Daily Papers

General-purpose robots require decision-making models that generalize across diverse tasks and environments. Recent works build robot foundation models by extending multimodal large language models (MLLMs) with action outputs, creating vision-language-action (VLA) systems. These efforts are motivated by the intuition that MLLMs' large-scale language and image pretraining can be effectively transferred to the action output modality. In this work, we explore an alternative paradigm of using large-scale video pretraining as a primary modality for building robot foundation models. Unlike static images and language, videos capture spatio-temporal sequences of states and actions in the physical world that are naturally aligned with robotic behavior. We curate an internet-scale video dataset of human activities and task demonstrations, and train, for the first time at a foundation-model scale, an open video model for generative robotics planning. The model produces zero-shot video plans for novel scenes and tasks, which we post-process to extract executable robot actions. We evaluate task-level generalization through third-party selected tasks in the wild and real-robot experiments, demonstrating successful physical execution. Together, these results show robust instruction following, strong generalization, and real-world feasibility. We release both the model and dataset to support open, reproducible video-based robot learning. Our website is available at https://www.boyuan.space/large-video-planner/.

Long-form video content constitutes a significant portion of internet traffic, making automated video summarization an essential research problem. However, existing video summarization datasets are notably limited in their size, constraining the effectiveness of state-of-the-art methods for generalization. Our work aims to overcome this limitation by capitalizing on the abundance of long-form videos with dense speech-to-video alignment and the remarkable capabilities of recent large language models (LLMs) in summarizing long text. We introduce an automated and scalable pipeline for generating a large-scale video summarization dataset using LLMs as Oracle summarizers. By leveraging the generated dataset, we analyze the limitations of existing approaches and propose a new video summarization model that effectively addresses them. To facilitate further research in the field, our work also presents a new benchmark dataset that contains 1200 long videos each with high-quality summaries annotated by professionals. Extensive experiments clearly indicate that our proposed approach sets a new state-of-the-art in video summarization across several benchmarks.

Large-scale video-text pretraining achieves strong performance but depends on noisy, synthetic captions with limited semantic coverage, often overlooking implicit world knowledge such as object motion, 3D geometry, and physical cues. In contrast, masked video modeling (MVM) directly exploits spatiotemporal structures but trails text-supervised methods on general tasks. We find this gap arises from overlooked architectural issues: pixel-level reconstruction struggles with convergence and its low-level requirement often conflicts with semantics, while latent prediction often encourages shortcut learning. To address these, we disentangle the traditional encoder-decoder design into an Encoder-Predictor-Decoder (EPD) framework, where the predictor acts as a latent world model, and propose InternVideo-Next, a two-stage pretraining scheme that builds a semantically consistent yet detail-preserving latent space for this world model. First, conventional linear decoder in pixel MVM enforces the predictor output latent to be linearly projected to, thus separable in pixel space, causing the conflict with semantic abstraction. Our Stage 1 proposes a conditional diffusion decoder and injects reliable image-level semantic priors to enhance semantics and convergence, thus bridging pixel-level fidelity with high-level semantic abstraction. Stage 2 further learns world knowledge by predicting frozen Stage 1 targets within this space, mitigating shortcut learning. Trained on public, unlabeled videos, InternVideo-Next achieves state-of-the-art results across benchmarks and provides a scalable path toward general video representation learning.

This paper presents RynnVLA-001, a vision-language-action(VLA) model built upon large-scale video generative pretraining from human demonstrations. We propose a novel two-stage pretraining methodology. The first stage, Ego-Centric Video Generative Pretraining, trains an Image-to-Video model on 12M ego-centric manipulation videos to predict future frames conditioned on an initial frame and a language instruction. The second stage, Human-Centric Trajectory-Aware Modeling, extends this by jointly predicting future keypoint trajectories, thereby effectively bridging visual frame prediction with action prediction. Furthermore, to enhance action representation, we propose ActionVAE, a variational autoencoder that compresses sequences of actions into compact latent embeddings, reducing the complexity of the VLA output space. When finetuned on the same downstream robotics datasets, RynnVLA-001 achieves superior performance over state-of-the-art baselines, demonstrating that the proposed pretraining strategy provides a more effective initialization for VLA models.

Video-language pre-trained models have shown remarkable success in guiding video question-answering (VideoQA) tasks. However, due to the length of video sequences, training large-scale video-based models incurs considerably higher costs than training image-based ones. This motivates us to leverage the knowledge from image-based pretraining, despite the obvious gaps between image and video domains. To bridge these gaps, in this paper, we propose Tem-Adapter, which enables the learning of temporal dynamics and complex semantics by a visual Temporal Aligner and a textual Semantic Aligner. Unlike conventional pretrained knowledge adaptation methods that only concentrate on the downstream task objective, the Temporal Aligner introduces an extra language-guided autoregressive task aimed at facilitating the learning of temporal dependencies, with the objective of predicting future states based on historical clues and language guidance that describes event progression. Besides, to reduce the semantic gap and adapt the textual representation for better event description, we introduce a Semantic Aligner that first designs a template to fuse question and answer pairs as event descriptions and then learns a Transformer decoder with the whole video sequence as guidance for refinement. We evaluate Tem-Adapter and different pre-train transferring methods on two VideoQA benchmarks, and the significant performance improvement demonstrates the effectiveness of our method.

Non-human primates are our closest living relatives, and analyzing their behavior is central to research in cognition, evolution, and conservation. Computer vision could greatly aid this research, but existing methods often rely on human-centric pretrained models and focus on single datasets, which limits generalization. We address this limitation by shifting from a model-centric to a data-centric approach and introduce PriVi, a large-scale primate-centric video pretraining dataset. PriVi contains 424 hours of curated video, combining 174 hours from behavioral research across 11 settings with 250 hours of diverse web-sourced footage, assembled through a scalable data curation pipeline. We pretrain V-JEPA on PriVi to learn primate-specific representations and evaluate it using a lightweight frozen classifier. Across four benchmark datasets, ChimpACT, BaboonLand, PanAf500, and ChimpBehave, our approach consistently outperforms prior work, including fully finetuned baselines, and scales favorably with fewer labels. These results demonstrate that primate-centric pretraining substantially improves data efficiency and generalization, making it a promising approach for low-label applications. Code, models, and the majority of the dataset will be made available.

We introduce Human-like Video Models (HVM-1), large-scale video models pretrained with nearly 5000 hours of curated human-like video data (mostly egocentric, temporally extended, continuous video recordings), using the spatiotemporal masked autoencoder (ST-MAE) algorithm. We release two 633M parameter models trained at spatial resolutions of 224x224 and 448x448 pixels. We evaluate the performance of these models in downstream few-shot video and image recognition tasks and compare them against a model pretrained with 1330 hours of short action-oriented video clips from YouTube (Kinetics-700). HVM-1 models perform competitively against the Kinetics-700 pretrained model in downstream evaluations despite substantial qualitative differences between the spatiotemporal characteristics of the corresponding pretraining datasets. HVM-1 models also learn more accurate and more robust object representations compared to models pretrained with the image-based MAE algorithm on the same data, demonstrating the potential benefits of learning to predict temporal regularities in natural videos for learning better object representations.

Large language models (LLMs) have demonstrated that large-scale pretraining enables systems to adapt rapidly to new problems with little supervision in the language domain. This success, however, has not translated as effectively to the visual domain, where models, including LLMs, continue to struggle with compositional understanding, sample efficiency, and general-purpose problem-solving. We investigate Video Diffusion Models (VDMs) as a promising direction for bridging this gap. Pretraining on spatiotemporal data endows these models with strong inductive biases for structure and dynamics, which we hypothesize can support broad task adaptability. To test this, we design a controlled evaluation in which both a pretrained LLM and a pretrained VDM are equipped with lightweight adapters and presented with tasks in their natural modalities. Across benchmarks including ARC-AGI, ConceptARC, visual games, route planning, and cellular automata, VDMs demonstrate higher data efficiency than their language counterparts. Taken together, our results indicate that video pretraining offers inductive biases that support progress toward visual foundation models.

Video-text retrieval plays an essential role in multi-modal research and has been widely used in many real-world web applications. The CLIP (Contrastive Language-Image Pre-training), an image-language pre-training model, has demonstrated the power of visual concepts learning from web collected image-text datasets. In this paper, we propose a CLIP4Clip model to transfer the knowledge of the CLIP model to video-language retrieval in an end-to-end manner. Several questions are investigated via empirical studies: 1) Whether image feature is enough for video-text retrieval? 2) How a post-pretraining on a large-scale video-text dataset based on the CLIP affect the performance? 3) What is the practical mechanism to model temporal dependency between video frames? And 4) The Hyper-parameters sensitivity of the model on video-text retrieval task. Extensive experimental results present that the CLIP4Clip model transferred from the CLIP can achieve SOTA results on various video-text retrieval datasets, including MSR-VTT, MSVC, LSMDC, ActivityNet, and DiDeMo. We release our code at https://github.com/ArrowLuo/CLIP4Clip.

Video Large Language Models (Video LLMs) have achieved significant success by leveraging a two-stage paradigm: pretraining on large-scale video-text data for vision-language alignment, followed by supervised fine-tuning (SFT) for task-specific capabilities. However, existing approaches struggle with temporal reasoning due to weak temporal correspondence in the data and reliance on the next-token prediction paradigm during training. To address these limitations, we propose TEMPLE (TEMporal Preference Learning), a systematic framework that enhances Video LLMs' temporal reasoning capabilities through Direct Preference Optimization (DPO). To facilitate this, we introduce an automated preference data generation pipeline that systematically constructs preference pairs by selecting videos that are rich in temporal information, designing video-specific perturbation strategies, and finally evaluating model responses on clean and perturbed video inputs. Our temporal alignment features two key innovations: curriculum learning which that progressively increases perturbation difficulty to improve model robustness and adaptability; and "Pre-SFT Alignment'', applying preference optimization before instruction tuning to prioritize fine-grained temporal comprehension. Extensive experiments demonstrate that our approach consistently improves Video LLM performance across multiple benchmarks with a relatively small set of self-generated DPO data. We further analyze the transferability of DPO data across architectures and the role of difficulty scheduling in optimization. Our findings highlight our TEMPLE as a scalable and efficient complement to SFT-based methods, paving the way for developing reliable Video LLMs. Code is available at https://github.com/lscpku/TEMPLE.

Multi-modal Large Language Models (MLLMs) have demonstrated their ability to perceive objects in still images, but their application in video-related tasks, such as object tracking, remains understudied. This lack of exploration is primarily due to two key challenges. Firstly, extensive pretraining on large-scale video datasets is required to equip MLLMs with the capability to perceive objects across multiple frames and understand inter-frame relationships. Secondly, processing a large number of frames within the context window of Large Language Models (LLMs) can impose a significant computational burden. To address the first challenge, we introduce ElysiumTrack-1M, a large-scale video dataset supported for three tasks: Single Object Tracking (SOT), Referring Single Object Tracking (RSOT), and Video Referring Expression Generation (Video-REG). ElysiumTrack-1M contains 1.27 million annotated video frames with corresponding object boxes and descriptions. Leveraging this dataset, we conduct training of MLLMs and propose a token-compression model T-Selector to tackle the second challenge. Our proposed approach, Elysium: Exploring Object-level Perception in Videos via MLLM, is an end-to-end trainable MLLM that attempts to conduct object-level tasks in videos without requiring any additional plug-in or expert models. All codes and datasets are available at https://github.com/Hon-Wong/Elysium.

Generating animatable human avatars from a single image is essential for various digital human modeling applications. Existing 3D reconstruction methods often struggle to capture fine details in animatable models, while generative approaches for controllable animation, though avoiding explicit 3D modeling, suffer from viewpoint inconsistencies in extreme poses and computational inefficiencies. In this paper, we address these challenges by leveraging the power of generative models to produce detailed multi-view canonical pose images, which help resolve ambiguities in animatable human reconstruction. We then propose a robust method for 3D reconstruction of inconsistent images, enabling real-time rendering during inference. Specifically, we adapt a transformer-based video generation model to generate multi-view canonical pose images and normal maps, pretraining on a large-scale video dataset to improve generalization. To handle view inconsistencies, we recast the reconstruction problem as a 4D task and introduce an efficient 3D modeling approach using 4D Gaussian Splatting. Experiments demonstrate that our method achieves photorealistic, real-time animation of 3D human avatars from in-the-wild images, showcasing its effectiveness and generalization capability.

To promote the development of Vision-Language Pre-training (VLP) and multimodal Large Language Model (LLM) in the Chinese community, we firstly release the largest public Chinese high-quality video-language dataset named Youku-mPLUG, which is collected from Youku, a well-known Chinese video-sharing website, with strict criteria of safety, diversity, and quality. Youku-mPLUG contains 10 million Chinese video-text pairs filtered from 400 million raw videos across a wide range of 45 diverse categories for large-scale pre-training. In addition, to facilitate a comprehensive evaluation of video-language models, we carefully build the largest human-annotated Chinese benchmarks covering three popular video-language tasks of cross-modal retrieval, video captioning, and video category classification. Youku-mPLUG can enable researchers to conduct more in-depth multimodal research and develop better applications in the future. Furthermore, we release popular video-language pre-training models, ALPRO and mPLUG-2, and our proposed modularized decoder-only model mPLUG-video pre-trained on Youku-mPLUG. Experiments show that models pre-trained on Youku-mPLUG gain up to 23.1% improvement in video category classification. Besides, mPLUG-video achieves a new state-of-the-art result on these benchmarks with 80.5% top-1 accuracy in video category classification and 68.9 CIDEr score in video captioning, respectively. Finally, we scale up mPLUG-video based on the frozen Bloomz with only 1.7% trainable parameters as Chinese multimodal LLM, and demonstrate impressive instruction and video understanding ability. The zero-shot instruction understanding experiment indicates that pretraining with Youku-mPLUG can enhance the ability to comprehend overall and detailed visual semantics, recognize scene text, and leverage open-domain knowledge.

Video-Language Pretraining (VLP), which aims to learn transferable representation to advance a wide range of video-text downstream tasks, has recently received increasing attention. Best performing works rely on large-scale, 3rd-person video-text datasets, such as HowTo100M. In this work, we exploit the recently released Ego4D dataset to pioneer Egocentric VLP along three directions. (i) We create EgoClip, a 1st-person video-text pretraining dataset comprising 3.8M clip-text pairs well-chosen from Ego4D, covering a large variety of human daily activities. (ii) We propose a novel pretraining objective, dubbed EgoNCE, which adapts video-text contrastive learning to the egocentric domain by mining egocentric-aware positive and negative samples. (iii) We introduce EgoMCQ, a development benchmark that is close to EgoClip and hence can support effective validation and fast exploration of our design decisions in EgoClip and EgoNCE. Furthermore, we demonstrate strong performance on five egocentric downstream tasks across three datasets: video-text retrieval on EPIC-KITCHENS-100; action recognition on Charades-Ego; natural language query, moment query, and object state change classification on Ego4D challenge benchmarks. The dataset and code are available at https://github.com/showlab/EgoVLP.

Surgical practice involves complex visual interpretation, procedural skills, and advanced medical knowledge, making surgical vision-language pretraining (VLP) particularly challenging due to this complexity and the limited availability of annotated data. To address the gap, we propose OphCLIP, a hierarchical retrieval-augmented vision-language pretraining framework specifically designed for ophthalmic surgical workflow understanding. OphCLIP leverages the OphVL dataset we constructed, a large-scale and comprehensive collection of over 375K hierarchically structured video-text pairs with tens of thousands of different combinations of attributes (surgeries, phases/operations/actions, instruments, medications, as well as more advanced aspects like the causes of eye diseases, surgical objectives, and postoperative recovery recommendations, etc). These hierarchical video-text correspondences enable OphCLIP to learn both fine-grained and long-term visual representations by aligning short video clips with detailed narrative descriptions and full videos with structured titles, capturing intricate surgical details and high-level procedural insights, respectively. Our OphCLIP also designs a retrieval-augmented pretraining framework to leverage the underexplored large-scale silent surgical procedure videos, automatically retrieving semantically relevant content to enhance the representation learning of narrative videos. Evaluation across 11 datasets for phase recognition and multi-instrument identification shows OphCLIP's robust generalization and superior performance.

The task of multimodal learning has seen a growing interest recently as it allows for training neural architectures based on different modalities such as vision, text, and audio. One challenge in training such models is that they need to jointly learn semantic concepts and their relationships across different input representations. Capsule networks have been shown to perform well in context of capturing the relation between low-level input features and higher-level concepts. However, capsules have so far mainly been used only in small-scale fully supervised settings due to the resource demand of conventional routing algorithms. We present a new multimodal capsule network that allows us to leverage the strength of capsules in the context of a multimodal learning framework on large amounts of video data. To adapt the capsules to large-scale input data, we propose a novel routing by self-attention mechanism that selects relevant capsules which are then used to generate a final joint multimodal feature representation. This allows not only for robust training with noisy video data, but also to scale up the size of the capsule network compared to traditional routing methods while still being computationally efficient. We evaluate the proposed architecture by pretraining it on a large-scale multimodal video dataset and applying it on four datasets in two challenging downstream tasks. Results show that the proposed multimodal capsule network is not only able to improve results compared to other routing techniques, but also achieves competitive performance on the task of multimodal learning.

Segmenting long videos into chapters enables users to quickly navigate to the information of their interest. This important topic has been understudied due to the lack of publicly released datasets. To address this issue, we present VidChapters-7M, a dataset of 817K user-chaptered videos including 7M chapters in total. VidChapters-7M is automatically created from videos online in a scalable manner by scraping user-annotated chapters and hence without any additional manual annotation. We introduce the following three tasks based on this data. First, the video chapter generation task consists of temporally segmenting the video and generating a chapter title for each segment. To further dissect the problem, we also define two variants of this task: video chapter generation given ground-truth boundaries, which requires generating a chapter title given an annotated video segment, and video chapter grounding, which requires temporally localizing a chapter given its annotated title. We benchmark both simple baselines and state-of-the-art video-language models for these three tasks. We also show that pretraining on VidChapters-7M transfers well to dense video captioning tasks in both zero-shot and finetuning settings, largely improving the state of the art on the YouCook2 and ViTT benchmarks. Finally, our experiments reveal that downstream performance scales well with the size of the pretraining dataset. Our dataset, code, and models are publicly available at https://antoyang.github.io/vidchapters.html.

We introduce PlayerOne, the first egocentric realistic world simulator, facilitating immersive and unrestricted exploration within vividly dynamic environments. Given an egocentric scene image from the user, PlayerOne can accurately construct the corresponding world and generate egocentric videos that are strictly aligned with the real scene human motion of the user captured by an exocentric camera. PlayerOne is trained in a coarse-to-fine pipeline that first performs pretraining on large-scale egocentric text-video pairs for coarse-level egocentric understanding, followed by finetuning on synchronous motion-video data extracted from egocentric-exocentric video datasets with our automatic construction pipeline. Besides, considering the varying importance of different components, we design a part-disentangled motion injection scheme, enabling precise control of part-level movements. In addition, we devise a joint reconstruction framework that progressively models both the 4D scene and video frames, ensuring scene consistency in the long-form video generation. Experimental results demonstrate its great generalization ability in precise control of varying human movements and worldconsistent modeling of diverse scenarios. It marks the first endeavor into egocentric real-world simulation and can pave the way for the community to delve into fresh frontiers of world modeling and its diverse applications.

Large-scale pretrained transformers have created milestones in text (GPT-3) and text-to-image (DALL-E and CogView) generation. Its application to video generation is still facing many challenges: The potential huge computation cost makes the training from scratch unaffordable; The scarcity and weak relevance of text-video datasets hinder the model understanding complex movement semantics. In this work, we present 9B-parameter transformer CogVideo, trained by inheriting a pretrained text-to-image model, CogView2. We also propose multi-frame-rate hierarchical training strategy to better align text and video clips. As (probably) the first open-source large-scale pretrained text-to-video model, CogVideo outperforms all publicly available models at a large margin in machine and human evaluations.

In this work, we introduce Vid2Seq, a multi-modal single-stage dense event captioning model pretrained on narrated videos which are readily-available at scale. The Vid2Seq architecture augments a language model with special time tokens, allowing it to seamlessly predict event boundaries and textual descriptions in the same output sequence. Such a unified model requires large-scale training data, which is not available in current annotated datasets. We show that it is possible to leverage unlabeled narrated videos for dense video captioning, by reformulating sentence boundaries of transcribed speech as pseudo event boundaries, and using the transcribed speech sentences as pseudo event captions. The resulting Vid2Seq model pretrained on the YT-Temporal-1B dataset improves the state of the art on a variety of dense video captioning benchmarks including YouCook2, ViTT and ActivityNet Captions. Vid2Seq also generalizes well to the tasks of video paragraph captioning and video clip captioning, and to few-shot settings. Our code is publicly available at https://antoyang.github.io/vid2seq.html.

Self-supervised learning has emerged as a powerful paradigm for label-free model pretraining, particularly in the video domain, where manual annotation is costly and time-intensive. However, existing self-supervised approaches employ diverse experimental setups, making direct comparisons challenging due to the absence of a standardized benchmark. In this work, we establish a unified benchmark that enables fair comparisons across different methods. Additionally, we systematically investigate five critical aspects of self-supervised learning in videos: (1) dataset size, (2) model complexity, (3) data distribution, (4) data noise, and (5) feature representations. To facilitate this study, we evaluate six self-supervised learning methods across six network architectures, conducting extensive experiments on five benchmark datasets and assessing performance on two distinct downstream tasks. Our analysis reveals key insights into the interplay between pretraining strategies, dataset characteristics, pretext tasks, and model architectures. Furthermore, we extend these findings to Video Foundation Models (ViFMs), demonstrating their relevance in large-scale video representation learning. Finally, leveraging these insights, we propose a novel approach that significantly reduces training data requirements while surpassing state-of-the-art methods that rely on 10% more pretraining data. We believe this work will guide future research toward a deeper understanding of self-supervised video representation learning and its broader implications.

Imitation learning has emerged as a promising approach towards building generalist robots. However, scaling imitation learning for large robot foundation models remains challenging due to its reliance on high-quality expert demonstrations. Meanwhile, large amounts of video data depicting a wide range of environments and diverse behaviors are readily available. This data provides a rich source of information about real-world dynamics and agent-environment interactions. Leveraging this data directly for imitation learning, however, has proven difficult due to the lack of action annotation required for most contemporary methods. In this work, we present Unified World Models (UWM), a framework that allows for leveraging both video and action data for policy learning. Specifically, a UWM integrates an action diffusion process and a video diffusion process within a unified transformer architecture, where independent diffusion timesteps govern each modality. We show that by simply controlling each diffusion timestep, UWM can flexibly represent a policy, a forward dynamics, an inverse dynamics, and a video generator. Through simulated and real-world experiments, we show that: (1) UWM enables effective pretraining on large-scale multitask robot datasets with both dynamics and action predictions, resulting in more generalizable and robust policies than imitation learning, (2) UWM naturally facilitates learning from action-free video data through independent control of modality-specific diffusion timesteps, further improving the performance of finetuned policies. Our results suggest that UWM offers a promising step toward harnessing large, heterogeneous datasets for scalable robot learning, and provides a simple unification between the often disparate paradigms of imitation learning and world modeling. Videos and code are available at https://weirdlabuw.github.io/uwm/.

Recent advancements in visual generation technologies have markedly increased the scale and availability of video datasets, which are crucial for training effective video generation models. However, a significant lack of high-quality, human-centric video datasets presents a challenge to progress in this field. To bridge this gap, we introduce OpenHumanVid, a large-scale and high-quality human-centric video dataset characterized by precise and detailed captions that encompass both human appearance and motion states, along with supplementary human motion conditions, including skeleton sequences and speech audio. To validate the efficacy of this dataset and the associated training strategies, we propose an extension of existing classical diffusion transformer architectures and conduct further pretraining of our models on the proposed dataset. Our findings yield two critical insights: First, the incorporation of a large-scale, high-quality dataset substantially enhances evaluation metrics for generated human videos while preserving performance in general video generation tasks. Second, the effective alignment of text with human appearance, human motion, and facial motion is essential for producing high-quality video outputs. Based on these insights and corresponding methodologies, the straightforward extended network trained on the proposed dataset demonstrates an obvious improvement in the generation of human-centric videos. Project page https://fudan-generative-vision.github.io/OpenHumanVid

To perform automatic family audio analysis, past studies have collected recordings using phone, video, or audio-only recording devices like LENA, investigated supervised learning methods, and used or fine-tuned general-purpose embeddings learned from large pretrained models. In this study, we advance the audio component of a new infant wearable multi-modal device called LittleBeats (LB) by learning family audio representation via wav2vec 2.0 (W2V2) pertaining. We show given a limited number of labeled LB home recordings, W2V2 pretrained using 1k-hour of unlabeled home recordings outperforms oracle W2V2 pretrained on 52k-hour unlabeled audio in terms of parent/infant speaker diarization (SD) and vocalization classifications (VC) at home. Extra relevant external unlabeled and labeled data further benefit W2V2 pretraining and fine-tuning. With SpecAug and environmental speech corruptions, we obtain 12% relative gain on SD and moderate boost on VC. Code and model weights are available.

Computer vision for animal behavior offers promising tools to aid research in ecology, cognition, and to support conservation efforts. Video camera traps allow for large-scale data collection, but high labeling costs remain a bottleneck to creating large-scale datasets. We thus need data-efficient learning approaches. In this work, we show that we can utilize self-supervised learning to considerably improve action recognition on primate behavior. On two datasets of great ape behavior (PanAf and ChimpACT), we outperform published state-of-the-art action recognition models by 6.1 %pt. accuracy and 6.3 %pt. mAP, respectively. We achieve this by utilizing a pretrained V-JEPA model and applying domain-adaptive pretraining (DAP), i.e. continuing the pretraining with in-domain data. We show that most of the performance gain stems from the DAP. Our method promises great potential for improving the recognition of animal behavior, as DAP does not require labeled samples. Code is available at https://github.com/ecker-lab/dap-behavior

Procedural video representation learning is an active research area where the objective is to learn an agent which can anticipate and forecast the future given the present video input, typically in conjunction with textual annotations. Prior works often rely on large-scale pretraining of visual encoders and prediction models with language supervision. However, the necessity and effectiveness of extending compute intensive pretraining to learn video clip sequences with noisy text supervision have not yet been fully validated by previous works. In this work, we show that a strong off-the-shelf frozen pretrained visual encoder, along with a well designed prediction model, can achieve state-of-the-art (SoTA) performance in forecasting and procedural planning without the need for pretraining the prediction model, nor requiring additional supervision from language or ASR. Instead of learning representations from pixel space, our method utilizes the latent embedding space of publicly available vision encoders. By conditioning on frozen clip-level embeddings from observed steps to predict the actions of unseen steps, our prediction model is able to learn robust representations for forecasting through iterative denoising - leveraging the recent advances in diffusion transformers (Peebles & Xie, 2023). Empirical studies over a total of five procedural learning tasks across four datasets (NIV, CrossTask, COIN and Ego4D-v2) show that our model advances the strong baselines in long-horizon action anticipation (+2.6% in Verb ED@20, +3.1% in Noun ED@20), and significantly improves the SoTA in step forecasting (+5.0%), task classification (+3.8%), and procedure planning tasks (up to +2.28% in success rate, +3.39% in mAcc, and +0.90% in mIoU).

Recent advances in pretraining general foundation models have significantly improved performance across diverse downstream tasks. While autoregressive (AR) generative models like GPT have revolutionized NLP, most visual generative pretraining methods still rely on BERT-style masked modeling, which often disregards the temporal information essential for video analysis. The few existing autoregressive visual pretraining methods suffer from issues such as inaccurate semantic localization and poor generation quality, leading to poor semantics. In this work, we propose NExT-Vid, a novel autoregressive visual generative pretraining framework that utilizes masked next-frame prediction to jointly model images and videos. NExT-Vid introduces a context-isolated autoregressive predictor to decouple semantic representation from target decoding, and a conditioned flow-matching decoder to enhance generation quality and diversity. Through context-isolated flow-matching pretraining, our approach achieves strong representations. Extensive experiments on large-scale pretrained models demonstrate that our proposed method consistently outperforms previous generative pretraining methods for visual representation learning via attentive probing in downstream classification.

Video-text retrieval is a class of cross-modal representation learning problems, where the goal is to select the video which corresponds to the text query between a given text query and a pool of candidate videos. The contrastive paradigm of vision-language pretraining has shown promising success with large-scale datasets and unified transformer architecture, and demonstrated the power of a joint latent space. Despite this, the intrinsic divergence between the visual domain and textual domain is still far from being eliminated, and projecting different modalities into a joint latent space might result in the distorting of the information inside the single modality. To overcome the above issue, we present a novel mechanism for learning the translation relationship from a source modality space S to a target modality space T without the need for a joint latent space, which bridges the gap between visual and textual domains. Furthermore, to keep cycle consistency between translations, we adopt a cycle loss involving both forward translations from S to the predicted target space T', and backward translations from T' back to S. Extensive experiments conducted on MSR-VTT, MSVD, and DiDeMo datasets demonstrate the superiority and effectiveness of our LaT approach compared with vanilla state-of-the-art methods.

Video editing using diffusion models has achieved remarkable results in generating high-quality edits for videos. However, current methods often rely on large-scale pretraining, limiting flexibility for specific edits. First-frame-guided editing provides control over the first frame, but lacks flexibility over subsequent frames. To address this, we propose a mask-based LoRA (Low-Rank Adaptation) tuning method that adapts pretrained Image-to-Video (I2V) models for flexible video editing. Our approach preserves background regions while enabling controllable edits propagation. This solution offers efficient and adaptable video editing without altering the model architecture. To better steer this process, we incorporate additional references, such as alternate viewpoints or representative scene states, which serve as visual anchors for how content should unfold. We address the control challenge using a mask-driven LoRA tuning strategy that adapts a pre-trained image-to-video model to the editing context. The model must learn from two distinct sources: the input video provides spatial structure and motion cues, while reference images offer appearance guidance. A spatial mask enables region-specific learning by dynamically modulating what the model attends to, ensuring that each area draws from the appropriate source. Experimental results show our method achieves superior video editing performance compared to state-of-the-art methods.

Recently, image-based Large Multimodal Models (LMMs) have made significant progress in video question-answering (VideoQA) using a frame-wise approach by leveraging large-scale pretraining in a zero-shot manner. Nevertheless, these models need to be capable of finding relevant information, extracting it, and answering the question simultaneously. Currently, existing methods perform all of these steps in a single pass without being able to adapt if insufficient or incorrect information is collected. To overcome this, we introduce a modular multi-LMM agent framework based on several agents with different roles, instructed by a Planner agent that updates its instructions using shared feedback from the other agents. Specifically, we propose TraveLER, a method that can create a plan to "Traverse" through the video, ask questions about individual frames to "Locate" and store key information, and then "Evaluate" if there is enough information to answer the question. Finally, if there is not enough information, our method is able to "Replan" based on its collected knowledge. Through extensive experiments, we find that the proposed TraveLER approach improves performance on several VideoQA benchmarks without the need to fine-tune on specific datasets. Our code is available at https://github.com/traveler-framework/TraveLER.

Video Temporal Grounding (VTG), which aims to ground target clips from videos (such as consecutive intervals or disjoint shots) according to custom language queries (e.g., sentences or words), is key for video browsing on social media. Most methods in this direction develop taskspecific models that are trained with type-specific labels, such as moment retrieval (time interval) and highlight detection (worthiness curve), which limits their abilities to generalize to various VTG tasks and labels. In this paper, we propose to Unify the diverse VTG labels and tasks, dubbed UniVTG, along three directions: Firstly, we revisit a wide range of VTG labels and tasks and define a unified formulation. Based on this, we develop data annotation schemes to create scalable pseudo supervision. Secondly, we develop an effective and flexible grounding model capable of addressing each task and making full use of each label. Lastly, thanks to the unified framework, we are able to unlock temporal grounding pretraining from large-scale diverse labels and develop stronger grounding abilities e.g., zero-shot grounding. Extensive experiments on three tasks (moment retrieval, highlight detection and video summarization) across seven datasets (QVHighlights, Charades-STA, TACoS, Ego4D, YouTube Highlights, TVSum, and QFVS) demonstrate the effectiveness and flexibility of our proposed framework. The codes are available at https://github.com/showlab/UniVTG.

Video understanding requires reasoning at multiple spatiotemporal resolutions -- from short fine-grained motions to events taking place over longer durations. Although transformer architectures have recently advanced the state-of-the-art, they have not explicitly modelled different spatiotemporal resolutions. To this end, we present Multiview Transformers for Video Recognition (MTV). Our model consists of separate encoders to represent different views of the input video with lateral connections to fuse information across views. We present thorough ablation studies of our model and show that MTV consistently performs better than single-view counterparts in terms of accuracy and computational cost across a range of model sizes. Furthermore, we achieve state-of-the-art results on six standard datasets, and improve even further with large-scale pretraining. Code and checkpoints are available at: https://github.com/google-research/scenic/tree/main/scenic/projects/mtv.

In recent years, considerable progress on the task of text-video retrieval has been achieved by leveraging large-scale pretraining on visual and audio datasets to construct powerful video encoders. By contrast, despite the natural symmetry, the design of effective algorithms for exploiting large-scale language pretraining remains under-explored. In this work, we are the first to investigate the design of such algorithms and propose a novel generalized distillation method, TeachText, which leverages complementary cues from multiple text encoders to provide an enhanced supervisory signal to the retrieval model. Moreover, we extend our method to video side modalities and show that we can effectively reduce the number of used modalities at test time without compromising performance. Our approach advances the state of the art on several video retrieval benchmarks by a significant margin and adds no computational overhead at test time. Last but not least, we show an effective application of our method for eliminating noise from retrieval datasets. Code and data can be found at https://www.robots.ox.ac.uk/~vgg/research/teachtext/.

World models simulate future states of the world in response to different actions. They facilitate interactive content creation and provides a foundation for grounded, long-horizon reasoning. Current foundation models do not fully meet the capabilities of general world models: large language models (LLMs) are constrained by their reliance on language modality and their limited understanding of the physical world, while video models lack interactive action control over the world simulations. This paper makes a step towards building a general world model by introducing Pandora, a hybrid autoregressive-diffusion model that simulates world states by generating videos and allows real-time control with free-text actions. Pandora achieves domain generality, video consistency, and controllability through large-scale pretraining and instruction tuning. Crucially, Pandora bypasses the cost of training-from-scratch by integrating a pretrained LLM (7B) and a pretrained video model, requiring only additional lightweight finetuning. We illustrate extensive outputs by Pandora across diverse domains (indoor/outdoor, natural/urban, human/robot, 2D/3D, etc.). The results indicate great potential of building stronger general world models with larger-scale training.

In this paper, we propose VideoLLaMA3, a more advanced multimodal foundation model for image and video understanding. The core design philosophy of VideoLLaMA3 is vision-centric. The meaning of "vision-centric" is two-fold: the vision-centric training paradigm and vision-centric framework design. The key insight of our vision-centric training paradigm is that high-quality image-text data is crucial for both image and video understanding. Instead of preparing massive video-text datasets, we focus on constructing large-scale and high-quality image-text datasets. VideoLLaMA3 has four training stages: 1) vision-centric alignment stage, which warms up the vision encoder and projector; 2) vision-language pretraining stage, which jointly tunes the vision encoder, projector, and LLM with large-scale image-text data covering multiple types (including scene images, documents, charts) as well as text-only data. 3) multi-task fine-tuning stage, which incorporates image-text SFT data for downstream tasks and video-text data to establish a foundation for video understanding. 4) video-centric fine-tuning, which further improves the model's capability in video understanding. As for the framework design, to better capture fine-grained details in images, the pretrained vision encoder is adapted to encode images of varying sizes into vision tokens with corresponding numbers, rather than a fixed number of tokens. For video inputs, we reduce the number of vision tokens according to their similarity so that the representation of videos will be more precise and compact. Benefit from vision-centric designs, VideoLLaMA3 achieves compelling performances in both image and video understanding benchmarks.

We introduce Matrix-Game, an interactive world foundation model for controllable game world generation. Matrix-Game is trained using a two-stage pipeline that first performs large-scale unlabeled pretraining for environment understanding, followed by action-labeled training for interactive video generation. To support this, we curate Matrix-Game-MC, a comprehensive Minecraft dataset comprising over 2,700 hours of unlabeled gameplay video clips and over 1,000 hours of high-quality labeled clips with fine-grained keyboard and mouse action annotations. Our model adopts a controllable image-to-world generation paradigm, conditioned on a reference image, motion context, and user actions. With over 17 billion parameters, Matrix-Game enables precise control over character actions and camera movements, while maintaining high visual quality and temporal coherence. To evaluate performance, we develop GameWorld Score, a unified benchmark measuring visual quality, temporal quality, action controllability, and physical rule understanding for Minecraft world generation. Extensive experiments show that Matrix-Game consistently outperforms prior open-source Minecraft world models (including Oasis and MineWorld) across all metrics, with particularly strong gains in controllability and physical consistency. Double-blind human evaluations further confirm the superiority of Matrix-Game, highlighting its ability to generate perceptually realistic and precisely controllable videos across diverse game scenarios. To facilitate future research on interactive image-to-world generation, we will open-source the Matrix-Game model weights and the GameWorld Score benchmark at https://github.com/SkyworkAI/Matrix-Game.

Existing approaches for video moment retrieval and highlight detection are not able to align text and video features efficiently, resulting in unsatisfying performance and limited production usage. To address this, we propose a novel architecture that utilizes recent foundational video models designed for such alignment. Combined with the introduced Saliency-Guided Cross Attention mechanism and a hybrid DETR architecture, our approach significantly enhances performance in both moment retrieval and highlight detection tasks. For even better improvement, we developed InterVid-MR, a large-scale and high-quality dataset for pretraining. Using it, our architecture achieves state-of-the-art results on the QVHighlights, Charades-STA and TACoS benchmarks. The proposed approach provides an efficient and scalable solution for both zero-shot and fine-tuning scenarios in video-language tasks.

Video is an increasingly prominent and information-dense medium, yet it poses substantial challenges for language models. A typical video consists of a sequence of shorter segments, or shots, that collectively form a coherent narrative. Each shot is analogous to a word in a sentence where multiple data streams of information (such as visual and auditory data) must be processed simultaneously. Comprehension of the entire video requires not only understanding the visual-audio information of each shot but also requires that the model links the ideas between each shot to generate a larger, all-encompassing story. Despite significant progress in the field, current works often overlook videos' more granular shot-by-shot semantic information. In this project, we propose a family of efficient large language vision models (LLVMs) to boost video summarization and captioning called Shotluck Holmes. By leveraging better pretraining and data collection strategies, we extend the abilities of existing small LLVMs from being able to understand a picture to being able to understand a sequence of frames. Specifically, we show that Shotluck Holmes achieves better performance than state-of-the-art results on the Shot2Story video captioning and summary task with significantly smaller and more computationally efficient models.

Popular video training methods mainly operate on a fixed number of tokens sampled from a predetermined spatiotemporal grid, resulting in sub-optimal accuracy-computation trade-offs due to inherent video redundancy. They also lack adaptability to varying computational budgets for downstream tasks, hindering applications of the most competitive model in real-world scenes. We thus propose a new test setting, Token Optimization, for maximized input information across budgets, which optimizes the size-limited set of input tokens through token selection from more suitably sampled videos. To this end, we propose a novel augmentation tool termed Flux. By making the sampling grid flexible and leveraging token selection, it is easily adopted in most popular video training frameworks, boosting model robustness with nearly no additional cost. We integrate Flux in large-scale video pre-training, and the resulting FluxViT establishes new state-of-the-art results across extensive tasks at standard costs. Notably, with 1/4 tokens only, it can still match the performance of previous state-of-the-art models with Token Optimization, yielding nearly 90\% savings. All models and data are available at https://github.com/OpenGVLab/FluxViT.

Learning a generalist embodied agent capable of completing multiple tasks poses challenges, primarily stemming from the scarcity of action-labeled robotic datasets. In contrast, a vast amount of human videos exist, capturing intricate tasks and interactions with the physical world. Promising prospects arise for utilizing actionless human videos for pre-training and transferring the knowledge to facilitate robot policy learning through limited robot demonstrations. In this paper, we introduce a novel framework that leverages a unified discrete diffusion to combine generative pre-training on human videos and policy fine-tuning on a small number of action-labeled robot videos. We start by compressing both human and robot videos into unified video tokens. In the pre-training stage, we employ a discrete diffusion model with a mask-and-replace diffusion strategy to predict future video tokens in the latent space. In the fine-tuning stage, we harness the imagined future videos to guide low-level action learning trained on a limited set of robot data. Experiments demonstrate that our method generates high-fidelity future videos for planning and enhances the fine-tuned policies compared to previous state-of-the-art approaches with superior generalization ability. Our project website is available at https://video-diff.github.io/.

Generative pre-trained models have demonstrated remarkable effectiveness in language and vision domains by learning useful representations. In this paper, we extend the scope of this effectiveness by showing that visual robot manipulation can significantly benefit from large-scale video generative pre-training. We introduce GR-1, a straightforward GPT-style model designed for multi-task language-conditioned visual robot manipulation. GR-1 takes as inputs a language instruction, a sequence of observation images, and a sequence of robot states. It predicts robot actions as well as future images in an end-to-end manner. Thanks to a flexible design, GR-1 can be seamlessly finetuned on robot data after pre-trained on a large-scale video dataset. We perform extensive experiments on the challenging CALVIN benchmark and a real robot. On CALVIN benchmark, our method outperforms state-of-the-art baseline methods and improves the success rate from 88.9% to 94.9%. In the setting of zero-shot unseen scene generalization, GR-1 improves the success rate from 53.3% to 85.4%. In real robot experiments, GR-1 also outperforms baseline methods and shows strong potentials in generalization to unseen scenes and objects. We provide inaugural evidence that a unified GPT-style transformer, augmented with large-scale video generative pre-training, exhibits remarkable generalization to multi-task visual robot manipulation. Project page: https://GR1-Manipulation.github.io

We study joint video and language (VL) pre-training to enable cross-modality learning and benefit plentiful downstream VL tasks. Existing works either extract low-quality video features or learn limited text embedding, while neglecting that high-resolution videos and diversified semantics can significantly improve cross-modality learning. In this paper, we propose a novel High-resolution and Diversified VIdeo-LAnguage pre-training model (HD-VILA) for many visual tasks. In particular, we collect a large dataset with two distinct properties: 1) the first high-resolution dataset including 371.5k hours of 720p videos, and 2) the most diversified dataset covering 15 popular YouTube categories. To enable VL pre-training, we jointly optimize the HD-VILA model by a hybrid Transformer that learns rich spatiotemporal features, and a multimodal Transformer that enforces interactions of the learned video features with diversified texts. Our pre-training model achieves new state-of-the-art results in 10 VL understanding tasks and 2 more novel text-to-visual generation tasks. For example, we outperform SOTA models with relative increases of 40.4% R@1 in zero-shot MSR-VTT text-to-video retrieval task and 55.4% in high-resolution dataset LSMDC. The learned VL embedding is also effective in generating visually pleasing and semantically relevant results in text-to-visual editing and super-resolution tasks.

Large-scale video-language pre-training has made remarkable strides in advancing video-language understanding tasks. However, the heavy computational burden of video encoding remains a formidable efficiency bottleneck, particularly for long-form videos. These videos contain massive visual tokens due to their inherent 3D properties and spatiotemporal redundancy, making it challenging to capture complex temporal and spatial relationships. To tackle this issue, we propose an efficient method called TEmporal-Spatial Token Aggregation (TESTA). TESTA condenses video semantics by adaptively aggregating similar frames, as well as similar patches within each frame. TESTA can reduce the number of visual tokens by 75% and thus accelerate video encoding. Building upon TESTA, we introduce a pre-trained video-language model equipped with a divided space-time token aggregation module in each video encoder block. We evaluate our model on five datasets for paragraph-to-video retrieval and long-form VideoQA tasks. Experimental results show that TESTA improves computing efficiency by 1.7 times, and achieves significant performance gains from its scalability in processing longer input frames, e.g., +13.7 R@1 on QuerYD and +6.5 R@1 on Condensed Movie.

This report presents Wan, a comprehensive and open suite of video foundation models designed to push the boundaries of video generation. Built upon the mainstream diffusion transformer paradigm, Wan achieves significant advancements in generative capabilities through a series of innovations, including our novel VAE, scalable pre-training strategies, large-scale data curation, and automated evaluation metrics. These contributions collectively enhance the model's performance and versatility. Specifically, Wan is characterized by four key features: Leading Performance: The 14B model of Wan, trained on a vast dataset comprising billions of images and videos, demonstrates the scaling laws of video generation with respect to both data and model size. It consistently outperforms the existing open-source models as well as state-of-the-art commercial solutions across multiple internal and external benchmarks, demonstrating a clear and significant performance superiority. Comprehensiveness: Wan offers two capable models, i.e., 1.3B and 14B parameters, for efficiency and effectiveness respectively. It also covers multiple downstream applications, including image-to-video, instruction-guided video editing, and personal video generation, encompassing up to eight tasks. Consumer-Grade Efficiency: The 1.3B model demonstrates exceptional resource efficiency, requiring only 8.19 GB VRAM, making it compatible with a wide range of consumer-grade GPUs. Openness: We open-source the entire series of Wan, including source code and all models, with the goal of fostering the growth of the video generation community. This openness seeks to significantly expand the creative possibilities of video production in the industry and provide academia with high-quality video foundation models. All the code and models are available at https://github.com/Wan-Video/Wan2.1.

Pre-training video transformers on extra large-scale datasets is generally required to achieve premier performance on relatively small datasets. In this paper, we show that video masked autoencoders (VideoMAE) are data-efficient learners for self-supervised video pre-training (SSVP). We are inspired by the recent ImageMAE and propose customized video tube masking with an extremely high ratio. This simple design makes video reconstruction a more challenging self-supervision task, thus encouraging extracting more effective video representations during this pre-training process. We obtain three important findings on SSVP: (1) An extremely high proportion of masking ratio (i.e., 90% to 95%) still yields favorable performance of VideoMAE. The temporally redundant video content enables a higher masking ratio than that of images. (2) VideoMAE achieves impressive results on very small datasets (i.e., around 3k-4k videos) without using any extra data. (3) VideoMAE shows that data quality is more important than data quantity for SSVP. Domain shift between pre-training and target datasets is an important issue. Notably, our VideoMAE with the vanilla ViT can achieve 87.4% on Kinetics-400, 75.4% on Something-Something V2, 91.3% on UCF101, and 62.6% on HMDB51, without using any extra data. Code is available at https://github.com/MCG-NJU/VideoMAE.

We present HERO, a novel framework for large-scale video+language omni-representation learning. HERO encodes multimodal inputs in a hierarchical structure, where local context of a video frame is captured by a Cross-modal Transformer via multimodal fusion, and global video context is captured by a Temporal Transformer. In addition to standard Masked Language Modeling (MLM) and Masked Frame Modeling (MFM) objectives, we design two new pre-training tasks: (i) Video-Subtitle Matching (VSM), where the model predicts both global and local temporal alignment; and (ii) Frame Order Modeling (FOM), where the model predicts the right order of shuffled video frames. HERO is jointly trained on HowTo100M and large-scale TV datasets to gain deep understanding of complex social dynamics with multi-character interactions. Comprehensive experiments demonstrate that HERO achieves new state of the art on multiple benchmarks over Text-based Video/Video-moment Retrieval, Video Question Answering (QA), Video-and-language Inference and Video Captioning tasks across different domains. We also introduce two new challenging benchmarks How2QA and How2R for Video QA and Retrieval, collected from diverse video content over multimodalities.

Bimanual robotic manipulation, which involves the coordinated control of two robotic arms, is foundational for solving challenging tasks. Despite recent progress in general-purpose manipulation, data scarcity and embodiment heterogeneity remain serious obstacles to further scaling up in bimanual settings. In this paper, we introduce Video Diffusion for Action Reasoning (Vidar), a two-stage framework that leverages large-scale, diffusion-based video pre-training and a novel masked inverse dynamics model for action prediction. We pre-train the video diffusion model on 750K multi-view videos from three real-world bimanual robot platforms, utilizing a unified observation space that encodes robot, camera, task, and scene contexts. Our masked inverse dynamics model learns masks to extract action-relevant information from generated trajectories without requiring pixel-level labels, and the masks can effectively generalize to unseen backgrounds. Our experiments demonstrate that with only 20 minutes of human demonstrations on an unseen robot platform (only 1% of typical data requirements), Vidar generalizes to unseen tasks and backgrounds with strong semantic understanding, surpassing state-of-the-art methods. Our findings highlight the potential of video foundation models, coupled with masked action prediction, to enable scalable and generalizable robotic manipulation in diverse real-world settings.

Pre-training on large-scale video data has become a common recipe for learning transferable spatiotemporal representations in recent years. Despite some progress, existing methods are mostly limited to highly curated datasets (e.g., K400) and exhibit unsatisfactory out-of-the-box representations. We argue that it is due to the fact that they only capture pixel-level knowledge rather than spatiotemporal semantics, which hinders further progress in video understanding. Inspired by the great success of image-text pre-training (e.g., CLIP), we take the first step to exploit language semantics to boost transferable spatiotemporal representation learning. We introduce a new pretext task, Turning to Video for Transcript Sorting (TVTS), which sorts shuffled ASR scripts by attending to learned video representations. We do not rely on descriptive captions and learn purely from video, i.e., leveraging the natural transcribed speech knowledge to provide noisy but useful semantics over time. Our method enforces the vision model to contextualize what is happening over time so that it can re-organize the narrative transcripts, and can seamlessly apply to large-scale uncurated video data in the real world. Our method demonstrates strong out-of-the-box spatiotemporal representations on diverse benchmarks, e.g., +13.6% gains over VideoMAE on SSV2 via linear probing. The code is available at https://github.com/TencentARC/TVTS.

We introduce Being-H0, a dexterous Vision-Language-Action model (VLA) trained on large-scale human videos. Existing VLAs struggle with complex manipulation tasks requiring high dexterity and generalize poorly to novel scenarios and tasks, primarily due to their reliance on synthetic data with significant sim-to-real gaps or teleoperated demonstrations lacking scale and diversity. To address this data bottleneck, we propose leveraging human hands as a foundation manipulator, capitalizing on the rich dexterity and scalability present in web data. Our approach centers on physical instruction tuning, a novel training paradigm that combines large-scale VLA pretraining from human videos, physical space alignment for 3D reasoning, and post-training adaptation for robotic tasks. Additionally, we introduce a part-level motion tokenization method which achieves millimeter-level reconstruction accuracy to model precise hand trajectories for action learning. To support our proposed paradigm, we further develop a comprehensive data curation pipeline that integrates heterogeneous sources -- including motion capture, VR, and RGB-only videos -- into a large-scale dataset with millions of motion-based instructional instances. We empirically show the excellence of Being-H0 in hand motion generation and instruction following, and it also scales well with model and data sizes. Importantly, we observe the expected gains of Being-H0 in real-world robotic manipulation as physical instruction tuning is applied. More details are available at https://beingbeyond.github.io/Being-H0.

In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models will be available at https://video-lavit.github.io.

Pre-training on Internet data has proven to be a key ingredient for broad generalization in many modern ML systems. What would it take to enable such capabilities in robotic reinforcement learning (RL)? Offline RL methods, which learn from datasets of robot experience, offer one way to leverage prior data into the robotic learning pipeline. However, these methods have a "type mismatch" with video data (such as Ego4D), the largest prior datasets available for robotics, since video offers observation-only experience without the action or reward annotations needed for RL methods. In this paper, we develop a system for leveraging large-scale human video datasets in robotic offline RL, based entirely on learning value functions via temporal-difference learning. We show that value learning on video datasets learns representations that are more conducive to downstream robotic offline RL than other approaches for learning from video data. Our system, called V-PTR, combines the benefits of pre-training on video data with robotic offline RL approaches that train on diverse robot data, resulting in value functions and policies for manipulation tasks that perform better, act robustly, and generalize broadly. On several manipulation tasks on a real WidowX robot, our framework produces policies that greatly improve over prior methods. Our video and additional details can be found at https://dibyaghosh.com/vptr/

Vision-Language-Action (VLA) models provide a promising paradigm for robot learning by integrating visual perception with language-guided policy learning. However, most existing approaches rely on 2D visual inputs to perform actions in 3D physical environments, creating a significant gap between perception and action grounding. To bridge this gap, we propose a Spatial-Aware VLA Pretraining paradigm that performs explicit alignment between visual space and physical space during pretraining, enabling models to acquire 3D spatial understanding before robot policy learning. Starting from pretrained vision-language models, we leverage large-scale human demonstration videos to extract 3D visual and 3D action annotations, forming a new source of supervision that aligns 2D visual observations with 3D spatial reasoning. We instantiate this paradigm with VIPA-VLA, a dual-encoder architecture that incorporates a 3D visual encoder to augment semantic visual representations with 3D-aware features. When adapted to downstream robot tasks, VIPA-VLA achieves significantly improved grounding between 2D vision and 3D action, resulting in more robust and generalizable robotic policies.

We introduce Latent Action Pretraining for general Action models (LAPA), an unsupervised method for pretraining Vision-Language-Action (VLA) models without ground-truth robot action labels. Existing Vision-Language-Action models require action labels typically collected by human teleoperators during pretraining, which significantly limits possible data sources and scale. In this work, we propose a method to learn from internet-scale videos that do not have robot action labels. We first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames, then pretrain a latent VLA model to predict these latent actions from observations and task descriptions, and finally finetune the VLA on small-scale robot manipulation data to map from latent to robot actions. Experimental results demonstrate that our method significantly outperforms existing techniques that train robot manipulation policies from large-scale videos. Furthermore, it outperforms the state-of-the-art VLA model trained with robotic action labels on real-world manipulation tasks that require language conditioning, generalization to unseen objects, and semantic generalization to unseen instructions. Training only on human manipulation videos also shows positive transfer, opening up the potential for leveraging web-scale data for robotics foundation model.

Large-scale pre-training using videos has proven effective for robot learning. However, the models pre-trained on such data can be suboptimal for robot learning due to the significant visual gap between human hands and those of different robots. To remedy this, we propose H2R, a simple data augmentation technique that detects human hand keypoints, synthesizes robot motions in simulation, and composites rendered robots into egocentric videos. This process explicitly bridges the visual gap between human and robot embodiments during pre-training. We apply H2R to augment large-scale egocentric human video datasets such as Ego4D and SSv2, replacing human hands with simulated robotic arms to generate robot-centric training data. Based on this, we construct and release a family of 1M-scale datasets covering multiple robot embodiments (UR5 with gripper/Leaphand, Franka) and data sources (SSv2, Ego4D). To verify the effectiveness of the augmentation pipeline, we introduce a CLIP-based image-text similarity metric that quantitatively evaluates the semantic fidelity of robot-rendered frames to the original human actions. We validate H2R across three simulation benchmarks: Robomimic, RLBench and PushT and real-world manipulation tasks with a UR5 robot equipped with Gripper and Leaphand end-effectors. H2R consistently improves downstream success rates, yielding gains of 5.0%-10.2% in simulation and 6.7%-23.3% in real-world tasks across various visual encoders and policy learning methods. These results indicate that H2R improves the generalization ability of robotic policies by mitigating the visual discrepancies between human and robot domains.

Human Activity Recognition (HAR) based on wearable inertial sensors plays a critical role in remote health monitoring. In patients with movement disorders, the ability to detect abnormal patient movements in their home environments can enable continuous optimization of treatments and help alert caretakers as needed. Machine learning approaches have been proposed for HAR tasks using Inertial Measurement Unit (IMU) data; however, most rely on application-specific labels and lack generalizability to data collected in different environments or populations. To address this limitation, we propose a new cross-modal self-supervised pretraining approach to learn representations from large-sale unlabeled IMU-video data and demonstrate improved generalizability in HAR tasks on out of distribution (OOD) IMU datasets, including a dataset collected from patients with Parkinson's disease. Specifically, our results indicate that the proposed cross-modal pretraining approach outperforms the current state-of-the-art IMU-video pretraining approach and IMU-only pretraining under zero-shot and few-shot evaluations. Broadly, our study provides evidence that in highly dynamic data modalities, such as IMU signals, cross-modal pretraining may be a useful tool to learn generalizable data representations. Our software is available at https://github.com/scheshmi/IMU-Video-OOD-HAR.

We study video-grounded dialogue generation, where a response is generated based on the dialogue context and the associated video. The primary challenges of this task lie in (1) the difficulty of integrating video data into pre-trained language models (PLMs) which presents obstacles to exploiting the power of large-scale pre-training; and (2) the necessity of taking into account the complementarity of various modalities throughout the reasoning process. Although having made remarkable progress in video-grounded dialogue generation, existing methods still fall short when it comes to integrating with PLMs in a way that allows information from different modalities to complement each other. To alleviate these issues, we first propose extracting pertinent information from videos and turning it into reasoning paths that are acceptable to PLMs. Additionally, we propose a multi-agent reinforcement learning method to collaboratively perform reasoning on different modalities (i.e., video and dialogue context). Empirical experiment results on two public datasets indicate that the proposed model can significantly outperform state-of-the-art models by large margins on both automatic and human evaluations.

We propose a novel approach for captioning and object grounding in video, where the objects in the caption are grounded in the video via temporally dense bounding boxes. We introduce the following contributions. First, we present a large-scale automatic annotation method that aggregates captions grounded with bounding boxes across individual frames into temporally dense and consistent bounding box annotations. We apply this approach on the HowTo100M dataset to construct a large-scale pre-training dataset, named HowToGround1M. We also introduce a Grounded Video Caption Generation model, dubbed GROVE, and pre-train the model on HowToGround1M. Second, we introduce a new dataset, called iGround, of 3500 videos with manually annotated captions and dense spatio-temporally grounded bounding boxes. This allows us to measure progress on this challenging problem, as well as to fine-tune our model on this small-scale but high-quality data. Third, we demonstrate that our approach achieves state-of-the-art results on the proposed iGround dataset compared to a number of baselines, as well as on the VidSTG and ActivityNet-Entities datasets. We perform extensive ablations that demonstrate the importance of pre-training using our automatically annotated HowToGround1M dataset followed by fine-tuning on the manually annotated iGround dataset and validate the key technical contributions of our model.

We present GR-2, a state-of-the-art generalist robot agent for versatile and generalizable robot manipulation. GR-2 is first pre-trained on a vast number of Internet videos to capture the dynamics of the world. This large-scale pre-training, involving 38 million video clips and over 50 billion tokens, equips GR-2 with the ability to generalize across a wide range of robotic tasks and environments during subsequent policy learning. Following this, GR-2 is fine-tuned for both video generation and action prediction using robot trajectories. It exhibits impressive multi-task learning capabilities, achieving an average success rate of 97.7% across more than 100 tasks. Moreover, GR-2 demonstrates exceptional generalization to new, previously unseen scenarios, including novel backgrounds, environments, objects, and tasks. Notably, GR-2 scales effectively with model size, underscoring its potential for continued growth and application. Project page: https://gr2-manipulation.github.io.

Recent researches have proven that pre-training on large-scale person images extracted from internet videos is an effective way in learning better representations for person re-identification. However, these researches are mostly confined to pre-training at the instance-level or single-video tracklet-level. They ignore the identity-invariance in images of the same person across different videos, which is a key focus in person re-identification. To address this issue, we propose a Cross-video Identity-cOrrelating pre-traiNing (CION) framework. Defining a noise concept that comprehensively considers both intra-identity consistency and inter-identity discrimination, CION seeks the identity correlation from cross-video images by modeling it as a progressive multi-level denoising problem. Furthermore, an identity-guided self-distillation loss is proposed to implement better large-scale pre-training by mining the identity-invariance within person images. We conduct extensive experiments to verify the superiority of our CION in terms of efficiency and performance. CION achieves significantly leading performance with even fewer training samples. For example, compared with the previous state-of-the-art~ISR, CION with the same ResNet50-IBN achieves higher mAP of 93.3\% and 74.3\% on Market1501 and MSMT17, while only utilizing 8\% training samples. Finally, with CION demonstrating superior model-agnostic ability, we contribute a model zoo named ReIDZoo to meet diverse research and application needs in this field. It contains a series of CION pre-trained models with spanning structures and parameters, totaling 32 models with 10 different structures, including GhostNet, ConvNext, RepViT, FastViT and so on. The code and models will be made publicly available at https://github.com/Zplusdragon/CION_ReIDZoo.

We present a large-scale study on unsupervised spatiotemporal representation learning from videos. With a unified perspective on four recent image-based frameworks, we study a simple objective that can easily generalize all these methods to space-time. Our objective encourages temporally-persistent features in the same video, and in spite of its simplicity, it works surprisingly well across: (i) different unsupervised frameworks, (ii) pre-training datasets, (iii) downstream datasets, and (iv) backbone architectures. We draw a series of intriguing observations from this study, e.g., we discover that encouraging long-spanned persistency can be effective even if the timespan is 60 seconds. In addition to state-of-the-art results in multiple benchmarks, we report a few promising cases in which unsupervised pre-training can outperform its supervised counterpart. Code is made available at https://github.com/facebookresearch/SlowFast

The correlation between the vision and text is essential for video moment retrieval (VMR), however, existing methods heavily rely on separate pre-training feature extractors for visual and textual understanding. Without sufficient temporal boundary annotations, it is non-trivial to learn universal video-text alignments. In this work, we explore multi-modal correlations derived from large-scale image-text data to facilitate generalisable VMR. To address the limitations of image-text pre-training models on capturing the video changes, we propose a generic method, referred to as Visual-Dynamic Injection (VDI), to empower the model's understanding of video moments. Whilst existing VMR methods are focusing on building temporal-aware video features, being aware of the text descriptions about the temporal changes is also critical but originally overlooked in pre-training by matching static images with sentences. Therefore, we extract visual context and spatial dynamic information from video frames and explicitly enforce their alignments with the phrases describing video changes (e.g. verb). By doing so, the potentially relevant visual and motion patterns in videos are encoded in the corresponding text embeddings (injected) so to enable more accurate video-text alignments. We conduct extensive experiments on two VMR benchmark datasets (Charades-STA and ActivityNet-Captions) and achieve state-of-the-art performances. Especially, VDI yields notable advantages when being tested on the out-of-distribution splits where the testing samples involve novel scenes and vocabulary.

Understanding long, real-world videos requires modeling of long-range visual dependencies. To this end, we explore video-first architectures, building on the common paradigm of transferring large-scale, image--text models to video via shallow temporal fusion. However, we expose two limitations to the approach: (1) decreased spatial capabilities, likely due to poor video--language alignment in standard video datasets, and (2) higher memory consumption, bottlenecking the number of frames that can be processed. To mitigate the memory bottleneck, we systematically analyze the memory/accuracy trade-off of various efficient methods: factorized attention, parameter-efficient image-to-video adaptation, input masking, and multi-resolution patchification. Surprisingly, simply masking large portions of the video (up to 75%) during contrastive pre-training proves to be one of the most robust ways to scale encoders to videos up to 4.3 minutes at 1 FPS. Our simple approach for training long video-to-text models, which scales to 1B parameters, does not add new architectural complexity and is able to outperform the popular paradigm of using much larger LLMs as an information aggregator over segment-based information on benchmarks with long-range temporal dependencies (YouCook2, EgoSchema).

Adapting large-scale image-text pre-training models, e.g., CLIP, to the video domain represents the current state-of-the-art for text-video retrieval. The primary approaches involve transferring text-video pairs to a common embedding space and leveraging cross-modal interactions on specific entities for semantic alignment. Though effective, these paradigms entail prohibitive computational costs, leading to inefficient retrieval. To address this, we propose a simple yet effective method, Global-Local Semantic Consistent Learning (GLSCL), which capitalizes on latent shared semantics across modalities for text-video retrieval. Specifically, we introduce a parameter-free global interaction module to explore coarse-grained alignment. Then, we devise a shared local interaction module that employs several learnable queries to capture latent semantic concepts for learning fine-grained alignment. Furthermore, an Inter-Consistency Loss (ICL) is devised to accomplish the concept alignment between the visual query and corresponding textual query, and an Intra-Diversity Loss (IDL) is developed to repulse the distribution within visual (textual) queries to generate more discriminative concepts. Extensive experiments on five widely used benchmarks (i.e., MSR-VTT, MSVD, DiDeMo, LSMDC, and ActivityNet) substantiate the superior effectiveness and efficiency of the proposed method. Remarkably, our method achieves comparable performance with SOTA as well as being nearly 220 times faster in terms of computational cost. Code is available at: https://github.com/zchoi/GLSCL.

Benefiting from large-scale pre-training of text-video pairs, current text-to-video (T2V) diffusion models can generate high-quality videos from the text description. Besides, given some reference images or videos, the parameter-efficient fine-tuning method, i.e. LoRA, can generate high-quality customized concepts, e.g., the specific subject or the motions from a reference video. However, combining the trained multiple concepts from different references into a single network shows obvious artifacts. To this end, we propose CustomTTT, where we can joint custom the appearance and the motion of the given video easily. In detail, we first analyze the prompt influence in the current video diffusion model and find the LoRAs are only needed for the specific layers for appearance and motion customization. Besides, since each LoRA is trained individually, we propose a novel test-time training technique to update parameters after combination utilizing the trained customized models. We conduct detailed experiments to verify the effectiveness of the proposed methods. Our method outperforms several state-of-the-art works in both qualitative and quantitative evaluations.

Recent advancements in large-scale pre-training of visual-language models on paired image-text data have demonstrated impressive generalization capabilities for zero-shot tasks. Building on this success, efforts have been made to adapt these image-based visual-language models, such as CLIP, for videos extending their zero-shot capabilities to the video domain. While these adaptations have shown promising results, they come at a significant computational cost and struggle with effectively modeling the crucial temporal aspects inherent to the video domain. In this study, we present EZ-CLIP, a simple and efficient adaptation of CLIP that addresses these challenges. EZ-CLIP leverages temporal visual prompting for seamless temporal adaptation, requiring no fundamental alterations to the core CLIP architecture while preserving its remarkable generalization abilities. Moreover, we introduce a novel learning objective that guides the temporal visual prompts to focus on capturing motion, thereby enhancing its learning capabilities from video data. We conducted extensive experiments on five different benchmark datasets, thoroughly evaluating EZ-CLIP for zero-shot learning and base-to-novel video action recognition, and also demonstrating its potential for few-shot generalization.Impressively, with a mere 5.2 million learnable parameters (as opposed to the 71.1 million in the prior best model), EZ-CLIP can be efficiently trained on a single GPU, outperforming existing approaches in several evaluations.

Recently, large-scale pre-trained language-image models like CLIP have shown extraordinary capabilities for understanding spatial contents, but naively transferring such models to video recognition still suffers from unsatisfactory temporal modeling capabilities. Existing methods insert tunable structures into or in parallel with the pre-trained model, which either requires back-propagation through the whole pre-trained model and is thus resource-demanding, or is limited by the temporal reasoning capability of the pre-trained structure. In this work, we present DiST, which disentangles the learning of spatial and temporal aspects of videos. Specifically, DiST uses a dual-encoder structure, where a pre-trained foundation model acts as the spatial encoder, and a lightweight network is introduced as the temporal encoder. An integration branch is inserted between the encoders to fuse spatio-temporal information. The disentangled spatial and temporal learning in DiST is highly efficient because it avoids the back-propagation of massive pre-trained parameters. Meanwhile, we empirically show that disentangled learning with an extra network for integration benefits both spatial and temporal understanding. Extensive experiments on five benchmarks show that DiST delivers better performance than existing state-of-the-art methods by convincing gaps. When pre-training on the large-scale Kinetics-710, we achieve 89.7% on Kinetics-400 with a frozen ViT-L model, which verifies the scalability of DiST. Codes and models can be found in https://github.com/alibaba-mmai-research/DiST.

One of the roadblocks for training generalist robotic models today is heterogeneity. Previous robot learning methods often collect data to train with one specific embodiment for one task, which is expensive and prone to overfitting. This work studies the problem of learning policy representations through heterogeneous pre-training on robot data across different embodiments and tasks at scale. We propose Heterogeneous Pre-trained Transformers (HPT), which pre-train a large, shareable trunk of a policy neural network to learn a task and embodiment agnostic shared representation. This general architecture aligns the specific proprioception and vision inputs from distinct embodiments to a short sequence of tokens and then processes such tokens to map to control robots for different tasks. Leveraging the recent large-scale multi-embodiment real-world robotic datasets as well as simulation, deployed robots, and human video datasets, we investigate pre-training policies across heterogeneity. We conduct experiments to investigate the scaling behaviors of training objectives, to the extent of 52 datasets. HPTs outperform several baselines and enhance the fine-tuned policy performance by over 20% on unseen tasks in multiple simulator benchmarks and real-world settings. See the project website (https://liruiw.github.io/hpt/) for code and videos.

Recently, large-scale pre-training methods like CLIP have made great progress in multi-modal research such as text-video retrieval. In CLIP, transformers are vital for modeling complex multi-modal relations. However, in the vision transformer of CLIP, the essential visual tokenization process, which produces discrete visual token sequences, generates many homogeneous tokens due to the redundancy nature of consecutive and similar frames in videos. This significantly increases computation costs and hinders the deployment of video retrieval models in web applications. In this paper, to reduce the number of redundant video tokens, we design a multi-segment token clustering algorithm to find the most representative tokens and drop the non-essential ones. As the frame redundancy occurs mostly in consecutive frames, we divide videos into multiple segments and conduct segment-level clustering. Center tokens from each segment are later concatenated into a new sequence, while their original spatial-temporal relations are well maintained. We instantiate two clustering algorithms to efficiently find deterministic medoids and iteratively partition groups in high dimensional space. Through this token clustering and center selection procedure, we successfully reduce computation costs by removing redundant visual tokens. This method further enhances segment-level semantic alignment between video and text representations, enforcing the spatio-temporal interactions of tokens from within-segment frames. Our method, coined as CenterCLIP, surpasses existing state-of-the-art by a large margin on typical text-video benchmarks, while reducing the training memory cost by 35\% and accelerating the inference speed by 14\% at the best case. The code is available at {https://github.com/mzhaoshuai/CenterCLIP}{{https://github.com/mzhaoshuai/CenterCLIP}}.

With the explosive growth of web videos and emerging large-scale vision-language pre-training models, e.g., CLIP, retrieving videos of interest with text instructions has attracted increasing attention. A common practice is to transfer text-video pairs to the same embedding space and craft cross-modal interactions with certain entities in specific granularities for semantic correspondence. Unfortunately, the intrinsic uncertainties of optimal entity combinations in appropriate granularities for cross-modal queries are understudied, which is especially critical for modalities with hierarchical semantics, e.g., video, text, etc. In this paper, we propose an Uncertainty-Adaptive Text-Video Retrieval approach, termed UATVR, which models each look-up as a distribution matching procedure. Concretely, we add additional learnable tokens in the encoders to adaptively aggregate multi-grained semantics for flexible high-level reasoning. In the refined embedding space, we represent text-video pairs as probabilistic distributions where prototypes are sampled for matching evaluation. Comprehensive experiments on four benchmarks justify the superiority of our UATVR, which achieves new state-of-the-art results on MSR-VTT (50.8%), VATEX (64.5%), MSVD (49.7%), and DiDeMo (45.8%). The code is available at https://github.com/bofang98/UATVR.

Although Large Language Models (LLMs) have shown promise for human-like conversations, they are primarily pre-trained on text data. Incorporating audio or video improves performance, but collecting large-scale multimodal data and pre-training multimodal LLMs is challenging. To this end, we propose a Fusion Low Rank Adaptation (FLoRA) technique that efficiently adapts a pre-trained unimodal LLM to consume new, previously unseen modalities via low rank adaptation. For device-directed speech detection, using FLoRA, the multimodal LLM achieves 22% relative reduction in equal error rate (EER) over the text-only approach and attains performance parity with its full fine-tuning (FFT) counterpart while needing to tune only a fraction of its parameters. Furthermore, with the newly introduced adapter dropout, FLoRA is robust to missing data, improving over FFT by 20% lower EER and 56% lower false accept rate. The proposed approach scales well for model sizes from 16M to 3B parameters.

This report introduces Kandinsky 5.0, a family of state-of-the-art foundation models for high-resolution image and 10-second video synthesis. The framework comprises three core line-up of models: Kandinsky 5.0 Image Lite - a line-up of 6B parameter image generation models, Kandinsky 5.0 Video Lite - a fast and lightweight 2B parameter text-to-video and image-to-video models, and Kandinsky 5.0 Video Pro - 19B parameter models that achieves superior video generation quality. We provide a comprehensive review of the data curation lifecycle - including collection, processing, filtering and clustering - for the multi-stage training pipeline that involves extensive pre-training and incorporates quality-enhancement techniques such as self-supervised fine-tuning (SFT) and reinforcement learning (RL)-based post-training. We also present novel architectural, training, and inference optimizations that enable Kandinsky 5.0 to achieve high generation speeds and state-of-the-art performance across various tasks, as demonstrated by human evaluation. As a large-scale, publicly available generative framework, Kandinsky 5.0 leverages the full potential of its pre-training and subsequent stages to be adapted for a wide range of generative applications. We hope that this report, together with the release of our open-source code and training checkpoints, will substantially advance the development and accessibility of high-quality generative models for the research community.

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.

Though pre-training vision-language models have demonstrated significant benefits in boosting video-text retrieval performance from large-scale web videos, fine-tuning still plays a critical role with manually annotated clips with start and end times, which requires considerable human effort. To address this issue, we explore an alternative cheaper source of annotations, single timestamps, for video-text retrieval. We initialise clips from timestamps in a heuristic way to warm up a retrieval model. Then a video clip editing method is proposed to refine the initial rough boundaries to improve retrieval performance. A student-teacher network is introduced for video clip editing. The teacher model is employed to edit the clips in the training set whereas the student model trains on the edited clips. The teacher weights are updated from the student's after the student's performance increases. Our method is model agnostic and applicable to any retrieval models. We conduct experiments based on three state-of-the-art retrieval models, COOT, VideoCLIP and CLIP4Clip. Experiments conducted on three video retrieval datasets, YouCook2, DiDeMo and ActivityNet-Captions show that our edited clips consistently improve retrieval performance over initial clips across all the three retrieval models.

Self-supervised pre-training recently demonstrates success on large-scale multimodal data, and state-of-the-art contrastive learning methods often enforce the feature consistency from cross-modality inputs, such as video/audio or video/text pairs. Despite its convenience to formulate and leverage in practice, such cross-modality alignment (CMA) is only a weak and noisy supervision, since two modalities can be semantically misaligned even they are temporally aligned. For example, even in the commonly adopted instructional videos, a speaker can sometimes refer to something that is not visually present in the current frame; and the semantic misalignment would only be more unpredictable for the raw videos from the internet. We conjecture that might cause conflicts and biases among modalities, and may hence prohibit CMA from scaling up to training with larger and more heterogeneous data. This paper first verifies our conjecture by observing that, even in the latest VATT pre-training using only instructional videos, there exist strong gradient conflicts between different CMA losses within the same video, audio, text triplet, indicating them as the noisy source of supervision. We then propose to harmonize such gradients, via two techniques: (i) cross-modality gradient realignment: modifying different CMA loss gradients for each sample triplet, so that their gradient directions are more aligned; and (ii) gradient-based curriculum learning: leveraging the gradient conflict information on an indicator of sample noisiness, to develop a curriculum learning strategy to prioritize training on less noisy sample triplets. Applying those techniques to pre-training VATT on the HowTo100M dataset, we consistently improve its performance on different downstream tasks. Moreover, we are able to scale VATT pre-training to more complicated non-narrative Youtube8M dataset to further improve the state-of-the-arts.

Video matting has traditionally been limited by the lack of high-quality ground-truth data. Most existing video matting datasets provide only human-annotated imperfect alpha and foreground annotations, which must be composited to background images or videos during the training stage. Thus, the generalization capability of previous methods in real-world scenarios is typically poor. In this work, we propose to solve the problem from two perspectives. First, we emphasize the importance of large-scale pre-training by pursuing diverse synthetic and pseudo-labeled segmentation datasets. We also develop a scalable synthetic data generation pipeline that can render diverse human bodies and fine-grained hairs, yielding around 200 video clips with a 3-second duration for fine-tuning. Second, we introduce a novel video matting approach that can effectively leverage the rich priors from pre-trained video diffusion models. This architecture offers two key advantages. First, strong priors play a critical role in bridging the domain gap between synthetic and real-world scenes. Second, unlike most existing methods that process video matting frame-by-frame and use an independent decoder to aggregate temporal information, our model is inherently designed for video, ensuring strong temporal consistency. We provide a comprehensive quantitative evaluation across three benchmark datasets, demonstrating our approach's superior performance, and present comprehensive qualitative results in diverse real-world scenes, illustrating the strong generalization capability of our method. The code is available at https://github.com/aim-uofa/GVM.

Image-based visual-language (I-VL) pre-training has shown great success for learning joint visual-textual representations from large-scale web data, revealing remarkable ability for zero-shot generalisation. This paper presents a simple but strong baseline to efficiently adapt the pre-trained I-VL model, and exploit its powerful ability for resource-hungry video understanding tasks, with minimal training. Specifically, we propose to optimise a few random vectors, termed as continuous prompt vectors, that convert video-related tasks into the same format as the pre-training objectives. In addition, to bridge the gap between static images and videos, temporal information is encoded with lightweight Transformers stacking on top of frame-wise visual features. Experimentally, we conduct extensive ablation studies to analyse the critical components. On 10 public benchmarks of action recognition, action localisation, and text-video retrieval, across closed-set, few-shot, and zero-shot scenarios, we achieve competitive or state-of-the-art performance to existing methods, despite optimising significantly fewer parameters.

In this paper, we propose a Vision-Audio-Language Omni-peRception pretraining model (VALOR) for multi-modal understanding and generation. Different from widely-studied vision-language pretraining models, VALOR jointly models relationships of vision, audio and language in an end-to-end manner. It contains three separate encoders for single modality representations, and a decoder for multimodal conditional text generation. We design two pretext tasks to pretrain VALOR model, including Multimodal Grouping Alignment (MGA) and Multimodal Grouping Captioning (MGC). MGA projects vision, language and audio to the same common space, building vision-language, audio-language and audiovisual-language alignment simultaneously. MGC learns how to generate text tokens in conditions of vision, audio or their both. To promote vision-audio-language pretraining research, we construct a large-scale high-quality tri-modality dataset named VALOR-1M, which contains 1M audiable videos with human annotated audiovisual captions. Extensive experiments show that VALOR can learn strong multimodal correlations and be generalized to various downstream tasks (e.g., retrieval, captioning and question answering), with different input modalities (e.g., vision-language, audio-language and audiovisual-language). VALOR achieves new state-of-the-art performances on series of public cross-modality benchmarks. Code and data are available at project page https://casia-iva-group.github.io/projects/VALOR.

Contrastive Language-Image Pre-training (CLIP) has been widely studied and applied in numerous applications. However, the emphasis on brief summary texts during pre-training prevents CLIP from understanding long descriptions. This issue is particularly acute regarding videos given that videos often contain abundant detailed contents. In this paper, we propose the VideoCLIP-XL (eXtra Length) model, which aims to unleash the long-description understanding capability of video CLIP models. Firstly, we establish an automatic data collection system and gather a large-scale VILD pre-training dataset with VIdeo and Long-Description pairs. Then, we propose Text-similarity-guided Primary Component Matching (TPCM) to better learn the distribution of feature space while expanding the long description capability. We also introduce two new tasks namely Detail-aware Description Ranking (DDR) and Hallucination-aware Description Ranking (HDR) for further understanding improvement. Finally, we construct a Long Video Description Ranking (LVDR) benchmark for evaluating the long-description capability more comprehensively. Extensive experimental results on widely-used text-video retrieval benchmarks with both short and long descriptions and our LVDR benchmark can fully demonstrate the effectiveness of our method.

Rapid advancements have been made in extending Large Language Models (LLMs) to Large Multi-modal Models (LMMs). However, extending input modality of LLMs to video data remains a challenging endeavor, especially for long videos. Due to insufficient access to large-scale high-quality video data and the excessive compression of visual features, current methods exhibit limitations in effectively processing long videos. In this paper, we introduce Kangaroo, a powerful Video LMM aimed at addressing these challenges. Confronted with issue of inadequate training data, we develop a data curation system to build a large-scale dataset with high-quality annotations for vision-language pre-training and instruction tuning. In addition, we design a curriculum training pipeline with gradually increasing resolution and number of input frames to accommodate long videos. Evaluation results demonstrate that, with 8B parameters, Kangaroo achieves state-of-the-art performance across a variety of video understanding benchmarks while exhibiting competitive results on others. Particularly, on benchmarks specialized for long videos, Kangaroo excels some larger models with over 10B parameters and proprietary models.

We present an architecture and a training recipe that adapts pre-trained open-world image models to localization in videos. Understanding the open visual world (without being constrained by fixed label spaces) is crucial for many real-world vision tasks. Contrastive pre-training on large image-text datasets has recently led to significant improvements for image-level tasks. For more structured tasks involving object localization applying pre-trained models is more challenging. This is particularly true for video tasks, where task-specific data is limited. We show successful transfer of open-world models by building on the OWL-ViT open-vocabulary detection model and adapting it to video by adding a transformer decoder. The decoder propagates object representations recurrently through time by using the output tokens for one frame as the object queries for the next. Our model is end-to-end trainable on video data and enjoys improved temporal consistency compared to tracking-by-detection baselines, while retaining the open-world capabilities of the backbone detector. We evaluate our model on the challenging TAO-OW benchmark and demonstrate that open-world capabilities, learned from large-scale image-text pre-training, can be transferred successfully to open-world localization across diverse videos.

Traditional video-to-audio generation techniques primarily focus on field-of-view (FoV) video and non-spatial audio, often missing the spatial cues necessary for accurately representing sound sources in 3D environments. To address this limitation, we introduce a novel task, 360V2SA, to generate spatial audio from 360-degree videos, specifically producing First-order Ambisonics (FOA) audio - a standard format for representing 3D spatial audio that captures sound directionality and enables realistic 3D audio reproduction. We first create Sphere360, a novel dataset tailored for this task that is curated from real-world data. We also design an efficient semi-automated pipeline for collecting and cleaning paired video-audio data. To generate spatial audio from 360-degree video, we propose a novel framework OmniAudio, which leverages self-supervised pre-training using both spatial audio data (in FOA format) and large-scale non-spatial data. Furthermore, OmniAudio features a dual-branch framework that utilizes both panoramic and FoV video inputs to capture comprehensive local and global information from 360-degree videos. Experimental results demonstrate that OmniAudio achieves state-of-the-art performance across both objective and subjective metrics on Sphere360. Code and datasets will be released at https://github.com/liuhuadai/OmniAudio. The demo page is available at https://OmniAudio-360V2SA.github.io.

Multimodal large-scale pretraining has shown impressive performance gains for unstructured data including language, image, audio, and video. Yet, the scenario most prominent in real-world applications is the existence of combination of structured (including tabular and time-series) and unstructured data, and this has so far been understudied. Towards this end, we propose LANISTR, a novel attention-based framework to learn from LANguage, Image, and STRuctured data. We introduce a new multimodal fusion module with a similarity-based multimodal masking loss that enables LANISTR to learn cross-modal relations from large-scale multimodal data with missing modalities during training and test time. On two publicly available challenging datasets, MIMIC-IV and Amazon Product Review, LANISTR achieves absolute improvements of 6.47% (AUROC) and up to 17.69% (accuracy), respectively, compared to the state-of-the-art multimodal models while showing superior generalization capabilities.

Long-context capability is critical for multi-modal foundation models. We introduce LongVILA, a full-stack solution for long-context vision-language models, including system, model training, and dataset development. On the system side, we introduce the first Multi-Modal Sequence Parallelism (MM-SP) system that enables long-context training and inference, enabling 2M context length training on 256 GPUs. MM-SP is also efficient, being 2.1x - 5.7x faster than Ring-Style Sequence Parallelism and 1.1x - 1.4x faster than Megatron-LM in text-only settings. Moreover, it seamlessly integrates with Hugging Face Transformers. For model training, we propose a five-stage pipeline comprising alignment, pre-training, context extension, and long-short joint supervised fine-tuning. Regarding datasets, we meticulously construct large-scale visual language pre-training datasets and long video instruction-following datasets to support our multi-stage training process. The full-stack solution extends the feasible frame number of VILA by a factor of 128 (from 8 to 1024 frames) and improves long video captioning score from 2.00 to 3.26 (1.6x), achieving 99.5% accuracy in 1400-frames video (274k context length) needle in a haystack. LongVILA-8B also demonstrates a consistent improvement in performance on long videos within the VideoMME benchmark as the video frames increase.

Recently, augmenting Vision-Language-Action models (VLAs) with world modeling has shown promise in improving robotic policy learning. However, it remains challenging to jointly predict next-state observations and action sequences because of the inherent difference between the two modalities. To address this, we propose DUal-STream diffusion (DUST), a world-model augmented VLA framework that handles the modality conflict and enhances the performance of VLAs across diverse tasks. Specifically, we propose a multimodal diffusion transformer architecture that explicitly maintains separate modality streams while still enabling cross-modal knowledge sharing. In addition, we introduce independent noise perturbations for each modality and a decoupled flow-matching loss. This design enables the model to learn the joint distribution in a bidirectional manner while avoiding the need for a unified latent space. Based on the decoupling of modalities during training, we also introduce a joint sampling method that supports test-time scaling, where action and vision tokens evolve asynchronously at different rates. Through experiments on simulated benchmarks such as RoboCasa and GR-1, DUST achieves up to 6% gains over baseline methods, while our test-time scaling approach provides an additional 2-5% boost. On real-world tasks with the Franka Research 3, DUST improves success rates by 13%, confirming its effectiveness beyond simulation. Furthermore, pre-training on action-free videos from BridgeV2 yields significant transfer gains on RoboCasa, underscoring DUST's potential for large-scale VLA pretraining.

We present Multiscale Vision Transformers (MViT) for video and image recognition, by connecting the seminal idea of multiscale feature hierarchies with transformer models. Multiscale Transformers have several channel-resolution scale stages. Starting from the input resolution and a small channel dimension, the stages hierarchically expand the channel capacity while reducing the spatial resolution. This creates a multiscale pyramid of features with early layers operating at high spatial resolution to model simple low-level visual information, and deeper layers at spatially coarse, but complex, high-dimensional features. We evaluate this fundamental architectural prior for modeling the dense nature of visual signals for a variety of video recognition tasks where it outperforms concurrent vision transformers that rely on large scale external pre-training and are 5-10x more costly in computation and parameters. We further remove the temporal dimension and apply our model for image classification where it outperforms prior work on vision transformers. Code is available at: https://github.com/facebookresearch/SlowFast

Predicting when to initiate speech in real-world environments remains a fundamental challenge for conversational agents. We introduce EgoSpeak, a novel framework for real-time speech initiation prediction in egocentric streaming video. By modeling the conversation from the speaker's first-person viewpoint, EgoSpeak is tailored for human-like interactions in which a conversational agent must continuously observe its environment and dynamically decide when to talk. Our approach bridges the gap between simplified experimental setups and complex natural conversations by integrating four key capabilities: (1) first-person perspective, (2) RGB processing, (3) online processing, and (4) untrimmed video processing. We also present YT-Conversation, a diverse collection of in-the-wild conversational videos from YouTube, as a resource for large-scale pretraining. Experiments on EasyCom and Ego4D demonstrate that EgoSpeak outperforms random and silence-based baselines in real time. Our results also highlight the importance of multimodal input and context length in effectively deciding when to speak.

Training generalist policies for robotic manipulation has shown great promise, as they enable language-conditioned, multi-task behaviors across diverse scenarios. However, evaluating these policies remains difficult because real-world testing is expensive, time-consuming, and labor-intensive. It also requires frequent environment resets and carries safety risks when deploying unproven policies on physical robots. Manually creating and populating simulation environments with assets for robotic manipulation has not addressed these issues, primarily due to the significant engineering effort required and the substantial sim-to-real gap, both in terms of physics and rendering. In this paper, we explore the use of action-conditional video generation models as a scalable way to learn world models for policy evaluation. We demonstrate how to incorporate action conditioning into existing pre-trained video generation models. This allows leveraging internet-scale in-the-wild online videos during the pre-training stage and alleviates the need for a large dataset of paired video-action data, which is expensive to collect for robotic manipulation. Our paper examines the effect of dataset diversity, pre-trained weights, and common failure cases for the proposed evaluation pipeline. Our experiments demonstrate that across various metrics, including policy ranking and the correlation between actual policy values and predicted policy values, these models offer a promising approach for evaluating policies without requiring real-world interactions.

In recent years, large-scale generative models for visual content (e.g., images, videos, and 3D objects/scenes) have made remarkable progress. However, training large-scale video generation models remains particularly challenging and resource-intensive due to cross-modal text-video alignment, the long sequences involved, and the complex spatiotemporal dependencies. To address these challenges, we present a training framework that optimizes four pillars: (i) data processing, (ii) model architecture, (iii) training strategy, and (iv) infrastructure for large-scale video generation models. These optimizations delivered significant efficiency gains and performance improvements across all stages of data preprocessing, video compression, parameter scaling, curriculum-based pretraining, and alignment-focused post-training. Our resulting model, MUG-V 10B, matches recent state-of-the-art video generators overall and, on e-commerce-oriented video generation tasks, surpasses leading open-source baselines in human evaluations. More importantly, we open-source the complete stack, including model weights, Megatron-Core-based large-scale training code, and inference pipelines for video generation and enhancement. To our knowledge, this is the first public release of large-scale video generation training code that exploits Megatron-Core to achieve high training efficiency and near-linear multi-node scaling, details are available in https://github.com/Shopee-MUG/MUG-V{our webpage}.

Vision-and-language navigation (VLN) requires an embodied agent to navigate in realistic 3D environments using natural language instructions. Existing VLN methods suffer from training on small-scale environments or unreasonable path-instruction datasets, limiting the generalization to unseen environments. There are massive house tour videos on YouTube, providing abundant real navigation experiences and layout information. However, these videos have not been explored for VLN before. In this paper, we propose to learn an agent from these videos by creating a large-scale dataset which comprises reasonable path-instruction pairs from house tour videos and pre-training the agent on it. To achieve this, we have to tackle the challenges of automatically constructing path-instruction pairs and exploiting real layout knowledge from raw and unlabeled videos. To address these, we first leverage an entropy-based method to construct the nodes of a path trajectory. Then, we propose an action-aware generator for generating instructions from unlabeled trajectories. Last, we devise a trajectory judgment pretext task to encourage the agent to mine the layout knowledge. Experimental results show that our method achieves state-of-the-art performance on two popular benchmarks (R2R and REVERIE). Code is available at https://github.com/JeremyLinky/YouTube-VLN

Vision-language pre-training has significantly elevated performance across a wide range of image-language applications. Yet, the pre-training process for video-related tasks demands exceptionally large computational and data resources, which hinders the progress of video-language models. This paper investigates a straightforward, highly efficient, and resource-light approach to adapting an existing image-language pre-trained model for dense video understanding. Our preliminary experiments reveal that directly fine-tuning pre-trained image-language models with multiple frames as inputs on video datasets leads to performance saturation or even a drop. Our further investigation reveals that it is largely attributed to the bias of learned high-norm visual features. Motivated by this finding, we propose a simple but effective pooling strategy to smooth the feature distribution along the temporal dimension and thus reduce the dominant impacts from the extreme features. The new model is termed Pooling LLaVA, or in short. achieves new state-of-the-art performance on modern benchmark datasets for both video question-answer and captioning tasks. Notably, on the recent popular Video ChatGPT benchmark, PLLaVA achieves a score of 3.48 out of 5 on average of five evaluated dimensions, exceeding the previous SOTA results from GPT4V (IG-VLM) by 9\%. On the latest multi-choice benchmark MVBench, PLLaVA achieves 58.1\% accuracy on average across 20 sub-tasks, 14.5\% higher than GPT4V (IG-VLM). Code is available at https://github.com/magic-research/PLLaVA.

We introduce VideoPrism, a general-purpose video encoder that tackles diverse video understanding tasks with a single frozen model. We pretrain VideoPrism on a heterogeneous corpus containing 36M high-quality video-caption pairs and 582M video clips with noisy parallel text (e.g., ASR transcripts). The pretraining approach improves upon masked autoencoding by global-local distillation of semantic video embeddings and a token shuffling scheme, enabling VideoPrism to focus primarily on the video modality while leveraging the invaluable text associated with videos. We extensively test VideoPrism on four broad groups of video understanding tasks, from web video question answering to CV for science, achieving state-of-the-art performance on 30 out of 33 video understanding benchmarks.

Despite the rapid progress of instruction-based image editing, its extension to video remains underexplored, primarily due to the prohibitive cost and complexity of constructing large-scale paired video editing datasets. To address this challenge, we introduce a low-cost pretraining strategy for instruction-based video editing that leverages in-context learning from unpaired video clips. We show that pretraining a foundation video generation model with this strategy endows it with general editing capabilities, such as adding, replacing, or deleting operations, according to input editing instructions. The pretrained model can then be efficiently refined with a small amount of high-quality paired editing data. Built upon HunyuanVideoT2V, our framework first pretrains on approximately 1M real video clips to learn basic editing concepts, and subsequently fine-tunes on fewer than 150k curated editing pairs to extend more editing tasks and improve the editing quality. Comparative experiments show that our method surpasses existing instruction-based video editing approaches in both instruction alignment and visual fidelity, achieving a 12\% improvement in editing instruction following and a 15\% improvement in editing quality.

Concepts involved in long-form videos such as people, objects, and their interactions, can be viewed as following an implicit prior. They are notably complex and continue to pose challenges to be comprehensively learned. In recent years, generative pre-training (GPT) has exhibited versatile capacities in modeling any kind of text content even visual locations. Can this manner work for learning long-form video prior? Instead of operating on pixel space, it is efficient to employ visual locations like bounding boxes and keypoints to represent key information in videos, which can be simply discretized and then tokenized for consumption by GPT. Due to the scarcity of suitable data, we create a new dataset called Storyboard20K from movies to serve as a representative. It includes synopses, shot-by-shot keyframes, and fine-grained annotations of film sets and characters with consistent IDs, bounding boxes, and whole body keypoints. In this way, long-form videos can be represented by a set of tokens and be learned via generative pre-training. Experimental results validate that our approach has great potential for learning long-form video prior. Code and data will be released at https://github.com/showlab/Long-form-Video-Prior.

Large text-to-video models trained on internet-scale data have demonstrated exceptional capabilities in generating high-fidelity videos from arbitrary textual descriptions. However, adapting these models to tasks with limited domain-specific data, such as animation or robotics videos, poses a significant computational challenge, since finetuning a pretrained large model can be prohibitively expensive. Inspired by how a small modifiable component (e.g., prompts, prefix-tuning) can adapt a large language model to perform new tasks without requiring access to the model weights, we investigate how to adapt a large pretrained text-to-video model to a variety of downstream domains and tasks without finetuning. In answering this question, we propose Video Adapter, which leverages the score function of a large pretrained video diffusion model as a probabilistic prior to guide the generation of a task-specific small video model. Our experiments show that Video Adapter is capable of incorporating the broad knowledge and preserving the high fidelity of a large pretrained video model in a task-specific small video model that is able to generate high-quality yet specialized videos on a variety of tasks such as animation, egocentric modeling, and modeling of simulated and real-world robotics data. More videos can be found on the website https://video-adapter.github.io/.

Scale is the primary factor for building a powerful foundation model that could well generalize to a variety of downstream tasks. However, it is still challenging to train video foundation models with billions of parameters. This paper shows that video masked autoencoder (VideoMAE) is a scalable and general self-supervised pre-trainer for building video foundation models. We scale the VideoMAE in both model and data with a core design. Specifically, we present a dual masking strategy for efficient pre-training, with an encoder operating on a subset of video tokens and a decoder processing another subset of video tokens. Although VideoMAE is very efficient due to high masking ratio in encoder, masking decoder can still further reduce the overall computational cost. This enables the efficient pre-training of billion-level models in video. We also use a progressive training paradigm that involves an initial pre-training on a diverse multi-sourced unlabeled dataset, followed by a post-pre-training on a mixed labeled dataset. Finally, we successfully train a video ViT model with a billion parameters, which achieves a new state-of-the-art performance on the datasets of Kinetics (90.0% on K400 and 89.9% on K600) and Something-Something (68.7% on V1 and 77.0% on V2). In addition, we extensively verify the pre-trained video ViT models on a variety of downstream tasks, demonstrating its effectiveness as a general video representation learner. The code and model is available at https://github.com/OpenGVLab/VideoMAEv2.

In recent years, the development of Large Language Models (LLMs) has significantly advanced, extending their capabilities to multimodal tasks through Multimodal Large Language Models (MLLMs). However, video understanding remains a challenging area due to the dynamic and information-dense nature of videos. Existing models struggle with the trade-off between spatial resolution and temporal coverage when processing video content. We present Keye-VL-1.5, which addresses fundamental challenges in video comprehension through three key innovations. First, we introduce a novel Slow-Fast video encoding strategy that dynamically allocates computational resources based on inter-frame similarity, processing key frames with significant visual changes at higher resolution (Slow pathway) while handling relatively static frames with increased temporal coverage at lower resolution (Fast pathway). Second, we implement a progressive four-stage pre-training methodology that systematically extends the model's context length from 8K to 128K tokens, enabling processing of longer videos and more complex visual content. Third, we develop a comprehensive post-training pipeline focusing on reasoning enhancement and human preference alignment, incorporating a 5-step chain-of-thought data construction process, iterative GSPO-based reinforcement learning with progressive prompt hinting for difficult cases, and alignment training. Through extensive evaluation on public benchmarks and rigorous internal human assessment, Keye-VL-1.5 demonstrates significant improvements over existing models, particularly excelling in video understanding tasks while maintaining competitive performance on general multimodal benchmarks.

Contrastive language-image pretraining has shown great success in learning visual-textual joint representation from web-scale data, demonstrating remarkable "zero-shot" generalization ability for various image tasks. However, how to effectively expand such new language-image pretraining methods to video domains is still an open problem. In this work, we present a simple yet effective approach that adapts the pretrained language-image models to video recognition directly, instead of pretraining a new model from scratch. More concretely, to capture the long-range dependencies of frames along the temporal dimension, we propose a cross-frame attention mechanism that explicitly exchanges information across frames. Such module is lightweight and can be plugged into pretrained language-image models seamlessly. Moreover, we propose a video-specific prompting scheme, which leverages video content information for generating discriminative textual prompts. Extensive experiments demonstrate that our approach is effective and can be generalized to different video recognition scenarios. In particular, under fully-supervised settings, our approach achieves a top-1 accuracy of 87.1% on Kinectics-400, while using 12 times fewer FLOPs compared with Swin-L and ViViT-H. In zero-shot experiments, our approach surpasses the current state-of-the-art methods by +7.6% and +14.9% in terms of top-1 accuracy under two popular protocols. In few-shot scenarios, our approach outperforms previous best methods by +32.1% and +23.1% when the labeled data is extremely limited. Code and models are available at https://aka.ms/X-CLIP

We present a framework for pre-training of 3D hand pose estimation from in-the-wild hand images sharing with similar hand characteristics, dubbed SimHand. Pre-training with large-scale images achieves promising results in various tasks, but prior methods for 3D hand pose pre-training have not fully utilized the potential of diverse hand images accessible from in-the-wild videos. To facilitate scalable pre-training, we first prepare an extensive pool of hand images from in-the-wild videos and design our pre-training method with contrastive learning. Specifically, we collect over 2.0M hand images from recent human-centric videos, such as 100DOH and Ego4D. To extract discriminative information from these images, we focus on the similarity of hands: pairs of non-identical samples with similar hand poses. We then propose a novel contrastive learning method that embeds similar hand pairs closer in the feature space. Our method not only learns from similar samples but also adaptively weights the contrastive learning loss based on inter-sample distance, leading to additional performance gains. Our experiments demonstrate that our method outperforms conventional contrastive learning approaches that produce positive pairs sorely from a single image with data augmentation. We achieve significant improvements over the state-of-the-art method (PeCLR) in various datasets, with gains of 15% on FreiHand, 10% on DexYCB, and 4% on AssemblyHands. Our code is available at https://github.com/ut-vision/SiMHand.

Vision-language pre-training (VLP) offers unique advantages for surgery by aligning language with surgical videos, enabling workflow understanding and transfer across tasks without relying on expert-labeled datasets. However, progress in surgical VLP remains constrained by the limited scale, procedural diversity, semantic quality, and hierarchical structure of existing datasets. In this work, we present SurgLaVi, the largest and most diverse surgical vision-language dataset to date, comprising nearly 240k clip-caption pairs from more than 200 procedures, and comprising hierarchical levels at phase-, step-, and task-level. At the core of SurgLaVi lies a fully automated pipeline that systematically generates fine-grained transcriptions of surgical videos and segments them into coherent procedural units. To ensure high-quality annotations, it applies dual-modality filtering to remove irrelevant and noisy samples. Within this framework, the resulting captions are enriched with contextual detail, producing annotations that are both semantically rich and easy to interpret. To ensure accessibility, we release SurgLaVi-eta, an open-source derivative of 113k clip-caption pairs constructed entirely from public data, which is over four times larger than existing surgical VLP datasets. To demonstrate the value of SurgLaVi datasets, we introduce SurgCLIP, a CLIP-style video-text contrastive framework with dual encoders, as a representative base model. SurgCLIP achieves consistent improvements across phase, step, action, and tool recognition, surpassing prior state-of-the-art methods, often by large margins. These results validate that large-scale, semantically rich, and hierarchically structured datasets directly translate into stronger and more generalizable representations, establishing SurgLaVi as a key resource for developing surgical foundation models.

Procedural videos, exemplified by recipe demonstrations, are instrumental in conveying step-by-step instructions. However, understanding such videos is challenging as it involves the precise localization of steps and the generation of textual instructions. Manually annotating steps and writing instructions is costly, which limits the size of current datasets and hinders effective learning. Leveraging large but noisy video-transcript datasets for pre-training can boost performance but demands significant computational resources. Furthermore, transcripts contain irrelevant content and differ in style from human-written instructions. To mitigate these issues, we propose a novel technique, Sieve-&-Swap, to automatically generate high-quality training data for the recipe domain: (i) Sieve: filters irrelevant transcripts and (ii) Swap: acquires high-quality text by replacing transcripts with human-written instruction from a text-only recipe dataset. The resulting dataset is three orders of magnitude smaller than current web-scale datasets but enables efficient training of large-scale models. Alongside Sieve-&-Swap, we propose Procedure Transformer (ProcX), a model for end-to-end step localization and instruction generation for procedural videos. When pre-trained on our curated dataset, this model achieves state-of-the-art performance on YouCook2 and Tasty while using a fraction of the training data. We have released code and dataset.

Generating videos from text is a challenging task due to its high computational requirements for training and infinite possible answers for evaluation. Existing works typically experiment on simple or small datasets, where the generalization ability is quite limited. In this work, we propose GODIVA, an open-domain text-to-video pretrained model that can generate videos from text in an auto-regressive manner using a three-dimensional sparse attention mechanism. We pretrain our model on Howto100M, a large-scale text-video dataset that contains more than 136 million text-video pairs. Experiments show that GODIVA not only can be fine-tuned on downstream video generation tasks, but also has a good zero-shot capability on unseen texts. We also propose a new metric called Relative Matching (RM) to automatically evaluate the video generation quality. Several challenges are listed and discussed as future work.

Recovering 3D structures with open-vocabulary scene understanding from 2D images is a fundamental but daunting task. Recent developments have achieved this by performing per-scene optimization with embedded language information. However, they heavily rely on the calibrated dense-view reconstruction paradigm, thereby suffering from severe rendering artifacts and implausible semantic synthesis when limited views are available. In this paper, we introduce a novel generative framework, coined LangScene-X, to unify and generate 3D consistent multi-modality information for reconstruction and understanding. Powered by the generative capability of creating more consistent novel observations, we can build generalizable 3D language-embedded scenes from only sparse views. Specifically, we first train a TriMap video diffusion model that can generate appearance (RGBs), geometry (normals), and semantics (segmentation maps) from sparse inputs through progressive knowledge integration. Furthermore, we propose a Language Quantized Compressor (LQC), trained on large-scale image datasets, to efficiently encode language embeddings, enabling cross-scene generalization without per-scene retraining. Finally, we reconstruct the language surface fields by aligning language information onto the surface of 3D scenes, enabling open-ended language queries. Extensive experiments on real-world data demonstrate the superiority of our LangScene-X over state-of-the-art methods in terms of quality and generalizability. Project Page: https://liuff19.github.io/LangScene-X.

Building video-language foundation models is costly and difficult due to the redundant nature of video data and the lack of high-quality video-language datasets. In this paper, we propose an efficient framework to harvest video foundation models from image ones. Our method is intuitively simple by randomly dropping input video patches and masking out input text during the post-pretraining procedure. The patch dropping boosts the training efficiency significantly and text masking enforces the learning of cross-modal fusion. We conduct extensive experiments to validate the effectiveness of our method on a wide range of video-language downstream tasks including various zero-shot tasks, video question answering, and video-text retrieval. Despite its simplicity, our method achieves state-of-the-art performances, which are comparable to some heavily pretrained video foundation models. Our method is extremely efficient and can be trained in less than one day on 8 GPUs, requiring only WebVid-10M as pretraining data. We hope our method can serve as a simple yet strong counterpart for prevalent video foundation models, provide useful insights when building them, and make large pretrained models more accessible and sustainable. This is part of the InternVideo project https://github.com/OpenGVLab/InternVideo.

The automatic generation of stylized co-speech gestures has recently received increasing attention. Previous systems typically allow style control via predefined text labels or example motion clips, which are often not flexible enough to convey user intent accurately. In this work, we present GestureDiffuCLIP, a neural network framework for synthesizing realistic, stylized co-speech gestures with flexible style control. We leverage the power of the large-scale Contrastive-Language-Image-Pre-training (CLIP) model and present a novel CLIP-guided mechanism that extracts efficient style representations from multiple input modalities, such as a piece of text, an example motion clip, or a video. Our system learns a latent diffusion model to generate high-quality gestures and infuses the CLIP representations of style into the generator via an adaptive instance normalization (AdaIN) layer. We further devise a gesture-transcript alignment mechanism that ensures a semantically correct gesture generation based on contrastive learning. Our system can also be extended to allow fine-grained style control of individual body parts. We demonstrate an extensive set of examples showing the flexibility and generalizability of our model to a variety of style descriptions. In a user study, we show that our system outperforms the state-of-the-art approaches regarding human likeness, appropriateness, and style correctness.

Recent advances in medical image segmentation have been driven by deep learning; however, most existing methods remain limited by modality-specific designs and exhibit poor adaptability to dynamic medical imaging scenarios. The Segment Anything Model 2 (SAM2) and its related variants, which introduce a streaming memory mechanism for real-time video segmentation, present new opportunities for prompt-based, generalizable solutions. Nevertheless, adapting these models to medical video scenarios typically requires large-scale datasets for retraining or transfer learning, leading to high computational costs and the risk of catastrophic forgetting. To address these challenges, we propose DD-SAM2, an efficient adaptation framework for SAM2 that incorporates a Depthwise-Dilated Adapter (DD-Adapter) to enhance multi-scale feature extraction with minimal parameter overhead. This design enables effective fine-tuning of SAM2 on medical videos with limited training data. Unlike existing adapter-based methods focused solely on static images, DD-SAM2 fully exploits SAM2's streaming memory for medical video object tracking and segmentation. Comprehensive evaluations on TrackRad2025 (tumor segmentation) and EchoNet-Dynamic (left ventricle tracking) datasets demonstrate superior performance, achieving Dice scores of 0.93 and 0.97, respectively. To the best of our knowledge, this work provides an initial attempt at systematically exploring adapter-based SAM2 fine-tuning for medical video segmentation and tracking. Code, datasets, and models will be publicly available at https://github.com/apple1986/DD-SAM2.

In this paper, we show that useful video representations can be learned from synthetic videos and natural images, without incorporating natural videos in the training. We propose a progression of video datasets synthesized by simple generative processes, that model a growing set of natural video properties (e.g. motion, acceleration, and shape transformations). The downstream performance of video models pre-trained on these generated datasets gradually increases with the dataset progression. A VideoMAE model pre-trained on our synthetic videos closes 97.2% of the performance gap on UCF101 action classification between training from scratch and self-supervised pre-training from natural videos, and outperforms the pre-trained model on HMDB51. Introducing crops of static images to the pre-training stage results in similar performance to UCF101 pre-training and outperforms the UCF101 pre-trained model on 11 out of 14 out-of-distribution datasets of UCF101-P. Analyzing the low-level properties of the datasets, we identify correlations between frame diversity, frame similarity to natural data, and downstream performance. Our approach provides a more controllable and transparent alternative to video data curation processes for pre-training.

Pre-trained video models learn powerful priors for generating high-quality, temporally coherent content. While these models excel at temporal coherence, their dynamics are often constrained by the continuous nature of their training data. We hypothesize that by injecting the rich and unconstrained content diversity from image data into this coherent temporal framework, we can generate image sets that feature both natural transitions and a far more expansive dynamic range. To this end, we introduce iMontage, a unified framework designed to repurpose a powerful video model into an all-in-one image generator. The framework consumes and produces variable-length image sets, unifying a wide array of image generation and editing tasks. To achieve this, we propose an elegant and minimally invasive adaptation strategy, complemented by a tailored data curation process and training paradigm. This approach allows the model to acquire broad image manipulation capabilities without corrupting its invaluable original motion priors. iMontage excels across several mainstream many-in-many-out tasks, not only maintaining strong cross-image contextual consistency but also generating scenes with extraordinary dynamics that surpass conventional scopes. Find our homepage at: https://kr1sjfu.github.io/iMontage-web/.

Large language models, trained on extensive corpora, successfully unify diverse linguistic tasks within a single generative framework. Inspired by this, recent works like Large Vision Model (LVM) extend this paradigm to vision by organizing tasks into sequential visual sentences, where visual prompts serve as the context to guide outputs. However, such modeling requires task-specific pre-training across modalities and sources, which is costly and limits scalability to unseen tasks. Given that pre-trained video generation models inherently capture temporal sequence dependencies, we explore a more unified and scalable alternative: can a pre-trained video generation model adapt to diverse image and video tasks? To answer this, we propose UniVid, a framework that fine-tunes a video diffusion transformer to handle various vision tasks without task-specific modifications. Tasks are represented as visual sentences, where the context sequence defines both the task and the expected output modality. We evaluate the generalization of UniVid from two perspectives: (1) cross-modal inference with contexts composed of both images and videos, extending beyond LVM's uni-modal setting; (2) cross-source tasks from natural to annotated data, without multi-source pre-training. Despite being trained solely on natural video data, UniVid generalizes well in both settings. Notably, understanding and generation tasks can easily switch by simply reversing the visual sentence order in this paradigm. These findings highlight the potential of pre-trained video generation models to serve as a scalable and unified foundation for vision modeling. Our code will be released at https://github.com/CUC-MIPG/UniVid.

Large-scale pretrained image-text models have shown incredible zero-shot performance in a handful of tasks, including video ones such as action recognition and text-to-video retrieval. However, these models have not been adapted to video, mainly because they do not account for the time dimension but also because video frames are different from the typical images (e.g., containing motion blur, and less sharpness). In this paper, we present a fine-tuning strategy to refine these large-scale pretrained image-text models for zero-shot video understanding tasks. We show that by carefully adapting these models we obtain considerable improvements on two zero-shot Action Recognition tasks and three zero-shot Text-to-video Retrieval tasks. The code is available at https://github.com/bryant1410/fitclip

Our objective in this work is video-text retrieval - in particular a joint embedding that enables efficient text-to-video retrieval. The challenges in this area include the design of the visual architecture and the nature of the training data, in that the available large scale video-text training datasets, such as HowTo100M, are noisy and hence competitive performance is achieved only at scale through large amounts of compute. We address both these challenges in this paper. We propose an end-to-end trainable model that is designed to take advantage of both large-scale image and video captioning datasets. Our model is an adaptation and extension of the recent ViT and Timesformer architectures, and consists of attention in both space and time. The model is flexible and can be trained on both image and video text datasets, either independently or in conjunction. It is trained with a curriculum learning schedule that begins by treating images as 'frozen' snapshots of video, and then gradually learns to attend to increasing temporal context when trained on video datasets. We also provide a new video-text pretraining dataset WebVid-2M, comprised of over two million videos with weak captions scraped from the internet. Despite training on datasets that are an order of magnitude smaller, we show that this approach yields state-of-the-art results on standard downstream video-retrieval benchmarks including MSR-VTT, MSVD, DiDeMo and LSMDC.

Video understanding represents the most challenging frontier in computer vision, requiring models to reason about complex spatiotemporal relationships, long-term dependencies, and multimodal evidence. The recent emergence of Video-Large Multimodal Models (Video-LMMs), which integrate visual encoders with powerful decoder-based language models, has demonstrated remarkable capabilities in video understanding tasks. However, the critical phase that transforms these models from basic perception systems into sophisticated reasoning engines, post-training, remains fragmented across the literature. This survey provides the first comprehensive examination of post-training methodologies for Video-LMMs, encompassing three fundamental pillars: supervised fine-tuning (SFT) with chain-of-thought, reinforcement learning (RL) from verifiable objectives, and test-time scaling (TTS) through enhanced inference computation. We present a structured taxonomy that clarifies the roles, interconnections, and video-specific adaptations of these techniques, addressing unique challenges such as temporal localization, spatiotemporal grounding, long video efficiency, and multimodal evidence integration. Through systematic analysis of representative methods, we synthesize key design principles, insights, and evaluation protocols while identifying critical open challenges in reward design, scalability, and cost-performance optimization. We further curate essential benchmarks, datasets, and metrics to facilitate rigorous assessment of post-training effectiveness. This survey aims to provide researchers and practitioners with a unified framework for advancing Video-LMM capabilities. Additional resources and updates are maintained at: https://github.com/yunlong10/Awesome-Video-LMM-Post-Training

Video large language models (Vid-LLMs), which excel in diverse video-language tasks, can be effectively constructed by adapting image-pretrained vision-language models (VLMs). However, this adaptation remains challenging, as it requires processing dense and temporally extended visual inputs that exceed the capacity of image-based models. This paper identifies the perception bottleneck and token overload as key challenges in extending image-based VLMs to the video domain. To address these issues, we propose D-CoDe, a training-free adaptation framework that incorporates dynamic compression and question decomposition. Specifically, dynamic compression alleviates the perception bottleneck through adaptive selection of representative frames and content-aware aggregation of spatial tokens, thereby reducing redundancy while preserving informative content. In parallel, question decomposition mitigates token overload by reformulating the original query into sub-questions, guiding the model to focus on distinct aspects of the video and enabling more comprehensive understanding. Experiments demonstrate that D-CoDe effectively improves video understanding across various benchmarks. Furthermore, strong performance on the challenging long-video benchmark highlights the potential of D-CoDe in handling complex video-language tasks. Code is available at https://github.com/hukcc/D-CoDe.

In this work, we introduce long-video masked-embedding autoencoders (LV-MAE), a self-supervised learning framework for long video representation. Our approach treats short- and long-span dependencies as two separate tasks. Such decoupling allows for a more intuitive video processing where short-span spatiotemporal primitives are first encoded and are then used to capture long-range dependencies across consecutive video segments. To achieve this, we leverage advanced off-the-shelf multimodal encoders to extract representations from short segments within the long video, followed by pre-training a masked-embedding autoencoder capturing high-level interactions across segments. LV-MAE is highly efficient to train and enables the processing of much longer videos by alleviating the constraint on the number of input frames. Furthermore, unlike existing methods that typically pre-train on short-video datasets, our approach offers self-supervised pre-training using long video samples (e.g., 20+ minutes video clips) at scale. Using LV-MAE representations, we achieve state-of-the-art results on three long-video benchmarks -- LVU, COIN, and Breakfast -- employing only a simple classification head for either attentive or linear probing. Finally, to assess LV-MAE pre-training and visualize its reconstruction quality, we leverage the video-language aligned space of short video representations to monitor LV-MAE through video-text retrieval.

This paper revisits the standard pretrain-then-finetune paradigm used in computer vision for visual recognition tasks. Typically, state-of-the-art foundation models are pretrained using large scale (weakly) supervised datasets with billions of images. We introduce an additional pre-pretraining stage that is simple and uses the self-supervised MAE technique to initialize the model. While MAE has only been shown to scale with the size of models, we find that it scales with the size of the training dataset as well. Thus, our MAE-based pre-pretraining scales with both model and data size making it applicable for training foundation models. Pre-pretraining consistently improves both the model convergence and the downstream transfer performance across a range of model scales (millions to billions of parameters), and dataset sizes (millions to billions of images). We measure the effectiveness of pre-pretraining on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our largest model achieves new state-of-the-art results on iNaturalist-18 (91.3%), 1-shot ImageNet-1k (62.1%), and zero-shot transfer on Food-101 (96.0%). Our study reveals that model initialization plays a significant role, even for web-scale pretraining with billions of images.

Recent studies have revealed that selecting informative and relevant video frames can significantly improve the performance of Video Large Language Models (Video-LLMs). Current methods, such as reducing inter-frame redundancy, employing separate models for image-text relevance assessment, or utilizing temporal video grounding for event localization, substantially adopt unsupervised learning paradigms, whereas they struggle to address the complex scenarios in long video understanding. We propose Instructed Temporal Grounding for Videos (VideoITG), featuring customized frame sampling aligned with user instructions. The core of VideoITG is the VidThinker pipeline, an automated annotation framework that explicitly mimics the human annotation process. First, it generates detailed clip-level captions conditioned on the instruction; then, it retrieves relevant video segments through instruction-guided reasoning; finally, it performs fine-grained frame selection to pinpoint the most informative visual evidence. Leveraging VidThinker, we construct the VideoITG-40K dataset, containing 40K videos and 500K instructed temporal grounding annotations. We then design a plug-and-play VideoITG model, which takes advantage of visual language alignment and reasoning capabilities of Video-LLMs, for effective frame selection in a discriminative manner. Coupled with Video-LLMs, VideoITG achieves consistent performance improvements across multiple multimodal video understanding benchmarks, showing its superiority and great potentials for video understanding.

The success of Multimodal Large Language Models (MLLMs) in the image domain has garnered wide attention from the research community. Drawing on previous successful experiences, researchers have recently explored extending the success to the video understanding realms. Apart from training from scratch, an efficient way is to utilize the pre-trained image-LLMs, leading to two mainstream approaches, i.e. zero-shot inference and further fine-tuning with video data. In this work, our study of these approaches harvests an effective data augmentation method. We first make a deeper inspection of the zero-shot inference way and identify two limitations, i.e. limited generalization and lack of temporal understanding capabilities. Thus, we further investigate the fine-tuning approach and find a low learning efficiency when simply using all the video data samples, which can be attributed to a lack of instruction diversity. Aiming at this issue, we develop a method called T2Vid to synthesize video-like samples to enrich the instruction diversity in the training corpus. Integrating these data enables a simple and efficient training scheme, which achieves performance comparable to or even superior to using full video datasets by training with just 15% the sample size. Meanwhile, we find that the proposed scheme can boost the performance of long video understanding without training with long video samples. We hope our study will spark more thinking about using MLLMs for video understanding and curation of high-quality data. The code is released at https://github.com/xjtupanda/T2Vid.

Image-Language Foundation Models (ILFM) have demonstrated remarkable success in image-text understanding/generation tasks, providing transferable multimodal representations that generalize across diverse downstream image-based tasks. The advancement of video-text research has spurred growing interest in extending image-based models to the video domain. This paradigm, known as image-to-video transfer learning, succeeds in alleviating the substantial data and computational requirements associated with training video-language foundation models from scratch for video-text learning. This survey provides the first comprehensive review of this emerging field, which begins by summarizing the widely used ILFM and their capabilities. We then systematically classify existing image-to-video transfer learning strategies into two categories: frozen features and modified features, depending on whether the original representations from ILFM are preserved or undergo modifications. Building upon the task-specific nature of image-to-video transfer, this survey methodically elaborates these strategies and details their applications across a spectrum of video-text learning tasks, ranging from fine-grained (e.g., spatio-temporal video grounding) to coarse-grained (e.g., video question answering). We further present a detailed experimental analysis to investigate the efficacy of different image-to-video transfer learning paradigms on a range of downstream video understanding tasks. Finally, we identify prevailing challenges and highlight promising directions for future research. By offering a comprehensive and structured overview, this survey aims to establish a structured roadmap for advancing video-text learning based on existing ILFM, and to inspire future research directions in this rapidly evolving domain.

Recent video and language pretraining frameworks lack the ability to generate sentences. We present Multimodal Video Generative Pretraining (MV-GPT), a new pretraining framework for learning from unlabelled videos which can be effectively used for generative tasks such as multimodal video captioning. Unlike recent video-language pretraining frameworks, our framework trains both a multimodal video encoder and a sentence decoder jointly. To overcome the lack of captions in unlabelled videos, we leverage the future utterance as an additional text source and propose a bidirectional generation objective -- we generate future utterances given the present mulitmodal context, and also the present utterance given future observations. With this objective, we train an encoder-decoder model end-to-end to generate a caption from raw pixels and transcribed speech directly. Our model achieves state-of-the-art performance for multimodal video captioning on four standard benchmarks, as well as for other video understanding tasks such as VideoQA, video retrieval and action classification.

Large Multimodal Models (LMMs) have demonstrated impressive performance in short video understanding tasks but face great challenges when applied to long video understanding. In contrast, Large Language Models (LLMs) exhibit outstanding capabilities in modeling long texts. Existing work attempts to address this issue by introducing long video-text pairs during training. However, these approaches require substantial computational and data resources. In this paper, we tackle the challenge of long video understanding from the perspective of context windows, aiming to apply LMMs to long video tasks without retraining on long video datasets. We first conduct an in-depth analysis of why pretrained LMMs struggle to understand lengthy video content, identifying that discrepancies between visual and language modalities lead to different context windows for visual and language tokens, making it difficult to directly extend the visual tokens to match the language context window. Based on this, we propose to adapt LMMs for long video understanding tasks by extending the visual context window, eliminating the need for retraining on large scalelong video datasets. To further mitigate the significant memory consumption caused by long sequences, we introduce a progressive pooling inference strategy that selectively adjusts the spatial resolution of frame embeddings, reducing the number of visual tokens while retaining important spatial information. Across multiple long video understanding benchmarks, our method consistently improves the performance as the number of video frames increases. On the MLVU benchmark, our method outperforms GPT-4o, even though our model size is only 7B. Additionally, in the 256-frame setting, our method reduces memory usage by approximately 45% compared to the baseline, without introducing any performance loss.

We introduce InternVideo2, a new video foundation model (ViFM) that achieves the state-of-the-art performance in action recognition, video-text tasks, and video-centric dialogue. Our approach employs a progressive training paradigm that unifies the different self- or weakly-supervised learning frameworks of masked video token reconstruction, cross-modal contrastive learning, and next token prediction. Different training stages would guide our model to capture different levels of structure and semantic information through different pretext tasks. At the data level, we prioritize the spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. We scale both data and model size for our InternVideo2. Through extensive experiments, we validate our designs and demonstrate the state-of-the-art performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related captioning, dialogue, and long video understanding benchmarks, highlighting its ability to reason and comprehend long temporal contexts. Code and models are available at https://github.com/OpenGVLab/InternVideo2/.