CVPR 2026 Papers

Recently, unified multimodal models (UMMs) have made remarkable progress in integrating visual understanding and generation, demonstrating strong potential for complex text-to-image (T2I) tasks. Despite their theoretical promise, a persistent capability gap exists: UMMs typically exhibit superior visual understanding but comparatively weaker generative capabilities. This discrepancy arises largely from the intrinsic decoupling between the understanding and generation processes. While a UMM can accurately interpret fine-grained visual details, it often struggles to produce semantically coherent images from complex textual prompts. To address this challenge, we explore UMMs' internal understanding capability to enhance generation quality. We propose a token-level intrinsic text-image alignment reward mechanism, GvU, enabling the UMM to act simultaneously as teacher and student: it evaluates its own outputs using the understanding branch to guide the generations accordingly. Building upon this, we design a self-supervised reinforcement learning framework, allowing UMMs to iteratively improve their generation quality through understanding-based intrinsic reward signals--without reliance on external supervision. Experimental results show that our method substantially boosts UMMs' generation, which in turn strengthens their fine-grained visual understanding, narrowing the capability gap between UMMs' visual understanding and generation. The project page is https://matrix0721.github.io/gvu.github.io/.

Recent work has shown that eliciting Large Language Models (LLMs) to generate reasoning traces in natural language before answering the user's request can significantly improve their performance across tasks. This approach has been extended to multimodal LLMs, where the models can produce chains-of-thoughts (CoT) about the content of input images and videos. For video inputs, prior works use complex multi-step pipelines that extract and include relevant frames from videos in the CoT, or produce simpler single-stage reasoning traces at the expense of poor temporal grounding. Here, we propose the first video LLMs with single-stage reasoning that includes explicit references to relevant frames, thereby reducing temporal inconsistencies in the reasoning process. Our approach is simple, unified, and self-contained, employing a single-stage inference to handle complex video understanding tasks without relying on auxiliary modules for frame selection or caption generation. For this, we first create CoF-Data, a large dataset of diverse questions, answers, and corresponding frame-grounded reasoning traces from both natural and synthetic videos, spanning various topics and tasks. Our models, obtained by fine-tuning video LLMs on this chain-of-frames (CoF) data, generate reasoning traces that accurately identify key frames to answer given questions. In turn, this consistently improves performance across multiple video understanding benchmarks. Surprisingly, we find that synthetic data alone, despite being out-of-distribution with respect to these real-world benchmarks, provides a significant boost in model accuracy.

Despite remarkable advancements in multimodal large language models (MLLMs), their fine-grained visual understanding is constrained by a primary reliance on sparse textual supervision. Existing efforts to introduce visual supervision typically do so during post-training, when visual representations have already been largely fixed, causing such signals to act mainly as auxiliary constraints rather than as a primary force for shaping perceptual features. In this paper, we aim to fundamentally reshape the model's perceptual backbone by incorporating vision supervision directly into the pre-training stage. We observe that pixel-level image patches and textual tokens naturally coexist in a shared, raw high-dimensional space characterized by an inherent input symmetry. Leveraging this insight, we propose UVU, a novel vision-language unified autoregressive framework that eschews vector quantization. It uniquely employs continuous visual encoding for lossless representation of visual inputs and proposes a large-scale iterative hierarchical clustering algorithm to construct a pixel-level visual codebook, thereby extending the vocabulary for unified supervision and enabling autoregressive generation of pixel-level image tokens alongside textual tokens. UVU effectively synergizes pixel-level visual perception with semantic-level visual understanding, internalizing visual reconstruction capabilities and unlocking the facilitative role of visual supervision in enhancing understanding in the pre-training stage. Extensive experiments across multiple tasks demonstrate that MLLMs are capable of achieving superior multimodal understanding performance under the supervised learning paradigm of UVU.

We introduce Mistake Attribution (MATT), a new task for fine-grained understanding of human mistakes in egocentric videos. While prior work detects whether a mistake occurs, MATT attributes the mistake to what part of the instruction is violated (semantic role), when in the video the deviation becomes irreversible (the Point-of-No-Return, PNR), and where the mistake appears in the PNR frame. We develop MisEngine, a data engine that automatically constructs mistake samples from existing datasets with attribution-rich annotations. Applied to large egocentric corpora, MisEngine yields EPIC-KITCHENS-M and Ego4D-M---two datasets up to two orders of magnitude larger than prior mistake datasets.We then present MisFormer, a unified attention-based model for mistake attribution across semantic, temporal, and spatial dimensions, trained with MisEngine supervision. A human study demonstrates the ecological validity of our MisEngine-constructed mistake samples, confirming that EPIC-KITCHENS-M and Ego4D-M can serve as reliable benchmarks for mistake understanding. Experiments on both our datasets and prior benchmarks show that MisFormer, as a single unified model, outperforms task-specific SOTA methods by at least 6.66%, 21.81%, 18.7%, and 3.00% in video-language understanding, temporal localization, hand-object interaction, and mistake detection, respectively. Project page: https://yayuanli.github.io/MATT/

Video Novel View Synthesis (VNVS) aims to render arbitrary novel viewpoints of dynamic scenes from a single-view video, but its algorithmic training faces a major challenge: the lack of large-scale multi-view video datasets. Prior methods often train on monocular data by framing it as an inpainting task, which typically leads to a training-inference gap and visual artifacts. While synthetic multi-view data can partially alleviate the data scarcity issue, its high acquisition costs and limited diversity restrict scalability. To address these problems, we propose Scaling4D, a novel strategy that theoretically bridges the training-inference gap while leveraging large-scale monocular videos for training. Specifically, we take a higher-level perspective on the problem, reformulating VNVS into a general correspondence-guided generation task. Furthermore, in conjunction with extensive real-world data, we establish a synthetic data pipeline integrated with our training strategy to enhance precision. Qualitative and quantitative results demonstrate a positive correlation between performance and training data volume, confirming the scalability.

While neural video codecs (NVCs) have demonstrated superior compression ratio, their prohibitive computational complexity remains a critical barrier to real-world deployment. This paper introduces a chunk-based coding framework designed to significantly improve the rate-distortion-complexity trade-off. Instead of processing frames sequentially, our approach encodes a chunk of multiple frames into a single compact latent representation and decodes them simultaneously. This is enabled by cross-frame interaction modules for joint spatial-temporal modeling and frame-specific decoders for parallel reconstruction. This paradigm not only dramatically enhances coding throughput but also facilitates more effective modeling of long-term temporal correlations. To further boost speed, we propose a streamlined entropy coding mechanism that consolidates bit-stream interactions into a single step, substantially reducing decoding overhead. Building on these innovations, we present DCVC-UF (Ultra-Fast), a new NVC that sets a new SOTA in performance. Our experiments show that DCVC-UF can achieve ultra-fast encoding and decoding speeds, significantly outperforming previous leading codecs. DCVC-UF serves as a notable landmark in the journey of NVC evolution. The code is at https://github.com/microsoft/DCVC.

The recovery of training data from generative models ("model inversion") has been extensively studied for diffusion models in the data domain as a memorization/overfitting phenomenon. Latent diffusion models (LDMs), which operate on the latent codes from encoder/decoder pairs, have been robust to prior inversion methods. In this work we describe two key findings: (1) the diffusion model exhibits non-uniform memorization across latent codes, tending to overfit samples located in high-distortion regions of the decoder pullback metric; (2) even within a single latent code, memorization contributions are unequal across representation dimensions. Our proposed method to ranks latent dimensions by their contribution to the decoder pullback metric, which in turn identifies dimensions that contribute to memorization. For score-based membership inference, a sub-task of model inversion, we find that removing less-memorizing dimensions improves performance on all tested methods and datasets, with average AUROC gains of 1-4% and substantial increases in TPR@1%FPR (1-32%) across diverse datasets including CIFAR-10, CelebA, ImageNet-1K, Pokemon, MS-COCO, and Flickr. Our results highlight the overlooked influence of the auto-encoder geometry on LDM memorization and provide a new perspective for analyzing privacy risks in diffusion-based generative models.

Video comprises the vast majority of bits that are generated daily, and is the primary signal driving current innovations in robotics, remote sensing, and wearable technology. Yet, the most powerful video understanding models are too expensive for the resource-constrained platforms used in these applications. One approach is to offload inference to the cloud; this gives access to GPUs capable of processing high-resolution videos in real time. But even with reliable, high-bandwidth communication channels, the combined latency of video encoding, model inference, and round-trip communication prohibits use for certain real-time applications. The alternative is to use fully local inference; but this places extreme constraints on computational and power costs, requiring smaller models and lower resolution, leading to degraded accuracy. To address these challenges, we propose DeDelayed, a real-time inference system that divides computation between a remote model operating on delayed video frames and a local model with access to the current frame. The remote model is trained to make predictions on anticipated future frames, which the local model incorporates into its prediction for the current frame. The local and remote models are jointly optimized with an autoencoder that limits the transmission bitrate required by the available downlink communication channel. We evaluate DeDelayed on the task of real-time streaming video segmentation using the BDD100k driving dataset. For a round trip delay of 100 ms, DeDelayed improves performance by 6.4 mIoU compared to fully local inference and 9.8 mIoU compared to remote inference---an equivalent improvement to using a model ten times larger. We release our training code, pretrained models, and python library at https://github.com/InterDigitalInc/dedelayed .

Aligning egocentric video with wearable sensors has shown promise for human action recognition, but faces practical limitations in user discomfort, privacy concerns, and scalability. We explore exocentric video with ambient sensors as a non-intrusive, scalable alternative. However, the Global Alignment approach prevalent in egocentric-wearable settings fails in this new setting due to two problems: (P1) inability to capture local details such as subtle motions, and (P2) over-reliance on modality-invariant temporal features that distort negative relationships. To resolve these problems, we propose DETACH, a decomposed spatio-temporal alignment framework. By decomposing both modalities into spatial-temporal components, we preserve subtle temporal cues in videos from spatial features, and convert implicit sensor channel activations into explicit spatial prototypes via online clustering, thereby establishing cross-modal spatial grounding. To avoid over-reliance on temporal features, a spatial-temporal weighted contrastive loss leverages this grounding for fine-grained temporal alignment, prioritizing hard negatives and suppressing false negatives. Comprehensive experiments with downstream tasks on Opportunity++ and HWU-USP datasets demonstrate improvements of up to 30% in F1 and 50% in mAP over adapted egocentric-wearable baselines.

Powerful 3D representations such as DUSt3R's invariant point maps, which encode 3D shape and camera parameters, have significantly advanced feed-forward 3D reconstruction. While point maps assume static scenes, Dynamic Point Maps (DPMs) extend the concept to dynamic 3D content by also representing scene motion. However, DPMs have so far been limited to image pairs and, like DUSt3R, require post-processing via optimisation when more than two views are involved. We argue that DPMs are more useful when applied to videos and introduce V-DPM to demonstrate this. First, we show how to set up DPMs for videos to optimise representational power, facilitate neural prediction, and enable reuse of pretrained models. Second, we implement these ideas on top of VGGT, a recent powerful 3D reconstructor. Although VGGT was trained on static scenes, we show that a modest amount of synthetic data suffices to adapt it into an effective V-DPM predictor. This yields state-of-the-art 3D and 4D reconstruction in dynamic settings. In particular, unlike recent dynamic extensions of VGGT such as P3, DPMs recover not only dynamic depth but also the 3D motion of every point in the scene.

Video Large Language Models (Video-LLMs) are improving rapidly, yet current Video Question Answering (VideoQA) benchmarks often admit single-cue shortcuts, under-testing reasoning that must integrate evidence across time. We introduce HERBench, a benchmark designed to make multi-evidence integration unavoidable: each question requires at least three non-overlapping cues drawn from distinct video segments. HERBench contains 26,806 five-way multiple-choice questions across 12 compositional tasks. To make evidential demand measurable, we introduce the Minimum Required Frame-Set (MRFS), the smallest number of frames a model must fuse to answer correctly, and show that HERBench imposes higher evidential demand than prior benchmarks. Evaluating 13 state-of-the-art Video-LLMs yields only 31-42% accuracy, only modestly above the 20% random-guess baseline. We disentangle this failure into two critical bottlenecks: (1) a retrieval deficit, where frame selectors overlook key evidence, and (2) a fusion deficit, where models fail to integrate information even when all necessary evidence is provided. HERBench thus provides a principled benchmark for studying robust multi-evidence video understanding.

Recent advances in trajectory-controllable video generation have achieved remarkable progress. Previous methods mainly use adapter-based architectures for precise motion control along predefined trajectories.However, all these methods rely on a multi-step denoising process, leading to substantial time redundancy and computational overhead.While existing video distillation methods successfully distill multi-step generators into few-step, directly applying these approaches to trajectory-controllable video generation results in noticeable degradation in both video quality and trajectory accuracy.To bridge this gap, we introduce FlashMotion, a novel training framework designed for few-step trajectory-controllable video generation.We first train a trajectory adapter on a multi-step video generator for precise trajectory control.Then, we distill the generator into a few-step version to accelerate video generation.Finally, we finetune the adapter using a hybrid strategy that combines diffusion and adversarial objectives, aligning it with the few-step generator to produce high-quality, trajectory-accurate videos.For evaluation, we introduce FlashBench, a benchmark for long-sequence trajectory-controllable video generation that measures both video quality and trajectory accuracy across varying numbers of foreground objects. Experiments on two adapter architectures show that FlashMotion surpasses existing video distillation methods and previous multi-step models in both visual quality and trajectory consistency.

High-fidelity generative video editing has seen significant quality improvements by leveraging pre-trained video foundation models. However, their computational cost is a major bottleneck, as they are often designed to inefficiently process the full video context regardless of the inpainting mask's size, even for sparse, localized edits. In this paper, we introduce EditCtrl, an efficient video inpainting control framework that focuses computation only where it is needed. Our approach features a novel local video context module that operates solely on masked tokens, yielding a computational cost proportional to the edit size. This local-first generation is then guided by a lightweight temporal global context embedder that ensures video-wide context consistency with minimal overhead. Not only is EditCtrl 10xmore compute efficient than state-of-the-art generative editing methods, it even improves editing quality compared to methods designed with full-attention. Finally, we showcase how EditCtrl unlocks new capabilities, including multi-region editing with text prompts and real-time autoregressive content propagation.

Reconstructing 3D humans and scenes from monocular videos is a challenging task, particularly due to human motion, varying illumination, and dynamic scene shadows. While recent works have explored scene disentanglement by jointly modeling humans and their surrounding scenes, they often overlook illumination and shadow effects--resulting in inconsistent human appearance and degraded scene realism. To address this gap, we propose a photometrically consistent integration of human and scene reconstruction based on 3D Gaussian Splatting, with a key focus on modeling spatially-varying illumination and shadows. Central to our method is a learnable light volume that provides localized lighting cues to human Gaussians, enabling more realistic and consistent appearance synthesis. To further ensure accurate human geometry and alignment, we adopt a two-stage reconstruction strategy: we first optimize a human mesh and then anchor Gaussians to the refined surface. In addition, we introduce an implicit shadow estimation module that disentangles cast shadows from the scene, thus supporting plausible human shadow synthesis. Our framework also facilitates human relighting and compositing into novel scenes with contextually appropriate lighting. Quantitative and qualitative results demonstrate that our method achieves state-of-the-art performance, producing consistent appearances, realistic illumination, and enhanced overall scene realism.

Although current Video-LLMs achieve impressive performance in video understanding tasks, their autoregressive decoding efficiency remains constrained by the massive number of video tokens. Visual token pruning can partially ease this bottleneck, yet existing approaches still suffer from information loss and yield only modest acceleration in decoding. In this paper, we propose ParallelVLM, a training-free draft-then-verify speculative decoding framework that overcomes both mutual waiting and limited speedup-ratio problems between draft and target models in long-video settings. ParallelVLM features two parallelized stages that maximize hardware utilization and incorporates an Unbiased Verifier-Guided Pruning strategy to better align the draft and target models by eliminating the positional bias in attention-guided pruning. Extensive experiments demonstrate that ParallelVLM effectively expands the draft window by 1.6-1.8x with high accepted lengths, and accelerates various video understanding benchmarks by 3.36x on LLaVA-Onevision-72B and 2.42x on Qwen2.5-VL-32B compared with vanilla autoregressive decoding.

Reconstructing High Dynamic Range (HDR) videos from sequences of alternating-exposure Low Dynamic Range (LDR) frames remains highly challenging, especially under dynamic scenes where cross-exposure inconsistencies and complex motion make inter-frame alignment difficult, leading to ghosting and detail loss. Existing methods often suffer from inaccurate alignment, suboptimal feature aggregation, and degraded reconstruction quality in motion-dominated regions. To address these challenges, we propose F^2HDR, a two-stage HDR video reconstruction framework that robustly perceives inter-frame motion and restores fine details in complex dynamic scenarios. The proposed framework integrates a flow adapter that adapts generic optical flow for robust cross-exposure alignment, a physical motion modeling to identify salient motion regions, and a motion-aware refinement network that aggregates complementary information while removing ghosting and noise. Extensive experiments demonstrate that F^2HDR achieves state-of-the-art performance on real-world HDR video benchmarks, producing ghost-free and high-fidelity results under large motion and exposure variations. Our code is available at https://github.com/wei1895/F2HDR.

Multimodal Large Language Models (MLLMs) have made great progress in video understanding tasks. However, when it comes to understanding complex or lengthy videos, MLLMs tend to overlook details or produce hallucinations. To alleviate these issues, recent work has attempted to leverage reinforcement learning (RL) to boost models' deep linguistic reasoning of complex videos. But these methods have two main problems: First, the RL framework they used has unstable training, high training costs, and is difficult to train satisfactory video reasoning models; Second, the linguistic reasoning process is difficult to guarantee the reliability of visual information. To alleviate these problems, we propose to use multimodal elements for reasoning, and we design a novel framework to build and enhance versatile video reasoning capabilities on MLLMs. We carefully design a multi-task cold start and multi-task reinforcement learning to improve the model's visual perception and proficiency in multiple capabilities. In the inference phase, we leverage multimodal reasoning and dynamic sampling to further improve the performance. We verified the efficiency of the framework on a base MLLM (Qwen2-VL-7B-Base). Through cold-start with 3k data and reinforcement learning training with 5k data, combined with inference design, our final model significantly outperforms the base model on seven public video benchmarks, even surpassing and approaching the state-of-the-art Instruct Models such as Qwen2.5-VL-7B-Instruct trained with large-scale data. Our code will be available at https://github.com/Wang-Xiaodong1899/VideoReasoner.

We present a method for harmonizing the lighting of a foreground video to match a target background scene, adjusting shadows, color tone, and illumination intensity (relightful harmonization). Unlike images, acquiring labeled data for videos, where identical motions are recorded under different lighting conditions, is practically infeasible and non-scalable. While one way to create such paired data is to apply existing image-based harmonization models frame by frame to a video, the resulting outputs often suffer from significant temporal jitters. We overcome this problem by introducing a novel lighting deflickering model that can stabilize the global and local lighting flickering artifacts. Our video diffusion model learns from these upgraded deflickered data with a volume of real and synthetic videos to generate high-quality video harmonization results. We further propose an asymmetric alpha mask conditioning technique to learn the clean boundaries from real videos. Experiments demonstrate that our model achieves strong temporal coherence, naturalness, cleaner boundaries, and physically meaningful lighting behavior, while maintaining strong relighting expressiveness compared to prior image-based and video-based harmonization methods.

We present EvoID, a novel framework that reformulates Identity-Preserving Video Generation as a self-evolving process through Reinforcement Learning. Moving beyond the static paradigm of imitation learning, EvoID enables a generative model to actively learn and optimize the complex trade-offs between identity fidelity, motion naturalness, and temporal coherence. At the heart of our EvoID is a dynamic, dual-path reward mechanism, which acts as an intrinsic critic by adaptively combining objective metric indicators and MLLM-based holistic quality assessment. This allows the model to "evolve" its generation strategy, focusing on different aspects of quality at different stages of training. To ensure stable and coherent evolution, we anchor the exploring Student model with a frozen Teacher, preserving robust world priors while allowing for creative refinement when generating videos. Extensive experiments demonstrate the superiority of our proposal, and EvoID achieves the total score of 0.687 on the Human-Domain of OpenS2V-Eval dataset, surpassing 0.658 of the open-source VACE and 0.653 of the commercial Hailuo. Moreover, EvoID also obtains a new record of 0.718 on our newly minted MLLM-based metric, prioritizing human perception and more comprehensively reflecting video quality.