring simultaneous recognition and localization (e.g., object detection). The encoder-decoder architectures are proposed to resolve this by applying a decoder network onto a backbone model designed for classification tasks. In this paper, we argue encoder-decoder architecture is ineffective in generating strong multi-scale features because of the scale-decreased backbone. We propose SpineNet, a backbone with scale-permuted intermediate features and cross-scale connections that is learned on an object detection task by Neural Architecture Search. Using similar building blocks, SpineNet models outperform ResNet-FPN models by ~3% AP at various scales while using 10-20% fewer FLOPs. In particular, SpineNet-190 achieves 52.5% AP with a MaskR-CNN detector and achieves 52.1% AP with a RetinaNet detector on COCO for a single model without test-time augmentation, significantly outperforms prior art of detectors. SpineNet can transfer to classification tasks, achieving 5% top-1 accuracy improvement on a challenging iNaturalist fine-grained dataset. Code is at: https://github.com/tensorflow/tpu/tree/master/models/official/detection.

We consider the training of the first layer of vision models and notice the clear relationship between pixel values and gradient update magnitudes: the gradients arriving at the weights of a first layer are by definition directly proportional to (normalized) input pixel values. Thus, an image with low contrast has a smaller impact on learning than an image with higher contrast, and a very bright or very dark image has a stronger impact on the weights than an image with moderate brightness. In this work, we propose performing gradient descent on the embeddings produced by the first layer of the model. However, switching to discrete inputs with an embedding layer is not a reasonable option for vision models. Thus, we propose the conceptual procedure of (i) a gradient descent step on first layer activations to construct an activation proposal, and (ii) finding the optimal weights of the first layer, i.e., those weights which minimize the squared distance to the activation proposal. We provide a closed form solution of the procedure and adjust it for robust stochastic training while computing everything efficiently. Empirically, we find that TrAct (Training Activations) speeds up training by factors between 1.25x and 4x while requiring only a small computational overhead. We demonstrate the utility of TrAct with different optimizers for a range of different vision models including convolutional and transformer architectures.

As Vision-Language Models (VLMs) are increasingly deployed in split-DNN configurations--with visual encoders (e.g., ResNet, ViT) operating on user devices and sending intermediate features to the cloud--there is a growing privacy risk from semantic information leakage. Existing approaches to reconstructing images from these intermediate features often result in blurry, semantically ambiguous images. To directly address semantic leakage, we propose CapRecover, a cross-modality inversion framework that recovers high-level semantic content, such as labels or captions, directly from intermediate features without image reconstruction. We evaluate CapRecover on multiple datasets and victim models, demonstrating strong performance in semantic recovery. Specifically, CapRecover achieves up to 92.71% Top-1 label accuracy on CIFAR-10 and generates fluent captions from ResNet50 features on COCO2017 with ROUGE-L scores up to 0.52. Our analysis further reveals that deeper convolutional layers encode significantly more semantic information compared to shallow layers. To mitigate semantic leakage, we introduce a simple yet effective protection method: adding random noise to intermediate features at each layer and removing the noise in the next layer. Experimental results show that this approach prevents semantic leakage without additional training costs. Our code is available at https://jus1mple.github.io/Image2CaptionAttack.

Automatic radiology report generation is a promising application of multimodal deep learning, aiming to reduce reporting workload and improve consistency. However, current state-of-the-art (SOTA) systems - such as Multimodal AI for Radiology Applications (MAIRA-2) and Medical Pathways Language Model-Multimodal (MedPaLM-M) - depend on large-scale multimodal training, clinical metadata, and multiple imaging views, making them resource-intensive and inaccessible for most settings. We introduce a compact image-to-text architecture that generates the Findings section of chest X-ray reports from a single frontal image. The model combines a frozen Self-Distillation with No Labels v3 (DINOv3) Vision Transformer (ViT) encoder with a Generative Pre-trained Transformer 2 (GPT-2) decoder enhanced by layer-wise anatomical attention. This mechanism integrates lung and heart segmentation masks through hierarchical Gaussian smoothing, biasing attention toward clinically relevant regions without adding trainable parameters. Evaluated on the official Medical Information Mart for Intensive Care-Chest X-ray (MIMIC-CXR) dataset using Chest Radiograph Expert (CheXpert) and Radiology Graph (RadGraph) metrics, our approach achieved substantial gains: CheXpert Macro-F1 for five key pathologies increased by 168% (0.083 -> 0.238) and Micro-F1 by 146% (0.137 -> 0.337), while broader performance across 14 observations improved by 86% (0.170 -> 0.316). Structural coherence also improved, with RadGraph F1 rising by 9.7%. Despite its small size and purely image-conditioned design, the model demonstrates that decoder-level anatomical guidance improves spatial grounding and enhances coherence in clinically relevant regions. The source code is publicly available at: https://github.com/devMuniz02/UDEM-CXR-Reporting-Thesis-2025.

Layer compositing is one of the most popular image editing workflows among both amateurs and professionals. Motivated by the success of diffusion models, we explore layer compositing from a layered image generation perspective. Instead of generating an image, we propose to generate background, foreground, layer mask, and the composed image simultaneously. To achieve layered image generation, we train an autoencoder that is able to reconstruct layered images and train diffusion models on the latent representation. One benefit of the proposed problem is to enable better compositing workflows in addition to the high-quality image output. Another benefit is producing higher-quality layer masks compared to masks produced by a separate step of image segmentation. Experimental results show that the proposed method is able to generate high-quality layered images and initiates a benchmark for future work.

Vision-language models (VLMs) remain unreliable when predictions require fine-grained visual evidence. We identify a previously overlooked cause: spectral response rigidity. Despite substantial frequency variation across images and tasks, pretrained vision encoders exhibit persistent, encoder-specific layerwise spectral profiles that change only marginally under downstream fine-tuning. Since pretrained vision encoders only receive images, they cannot adapt spectral extraction to the evidence required by the current query. We therefore propose HAFI-VLM, which introduces a task-conditioned frequency pathway while preserving the pretrained semantic representation. Hierarchical Adaptive Frequency Injection (HAFI) retrieves complementary low-, mid-, and high-frequency evidence at multiple encoder depths using text-modulated, spatially aligned cross-attention. A Visual Enrichment Layer Adapter further recalibrates shallow LLM attention to effectively utilize the enriched visual tokens. Experiments on LLaVA-1.5 and Qwen2.5-VL demonstrate consistent improvements in general VQA, text-rich understanding, and hallucination robustness, outperforming representation-level enhancement methods and most resolution- or cropping-based approaches without additional high-resolution encoding. Mechanistic analyses show that HAFI restores task-dependent spectral allocation while retaining semantic attention, establishing frequency enrichment as a distinct and effective route for improving VLM perception.

A central objective in computer vision is to design models with appropriate 2-D inductive bias. Desiderata for 2D inductive bias include two-dimensional position awareness, dynamic spatial locality, and translation and permutation invariance. To address these goals, we leverage an expressive variation of the multidimensional State Space Model (SSM). Our approach introduces efficient parameterization, accelerated computation, and a suitable normalization scheme. Empirically, we observe that incorporating our layer at the beginning of each transformer block of Vision Transformers (ViT) significantly enhances performance for multiple ViT backbones and across datasets. The new layer is effective even with a negligible amount of additional parameters and inference time. Ablation studies and visualizations demonstrate that the layer has a strong 2-D inductive bias. For example, vision transformers equipped with our layer exhibit effective performance even without positional encoding

Multi-layer perceptrons (MLP) have proven to be effective scene encoders when combined with higher-dimensional projections of the input, commonly referred to as positional encoding. However, scenes with a wide frequency spectrum remain a challenge: choosing high frequencies for positional encoding introduces noise in low structure areas, while low frequencies result in poor fitting of detailed regions. To address this, we propose a progressive positional encoding, exposing a hierarchical MLP structure to incremental sets of frequency encodings. Our model accurately reconstructs scenes with wide frequency bands and learns a scene representation at progressive level of detail without explicit per-level supervision. The architecture is modular: each level encodes a continuous implicit representation that can be leveraged separately for its respective resolution, meaning a smaller network for coarser reconstructions. Experiments on several 2D and 3D datasets show improvements in reconstruction accuracy, representational capacity and training speed compared to baselines.

Current vision models typically maintain a fixed correspondence between their representation structure and image space. Each layer comprises a set of tokens arranged "on-the-grid," which biases patches or tokens to encode information at a specific spatio(-temporal) location. In this work we present Moving Off-the-Grid (MooG), a self-supervised video representation model that offers an alternative approach, allowing tokens to move "off-the-grid" to better enable them to represent scene elements consistently, even as they move across the image plane through time. By using a combination of cross-attention and positional embeddings we disentangle the representation structure and image structure. We find that a simple self-supervised objective--next frame prediction--trained on video data, results in a set of latent tokens which bind to specific scene structures and track them as they move. We demonstrate the usefulness of MooG's learned representation both qualitatively and quantitatively by training readouts on top of the learned representation on a variety of downstream tasks. We show that MooG can provide a strong foundation for different vision tasks when compared to "on-the-grid" baselines.

Pre-trained encoders are widely employed in dense prediction tasks for their capability to effectively extract visual features from images. The decoder subsequently processes these features to generate pixel-level predictions. However, due to structural differences and variations in input data, only encoders benefit from pre-learned representations from vision benchmarks such as image classification and self-supervised learning, while decoders are typically trained from scratch. In this paper, we introduce timesNet, which facilitates a "pre-trained encoder times pre-trained decoder" collaboration through three innovative designs. timesNet enables the direct utilization of pre-trained models within the decoder, integrating pre-learned representations into the decoding process to enhance performance in dense prediction tasks. By simply coupling the pre-trained encoder and pre-trained decoder, timesNet distinguishes itself as a highly promising approach. Remarkably, it achieves this without relying on decoding-specific structures or task-specific algorithms. Despite its streamlined design, timesNet outperforms advanced methods in tasks such as monocular depth estimation and semantic segmentation, achieving state-of-the-art performance particularly in monocular depth estimation. and semantic segmentation, achieving state-of-the-art results, especially in monocular depth estimation. embedding algorithms. Despite its streamlined design, timesNet outperforms advanced methods in tasks such as monocular depth estimation and semantic segmentation, achieving state-of-the-art performance particularly in monocular depth estimation.

After their initial success in natural language processing, transformer architectures have rapidly gained traction in computer vision, providing state-of-the-art results for tasks such as image classification, detection, segmentation, and video analysis. We offer three insights based on simple and easy to implement variants of vision transformers. (1) The residual layers of vision transformers, which are usually processed sequentially, can to some extent be processed efficiently in parallel without noticeably affecting the accuracy. (2) Fine-tuning the weights of the attention layers is sufficient to adapt vision transformers to a higher resolution and to other classification tasks. This saves compute, reduces the peak memory consumption at fine-tuning time, and allows sharing the majority of weights across tasks. (3) Adding MLP-based patch pre-processing layers improves Bert-like self-supervised training based on patch masking. We evaluate the impact of these design choices using the ImageNet-1k dataset, and confirm our findings on the ImageNet-v2 test set. Transfer performance is measured across six smaller datasets.

The fully convolutional network (FCN) with an encoder-decoder architecture has been the standard paradigm for semantic segmentation. The encoder-decoder architecture utilizes an encoder to capture multilevel feature maps, which are incorporated into the final prediction by a decoder. As the context is crucial for precise segmentation, tremendous effort has been made to extract such information in an intelligent fashion, including employing dilated/atrous convolutions or inserting attention modules. However, these endeavors are all based on the FCN architecture with ResNet or other backbones, which cannot fully exploit the context from the theoretical concept. By contrast, we introduce the Swin Transformer as the backbone to extract the context information and design a novel decoder of densely connected feature aggregation module (DCFAM) to restore the resolution and produce the segmentation map. The experimental results on two remotely sensed semantic segmentation datasets demonstrate the effectiveness of the proposed scheme.Code is available at https://github.com/WangLibo1995/GeoSeg

We release MiDaS v3.1 for monocular depth estimation, offering a variety of new models based on different encoder backbones. This release is motivated by the success of transformers in computer vision, with a large variety of pretrained vision transformers now available. We explore how using the most promising vision transformers as image encoders impacts depth estimation quality and runtime of the MiDaS architecture. Our investigation also includes recent convolutional approaches that achieve comparable quality to vision transformers in image classification tasks. While the previous release MiDaS v3.0 solely leverages the vanilla vision transformer ViT, MiDaS v3.1 offers additional models based on BEiT, Swin, SwinV2, Next-ViT and LeViT. These models offer different performance-runtime tradeoffs. The best model improves the depth estimation quality by 28% while efficient models enable downstream tasks requiring high frame rates. We also describe the general process for integrating new backbones. A video summarizing the work can be found at https://youtu.be/UjaeNNFf9sE and the code is available at https://github.com/isl-org/MiDaS.

Vision-Language Models (VLMs) are a new family of models that align image content with natural language. Existing approaches typically fuse either (a) early: by mixing tokens/features inside the encoders, or (b) late: by comparing pooled embeddings. Many methods also tie fusion to an autoregressive decoder. However, the hidden states of both modalities already carry rich, modality-specific structure (spatial layout in vision; syntax and semantics in text), so directly aligning these states is a natural way to match what the two modalities "think". We propose a lightweight fusion module: a few cross-only, bidirectional attention layers placed near the top of both encoders. Each layer projects the vision and text encoder hidden-state sequences into a shared space, attends across modalities, and sends gated residual updates back, with simple stabilizers to improve alignment. The encoders remain non-causal and strong for understanding, while generation stays cleanly decoupled via an optional decoder. Across standard retrieval, VQA, and visual reasoning benchmarks, BRIDGE outperforms comparable VLMs while preserving the bi-encoder efficiency of contrastive models. We make our code publicly available at https://github.com/jfeinashley/BRIDGE.

Spatial pyramid pooling module or encode-decoder structure are used in deep neural networks for semantic segmentation task. The former networks are able to encode multi-scale contextual information by probing the incoming features with filters or pooling operations at multiple rates and multiple effective fields-of-view, while the latter networks can capture sharper object boundaries by gradually recovering the spatial information. In this work, we propose to combine the advantages from both methods. Specifically, our proposed model, DeepLabv3+, extends DeepLabv3 by adding a simple yet effective decoder module to refine the segmentation results especially along object boundaries. We further explore the Xception model and apply the depthwise separable convolution to both Atrous Spatial Pyramid Pooling and decoder modules, resulting in a faster and stronger encoder-decoder network. We demonstrate the effectiveness of the proposed model on PASCAL VOC 2012 and Cityscapes datasets, achieving the test set performance of 89.0\% and 82.1\% without any post-processing. Our paper is accompanied with a publicly available reference implementation of the proposed models in Tensorflow at https://github.com/tensorflow/models/tree/master/research/deeplab.

Visual encoders are fundamental components in vision-language models (VLMs), each showcasing unique strengths derived from various pre-trained visual foundation models. To leverage the various capabilities of these encoders, recent studies incorporate multiple encoders within a single VLM, leading to a considerable increase in computational cost. In this paper, we present Mixture-of-Visual-Encoder Knowledge Distillation (MoVE-KD), a novel framework that distills the unique proficiencies of multiple vision encoders into a single, efficient encoder model. Specifically, to mitigate conflicts and retain the unique characteristics of each teacher encoder, we employ low-rank adaptation (LoRA) and mixture-of-experts (MoEs) to selectively activate specialized knowledge based on input features, enhancing both adaptability and efficiency. To regularize the KD process and enhance performance, we propose an attention-based distillation strategy that adaptively weighs the different visual encoders and emphasizes valuable visual tokens, reducing the burden of replicating comprehensive but distinct features from multiple teachers. Comprehensive experiments on popular VLMs, such as LLaVA and LLaVA-NeXT, validate the effectiveness of our method. The code will be released.

As the key component in multimodal large language models (MLLMs), the ability of the visual encoder greatly affects MLLM's understanding on diverse image content. Although some large-scale pretrained vision encoders such as vision encoders in CLIP and DINOv2 have brought promising performance, we found that there is still no single vision encoder that can dominate various image content understanding, e.g., the CLIP vision encoder leads to outstanding results on general image understanding but poor performance on document or chart content. To alleviate the bias of CLIP vision encoder, we first delve into the inherent behavior of different pre-trained vision encoders and then propose the MoVA, a powerful and novel MLLM, adaptively routing and fusing task-specific vision experts with a coarse-to-fine mechanism. In the coarse-grained stage, we design a context-aware expert routing strategy to dynamically select the most suitable vision experts according to the user instruction, input image, and expertise of vision experts. This benefits from the powerful model function understanding ability of the large language model (LLM) equipped with expert-routing low-rank adaptation (LoRA). In the fine-grained stage, we elaborately conduct the mixture-of-vision-expert adapter (MoV-Adapter) to extract and fuse task-specific knowledge from various experts. This coarse-to-fine paradigm effectively leverages representations from experts based on multimodal context and model expertise, further enhancing the generalization ability. We conduct extensive experiments to evaluate the effectiveness of the proposed approach. Without any bells and whistles, MoVA can achieve significant performance gains over current state-of-the-art methods in a wide range of challenging multimodal benchmarks. Codes and models will be available at https://github.com/TempleX98/MoVA.

Complex 3D scene understanding has gained increasing attention, with scene encoding strategies playing a crucial role in this success. However, the optimal scene encoding strategies for various scenarios remain unclear, particularly compared to their image-based counterparts. To address this issue, we present a comprehensive study that probes various visual encoding models for 3D scene understanding, identifying the strengths and limitations of each model across different scenarios. Our evaluation spans seven vision foundation encoders, including image-based, video-based, and 3D foundation models. We evaluate these models in four tasks: Vision-Language Scene Reasoning, Visual Grounding, Segmentation, and Registration, each focusing on different aspects of scene understanding. Our evaluations yield key findings: DINOv2 demonstrates superior performance, video models excel in object-level tasks, diffusion models benefit geometric tasks, and language-pretrained models show unexpected limitations in language-related tasks. These insights challenge some conventional understandings, provide novel perspectives on leveraging visual foundation models, and highlight the need for more flexible encoder selection in future vision-language and scene-understanding tasks.

Recent advancements in large generative models, particularly diffusion-based methods, have significantly enhanced the capabilities of image editing. However, achieving precise control over image composition tasks remains a challenge. Layered representations, which allow for independent editing of image components, are essential for user-driven content creation, yet existing approaches often struggle to decompose image into plausible layers with accurately retained transparent visual effects such as shadows and reflections. We propose LayerDecomp, a generative framework for image layer decomposition which outputs photorealistic clean backgrounds and high-quality transparent foregrounds with faithfully preserved visual effects. To enable effective training, we first introduce a dataset preparation pipeline that automatically scales up simulated multi-layer data with synthesized visual effects. To further enhance real-world applicability, we supplement this simulated dataset with camera-captured images containing natural visual effects. Additionally, we propose a consistency loss which enforces the model to learn accurate representations for the transparent foreground layer when ground-truth annotations are not available. Our method achieves superior quality in layer decomposition, outperforming existing approaches in object removal and spatial editing tasks across several benchmarks and multiple user studies, unlocking various creative possibilities for layer-wise image editing. The project page is https://rayjryang.github.io/LayerDecomp.

Multimodal large language models (MLLMs) extend LLMs with visual understanding through a three-stage pipeline: multimodal preprocessing, vision encoding, and LLM inference. While these stages enhance capability, they introduce significant system bottlenecks. First, multimodal preprocessing-especially video decoding-often dominates Time-to-First-Token (TTFT). Most systems rely on CPU-based decoding, which severely limits throughput, while existing GPU-based approaches prioritize throughput-oriented parallelism and fail to meet the latency-sensitive requirements of MLLM inference. Second, the vision encoder is a standalone, compute-intensive stage that produces visual embeddings and cannot be co-batched with LLM prefill or decoding. This heterogeneity forces inter-stage blocking and increases token-generation latency. Even when deployed on separate GPUs, these stages underutilize available compute and memory resources, reducing overall utilization and constraining system throughput. To address these challenges, we present FlashCodec and UnifiedServe, two complementary designs that jointly optimize the end-to-end MLLM pipeline. FlashCodec accelerates the multimodal preprocessing stage through collaborative multi-GPU video decoding, reducing decoding latency while preserving high throughput. UnifiedServe optimizes the vision-to-text and inference stages using a logically decoupled their execution to eliminate inter-stage blocking, yet physically sharing GPU resources to maximize GPU system utilization. By carefully orchestrating execution across stages and minimizing interference, UnifiedServe Together, our proposed framework forms an end-to-end optimized stack that can serve up to 3.0times more requests or enforce 1.5times tighter SLOs, while achieving up to 4.4times higher throughput compared to state-of-the-art systems.

We present an unsupervised visual feature learning algorithm driven by context-based pixel prediction. By analogy with auto-encoders, we propose Context Encoders -- a convolutional neural network trained to generate the contents of an arbitrary image region conditioned on its surroundings. In order to succeed at this task, context encoders need to both understand the content of the entire image, as well as produce a plausible hypothesis for the missing part(s). When training context encoders, we have experimented with both a standard pixel-wise reconstruction loss, as well as a reconstruction plus an adversarial loss. The latter produces much sharper results because it can better handle multiple modes in the output. We found that a context encoder learns a representation that captures not just appearance but also the semantics of visual structures. We quantitatively demonstrate the effectiveness of our learned features for CNN pre-training on classification, detection, and segmentation tasks. Furthermore, context encoders can be used for semantic inpainting tasks, either stand-alone or as initialization for non-parametric methods.

Representation autoencoders that reuse frozen pretrained vision encoders as visual tokenizers have achieved strong reconstruction and generation quality. However, existing methods universally extract features from only the last encoder layer, discarding the rich hierarchical information distributed across intermediate layers. We show that low-level visual details survive in the last layer merely as attenuated residuals after multiple layers of semantic abstraction, and that explicitly fusing multi-layer features can substantially recover this lost information. We propose DRoRAE (Depth-Routed Representation AutoEncoder), a lightweight fusion module that adaptively aggregates all encoder layers via energy-constrained routing and incremental correction, producing an enriched latent compatible with a frozen pretrained decoder. A three-phase decoupled training strategy first learns the fusion under the implicit distributional constraint of the frozen decoder, then fine-tunes the decoder to fully exploit the enriched representation. On ImageNet-256, DRoRAE reduces rFID from 0.57 to 0.29 and improves generation FID from 1.74 to 1.65 (with AutoGuidance), with gains also transferring to text-to-image synthesis. Furthermore, we uncover a log-linear scaling law (R^2{=}0.86) between fusion capacity and reconstruction quality, identifying representation richness as a new, predictably scalable dimension for visual tokenizers analogous to vocabulary size in NLP.

Pre-trained vision encoders like DINOv2 have demonstrated exceptional performance on unimodal tasks. However, we observe that their feature representations are poorly aligned across different modalities. For instance, the feature embedding for an RGB image and its corresponding depth map of the same scene exhibit a cosine similarity that is nearly identical to that of two random, unrelated images. To address this, we propose the Omnivorous Vision Encoder, a novel framework that learns a modality-agnostic feature space. We train the encoder with a dual objective: first, to maximize the feature alignment between different modalities of the same scene; and second, a distillation objective that anchors the learned representations to the output of a fully frozen teacher such as DINOv2. The resulting student encoder becomes "omnivorous" by producing a consistent, powerful embedding for a given scene, regardless of the input modality (RGB, Depth, Segmentation, etc.). This approach enables robust cross-modal understanding while retaining the discriminative semantics of the original foundation model.

Vision Transformers (ViTs) lack the hierarchical inductive biases inherent to Convolutional Neural Networks (CNNs), theoretically allowing them to maintain high-dimensional representations throughout all layers. However, recent observations suggest ViTs often spontaneously manifest a "U-shaped" entropy profile-compressing information in middle layers before expanding it for the final classification. In this work, we demonstrate that this "Inductive Bottleneck" is not an architectural artifact, but a data-dependent adaptation. By analyzing the layer-wise Effective Encoding Dimension (EED) of DINO-trained ViTs across datasets of varying compositional complexity (UC Merced, Tiny ImageNet, and CIFAR-100), we show that the depth of the bottleneck correlates strongly with the semantic abstraction required by the task. We find that while texture-heavy datasets preserve high-rank representations throughout, object-centric datasets drive the network to dampen high-frequency information in middle layers, effectively "learning" a bottleneck to isolate semantic features.

Deep encoder-decoder based CNNs have advanced image inpainting methods for hole filling. While existing methods recover structures and textures step-by-step in the hole regions, they typically use two encoder-decoders for separate recovery. The CNN features of each encoder are learned to capture either missing structures or textures without considering them as a whole. The insufficient utilization of these encoder features limit the performance of recovering both structures and textures. In this paper, we propose a mutual encoder-decoder CNN for joint recovery of both. We use CNN features from the deep and shallow layers of the encoder to represent structures and textures of an input image, respectively. The deep layer features are sent to a structure branch and the shallow layer features are sent to a texture branch. In each branch, we fill holes in multiple scales of the CNN features. The filled CNN features from both branches are concatenated and then equalized. During feature equalization, we reweigh channel attentions first and propose a bilateral propagation activation function to enable spatial equalization. To this end, the filled CNN features of structure and texture mutually benefit each other to represent image content at all feature levels. We use the equalized feature to supplement decoder features for output image generation through skip connections. Experiments on the benchmark datasets show the proposed method is effective to recover structures and textures and performs favorably against state-of-the-art approaches.

This paper presents a family of advanced vision encoder, named OpenVision 3, that learns a single, unified visual representation that can serve both image understanding and image generation. Our core architecture is simple: we feed VAE-compressed image latents to a ViT encoder and train its output to support two complementary roles. First, the encoder output is passed to the ViT-VAE decoder to reconstruct the original image, encouraging the representation to capture generative structure. Second, the same representation is optimized with contrastive learning and image-captioning objectives, strengthening semantic features. By jointly optimizing reconstruction- and semantics-driven signals in a shared latent space, the encoder learns representations that synergize and generalize well across both regimes. We validate this unified design through extensive downstream evaluations with the encoder frozen. For multimodal understanding, we plug the encoder into the LLaVA-1.5 framework: it performs comparably with a standard CLIP vision encoder (e.g., 62.4 vs 62.2 on SeedBench, and 83.7 vs 82.9 on POPE). For generation, we test it under the RAE framework: ours substantially surpasses the standard CLIP-based encoder (e.g., gFID: 1.89 vs 2.54 on ImageNet). We hope this work can spur future research on unified modeling.

Representation autoencoders (RAEs) reuse features from a pretrained visual encoder as reconstruction and diffusion latents, integrating strong visual representations into image generation. However, RAEs still need to decide which encoder layers form the shared latent space for the generator and pixel decoder. This choice involves a trade-off. Shallower layers tend to preserve fine pixel details better, while deeper layers tend to yield better generation metrics. A fixed heuristic layer fusion therefore couples two stages that benefit from different information. We introduce FuseReg, which replaces heuristic feature selection with training over random subsets of encoder layers. We theoretically analyze the underlying mechanism: subset sampling explicitly penalizes sensitivity to cross-layer disagreement. On ImageNet-256 with DINOv3-L, a single FuseReg decoder reconstructs from full, sparse, and single-layer fusions without retraining, achieving higher PSNR than decoders specialized to fixed fusions. This flexibility also benefits generation: decoder replacement alone reduces unguided gFID by 27% with an unchanged RAEv2 DiT-XL generator. The same regularization principle extends to diffusion training, with joint regularization of both stages reducing unguided gFID by 29% on DiT-Base. These results show that training downstream models for layer-fusion robustness narrows the reconstruction-generation gap without modifying the pretrained encoder.

Vision-Language (VL) models have gained significant research focus, enabling remarkable advances in multimodal reasoning. These architectures typically comprise a vision encoder, a Large Language Model (LLM), and a projection module that aligns visual features with the LLM's representation space. Despite their success, a critical limitation persists: the vision encoding process remains decoupled from user queries, often in the form of image-related questions. Consequently, the resulting visual features may not be optimally attuned to the query-specific elements of the image. To address this, we introduce QA-ViT, a Question Aware Vision Transformer approach for multimodal reasoning, which embeds question awareness directly within the vision encoder. This integration results in dynamic visual features focusing on relevant image aspects to the posed question. QA-ViT is model-agnostic and can be incorporated efficiently into any VL architecture. Extensive experiments demonstrate the effectiveness of applying our method to various multimodal architectures, leading to consistent improvement across diverse tasks and showcasing its potential for enhancing visual and scene-text understanding.

Vision encoders are a critical component of vision-language models, and scaling their capacity effectively improves performance. However, dense scaling increases compute cost and inference latency. Mixture-of-Experts (MoE) architectures offer a compelling alternative, having enabled efficient scaling in LLMs, yet the MoE design space for CLIP-style vision encoders remains underexplored at State-of-the-Art (SOTA) levels. In this work, we systematically study MoE designs for vision encoder scaling and find that fine-grained MoE topologies yield substantial gains over both dense and standard MoE counterparts. We further propose an auxiliary-loss-free balancing variant for better expert utilization, and design a specialized MoE kernel to mitigate inference latency overhead. To enhance video capabilities while preserving image knowledge, we introduce frame-level distillation paired with a novel freezing mechanism. We pretrain a series of Mixture-of-Experts Vision Encoders (MoE-ViE) across a range of sizes, all consistently outperforming their dense counterparts. Our largest model matches the zero-shot performance of a SOTA encoder 1.7x its size at 76% of its latency. When aligned with an LLM, MoE-ViE surpasses all compared encoders on image and video benchmarks, including those with up to 5x more activated parameters. Code is available at https://github.com/facebookresearch/moe_vie.

Recent advances in scene understanding benefit a lot from depth maps because of the 3D geometry information, especially in complex conditions (e.g., low light and overexposed). Existing approaches encode depth maps along with RGB images and perform feature fusion between them to enable more robust predictions. Taking into account that depth can be regarded as a geometry supplement for RGB images, a straightforward question arises: Do we really need to explicitly encode depth information with neural networks as done for RGB images? Based on this insight, in this paper, we investigate a new way to learn RGBD feature representations and present DFormerv2, a strong RGBD encoder that explicitly uses depth maps as geometry priors rather than encoding depth information with neural networks. Our goal is to extract the geometry clues from the depth and spatial distances among all the image patch tokens, which will then be used as geometry priors to allocate attention weights in self-attention. Extensive experiments demonstrate that DFormerv2 exhibits exceptional performance in various RGBD semantic segmentation benchmarks. Code is available at: https://github.com/VCIP-RGBD/DFormer.

Implicit neural representations store videos as neural networks and have performed well for various vision tasks such as video compression and denoising. With frame index or positional index as input, implicit representations (NeRV, E-NeRV, \etc) reconstruct video from fixed and content-agnostic embeddings. Such embedding largely limits the regression capacity and internal generalization for video interpolation. In this paper, we propose a Hybrid Neural Representation for Videos (HNeRV), where a learnable encoder generates content-adaptive embeddings, which act as the decoder input. Besides the input embedding, we introduce HNeRV blocks, which ensure model parameters are evenly distributed across the entire network, such that higher layers (layers near the output) can have more capacity to store high-resolution content and video details. With content-adaptive embeddings and re-designed architecture, HNeRV outperforms implicit methods in video regression tasks for both reconstruction quality (+4.7 PSNR) and convergence speed (16times faster), and shows better internal generalization. As a simple and efficient video representation, HNeRV also shows decoding advantages for speed, flexibility, and deployment, compared to traditional codecs~(H.264, H.265) and learning-based compression methods. Finally, we explore the effectiveness of HNeRV on downstream tasks such as video compression and video inpainting. We provide project page at https://haochen-rye.github.io/HNeRV, and Code at https://github.com/haochen-rye/HNeRV

In this paper, we propose a post-training quantization framework of large vision-language models (LVLMs) for efficient multi-modal inference. Conventional quantization methods sequentially search the layer-wise rounding functions by minimizing activation discretization errors, which fails to acquire optimal quantization strategy without considering cross-layer dependency. On the contrary, we mine the cross-layer dependency that significantly influences discretization errors of the entire vision-language model, and embed this dependency into optimal quantization strategy searching with low search cost. Specifically, we observe the strong correlation between the activation entropy and the cross-layer dependency concerning output discretization errors. Therefore, we employ the entropy as the proxy to partition blocks optimally, which aims to achieve satisfying trade-offs between discretization errors and the search cost. Moreover, we optimize the visual encoder to disentangle the cross-layer dependency for fine-grained decomposition of search space, so that the search cost is further reduced without harming the quantization accuracy. Experimental results demonstrate that our method compresses the memory by 2.78x and increase generate speed by 1.44x about 13B LLaVA model without performance degradation on diverse multi-modal reasoning tasks. Code is available at https://github.com/ChangyuanWang17/QVLM.

Although current deep learning methods have achieved impressive results for semantic segmentation, they incur high computational costs and have a huge number of parameters. For real-time applications, inference speed and memory usage are two important factors. To address the challenge, we propose a lightweight feature pyramid encoding network (FPENet) to make a good trade-off between accuracy and speed. Specifically, we use a feature pyramid encoding block to encode multi-scale contextual features with depthwise dilated convolutions in all stages of the encoder. A mutual embedding upsample module is introduced in the decoder to aggregate the high-level semantic features and low-level spatial details efficiently. The proposed network outperforms existing real-time methods with fewer parameters and improved inference speed on the Cityscapes and CamVid benchmark datasets. Specifically, FPENet achieves 68.0\% mean IoU on the Cityscapes test set with only 0.4M parameters and 102 FPS speed on an NVIDIA TITAN V GPU.

Vision transformers (ViTs) are quickly becoming the de-facto architecture for computer vision, yet we understand very little about why they work and what they learn. While existing studies visually analyze the mechanisms of convolutional neural networks, an analogous exploration of ViTs remains challenging. In this paper, we first address the obstacles to performing visualizations on ViTs. Assisted by these solutions, we observe that neurons in ViTs trained with language model supervision (e.g., CLIP) are activated by semantic concepts rather than visual features. We also explore the underlying differences between ViTs and CNNs, and we find that transformers detect image background features, just like their convolutional counterparts, but their predictions depend far less on high-frequency information. On the other hand, both architecture types behave similarly in the way features progress from abstract patterns in early layers to concrete objects in late layers. In addition, we show that ViTs maintain spatial information in all layers except the final layer. In contrast to previous works, we show that the last layer most likely discards the spatial information and behaves as a learned global pooling operation. Finally, we conduct large-scale visualizations on a wide range of ViT variants, including DeiT, CoaT, ConViT, PiT, Swin, and Twin, to validate the effectiveness of our method.

Rich feature representations derived from CLIP-ViT have been widely utilized in AI-generated image detection. While most existing methods primarily leverage features from the final layer, we systematically analyze the contributions of layer-wise features to this task. Our study reveals that earlier layers provide more localized and generalizable features, often surpassing the performance of final-layer features in detection tasks. Moreover, we find that different layers capture distinct aspects of the data, each contributing uniquely to AI-generated image detection. Motivated by these findings, we introduce a novel adaptive method, termed MoLD, which dynamically integrates features from multiple ViT layers using a gating-based mechanism. Extensive experiments on both GAN- and diffusion-generated images demonstrate that MoLD significantly improves detection performance, enhances generalization across diverse generative models, and exhibits robustness in real-world scenarios. Finally, we illustrate the scalability and versatility of our approach by successfully applying it to other pre-trained ViTs, such as DINOv2.

A significant weakness of most current deep Convolutional Neural Networks is the need to train them using vast amounts of manu- ally labelled data. In this work we propose a unsupervised framework to learn a deep convolutional neural network for single view depth predic- tion, without requiring a pre-training stage or annotated ground truth depths. We achieve this by training the network in a manner analogous to an autoencoder. At training time we consider a pair of images, source and target, with small, known camera motion between the two such as a stereo pair. We train the convolutional encoder for the task of predicting the depth map for the source image. To do so, we explicitly generate an inverse warp of the target image using the predicted depth and known inter-view displacement, to reconstruct the source image; the photomet- ric error in the reconstruction is the reconstruction loss for the encoder. The acquisition of this training data is considerably simpler than for equivalent systems, requiring no manual annotation, nor calibration of depth sensor to camera. We show that our network trained on less than half of the KITTI dataset (without any further augmentation) gives com- parable performance to that of the state of art supervised methods for single view depth estimation.

Reconstructing a 3D object from a 2D image is a well-researched vision problem, with many kinds of deep learning techniques having been tried. Most commonly, 3D convolutional approaches are used, though previous work has shown state-of-the-art methods using 2D convolutions that are also significantly more efficient to train. With the recent rise of transformers for vision tasks, often outperforming convolutional methods, along with some earlier attempts to use transformers for 3D object reconstruction, we set out to use visual transformers in place of convolutions in existing efficient, high-performing techniques for 3D object reconstruction in order to achieve superior results on the task. Using a transformer-based encoder and decoder to predict 3D structure from 2D images, we achieve accuracy similar or superior to the baseline approach. This study serves as evidence for the potential of visual transformers in the task of 3D object reconstruction.

Pretrained visual representations support image generation, but may not fully preserve the fine-grained details needed for faithful reconstruction. Meanwhile, intermediate encoder layers contain complementary visual details, but learning to fuse them for reconstruction can produce a latent distribution that is difficult to model. Existing fusion methods require empirical tuning of layer selection or staged optimization of fusion and decoding, increasing configuration effort or training complexity. We introduce HiRAE (Hierarchical Representation Autoencoder), which learns a hierarchical fusion framework over the full encoder hierarchy to improve reconstruction fidelity while maintaining compatibility with generative modeling. HiRAE groups encoder layers by depth and learns residual corrections to the deepest representation. Group-wise norm caps bound these corrections relative to the deep anchor, with tighter budgets for shallower groups. Our HiRAE-24 preserves the latent token count and channel dimension. On ImageNet-256, HiRAE-24 reduces reconstruction FID from 0.299 to 0.209 relative to RAEv2 while maintaining competitive guided generation quality. For text-to-image generation, HiRAE-24 improves alignment over RAEv2 on GenEval, DPG-Bench, and GenAI-Bench both before and after supervised fine-tuning. Under the same generator-training and evaluation protocol, post-fine-tuning GenEval increases from 84.86 to 87.70.

Prior research on deep video compression (DVC) for machine tasks typically necessitates training a unique codec for each specific task, mandating a dedicated decoder per task. In contrast, traditional video codecs employ a flexible encoder controller, enabling the adaptation of a single codec to different tasks through mechanisms like mode prediction. Drawing inspiration from this, we introduce an innovative encoder controller for deep video compression for machines. This controller features a mode prediction and a Group of Pictures (GoP) selection module. Our approach centralizes control at the encoding stage, allowing for adaptable encoder adjustments across different tasks, such as detection and tracking, while maintaining compatibility with a standard pre-trained DVC decoder. Empirical evidence demonstrates that our method is applicable across multiple tasks with various existing pre-trained DVCs. Moreover, extensive experiments demonstrate that our method outperforms previous DVC by about 25% bitrate for different tasks, with only one pre-trained decoder.

We propose Dual PatchNorm: two Layer Normalization layers (LayerNorms), before and after the patch embedding layer in Vision Transformers. We demonstrate that Dual PatchNorm outperforms the result of exhaustive search for alternative LayerNorm placement strategies in the Transformer block itself. In our experiments, incorporating this trivial modification, often leads to improved accuracy over well-tuned Vision Transformers and never hurts.

MLP-based architectures, which consist of a sequence of consecutive multi-layer perceptron blocks, have recently been found to reach comparable results to convolutional and transformer-based methods. However, most adopt spatial MLPs which take fixed dimension inputs, therefore making it difficult to apply them to downstream tasks, such as object detection and semantic segmentation. Moreover, single-stage designs further limit performance in other computer vision tasks and fully connected layers bear heavy computation. To tackle these problems, we propose ConvMLP: a hierarchical Convolutional MLP for visual recognition, which is a light-weight, stage-wise, co-design of convolution layers, and MLPs. In particular, ConvMLP-S achieves 76.8% top-1 accuracy on ImageNet-1k with 9M parameters and 2.4G MACs (15% and 19% of MLP-Mixer-B/16, respectively). Experiments on object detection and semantic segmentation further show that visual representation learned by ConvMLP can be seamlessly transferred and achieve competitive results with fewer parameters. Our code and pre-trained models are publicly available at https://github.com/SHI-Labs/Convolutional-MLPs.

We propose Unicoder-VL, a universal encoder that aims to learn joint representations of vision and language in a pre-training manner. Borrow ideas from cross-lingual pre-trained models, such as XLM and Unicoder, both visual and linguistic contents are fed into a multi-layer Transformer for the cross-modal pre-training, where three pre-trained tasks are employed, including Masked Language Modeling (MLM), Masked Object Classification (MOC) and Visual-linguistic Matching (VLM). The first two tasks learn context-aware representations for input tokens based on linguistic and visual contents jointly. The last task tries to predict whether an image and a text describe each other. After pretraining on large-scale image-caption pairs, we transfer Unicoder-VL to caption-based image-text retrieval and visual commonsense reasoning, with just one additional output layer. We achieve state-of-the-art or comparable results on both two tasks and show the powerful ability of the cross-modal pre-training.

This paper provides a simplification on OpenVision's architecture and loss design for enhancing its training efficiency. Following the prior vision-language pretraining works CapPa and AIMv2, as well as modern multimodal designs like LLaVA, our changes are straightforward: we remove the text encoder (and therefore the contrastive loss), retaining only the captioning loss as a purely generative training signal. We name this new version OpenVision 2. The initial results are promising: despite this simplification, OpenVision 2 competitively matches the original model's performance on a broad set of multimodal benchmarks while substantially cutting both training time and memory consumption. For example, with ViT-L/14, it reduces training time by about 1.5x (from 83h to 57h), and memory usage by about 1.8x (from 24.5GB to 13.8GB, equivalently allowing the maximum batch size to grow from 2k to 8k). This superior training efficiency also allows us to scale far beyond the largest vision encoder used in OpenVision, reaching more than 1 billion parameters. We hold a strong belief that this lightweight, generative-only paradigm is compelling for future vision encoder development in multimodal foundation models.

We propose a Deep Texture Encoding Network (Deep-TEN) with a novel Encoding Layer integrated on top of convolutional layers, which ports the entire dictionary learning and encoding pipeline into a single model. Current methods build from distinct components, using standard encoders with separate off-the-shelf features such as SIFT descriptors or pre-trained CNN features for material recognition. Our new approach provides an end-to-end learning framework, where the inherent visual vocabularies are learned directly from the loss function. The features, dictionaries and the encoding representation for the classifier are all learned simultaneously. The representation is orderless and therefore is particularly useful for material and texture recognition. The Encoding Layer generalizes robust residual encoders such as VLAD and Fisher Vectors, and has the property of discarding domain specific information which makes the learned convolutional features easier to transfer. Additionally, joint training using multiple datasets of varied sizes and class labels is supported resulting in increased recognition performance. The experimental results show superior performance as compared to state-of-the-art methods using gold-standard databases such as MINC-2500, Flickr Material Database, KTH-TIPS-2b, and two recent databases 4D-Light-Field-Material and GTOS. The source code for the complete system are publicly available.

Object hallucination in Large Vision-Language Models (LVLMs) significantly impedes their real-world applicability. As the primary component for accurately interpreting visual information, the choice of visual encoder is pivotal. We hypothesize that the diverse training paradigms employed by different visual encoders instill them with distinct inductive biases, which leads to their diverse hallucination performances. Existing benchmarks typically focus on coarse-grained hallucination detection and fail to capture the diverse hallucinations elaborated in our hypothesis. To systematically analyze these effects, we introduce VHBench-10, a comprehensive benchmark with approximately 10,000 samples for evaluating LVLMs across ten fine-grained hallucination categories. Our evaluations confirm encoders exhibit unique hallucination characteristics. Building on these insights and the suboptimality of simple feature fusion, we propose VisionWeaver, a novel Context-Aware Routing Network. It employs global visual features to generate routing signals, dynamically aggregating visual features from multiple specialized experts. Comprehensive experiments confirm the effectiveness of VisionWeaver in significantly reducing hallucinations and improving overall model performance.

Vision-Language Models (VLMs) encode images into lengthy sequences of visual tokens, leading to excessive computational overhead and limited inference efficiency. While prior efforts prune or merge tokens to address this issue, they often rely on special tokens (e.g., CLS) or require task-specific training, hindering scalability across architectures. In this paper, we propose HiPrune, a training-free and model-agnostic token Pruning framework that exploits the Hierarchical attention structure within vision encoders. We identify that middle layers attend to object-centric regions, while deep layers capture global contextual features. Based on this observation, HiPrune selects three types of informative tokens: (1) Anchor tokens with high attention in object-centric layers, (2) Buffer tokens adjacent to anchors for spatial continuity, and (3) Register tokens with strong attention in deep layers for global summarization. Our method requires no retraining and integrates seamlessly with any ViT-based VLM. Extensive experiments on LLaVA-1.5, LLaVA-NeXT, and Qwen2.5-VL demonstrate that HiPrune achieves state-of-the-art pruning performance, preserving up to 99.3% task accuracy with only 33.3% tokens, and maintaining 99.5% accuracy with just 11.1% tokens. Meanwhile, it reduces inference FLOPs and latency by up to 9times, showcasing strong generalization across models and tasks. Code is available at https://github.com/Danielement321/HiPrune.

Vision encoders typically generate a large number of visual tokens, providing information-rich representations but significantly increasing computational demands. This raises the question of whether all generated tokens are equally valuable or if some of them can be discarded to reduce computational costs without compromising quality. In this paper, we introduce a new method for determining feature utility based on the idea that less valuable features can be reconstructed from more valuable ones. We implement this concept by integrating an autoencoder with a Gumbel-Softmax selection mechanism, that allows identifying and retaining only the most informative visual tokens. To validate our approach, we compared the performance of the LLaVA-NeXT model, using features selected by our method with randomly selected features. We found that on OCR-based tasks, more than 50% of the visual context can be removed with minimal performance loss, whereas randomly discarding the same proportion of features significantly affects the model capabilities. Furthermore, in general-domain tasks, even randomly retaining only 30% of tokens achieves performance comparable to using the full set of visual tokens. Our results highlight a promising direction towards adaptive and efficient multimodal pruning that facilitates scalable and low-overhead inference without compromising performance.

Scene-agnostic visual inpainting remains very challenging despite progress in patch-based methods. Recently, Pathak et al. 2016 have introduced convolutional "context encoders" (CEs) for unsupervised feature learning through image completion tasks. With the additional help of adversarial training, CEs turned out to be a promising tool to complete complex structures in real inpainting problems. In the present paper we propose to push further this key ability by relying on perceptual reconstruction losses at training time. We show on a wide variety of visual scenes the merit of the approach for structural inpainting, and confirm it through a user study. Combined with the optimization-based refinement of Yang et al. 2016 with neural patches, our context encoder opens up new opportunities for prior-free visual inpainting.

The Position Embedding (PE) is critical for Vision Transformers (VTs) due to the permutation-invariance of self-attention operation. By analyzing the input and output of each encoder layer in VTs using reparameterization and visualization, we find that the default PE joining method (simply adding the PE and patch embedding together) operates the same affine transformation to token embedding and PE, which limits the expressiveness of PE and hence constrains the performance of VTs. To overcome this limitation, we propose a simple, effective, and robust method. Specifically, we provide two independent layer normalizations for token embeddings and PE for each layer, and add them together as the input of each layer's Muti-Head Self-Attention module. Since the method allows the model to adaptively adjust the information of PE for different layers, we name it as Layer-adaptive Position Embedding, abbreviated as LaPE. Extensive experiments demonstrate that LaPE can improve various VTs with different types of PE and make VTs robust to PE types. For example, LaPE improves 0.94% accuracy for ViT-Lite on Cifar10, 0.98% for CCT on Cifar100, and 1.72% for DeiT on ImageNet-1K, which is remarkable considering the negligible extra parameters, memory and computational cost brought by LaPE. The code is publicly available at https://github.com/Ingrid725/LaPE.

Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. "blindness" to certain image features, visual hallucination, etc. To address these issues, we study broadening the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs.

We present a novel masked image modeling (MIM) approach, context autoencoder (CAE), for self-supervised representation pretraining. We pretrain an encoder by making predictions in the encoded representation space. The pretraining tasks include two tasks: masked representation prediction - predict the representations for the masked patches, and masked patch reconstruction - reconstruct the masked patches. The network is an encoder-regressor-decoder architecture: the encoder takes the visible patches as input; the regressor predicts the representations of the masked patches, which are expected to be aligned with the representations computed from the encoder, using the representations of visible patches and the positions of visible and masked patches; the decoder reconstructs the masked patches from the predicted encoded representations. The CAE design encourages the separation of learning the encoder (representation) from completing the pertaining tasks: masked representation prediction and masked patch reconstruction tasks, and making predictions in the encoded representation space empirically shows the benefit to representation learning. We demonstrate the effectiveness of our CAE through superior transfer performance in downstream tasks: semantic segmentation, object detection and instance segmentation, and classification. The code will be available at https://github.com/Atten4Vis/CAE.

The recently developed and publicly available synthetic image generation methods and services make it possible to create extremely realistic imagery on demand, raising great risks for the integrity and safety of online information. State-of-the-art Synthetic Image Detection (SID) research has led to strong evidence on the advantages of feature extraction from foundation models. However, such extracted features mostly encapsulate high-level visual semantics instead of fine-grained details, which are more important for the SID task. On the contrary, shallow layers encode low-level visual information. In this work, we leverage the image representations extracted by intermediate Transformer blocks of CLIP's image-encoder via a lightweight network that maps them to a learnable forgery-aware vector space capable of generalizing exceptionally well. We also employ a trainable module to incorporate the importance of each Transformer block to the final prediction. Our method is compared against the state-of-the-art by evaluating it on 20 test datasets and exhibits an average +10.6% absolute performance improvement. Notably, the best performing models require just a single epoch for training (~8 minutes). Code available at https://github.com/mever-team/rine.

From extracting features to generating text, the outputs of large language models (LLMs) typically rely on their final layers, following the conventional wisdom that earlier layers capture only low-level cues. However, our analysis shows that intermediate layers can encode even richer representations, often improving performance on a wide range of downstream tasks. To explain and quantify these hidden-layer properties, we propose a unified framework of representation quality metrics based on information theory, geometry, and invariance to input perturbations. Our framework highlights how each model layer balances information compression and signal preservation, revealing why mid-depth embeddings can exceed the last layer's performance. Through extensive experiments on 32 text-embedding tasks and comparisons across model architectures (transformers, state-space models) and domains (language, vision), we demonstrate that intermediate layers consistently provide stronger features. These findings challenge the standard focus on final-layer embeddings and open new directions for model analysis and optimization, including strategic use of mid-layer representations for more robust and accurate AI systems.

Predicting salient regions in natural images requires the detection of objects that are present in a scene. To develop robust representations for this challenging task, high-level visual features at multiple spatial scales must be extracted and augmented with contextual information. However, existing models aimed at explaining human fixation maps do not incorporate such a mechanism explicitly. Here we propose an approach based on a convolutional neural network pre-trained on a large-scale image classification task. The architecture forms an encoder-decoder structure and includes a module with multiple convolutional layers at different dilation rates to capture multi-scale features in parallel. Moreover, we combine the resulting representations with global scene information for accurately predicting visual saliency. Our model achieves competitive and consistent results across multiple evaluation metrics on two public saliency benchmarks and we demonstrate the effectiveness of the suggested approach on five datasets and selected examples. Compared to state of the art approaches, the network is based on a lightweight image classification backbone and hence presents a suitable choice for applications with limited computational resources, such as (virtual) robotic systems, to estimate human fixations across complex natural scenes.

Medical image segmentation plays an important role in computer-aided diagnosis. Traditional convolution-based U-shape segmentation architectures are usually limited by the local receptive field. Existing vision transformers have been widely applied to diverse medical segmentation frameworks due to their superior capabilities of capturing global contexts. Despite the advantage, the real-world application of vision transformers is challenged by their non-linear self-attention mechanism, requiring huge computational costs. To address this issue, the selective state space model (SSM) Mamba has gained recognition for its adeptness in modeling long-range dependencies in sequential data, particularly noted for its efficient memory costs. In this paper, we propose MambaVesselNet++, a Hybrid CNN-Mamba framework for medical image segmentation. Our MambaVesselNet++ is comprised of a hybrid image encoder (Hi-Encoder) and a bifocal fusion decoder (BF-Decoder). In Hi-Encoder, we first devise the texture-aware layer to capture low-level semantic features by leveraging convolutions. Then, we utilize Mamba to effectively model long-range dependencies with linear complexity. The Bi-Decoder adopts skip connections to combine local and global information of the Hi-Encoder for the accurate generation of segmentation masks. Extensive experiments demonstrate that MambaVesselNet++ outperforms current convolution-based, transformer-based, and Mamba-based state-of-the-arts across diverse medical 2D, 3D, and instance segmentation tasks. The code is available at https://github.com/CC0117/MambaVesselNet.

Text-to-video generation has advanced rapidly, but existing methods typically output only the final composited video and lack editable layered representations, limiting their use in professional workflows. We propose LayerT2V, a unified multi-layer video generation framework that produces multiple semantically consistent outputs in a single inference pass: the full video, an independent background layer, and multiple foreground RGB layers with corresponding alpha mattes. Our key insight is that recent video generation backbones use high compression in both time and space, enabling us to serialize multiple layer representations along the temporal dimension and jointly model them on a shared generation trajectory. This turns cross-layer consistency into an intrinsic objective, improving semantic alignment and temporal coherence. To mitigate layer ambiguity and conditional leakage, we augment a shared DiT backbone with LayerAdaLN and layer-aware cross-attention modulation. LayerT2V is trained in three stages: alpha mask VAE adaptation, joint multi-layer learning, and multi-foreground extension. We also introduce VidLayer, the first large-scale dataset for multi-layer video generation. Extensive experiments demonstrate that LayerT2V substantially outperforms prior methods in visual fidelity, temporal consistency, and cross-layer coherence.

OpenAI's CLIP, released in early 2021, have long been the go-to choice of vision encoder for building multimodal foundation models. Although recent alternatives such as SigLIP have begun to challenge this status quo, to our knowledge none are fully open: their training data remains proprietary and/or their training recipes are not released. This paper fills this gap with OpenVision, a fully-open, cost-effective family of vision encoders that match or surpass the performance of OpenAI's CLIP when integrated into multimodal frameworks like LLaVA. OpenVision builds on existing works -- e.g., CLIPS for training framework and Recap-DataComp-1B for training data -- while revealing multiple key insights in enhancing encoder quality and showcasing practical benefits in advancing multimodal models. By releasing vision encoders spanning from 5.9M to 632.1M parameters, OpenVision offers practitioners a flexible trade-off between capacity and efficiency in building multimodal models: larger models deliver enhanced multimodal performance, while smaller versions enable lightweight, edge-ready multimodal deployments.