论文检索

Qi Wang, Zhipeng Zhang, Baao Xie, Xin Jin, Yunbo Wang, Shiyu Wang, Liaomo Zheng, Xiaokang Yang, Wenjun Zeng

Training visual reinforcement learning (RL) in practical scenarios presents a significant challenge, i.e., RL agents suffer from low sample efficiency in environments with variations. While various approaches have attempted to alleviate this issue by disentangled representation learning, these methods usually start learning from scratch without prior knowledge of the world. This paper, in contrast, tries to learn and understand underlying semantic variations from distracting videos via offline-to-online latent distillation and flexible disentanglement constraints. To enable effective cross-domain semantic knowledge transfer, we introduce an interpretable model-based RL framework, dubbed Disentangled World Models (DisWM). Specifically, we pretrain the action-free video prediction model offline with disentanglement regularization to extract semantic knowledge from distracting videos. The disentanglement capability of the pretrained model is then transferred to the world model through latent distillation. For finetuning in the online environment, we exploit the knowledge from the pretrained model and introduce a disentanglement constraint to the world model. During the adaptation phase, the incorporation of actions and rewards from online environment interactions enriches the diversity of the data, which in turn strengthens the disentangled representation learning. Experimental results validate the superiority of our approach on various benchmarks.

Kuo Wang, Quanlong Zheng, Junlin Xie, Yanhao Zhang, Jinguo Luo, Haonan Lu, Liang Lin, Fan Zhou, Guanbin Li

Video Multimodal Large Language Models (Video-MLLM) have achieved remarkable advancements in video understanding tasks. However, constrained by the context length limitation in the underlying LLMs, existing Video-MLLMs typically exhibit suboptimal performance on long video scenarios. To understand extended input frames, common solutions span token compression and streaming inference techniques, which sacrifice feature granularity or inference efficiency. Differently, to efficiently achieve comprehensive understanding of longer frame inputs, we draw ideas from MoE and propose a training-free approach Free-MoRef, which instantly multiplexes the context perception capabilities of Video-MLLMs within one inference pass. Specifically, Free-MoRef reconstructs the vision tokens into several short sequences as multi-references. Subsequently, we introduce MoRef-attention, which gathers clues from the multi-reference chunks in parallel to summarize unified query activations. After the shadow layers in LLMs, a reference fusion step is derived to compose a final mixed reasoning sequence with key tokens from parallel chunks, which compensates the cross-reference vision interactions that are neglected in MoRef-attention. By splitting and fusing the long vision token sequences, Free-MoRef achieves improved performance under much lower computing costs in reasoning multiplexed context length, demonstrating strong efficiency and effectiveness. Experiments on VideoMME, MLVU, LongVideoBench show that Free-MoRef achieves full perception of 2xto 8xlonger input frames without compression on a single A100 GPU while keeping instant responses, thereby bringing significant performance gains, even surpassing dedicatedly trained long-video-MLLMs. Codes are available at https://github.com/wkfdb/Free-MoRef

Kaining Ying, Hengrui Hu, Henghui Ding

This work addresses motion-guided few-shot video object segmentation (FSVOS), which aims to segment dynamic objects in videos based on a few annotated examples with the same motion patterns. Existing FSVOS datasets and methods typically focus on object categories, which are static attributes that ignore the rich temporal dynamics in videos, limiting their application in scenarios requiring motion understanding. To fill this gap, we introduce MOVE, a large-scale dataset specifically designed for motion-guided FSVOS. Based on MOVE, we comprehensively evaluate 6 state-of-the-art methods from 3 different related tasks across 2 experimental settings. Our results reveal that current methods struggle to address motion-guided FSVOS, prompting us to analyze the associated challenges and propose a baseline method, Decoupled Motion Appearance Network (DMA). Experiments demonstrate that our approach achieves superior performance in few shot motion understanding, establishing a solid foundation for future research in this direction.

Rundong Luo, Matthew Wallingford, Ali Fahardi, Noah Snavely, Wei-Chiu Ma

360deg videos have emerged as a promising medium to represent our dynamic visual world. Compared to the "tunnel vision" of standard cameras, their borderless field of view offers a more complete perspective of our surroundings. While existing video models excel at producing standard videos, their ability to generate full panoramic videos remains elusive. In this paper, we investigate the task of video-to-360deg generation: given a perspective video as input, our goal is to generate a full panoramic video that is consistent with the original video. Unlike conventional video generation tasks, the output's field of view is significantly larger, and the model is required to have a deep understanding of both the spatial layout of the scene and the dynamics of objects to maintain spatio-temporal consistency. To address these challenges, we first leverage the abundant 360deg videos available online and develop a high-quality data filtering pipeline to curate pairwise training data. We then carefully design a series of geometry- and motion-aware operations to facilitate the learning process and improve the quality of 360deg video generation. Experimental results demonstrate that our model can generate realistic and coherent 360deg videos from in-the-wild perspective video. In addition, we showcase its potential applications, including video stabilization, camera viewpoint control, and interactive visual question answering.

Yingxian Chen, Jiahui Liu, Ruidi Fan, Yanwei Li, Chirui Chang, Shizhen Zhao, Wilton W. T. Fok, Xiaojuan Qi, Yik-Chung Wu

Understanding abnormal events in videos is a vital and challenging task that has garnered significant attention in a wide range of applications. Although current video understanding Multi-modal Large Language Models (MLLMs) are capable of analyzing general videos, they often struggle to handle anomalies due to the spatial and temporal sparsity of abnormal events, where the redundant information always leads to suboptimal outcomes. To address these challenges, exploiting the representation and generalization capabilities of Vison Language Models (VLMs) and Large Language Models (LLMs), we propose VA-GPT, a novel MLLM designed for summarizing and localizing abnormal events in various videos. Our approach efficiently aligns effective tokens between visual encoders and LLMs through two key proposed modules: Spatial Effective Token Selection (SETS) and Temporal Effective Token Generation (TETG). These modules enable our model to effectively capture and analyze both spatial and temporal information associated with abnormal events, resulting in more accurate responses and interactions. Furthermore, we construct an instruction-following dataset specifically for fine-tuning video-anomaly-aware MLLMs, and introduce a cross-domain evaluation benchmark based on XD-Violence dataset. Our proposed method outperforms existing state-of-the-art methods on various benchmarks.

Yikang Zhou, Tao Zhang, Shilin Xu, Shihao Chen, Qianyu Zhou, Yunhai Tong, Shunping Ji, Jiangning Zhang, Lu Qi, Xiangtai Li

Recent advancements in multimodal large language models (MLLM) have shown a strong ability in visual perception, reasoning abilities, and vision-language understanding. However, the visual matching ability of MLLMs is rarely studied, despite finding the visual correspondence of objects is essential in computer vision. Our research reveals that the matching capabilities in recent MLLMs still exhibit systematic shortcomings, even with current strong MLLMs models, GPT-4o. In particular, we construct a Multimodal Visual Matching (MMVM) benchmark to fairly benchmark over 30 different MLLMs. The MMVM benchmark is built from 15 open-source datasets and Internet videos with manual annotation. In addition, we have designed an automatic annotation pipeline to generate the MMVM SFT dataset, including 220K visual matching data with reasoning annotation. To our knowledge, this is the first MLLMs dataset and benchmark for the MLLM community. Finally, we present CoLVA, a novel contrastive MLLM with two novel technical designs: fine-grained vision expert with object-level contrastive learning and instruction augmentation strategy. The former learns instance discriminative tokens, while the latter further improves instruction following ability. CoLVA-InternVL2-4B achieves an overall accuracy (OA) of 49.80% on the MMVM benchmark, surpassing GPT-4o and the best open-source MLLM, Qwen2VL-72B, by 7.15% and 11.72% OA, respectively. These results demonstrate the effectiveness of our MMVM SFT dataset and our novel technical designs. Code, benchmark, dataset, and models will be released.

Wenqi Zhang, Hang Zhang, Xin Li, Jiashuo Sun, Yongliang Shen, Weiming Lu, Deli Zhao, Yueting Zhuang, Lidong Bing

Compared to image-text pair data, interleaved corpora enable Vision-Language Models (VLMs) to understand the world more naturally like humans. However, such existing datasets are crawled from webpage, facing challenges like low knowledge density, loose image-text relations, and poor logical coherence between images. On the other hand, the internet hosts vast instructional videos (e.g., online geometry courses) that are widely used by humans to learn foundational subjects, yet these valuable resources remain underexplored in VLM training. In this paper, we introduce a high-quality multimodal textbook corpus with richer foundational knowledge for VLM pretraining. It collects over 2.5 years of instructional videos, totaling 22,000 class hours. We first use an LLM-proposed taxonomy to systematically gather instructional videos. Then we progressively extract and refine visual (keyframes), audio (ASR), and textual knowledge (OCR) from the videos, and organize as an image-text interleaved corpus based on temporal order. Compared to its counterparts, our video-centric textbook offers more coherent context, richer knowledge, and better image-text alignment. Experiments demonstrate its superb pretraining performance, particularly in knowledge- and reasoning-intensive tasks like ScienceQA and MathVista. Moreover, VLMs pre-trained on our textbook exhibit outstanding interleaved context awareness, leveraging visual and textual cues in their few-shot context for task solving. Code and dataset are available on https://multimodal-interleaved-textbook.github.io.

Yeon-Ji Song, Jaein Kim, Suhyung Choi, Jin-Hwa Kim, Byoung-Tak Zhang

Human perception involves decomposing complex multi-object scenes into time-static object appearance (i.e., size, shape, color) and time-varying object motion (i.e., position, velocity, acceleration). For machines to achieve human-like intelligence in real-world interactions, understanding these physical properties of objects is essential, forming the foundation for dynamic video prediction. While recent advancements in object-centric transformers have demonstrated potential in video prediction, they primarily focus on object appearance, often overlooking motion dynamics, which is crucial for modeling dynamic interactions and maintaining temporal consistency in complex environments. To address these limitations, we propose OCK, a dynamic video prediction model leveraging object-centric kinematics and object slots. We introduce a novel component named Object Kinematics that comprises explicit object motions, serving as an additional attribute beyond conventional appearance features to model dynamic scenes. The Object Kinematics are integrated into various OCK mechanisms, enabling spatiotemporal prediction of complex object interactions over long video sequences. Our model demonstrates superior performance in handling complex scenes with intricate object attributes and motions, highlighting its potential for applicability in vision-related dynamics learning tasks.

Yudong Liu, Jingwei Sun, Yueqian Lin, Jianyi Zhang, Jingyang Zhang, Ming Yin, Qinsi Wang, Hai Li, Yiran Chen

Vision language models (VLMs) demonstrate strong capabilities in jointly processing visual and textual data. However, they often incur substantial computational overhead due to redundant visual information, particularly in long-form video scenarios. Existing approaches predominantly focus on either vision token pruning, which may overlook spatio-temporal dependencies, or keyframe selection, which identifies informative frames but discards others, thus disrupting contextual continuity. In this work, we propose KVTP (Keyframe-oriented Vision Token Pruning), a novel framework that overcomes the drawbacks of token pruning and keyframe selection. By adaptively assigning pruning rates based on frame relevance to the query, KVTP effectively retains essential contextual information while significantly reducing redundant computation. To thoroughly evaluate the long-form video understanding capacities of VLMs, we curated and reorganized subsets from VideoMME, EgoSchema, and NextQA into a unified benchmark named SparseKV-QA that highlights real-world scenarios with sparse but crucial events. Our experiments with VLMs of various scales show that KVTP can reduce token usage by 80% without compromising spatiotemporal and contextual consistency, significantly cutting computation while maintaining the performance. These results demonstrate our approach's effectiveness in efficient long-video processing, facilitating more scalable VLM deployment.

Aoxiong Yin, Xu Tan, Kai Shen, Yichong Leng, Xinyu Zhou, Juncheng Li, Siliang Tang

Recent advancements in text-to-video (T2V) generation have been driven by two competing paradigms: autoregressive language models and diffusion models. However, each paradigm has intrinsic limitations: language models struggle with visual quality and error accumulation, while diffusion models lack semantic understanding and causal modeling. In this work, we propose LanDiff, a hybrid framework that synergizes the strengths of both paradigms through coarse-to-fine generation. Our architecture introduces three key innovations: (1) a semantic tokenizer that compresses 3D visual features into compact 1D discrete representations through efficient semantic compression, achieving a ~14,000xcompression ratio; (2) a language model that generates semantic tokens with high-level semantic relationships; (3) a streaming diffusion model that refines coarse semantics into high-fidelity videos. Experiments show that LanDiff, a 5B model, achieves a score of 85.43 on the VBench T2V benchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo. Furthermore, our model also achieves state-of-the-art performance in long video generation, surpassing other open-source models in this field. Our demo can be viewed at https://landiff.github.io/.

Haonan Wang, Qixiang Zhang, Lehan Wang, Xuanqi Huang, Xiaomeng Li

Decoding visual stimuli from neural activity is essential for understanding the human brain. While fMRI methods have successfully reconstructed static images, fMRI-tovideo reconstruction faces challenges due to the need for capturing spatiotemporal dynamics like motion and scene transitions. Recent approaches have improved semantic and perceptual alignment but struggle to integrate coarse fMRI data with detailed visual features. Inspired by the hierarchical organization of the visual system, we propose NEURONS, a novel framework that decouples learning into four correlated sub-tasks: key object segmentation, concept recognition, scene description, and blurry video reconstruction. This approach simulates the visual cortex's functional specialization, allowing the model to capture diverse video content. In the inference stage, NEURONS generates robust conditioning signals for a pre-trained textto-video diffusion model to reconstruct the videos. Extensive experiments demonstrate that NEURONS outperforms state-of-the-art baselines, achieving solid improvements in video consistency (26.6%) and semantic-level accuracy (19.1%). Notably, NEURONS shows a strong functional correlation with the visual cortex, highlighting its potential for brain-computer interfaces and clinical applications. Code and model weights are available at: https://github.com/xmed-lab/NEURONS.

Chi-Hsi Kung, Frangil Ramirez, Juhyung Ha, Yi-Ting Chen, David Crandall, Yi-Hsuan Tsai

Understanding a procedural activity requires modeling both how action steps transform the scene, and how evolving scene transformations can influence the sequence of action steps, even those that are accidental or erroneous. Existing work has studied procedure-aware video representations by modeling the temporal order of actions, but has not explicitly learned the state changes (scene transformations). In this work, we study procedure-aware video representation learning by incorporating state-change descriptions generated by Large Language Models (LLMs) as supervision signals for video encoders. Moreover, we generate state-change counterfactuals that simulate hypothesized failure outcomes, allowing models to learn by imagining unseen "What if" scenarios. This counterfactual reasoning facilitates the model's ability to understand the cause and effect of each step in an activity. We conduct extensive experiments on procedure-aware tasks, including temporal action segmentation, error detection, action phase classification, frame retrieval, multi-instance retrieval, and action recognition. Our results demonstrate the effectiveness of the proposed state-change descriptions and their counterfactuals, and achieve significant improvements on multiple tasks. Code is available at https://github.com/HCIS-Lab/counterfactual-video-pretrain.

Mingfeng Zha, Tianyu Li, Guoqing Wang, Peng Wang, Yangyang Wu, Yang Yang, Heng Tao Shen

Audio-visual segmentation (AVS) aims to segment objects in videos based on audio cues. Existing AVS methods are primarily designed to enhance interaction efficiency but pay limited attention to modality representation discrepancies and imbalances. To overcome this, we propose the implicit counterfactual framework (ICF) to achieve unbiased cross-modal understanding. Due to the lack of semantics, heterogeneous representations may lead to erroneous matches, especially in complex scenes with ambiguous visual content or interference from multiple audio sources. We introduce the multi-granularity implicit text (MIT) involving video-, segment- and frame-level as the bridge to establish the modality-shared space, reducing modality gaps and providing prior guidance. Visual content carries more information and typically dominates, thereby marginalizing audio features in the decision-making. To mitigate knowledge preference, we propose the semantic counterfactual (SC) to learn orthogonal representations in the latent space, generating diverse counterfactual samples, thus avoiding biases introduced by complex functional designs and explicit modifications of text structures or attributes. We further formulate the collaborative distribution-aware contrastive learning (CDCL), incorporating factual-counterfactual and inter-modality contrasts to align representations, promoting cohesion and decoupling. Extensive experiments on three public datasets validate that the proposed method achieves state-of-the-art performance.

Yiwu Zhong, Zhuoming Liu, Yin Li, Liwei Wang

Large language models (LLMs) have enabled the creation of multi-modal LLMs that exhibit strong comprehension of visual data such as images and videos. However, these models usually rely on extensive visual tokens from visual encoders, leading to high computational demands, which limits their applicability in resource-constrained environments and for long-context tasks. In this work, we propose a training-free adaptive inference method for multi-modal LLMs that can accommodate a broad range of efficiency requirements with a minimum performance drop. Our method consists of a) iterative token merging based on embedding similarity before LLMs, and b) progressive token pruning within LLM layers based on multi-modal importance. With a minimalist design, our method can be applied to both video and image LLMs. Extensive experiments on diverse video and image benchmarks demonstrate that our method substantially reduces computation load (e.g., a 7-fold reduction in FLOPs) while preserving the performance of video and image LLMs. Further, at a similar computational cost, our method outperforms the state-of-the-art methods in long video understanding (e.g., +4.6 on MLVU). Additionally, our in-depth analysis provides insights into token redundancy and LLM layer behaviors, offering guidance for future research in designing efficient multi-modal LLMs. Our code is available at https://github.com/LaVi-Lab/AIM.

Xiaolong Sun, Le Wang, Sanping Zhou, Liushuai Shi, Kun Xia, Mengnan Liu, Yabing Wang, Gang Hua

Video temporal grounding is a critical video understanding task, which aims to localize moments relevant to a language description. The challenge of this task lies in distinguishing relevant and irrelevant moments. Previous methods focused on learning continuous features exhibit weak differentiation between foreground and background features. In this paper, we propose a novel Moment-Quantization based Video Temporal Grounding method (MQVTG), which quantizes the input video into various discrete vectors to enhance the discrimination between relevant and irrelevant moments. Specifically, MQVTG maintains a learnable moment codebook, where each video moment matches a codeword. Considering the visual diversity, i.e., various visual expressions for the same moment, MQVTG treats moment-codeword matching as a clustering process without using discrete vectors, avoiding the loss of useful information from direct hard quantization. Additionally, we employ effective prior-initialization and joint-projection strategies to enhance the maintained moment codebook. With its simple implementation, the proposed method can be integrated into existing temporal grounding models as a plug-and-play component. Extensive experiments on six popular benchmarks demonstrate the effectiveness and generalizability of MQVTG, significantly outperforming state-of-the-art methods. Further qualitative analysis shows that our method effectively groups relevant features and separates irrelevant ones, aligning with our goal of enhancing discrimination. Code is available at https://github.com/TensorsSun/MQVTG.

Huachao Zhu, Zelong Liu, Zhichao Sun, Yuda Zou, Gui-Song Xia, Yongchao Xu

Recognizing out-of-distribution (OoD) objects on roads is crucial for safe driving. Most existing methods rely on segmentation models' uncertainty as anomaly scores, often resulting in false positives - especially at ambiguous regions like boundaries, where segmentation models inherently exhibit high uncertainty. Additionally, it is challenging to define a suitable threshold to generate anomaly masks, especially with the inconsistencies in predictions across consecutive frames. We propose DetSeg, a novel paradigm that helps incorporate object-level understanding. DetSeg first detects all objects in the open world and then suppresses in-distribution (ID) bounding boxes, leaving only OoD proposals. These proposals can either help previous methods eliminate false positives (DetSeg-\mathcal R ), or generate binary anomaly masks without complex threshold search when combined with a box-prompted segmentation module (DetSeg-\mathcal S ). Additionally, we introduce vanishing point guided Hungarian matching (VPHM) to smooth the prediction results within a video clip, mitigating abrupt variations of predictions between consecutive frames. Comprehensive experiments on various benchmarks demonstrate that DetSeg significantly improves performance, reducing the FPR\it _ 95 of previous methods by up to 37.45%, offering a more robust and practical solution. Code: https://github.com/huachao0124/DetSeg-official.

Björn Braun, Rayan Armani, Manuel Meier, Max Moebus, Christian Holz

Egocentric vision systems aim to understand the spatial surroundings and the wearer's behavior inside it, including motions, activities, and interactions. We argue that egocentric systems must additionally detect physiological states to capture a person's attention and situational responses, which are critical for context-aware behavior modeling. In this paper, we propose egoPPG, a novel vision task for egocentric systems to recover a person's cardiac activity to aid downstream vision tasks. We introduce PulseFormer, a method to extract heart rate as a key indicator of physiological state from the eye tracking cameras on unmodified egocentric vision systems. PulseFormer continuously estimates the photoplethysmogram (PPG) from areas around the eyes and fuses motion cues from the headset's inertial measurement unit to track HR values. We demonstrate egoPPG's downstream benefit for a key task on EgoExo4D, an existing egocentric dataset for which we find PulseFormer's estimates of HR to improve proficiency estimation by 14%. To train and validate PulseFormer, we collected a dataset of 13+ hours of eye tracking videos from Project Aria and contact-based PPG signals as well as an electrocardiogram (ECG) for ground-truth HR values. Similar to EgoExo4D, 25 participants performed diverse everyday activities such as office work, cooking, dancing, and exercising, which induced significant natural motion and HR variation (44-164 bpm). Our model robustly estimates HR (MAE=7.67 bpm) and captures patterns (r=0.85). Our results show how egocentric systems may unify environmental and physiological tracking to better understand users and that egoPPG as a complementary task provides meaningful augmentations for existing datasets and tasks. We release our code, dataset, and HR augmentations for EgoExo4D to inspire research on physiology-aware egocentric tasks.

Xindi Yang, Baolu Li, Yiming Zhang, Zhenfei Yin, Lei Bai, Liqian Ma, Zhiyong Wang, Jianfei Cai, Tien-Tsin Wong, Huchuan Lu 等

Video diffusion models (VDMs) have advanced significantly in recent years, enabling the generation of highly realistic videos and drawing the attention of the community in their potential as world simulators. However, despite their capabilities, VDMs often fail to produce physically plausible videos due to an inherent lack of understanding of physics, resulting in incorrect dynamics and event sequences. To address this limitation, we propose a novel two-stage image-to-video generation framework that explicitly incorporates physics with vision and language informed physical prior. In the first stage, we employ a Vision Language Model (VLM) as a coarse-grained motion planner, integrating chain-of-thought and physics-aware reasoning to predict a rough motion trajectories/changes that approximate real-world physical dynamics while ensuring the inter-frame consistency. In the second stage, we use the predicted motion trajectories/changes to guide the video generation of a VDM. As the predicted motion trajectories/changes are rough, noise is added during inference to provide freedom to the VDM in generating motion with more fine details. Extensive experimental results demonstrate that our framework can produce physically plausible motion, and comparative evaluations highlight the notable superiority of our approach over existing methods. More video results and code are available on our Project Page: https://madaoer.github.io/projects/physically_plausible_video_generation/.

Pulkit Kumar, Shuaiyi Huang, Matthew Walmer, Sai Saketh Rambhatla, Abhinav Shrivastava

Video understanding requires effective modeling of both motion and appearance information, particularly for few-shot action recognition. While recent advances in point tracking have been shown to improve few-shot action recognition, two fundamental challenges persist: selecting informative points to track and effectively modeling their motion patterns. We present Trokens, a novel approach that transforms trajectory points into semantic-aware relational tokens for action recognition. First, we introduce a semantic-aware sampling strategy to adaptively distribute tracking points based on object scale and semantic relevance. Second, we develop a motion modeling framework that captures both intra-trajectory dynamics through the Histogram of Oriented Displacements (HoD) and inter-trajectory relationships to model complex action patterns. Our approach effectively combines these trajectory tokens with semantic features to enhance appearance features with motion information, achieving state-of-the-art performance across six diverse few-shot action recognition benchmarks: Something-Something-V2 (both full and small splits), Kinetics, UCF101, HMDB51, and FineGym.

Han Wang, Yuxiang Nie, Yongjie Ye, Yanjie Wang, Shuai Li, Haiyang Yu, Jinghui Lu, Can Huang

The application of Large Vision-Language Models (LVLMs) for analyzing images and videos is an exciting and rapidly evolving field. In recent years, we've seen significant growth in high-quality image-text datasets for fine-tuning image understanding, but there is still a lack of comparable datasets for videos. Additionally, many VideoLLMs are extensions of single-image VLMs, which may not efficiently handle the complexities of longer videos. In this study, we introduce a large-scale synthetic dataset created from proprietary models, using carefully designed prompts to tackle a wide range of questions. We also explore a dynamic visual token compression architecture that strikes a balance between computational efficiency and performance. Our proposed Dynamic-VLM achieves state-of-the-art results across various video tasks and shows impressive generalization, setting new baselines in multi-image understanding. Notably, Dynamic-VLM delivers an absolute improvement of 2.7% over LLaVA-OneVision on VideoMME and 10.7% on MuirBench.