CVPR 2026 论文列表

Diffusion model (DM) based Video Super-Resolution (VSR) approaches achieve impressive perceptual quality. Diffusion model (DM) based Video Super-Resolution (VSR) approaches achieve impressive perceptual quality. However, existing DM-based VSR methods over-prioritize perceptual synthesis while neglecting fidelity gains from accurate alignment and sufficient compensation. In this paper, within the DM-based VSR pipeline, we revisit the role of alignment and compensation between adjacent video frames and reveal two crucial observations: (a) the feature domain is better suited than the pixel domain for information compensation due to its stronger spatial and temporal correlations, and (b) warping at an upscaled resolution better preserves high-frequency information, but this benefit is not necessarily monotonic. Therefore, we propose a novel Densely Guided diffusion model with Aligned Features for Video Super-Resolution (DGAF-VSR), with an Optical Guided Warping Module (OGWM) to maintain high-frequency details in the aligned features and a Feature-wise Temporal Condition Module (FTCM) to deliver dense guidance in the feature domain. Extensive experiments on synthetic and real-world datasets demonstrate that DGAF-VSR surpasses state-of-the-art methods in key aspects of VSR, including perceptual quality (35.82% DISTS reduction), fidelity (0.20 dB PSNR gain), and temporal consistency (30.37% tLPIPS reduction).

Vision Large Language Models (VLLMs) incur high computational costs due to their reliance on hundreds of visual tokens to represent images. While token pruning offers a promising solution for accelerating inference, this paper, however, identifies a key observation: in deeper layers (e.g., beyond the 20th), existing training-free pruning methods perform no better than random pruning. We hypothesize that this degradation is caused by "vanishing token information", where visual tokens progressively lose their salience with increasing network depth. To validate this hypothesis, we quantify a token's information content by measuring the change in the model output probabilities upon its removal. Using this proposed metric, our analysis of the information of visual tokens across layers reveals three key findings: (1) As layers deepen, the information of visual tokens gradually becomes uniform and eventually vanishes at an intermediate layer, which we term as "information horizon", beyond which the visual tokens become redundant;(2) The position of this horizon is not static; it extends deeper for visually intensive tasks, such as Optical Character Recognition (OCR), compared to more general tasks like Visual Question Answering (VQA);(3) This horizon is also strongly correlated with model capacity, as stronger VLLMs (e.g., Qwen2.5-VL) employ deeper visual tokens than weaker models (e.g., LLaVA-1.5).Based on our findings, we show that simple random pruning in deep layers efficiently balances performance and efficiency. Moreover, integrating random pruning consistently enhances existing methods. Using DivPrune with random pruning achieves state-of-the-art results, maintaining 96.9% of Qwen-2.5-VL-7B performance while pruning 50% of visual tokens. The code is available at https://github.com/YahongWang1/Information-Horizon.

Robotic Foundation Models (RFMs) hold great promise as generalist, end-to-end systems for robot control.Yet their ability to generalize across new environments, tasks, and embodiments remains limited.We argue that a major bottleneck lies in their foundations: most RFMs are built by fine-tuning internet-pretrained Vision-Language Models (VLMs).However, these VLMs are trained on 2D image-language tasks and lack the 3D spatial reasoning inherently required for embodied control in the 3D world.Bridging this gap directly with large-scale robotic data is costly and difficult to scale.Instead, we propose to enrich easy-to-collect non-robotic image data with 3D annotations and enhance a pretrained VLM with 3D understanding capabilities.Following this strategy, we train SPEAR-VLM, a 3D-aware VLM that infers object coordinates in 3D space from a single 2D image.Building on SPEAR-VLM, we introduce our main contribution, SPEAR-1: a robotic foundation model that integrates grounded 3D perception with language-instructed embodied control.Trained on ~45M frames from 24 Open X-Embodiment datasets, SPEAR-1 outperforms or matches state-of-the-art models such as \pi_0-FAST and \pi_ 0.5 , while it uses 20xfewer robot demonstrations.This carefully-engineered training strategy unlocks new VLM capabilities and as a consequence boosts the reliability of embodied control beyond what is achievable with only robotic data.We make our model weights and 3D-annotated datasets publicly available.

Human perception of social environments is inherently a multi-view synthesis problem, requiring the integration of complementary and often occluded information across space and time. However, existing benchmarks for Multimodal Large Language Models (MLLMs) are overwhelmingly predicated on a "sufficient-view" assumption, rewarding single-view pattern recognition while failing to evaluate cross-view fusion. To address this critical gap, we introduce CVBench, a large-scale, multi-task benchmark for cross-view human understanding. CVBench comprises 3,000 challenging questions across 12 spatial and temporal tasks, where every item is designed with verifiable single-view insufficiency, mandating that models synthesize disparate evidence to resolve ambiguities. Our comprehensive evaluation of state-of-the-art open and closed-source MLLMs reveals a substantial performance gap, with the best models falling nearly 50 points behind human performance. We identify a systemic failure mechanism across all models: a dominant "Single-View Bias", whereby models ignore conflicting evidence and default to the most confident but incorrect single-view prediction. This demonstrates that current MLLMs lack the fundamental mechanisms for geometric grounding, identity persistence, and true spatio-temporal fusion. CVBench provides a rigorous diagnostic framework to catalyze the development of next-generation, cross-view-aware architectures.

Evaluating the nuanced human-centric video understanding capabilities of Multimodal Large Language Models (MLLMs) remains a great challenge, as existing benchmarks often overlook the intricacies of emotion, behavior, and cross-modal alignment. We introduce HumanVBench, a comprehensive video benchmark designed to rigorously probe these capabilities across 16 fine-grained tasks. A cornerstone of our work is a novel and scalable benchmark construction methodology, featuring two automated pipelines that synthesize high-quality video annotations and challenging multiple-choice questions with minimal human labor. By leveraging state-of-the-art models for annotation and systematically converting model-induced errors into plausible distractors, our framework provides a generalizable "machine" for creating nuanced evaluation suites. Our extensive evaluation of 30 leading MLLMs on HumanVBench reveals critical deficiencies, particularly in perceiving subtle emotions and aligning speech with visual cues, with even top proprietary models falling short of human performance. We open-source HumanVBench and our synthesis pipelines to catalyze the development of more socially intelligent and capable video MLLMs.

We introduce MusicInfuser, an approach that aligns pre-trained text-to-video diffusion models to generate high-quality dance videos synchronized with specified music tracks. Rather than training a multimodal audio-video or audio-motion model from scratch, our method demonstrates how existing video diffusion models can be efficiently adapted to align with musical inputs. We propose a novel layer-wise adaptability criterion based on a guidance-inspired constructive influence function to select adaptable layers, significantly reducing training costs while preserving rich prior knowledge, even with limited, specialized datasets. Experiments show that MusicInfuser effectively bridges the gap between music and video, generating novel and diverse dance movements that respond dynamically to music. Furthermore, our framework generalizes well to unseen music tracks, longer video sequences, and unconventional subjects, outperforming baseline models in consistency and synchronization. All of this is achieved without requiring motion data, with training completed on a single GPU within a day.

Exploiting bi-directional context prediction has long been recognized as a key direction for improving compression efficiency in neural video coding. However, existing neural B-frame codecs still exhibit limited performance gains, particularly in high-resolution videos with large motion, where optical flow estimation becomes unreliable and balanced prediction fusion introduces distortions. To address these challenges, we present the first High-Resolution bi-directional neural video coding method, termed as HR-NVC, which non-uniformly integrates confidence-guided predictive cues from both temporal directions to achieve more reliable and efficient compression. Specifically, we propose Spatio-Temporal Anchored Motion Estimation, which introduces virtual anchor frames and low-resolution priors to significantly improve estimation robustness under large displacements. We further design a Hierarchical Motion Representation that converges multi-scale motion with temporal references, enabling compact and adaptive modeling of motion reliability across resolutions. Finally, a Bi-Contextual Asymmetric Harmonization module performs confidence-guided fusion of bidirectional references, effectively suppressing unreliable contexts and restoring structural consistency near occlusion and scene transition regions. Notably, our model is the first end-to-end-optimized video codec evaluated on 4K-resolution videos, establishing a new benchmark for higher-resolution NVC and achieving state-of-the-art performance among neural B-frame codecs.

Video large language models (Video-LLMs) have demonstrated strong capabilities in video understanding tasks. However, their practical deployment is still hindered by the inefficiency introduced by processing massive amounts of visual tokens. Although recent approaches achieve extremely low token retention ratios while maintaining accuracy comparable to full-token baselines, most of them perform compression only at the late stage of prefilling, leaving the efficiency of the vision encoder unoptimized. In this paper, we first show that vision encoding contributes a large portion to the time-to-first-token (TTFT). Therefore, instead of compressing visual tokens only after the vision encoder, performing compression inside the encoder still leaves substantial room for exploration. Based on this insight, we propose EarlyTom, a training-free token compression framework that performs early-stage visual token compression inside the vision encoder, enabling significantly better TTFT reduction and higher throughput. In addition, we introduce a decoupled spatial token selection strategy that improves the overall compression effectiveness. EarlyTom reduces TTFT by up to 2.65x and FLOPs by up to 61% on a single NVIDIA A100 GPU for the LLaVA-OneVision-7B model, while maintaining accuracy comparable to the full-token baseline. These improvements substantially enhance the practicality of deploying Video-LLMs in real-world production scenarios.

Diffusion Transformers (DiTs) achieve strong video generation performance but suffer from prohibitive computation cost due to dense spatiotemporal tokenization. Most existing works rely on uniform patchification, tokenizing non-overlapping spatiotemporal with a fixed patch size regardless of the underlying content. This content-agnostic tokenization results in substantial redundant computation, especially in visually simple or static areas. To address this inefficiency while preserving the video generation quality, we propose DynaPatch, a fine-grained dynamic patchification framework that adaptively selects patch sizes for each spatiotemporal region based on content complexity. A lightweight router predicts patch sizes directly from the latents encoded by 3D Variational Autoencoder (VAE), and is jointly optimized with the diffusion model through diffusion loss, an attention-guided saliency alignment loss, and a token-budget regularizer. Learnable patchify/unpatchify layers integrate seamlessly with standard DiT backbones, allowing flexible tokenization without architectural changes. Experiments demonstrate that DynaPatch can effectively reduce redundant computations while preserving fine details, achieving 1.3-1.8x acceleration with minimal quality degradation. On VBench, DynaPatch attains a Total Score of 83.42 at 30% token reduction, significantly outperforming prior patchification and token pruning approaches. These results indicate that content-aware patchification offers an effective direction for efficient and scalable video diffusion. Project page: https://shengli99.github.io/DynaPatch/.

Recent advances in point cloud In-Context Learning (ICL) have demonstrated strong multitask capabilities. Existing approaches typically adopt a Masked Point Modeling (MPM)-based paradigm for point cloud ICL. However, MPM-based methods directly predict the target point cloud from masked tokens without leveraging geometric priors, requiring the model to infer spatial structure and geometric details solely from token-level correlations via transformers. Additionally, these methods suffer from a training-inference objective mismatch, as the model learns to predict the target point cloud using target-side information that is unavailable at inference time. To address these challenges, we propose DeformPIC, a deformation-based framework for point cloud ICL. Unlike existing approaches that rely on masked reconstruction, DeformPIC learns to deform the query point cloud under task-specific guidance from prompts, enabling explicit geometric reasoning and consistent objectives. Extensive experiments demonstrate that DeformPIC consistently outperforms previous state-of-the-art methods, achieving reductions of 1.6, 1.8, and 4.7 points in average Chamfer Distance on reconstruction, denoising, and registration tasks, respectively. Furthermore, we introduce a new out-of-domain benchmark to evaluate generalization across unseen data distributions, where DeformPIC achieves state-of-the-art performance.

Video diffusion models achieve impressive visual fidelity but remain computationally prohibitive for real-time or interactive generation due to their sequential denoising process. Recent caching methods accelerate inference by reusing outputs across timesteps, typically estimating each new output from the first-order residual, which is the difference between adjacent model predictions.To mitigate the accumulated error in caching methods, we propose D2Cache, a training-free method that leverages the smoothness of second-order residual delta, which is temporal differences between consecutive first-order residuals, to predict future timesteps more accurately. We theoretically show that this second-order correction improves prediction accuracy and effectively suppresses cumulative errors. Moreover, D2Cache adaptively scales second-order deltas using error estimates derived from timestep embeddings, maintaining accuracy across varying cache intervals.Empirically, D2Cache outperforms the state-of-the-art TeaCache across four video diffusion models (Latte, Open-Sora, LTX-video, and Wan2.1) at comparable acceleration rates, showing even larger gains under higher acceleration settings.

Recent advancements in Video Large Language Models (VideoLLMs) have enabled strong performance across diverse multimodal video tasks. To reduce the high computational cost of processing dense video frames, efficiency-oriented methods such as frame selection have been widely adopted. While effective at minimizing redundancy, these methods often cause notable performance drops on tasks requiring temporal reasoning. Unlike humans, who can infer event progression from sparse visual cues, VideoLLMs frequently misinterpret temporal relations when intermediate frames are omitted. To address this limitation, we explore visual prompting (VP) as a lightweight yet effective way to enhance temporal understanding in VideoLLMs. Our analysis reveals that simply annotating each frame with explicit ordinal information helps the model perceive temporal continuity. This visual cue also supports frame-level referencing and mitigates positional ambiguity within a sparsely sampled sequence. Building on these insights, we introduce ViKey, a training-free framework that combines VP with a lightweight Keyword-Frame Mapping (KFM) module. KFM leverages frame indices as dictionary-like keys to link textual cues to the most relevant frames, providing explicit temporal anchors during inference. Despite its simplicity, our approach substantially improves temporal reasoning and, on some datasets, preserves dense-frame baseline performance with as few as 20% of frames.

Strand-level hair geometry reconstruction is a fundamental problem in virtual human modeling and the digitization of hairstyles. However, existing methods still suffer from a significant trade-off between accuracy and efficiency. Implicit neural representations can capture the global hair shape but often fail to preserve fine-grained strand details, while explicit optimization-based approaches achieve high-fidelity reconstructions at the cost of heavy computation and poor scalability. To address this issue, we propose EfficientMonoHair, a fast and accurate framework that combines the implicit neural network with multi-view geometric fusion for strand-level reconstruction from monocular video. Our method introduces a fusion-patch-based multi-view optimization that reduces the number of optimization iterations for point cloud direction, as well as a novel parallel hair-growing strategy that relaxes voxel occupancy constraints, allowing large-scale strand tracing to remain stable and robust even under inaccurate or noisy orientation fields. Extensive experiments on representative real-world hairstyles demonstrate that our method can robustly reconstruct high-fidelity strand geometries with accuracy. On synthetic benchmarks, our method achieves reconstruction quality comparable to state-of-the-art methods, while improving runtime efficiency by nearly an order of magnitude.

While diffusion models have achieved great success in the field of video generation, this progress is accompanied by a rapidly escalating computational burden. Among the existing acceleration methods, Feature Caching is popular due to its training-free property and considerable speedup performance, but it inevitably faces semantic and detail drop with further compression. Another widely adopted method, training-aware step-distillation, though successful in image generation, also faces drastic degradation in video generation with a few steps. Furthermore, the quality loss becomes more severe when simply applying training-free feature caching to the step-distilled models, due to the sparser sampling steps. This paper novelly introduces a distillation-compatible learnable feature caching mechanism for the first time. We employ a lightweight learnable neural predictor instead of traditional training-free heuristics for diffusion models, enabling a more accurate capture of the high-dimensional feature evolution process. Furthermore, we explore the challenges of highly compressed distillation on large-scale video models and propose a conservative Restricted MeanFlow approach to achieve more stable and lossless distillation. By undertaking these initiatives, we further push the acceleration boundaries to 11.8 times while preserving generation quality. Extensive experiments demonstrate the effectiveness of our method.

While neural video codecs (NVCs) have recently demonstrated superior performance over traditional codecs through end-to-end learning, existing approaches primarily focus on architectural enhancements and coding module design, with limited exploration into optimizing hierarchical structures--specifically, quality and reference configurations. Current hierarchical structure optimization methods face two major limitations: (1) insufficient content-adaptive optimization, and (2) disjointed handling of quality and reference structures. To overcome these challenges, we propose a novel NVC framework that introduces content-adaptive hierarchical structure optimization through a hierarchical hyperprior derived from the current frame. Our NVC integrates two key components: (1) a hierarchical hyperprior extracted from the original frame to enable content-aware adaptation of the hierarchical structure; and (2) an adaptor within the hierarchical hyperprior codec combined with a dual-reference scheme, guided by the hyperprior, to jointly optimize quality and reference structures. By leveraging this content-adaptive hierarchical structure, our NVC achieves state-of-the-art rate-distortion performance, outperforming the previous leading NVC method DCVC-FM with BD-rate reductions of 15.51% and 12.20% relative to VTM-23.4 low-delay B (LDB) under intra-period settings of -1 and 32, respectively.

Video temporal grounding, the task of localizing the start and end times of a natural language query in untrimmed video, requires capturing both global context and fine-grained temporal detail. This challenge is particularly pronounced in long videos, where existing methods often compromise temporal fidelity by over-downsampling or using fixed windows. We present HieraMamba, a hierarchical architecture that preserves temporal structure and semantic richness across scales. At its core are Anchor-MambaPooling (AMP) blocks, which utilize Mamba's selective scanning to produce compact anchor tokens summarizing video content across scales. We further introduce anchor-conditioned and segment-pooled contrastive losses-two complementary objectives that encourage anchors to retain local detail while remaining globally discriminative. HieraMamba sets a new state-of-the-art on Ego4D-NLQ, MAD, and TACoS, demonstrating precise, temporally faithful localization in long, untrimmed videos.

Referring Video Object Segmentation (RVOS) aims to segment target objects in videos based on natural language descriptions. However, fixed keyframe-based approaches that couple a vision language model with a separate propagation module often fail to capture rapidly changing spatiotemporal dynamics and to handle queries requiring multi-step reasoning, leading to sharp performance drops on motion-intensive and reasoning-oriented videos beyond static RVOS benchmarks. To address these limitations, we propose VIRST (Video-Instructed Reasoning Assistant for Spatio-Temporal Segmentation), an end-to-end framework that unifies global video reasoning and pixel-level mask prediction within a single model. VIRST bridges semantic and segmentation representations through the Spatio-Temporal Fusion (STF), which fuses segmentation-aware video features into the vision-language backbone, and employs the Temporal Dynamic Anchor Updater to maintain temporally adjacent anchor frames that provide stable temporal cues under large motion, occlusion, and reappearance. This unified design achieves state-of-the-art results across diverse RVOS benchmarks under realistic and challenging conditions, demonstrating strong generalization to both referring and reasoning oriented settings.

Achieving human-like spatial intelligence for vision-language models (VLMs) requires inferring 3D structures from 2D observations, recognizing object properties and relations in 3D space, and performing high-level spatial reasoning. In this paper, we propose a principled hierarchical framework that decomposes the learning of 3D spatial understanding in VLMs into four progressively complex levels, from geometric perception to abstract spatial reasoning. Guided by this framework, we construct an automated pipeline that processes approximately 5M images with over 45M objects to generate 3D spatial VQA pairs across diverse tasks and scenes for VLM supervised fine-tuning. We also develop an RGB-D VLM incorporating metric-scale point maps as auxiliary inputs to further enhance spatial understanding. Extensive experiments demonstrate that our approach achieves state-of-the-art performance on multiple spatial understanding and reasoning benchmarks, surpassing specialized spatial models and large proprietary systems such as Gemini-2.5-pro and GPT-5. Moreover, our analysis reveals clear dependencies among hierarchical task levels, offering new insights into how multi-level task design facilitates the emergence of 3D spatial intelligence.

We revisit video hallucination in multimodal large language models (Video-MLLMs) from a semantic aggregation perspective. While prior work attributes hallucinations to language priors, missing frames, or visual encoder biases, these explanations overlook errors arising during the aggregation of correct frame-level semantics into event-level interpretations. We term this phenomenon Semantic Aggregation Hallucination (SAH), which becomes increasingly prevalent in complex, multi-event video understanding tasks with rich temporal dependencies. To systematically study SAH, we introduce ELV-Halluc, the first benchmark designed for fine-grained evaluation of semantic aggregation errors. Our experiments reveal that SAH correlates with both semantic complexity and rapid semantic transitions. We further propose mitigation strategies: improved positional encoding preserves temporal structure, and reinforcement learning such as DPO enhances the model's ability to distinguish semantics within and across events. Using a curated 8K adversarial video-text pair dataset, our approach achieves consistent gains across benchmarks, including a 27.7% reduction in SAH rate on ELV-Halluc and Video-MME. Data and code are available at https://github.com/hlsv02/ELV-Halluc.