Video large language models (ViLLMs) excel in general video understanding, e.g., recognizing activities like talking and eating, but struggle with identity-aware comprehension, such as "Wilson is receiving chemotherapy" or "Tom is discussing with Sarah", limiting their applicability in smart healthcare and smart home environments. To address this limitation, we propose a one-shot learning framework PVChat, the first personalized ViLLM that enables subject-aware question answering (QA) from a single video for each subject. Our approach optimizes a Mixture-of-Heads (MoH) enhanced ViLLM on a synthetically augmented video-QA dataset, leveraging a progressive image-to-video learning strategy. Specifically, we introduce an automated augmentation pipeline that synthesizes identity-preserving positive samples and retrieves hard negatives from existing video corpora, generating a diverse training dataset with four QA types: existence, appearance, action, and location inquiries. To enhance subject-specific learning, we propose a ReLU Routing MoH attention mechanism, alongside two novel objectives: (1) Smooth Proximity Regularization for progressive learning through exponential distance scaling and (2) Head Activation Enhancement for balanced attention routing. Finally, we adopt a two-stage training strategy, transitioning from image pre-training to video fine-tuning, enabling a gradual learning process from static attributes to dynamic representations. We evaluate PVChat on diverse datasets covering medical scenarios, TV series, anime, and real-world footage, demonstrating its superiority in personalized feature understanding after learning from a single video, compared to state-of-the-art ViLLMs.
We introduce SlowFast-LLaVA-1.5 (abbreviated as SF-LLaVA-1.5), a family of video large language models (LLMs) offering a token-efficient solution for long-form video understanding. We incorporate the two-stream SlowFast mechanism into a streamlined training pipeline, and perform joint video-image training on a carefully curated data mixture of only publicly available datasets. Our primary focus is on highly efficient model scales (1B and 3B), demonstrating that even relatively small Video LLMs can achieve state-of-the-art performance on video understanding, meeting the demand for mobile-friendly models. Experimental results demonstrate that SF-LLaVA-1.5 achieves superior performance on a wide range of video and image tasks, with robust results at all model sizes (ranging from 1B to 7B). Notably, SF-LLaVA-1.5 achieves state-of-the-art results in long-form video understanding (e.g., LongVideoBench and MLVU) and excels at small scales across various video benchmarks.
Video Anomaly Detection (VAD) remains a fundamental yet formidable task in the video understanding community, with promising applications in areas such as information forensics and public safety protection. Due to the rarity and diversity of anomalies, existing methods only use easily collected regular events to model the inherent normality of normal spatial-temporal patterns in an unsupervised manner. Previous studies have shown that existing unsupervised VAD models are incapable of label-independent data offsets (e.g., scene changes) in real-world scenarios and may fail to respond to light anomalies due to the overgeneralization of deep neural networks. Inspired by causality learning, we argue that there exist causal factors that can adequately generalize the prototypical patterns of regular events and present significant deviations when anomalous instances occur. In this regard, we propose Causal Representation Consistency Learning (CRCL) to implicitly mine potential scene-robust causal variable in unsupervised video normality learning. Specifically, building on the structural causal models, we propose scene-debiasing learning and causality-inspired normality learning to strip away entangled scene bias in deep representations and learn causal video normality, respectively. Extensive experiments on benchmarks validate the superiority of our method over conventional deep representation learning. Moreover, ablation studies and extension validation show that the CRCL can cope with label-independent biases in multi-scene settings and maintain stable performance with only limited training data available.
We propose a new "Unbiased through Textual Description (UTD)" video benchmark based on unbiased subsets of existing video classification and retrieval datasets to enable a more robust assessment of video understanding capabilities. Namely, we tackle the problem that current video benchmarks may suffer from different representation biases, e.g., object bias or single-frame bias, where mere recognition of objects or utilization of only a single frame is sufficient for correct prediction. We leverage VLMs and LLMs to analyze and debias benchmarks from such representation biases. Specifically, we generate frame-wise textual descriptions of videos, filter them for specific information (e.g. only objects) and leverage them to examine representation biases across three dimensions: 1) concept bias - determining if a specific concept (e.g., objects) alone suffice for prediction; 2) temporal bias - assessing if temporal information contributes to prediction; and 3) common sense vs. dataset bias - evaluating whether zero-shot reasoning or dataset correlations contribute to prediction. We conduct a systematic analysis of 12 popular video classification and retrieval datasets and create new object-debiased test splits for these datasets. Moreover, we benchmark 30 state-of-the-art video models on original and debiased splits and analyze biases in the models. To facilitate the future development of more robust video understanding benchmarks and models, we release: "UTD-descriptions", a dataset with our rich structured descriptions for each dataset, and "UTD-splits", a dataset of object-debiased test splits.
Despite advanced token compression techniques, existing multimodal large language models (MLLMs) still struggle with hour-long video understanding. In this work, we propose Video-XL-Pro, an efficient method for extremely long video understanding, built upon Reconstructive Compression of Tokens (ReCoT), a learnable module that leverages self-supervised learning to generate comprehensive and compact video tokens. ReCoT introduces two key components: (i) Dynamic Token Synthesizer (DTS): DTS generates pseudo-video tokens from static image tokens by learning intra-token relationships, which are then used in masked video modeling. (ii) Semantic-Guided Masking (SGM): SGM adaptively masks redundant visual tokens to facilitate more effective reconstructive learning. To improve training efficiency in MLLMs fine-tuning, we introduce a video-specific dataset pruning strategy and design a simple yet Query-aware Selector that enables the model to precisely locate query-relevant video tokens. With only 3B parameters, Video-XL-Pro outperforms most 7B models trained on larger datasets across multiple long video understanding benchmarks. Moreover, it can process over 8K frames on a single A100 GPU while maintaining high-quality performance.
Most Video-Large Language Models (Video-LLMs) adopt an encoder-decoder framework, where a vision encoder extracts frame-wise features for processing by a language model. However, this approach incurs high computational costs, introduces resolution biases, and struggles to capture fine-grained multimodal interactions. To overcome these limitations, we propose ELVA, an encoder-free Video-LLM that directly models nuanced video-language interactions without relying on a vision encoder. ELVA employs token merging to construct a bottom-up hierarchical representation and incorporates a video guidance supervisor for direct spatiotemporal representation learning. Additionally, a hybrid-resolution mechanism strategically integrates high- and low-resolution frames as inputs to achieve an optimal balance between performance and efficiency. With only 7M publicly available video-text pairs, ELVA achieves performance on par with encoder-based Video-LLMs while reducing FLOPs by up to 95\% and inference latency by 92\%, offering a scalable and efficient solution for real-time video understanding.
Multimodal Large Language Models (MLLMs) have demonstrated impressive 2D image/video understanding capabilities. However, there are no publicly standardized benchmarks to assess the abilities of MLLMs in understanding the 4D objects (3D objects with temporal evolution over time). In this paper, we introduce 4D-Bench, the first benchmark to evaluate the capabilities of MLLMs in 4D object understanding, featuring tasks in 4D object Question Answering (4D object QA) and 4D object captioning. 4D-Bench provides 4D objects with diverse categories, high-quality annotations, and tasks necessitating multi-view spatial-temporal understanding, different from existing 2D image/video-based benchmarks. With 4D-Bench, we evaluate a wide range of open-source and closed-source MLLMs. The results from the 4D object captioning experiment indicate that MLLMs generally exhibit weaker temporal understanding compared to their appearance understanding, notably, while open-source models approach closed-source performance in appearance understanding, they show larger performance gaps in temporal understanding. 4D object QA yields surprising findings: even with simple single-object videos, MLLMs perform poorly, with state-of-the-art GPT-4o achieving only 63\% accuracy compared to the human baseline of 91\%. These findings highlight a substantial gap in 4D object understanding and the need for further advancements in MLLMs.
Large Vision-Language Models (LVLMs) have made significant progress in the field of video understanding recently. However, current benchmarks uniformly lean on text prompts for evaluation, which often necessitate complex referential language and fail to provide precise spatial and temporal references. This limitation diminishes the experience and efficiency of human-model interaction. To address this limitation, we propose the Video Visual Prompt Benchmark(V2P-Bench), a comprehensive benchmark specifically designed to evaluate LVLMs' video understanding capabilities in multimodal human-model interaction scenarios. V2P-Bench includes 980 unique videos and 1,172 QA pairs, covering 5 main tasks and 12 dimensions, facilitating instance-level fine-grained understanding aligned with human cognition. Benchmarking results reveal that even the most powerful models perform poorly on V2P-Bench (65.4% for GPT-4o and 67.9% for Gemini-1.5-Pro), significantly lower than the human experts' 88.3%, highlighting the current shortcomings of LVLMs in understanding video visual prompts. We hope V2P-Bench will serve as a foundation for advancing multimodal human-model interaction and video understanding evaluation. Project page: https://github.com/gaotiexinqu/V2P-Bench.
Monitoring wildlife is essential for ecology and ethology, especially in light of the increasing human impact on ecosystems. Camera traps have emerged as habitat-centric sensors enabling the study of wildlife populations at scale with minimal disturbance. However, the lack of annotated video datasets limits the development of powerful video understanding models needed to process the vast amount of fieldwork data collected. To advance research in wild animal behavior monitoring we present MammAlps, a multimodal and multi-view dataset of wildlife behavior monitoring from 9 camera-traps in the Swiss National Park. MammAlps contains over 14 hours of video with audio, 2D segmentation maps and 8.5 hours of individual tracks densely labeled for species and behavior. Based on 6135 single animal clips, we propose the first hierarchical and multimodal animal behavior recognition benchmark using audio, video and reference scene segmentation maps as inputs. Furthermore, we also propose a second ecology-oriented benchmark aiming at identifying activities, species, number of individuals and meteorological conditions from 397 multi-view and long-term ecological events, including false positive triggers. We advocate that both tasks are complementary and contribute to bridging the gap between machine learning and ecology. Code and data are available at: https://github.com/eceo-epfl/MammAlps
Natural language video localization (NLVL) is a crucial task in video understanding that aims to localize the target moment in videos specified by a given language description. Recently, a point-supervised paradigm has been presented to address this task, requiring only a single annotated frame within the target moment rather than complete temporal boundaries. Compared with the fully-supervised paradigm, it offers a balance between localization accuracy and annotation cost. However, due to the absence of complete annotation, it is challenging to align the video content with language descriptions, consequently hindering accurate moment prediction. To address this problem, we propose a new COllaborative Temporal consistEncy Learning (COTEL) framework that leverages the synergy between saliency detection and moment localization to strengthen the video-language alignment. Specifically, we first design a frame- and a segment-level Temporal Consistency Learning (TCL) module that models semantic alignment across frame saliencies and sentence-moment pairs. Then, we design a cross-consistency guidance scheme, including a Frame-level Consistency Guidance (FCG) and a Segment-level Consistency Guidance (SCG), that enables the two temporal consistency learning paths to reinforce each other mutually. Further, we introduce a Hierarchical Contrastive Alignment Loss (HCAL) to comprehensively align the video and text query. Extensive experiments on two benchmarks demonstrate that our method performs favorably against SoTA approaches. We will release all the source codes.
Videos, with their unique temporal dimension, demand precise grounded understanding, where answers are directly linked to visual, interpretable evidence. Despite significant breakthroughs in reasoning capabilities within Large Language Models, multi-modal reasoning - especially for videos - remains unexplored. In this work, we introduce VideoMind, a novel video-language agent designed for temporal-grounded video understanding. VideoMind incorporates two key innovations: (i) We identify essential capabilities for video temporal reasoning and develop a role-based agentic workflow, including a planner for coordinating different roles, a grounder for temporal localization, a verifier to assess temporal interval accuracy, and an answerer for question-answering. (ii) To efficiently integrate these diverse roles, we propose a novel Chain-of-LoRA strategy, enabling seamless role-switching via lightweight LoRA adaptors while avoiding the overhead of multiple models, thus balancing efficiency and flexibility. Extensive experiments on 14 public benchmarks, including 3 on grounded video question-answering (Grounded VideoQA), 6 on video temporal grounding (VTG), and 5 on general video question-answering (VideoQA), verify that our agent achieves state-of-the-art performance on diverse video understanding tasks, underscoring its effectiveness in advancing video agent and long-form temporal reasoning.
Existing research of video understanding still struggles to achieve in-depth comprehension and reasoning in complex videos, primarily due to the under-exploration of two key bottlenecks: fine-grained spatial-temporal perceptive understanding and cognitive-level video scene comprehension. This paper bridges the gap by presenting a novel solution. We first introduce a novel video Multimodal Large Language Model (MLLM), MotionEpic, which achieves fine-grained pixel-level spatial-temporal video grounding by integrating video spatial-temporal scene graph (STSG) representation. Building upon MotionEpic, we then develop a Video-of-Thought (VoT) reasoning framework. VoT inherits the Chain-of-Thought (CoT) core, breaking down a complex task into simpler and manageable sub-problems, and addressing them step-by-step from a low-level pixel perception to high-level cognitive interpretation. Extensive experiments across various complex video QA benchmarks demonstrate that our overall framework strikingly boosts existing state-of-the-art. To our knowledge, this is the first attempt at successfully implementing the CoT technique for achieving human-level video reasoning, where we show great potential in extending it to a wider range of video understanding scenarios. Project is open at https://haofei.vip/VoT
Video question answering (VideoQA) enables machines to extract and comprehend key information from videos through natural language interaction, which is a critical step towards achieving intelligence. However, the demand for a thorough understanding of videos and high computational costs still limit the widespread applications of VideoQA. To address it, we propose Agentic Keyframe Search (AKeyS), a simple yet powerful algorithm for identifying keyframes in the VideoQA task. It can effectively distinguish key information from redundant, irrelevant content by leveraging modern language agents to direct classical search algorithms. Specifically, we first segment the video and organize it as a tree structure. Then, AKeyS uses a language agent to estimate heuristics and movement costs while dynamically expanding nodes. Finally, the agent determines if sufficient keyframes have been collected based on termination conditions and provides answers. Extensive experiments on the EgoSchema and NExT-QA datasets show that AKeyS outperforms all previous methods with the highest keyframe searching efficiency, which means it can accurately identify key information and conduct effective visual reasoning with minimal computational overhead. For example, on the EgoSchema subset, it achieves 1.8% higher accuracy while processing only 43.5% of the frames compared to VideoTree. We believe that AKeyS represents a significant step towards building intelligent agents for video understanding. The code is publicly available at https://github.com/fansunqi/AKeyS.
Pre-trained on tremendous image-text pairs, vision-language models like CLIP have demonstrated promising zero-shot generalization across numerous image-based tasks. However, extending these capabilities to video tasks remains challenging due to limited labeled video data and high training costs. Recent video prompting methods attempt to adapt CLIP for video tasks by introducing learnable prompts, but they typically rely on a single static prompt for all video sequences, overlooking the diverse temporal dynamics and spatial variations that exist across frames. This limitation significantly hinders the model's ability to capture essential temporal information for effective video understanding. To address this, we propose an integrated Spatial-TempOral dynamic Prompting (STOP) model which consists of two complementary modules, the intra-frame spatial prompting and inter-frame temporal prompting. Our intra-frame spatial prompts are designed to adaptively highlight discriminative regions within each frame by leveraging intra-frame attention and temporal variation, allowing the model to focus on areas with substantial temporal dynamics and capture fine-grained spatial details. Additionally, to highlight the varying importance of frames for video understanding, we further introduce inter-frame temporal prompts, dynamically inserting prompts between frames with high temporal variance as measured by frame similarity. This enables the model to prioritize key frames and enhances its capacity to understand temporal dependencies across sequences. Extensive experiments on various video benchmarks demonstrate that STOP consistently achieves superior performance against state-of-the-art methods. The code is available at https://github.com/zhoujiahuan1991/CVPR2025-STOP.
Remote work and online courses have become important methods of knowledge dissemination, leading to a large number of document-based instructional videos. Unlike traditional video datasets, these videos mainly feature rich-text images and audio that are densely packed with information closely tied to the visual content, requiring advanced multimodal understanding capabilities. However, this domain remains underexplored due to dataset availability and its inherent complexity. In this paper, we introduce the DocVideoQA task and dataset for the first time, comprising 1454 videos across 23 categories with a total duration of about 828 hours. The dataset is annotated with 154k question-answer pairs generated manually and via GPT, assessing models' comprehension, temporal awareness, and modality integration capabilities. Initially, we establish a baseline using open-source MLLMs. Recognizing the challenges in modality comprehension for document-centric videos, we present DV-LLaMA, a robust video MLLM baseline. Our method enhances unimodal feature extraction with diverse instruction-tuning data and employs contrastive learning to strengthen modality integration. Through fine-tuning, the LLM is equipped with audio-visual capabilities, leading to significant improvements in document-centric video understanding. Extensive testing on the DocVideoQA dataset shows that DV-LLaMA significantly outperforms existing models. We'll release the code and dataset to facilitate future research.
In this work, we tackle action-scene hallucination in Video Large Language Models (Video-LLMs), where models incorrectly predict actions based on the scene context or scenes based on observed actions. We observe that existing Video-LLMs often suffer from action-scene hallucination due to two main factors. First, existing Video-LLMs intermingle spatial and temporal features by applying an attention operation across all tokens. Second, they use the standard Rotary Position Embedding (RoPE), which causes the text tokens to overemphasize certain types of tokens depending on their sequential orders. To address these issues, we introduce MASH-VLM, Mitigating Action-Scene Hallucination in Video-LLMs through disentangled spatial-temporal representations. Our approach includes two key innovations: (1) DST-attention, a novel attention mechanism that disentangles the spatial and temporal tokens within the LLM by using masked attention to restrict direct interactions between the spatial and temporal tokens; (2) Harmonic-RoPE, which extends the dimensionality of the positional IDs, allowing the spatial and temporal tokens to maintain balanced positions relative to the text tokens. To evaluate the action-scene hallucination in Video-LLMs, we introduce the UNSCENE benchmark with 1,320 videos and 4,078 QA pairs. Extensive experiments demonstrate that MASH-VLM achieves state-of-the-art results on the UNSCENE benchmark, as well as on existing video understanding benchmarks.
Long-Context Transformer Models (LCTMs) are vital for real-world applications but suffer high computational costs due to attention's quadratic complexity. Block-sparse attention mitigates this by focusing computation on critical regions, yet existing methods struggle with balancing accuracy and efficiency due to costly block importance measurements. In this paper, we introduce XAttention, a plug-and-play framework that dramatically accelerates long-context inference in Transformers models using sparse attention. XAttention's key innovation is the insight that the sum of antidiagonal values (i.e., from the lower-left to upper-right) in the attention matrix provides a powerful proxy for block importance. This allows for precise identification and pruning of non-essential blocks, resulting in high sparsity and dramatically accelerated inference. Across comprehensive evaluations on demanding long-context benchmarks-including RULER and LongBench for language, VideoMME for video understanding, and VBench for video generation. XAttention achieves accuracy comparable to full attention while delivering substantial computational gains. We demonstrate up to 13.5x acceleration in attention computation. These results underscore XAttention's ability to unlock the practical potential of block sparse attention, paving the way for scalable and efficient deployment of LCTMs in real-world applications. Code is available at https://github.com/mit-han-lab/x-attention.
Recent Multi-modal Large Language Models (MLLMs) have been challenged by the computational overhead resulting from massive video frames, often alleviated through compression strategies. However, the visual content is not equally contributed to user instructions, existing strategies (\eg, average pool) inevitably lead to the loss of potentially useful information. To tackle this, we propose the Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom), utilizing the instruction as a condition to guide the compression from both local and global levels. This encourages the compression to retain the maximum amount of user-focused information while reducing visual tokens to minimize computational burden. Specifically, the instruction condition is injected into the grouped visual tokens at the local level and the learnable tokens at the global level, and we conduct the attention mechanism to complete the conditional compression. From the hybrid-level compression, the instruction-relevant visual parts are highlighted while the temporal-spatial structure is also preserved for easier understanding of LLMs. To further unleash the potential of HICom, we introduce a new conditional pre-training stage with our proposed dataset HICom-248K. Experiments show that our HICom can obtain distinguished video understanding ability with fewer tokens, increasing the performance by 2.43\% average on three multiple-choice QA benchmarks and saving 78.8\% tokens compared with the SOTA method. The code is available at https://github.com/lntzm/HICom.
Synthetic videos nowadays is widely used to complement data scarcity and diversity of real-world videos. Current synthetic datasets primarily replicate real-world scenarios, leaving impossible, counterfactual and anti-reality video concepts underexplored. This work aims to answer two questions: 1) Can today's video generation models effectively follow prompts to create impossible video content? 2) Are today's video understanding models good enough for understanding impossible videos? To this end, we introduce IPV-Bench, a novel benchmark designed to evaluate and foster progress in video understanding and generation. IPV-Bench is underpinned by a comprehensive taxonomy, encompassing 4 domains, 14 categories. It features diverse scenes that defy physical, biological, geographical, or social laws. Based on the taxonomy, a prompt suite is constructed to evaluate video generation models, challenging their prompt following and creativity capabilities. In addition, a video benchmark is curated to assess Video-LLMs on their ability of understanding impossible videos, which particularly requires reasoning on temporal dynamics and world knowledge. Comprehensive evaluations reveal limitations and insights for future directions of video models, paving the way for next-generation video models.
Multimodal large language models (MLLMs) have made remarkable progress in either temporal or spatial localization. However, they struggle to perform spatio-temporal video grounding. This limitation stems from two major challenges. Firstly, it is difficult to extract accurate spatio-temporal information of each frame in the video. Secondly, the substantial number of visual tokens makes it challenging to precisely map visual tokens of each frame to their corresponding spatial coordinates. To address these issues, we introduce SpaceVLLM, a MLLM endowed with spatio-temporal video grounding capability. Specifically, we adopt a set of interleaved Spatio-Temporal Aware Queries to capture temporal perception and dynamic spatial information. Moreover, we propose a Query-Guided Space Decoder to establish a corresponding connection between the queries and spatial coordinates. Additionally, due to the lack of spatio-temporal datasets, we construct the Unified Spatio-Temporal Grounding (Uni-STG) dataset, comprising 480K instances across three tasks. This dataset fully exploits the potential of MLLM to simultaneously facilitate localization in both temporal and spatial dimensions. Extensive experiments demonstrate that SpaceVLLM achieves the state-of-the-art performance across 11 benchmarks covering temporal, spatial, spatio-temporal and video understanding tasks, highlighting the effectiveness of our approach. Our code, datasets and model will be released.