CVPR 2026 Papers

Referring video object segmentation (RVOS) aims to segment objects within a video according to natural language expressions. Unlike earlier works focusing on static single-object scenarios, recent studies address more complex motion scenes. Previous methods typically adopt a query-based, logically multi-stage pipeline to handle these scenarios. However, this paradigm learns trajectory consistency modeling and multimodal fusion from scratch, which often leads to trajectory inconsistencies and insufficient multimodal understanding. To address these limitations, we propose DeRVOS, a framework that decouples RVOS into two key branches: consistent trajectory generation and multimodal understanding. We extract temporally consistent object representations using a powerful pretrained instance trajectory generation model and perform cross-modal alignment via a unified multimodal encoder, enabling upstream modeling of trajectory consistency and vision-language understanding. This design reduces RVOS to the task of modeling the relationship between referring expressions and instance trajectories. To connect the two branches and enable efficient motion-aware semantic understanding, we introduce the Trajectory Alignment and Implicit Selection (TAIS) module, which progressively performs cross-frame multimodal alignment and motion-guided implicit trajectory selection. Extensive experiments demonstrate that DeRVOS achieves state-of-the-art results on both traditional RVOS benchmarks and the challenging MeViS dataset, surpassing LVLM-based methods by 4.7%.

With the advancement of multi-modal Large Language Models (LLMs), Video LLMs have been further developed to perform on holistic and specialized video understanding. However, existing works are limited to specialized video understanding tasks, failing to achieve a comprehensive and multi-grained video perception. To bridge this gap, we introduce UFVideo, the first Video LLM with unified multi-grained cooperative understanding capabilities. Specifically, we design unified visual-language guided alignment to flexibly handle video understanding across global, pixel and temporal scales within a single model. UFVideo dynamically encodes the visual and text inputs of different tasks and generates the textual response, temporal localization, or grounded mask. Additionally, to evaluate challenging multi-grained video understanding tasks, we construct the UFVideo-Bench consisting of three distinct collaborative tasks within the scales, which demonstrates UFVideo's flexibility and advantages over GPT-4o. Furthermore, we validate the effectiveness of our model across 9 public benchmarks covering various common video understanding tasks, providing valuable insights for future Video LLMs.

Dataset distillation (DD) aims to synthesize compact training sets that enable models to achieve high accuracy with significantly fewer samples. Recent diffusion-based DD methods commonly introduce semantic guidance through late-stage cross-attention, where textual prompts tend to dominate the generative process. Although this strategy enforces label relevance, it diminishes the contribution of visual latents, resulting in over-corrected samples that mirror prompt patterns rather than reflecting intrinsic visual features. To solve this problem, we introduce an Early Vision-Language Fusion (EVLF) method that aligns textual and visual embeddings at the transition between the encoder and the generative backbone. By incorporating a lightweight cross-attention module at this transition, the early representations simultaneously encode local textures and global semantic directions across the denoising process. Importantly, EVLF is plug-and-play and can be easily integrated into any diffusion-based dataset distillation pipeline with an encoder. It works across different denoiser architectures and sampling schedules without any task-specific modifications. Extensive experiments demonstrate that EVLF generates semantically faithful and visually coherent synthetic data, yielding consistent improvements in downstream classification accuracy across varied settings. Source code is available at https://github.com/wenqi-cai297/earlyfusion-for-dd .

Recent advances in Multimodal Large Language Models (MLLMs) have enabled unified multimodal understanding and generation. However, they still struggle with fine-grained text-image alignment, often failing to faithfully depict objects with correct attributes such as color, shape, and spatial relationships. To mitigate this issue, previous studies have explored preference optimization methods such as DPO and GRPO, but these approaches incur substantial computational cost, both in constructing preference data and in performing optimization. This has motivated self-improving preference optimization approaches, in which the MLLM autonomously generates its own training data, self-estimates preference feedback, and self-optimizes using the resulting self-constructed preference pairs. However, existing self-improving methods still overlook fine-grained, object-level semantics, allowing object hallucination to persist. To tackle this problem, we propose Object-centric Self-improving Preference Optimization (OSPO), a self-improving framework designed to enhance object-level text-image alignment. OSPO explicitly constructs object-centric preference data without relying on any external data or models. We also introduce a new approach that leverages attention-based object masks together with an object-weighted SimPO loss to enhance object-specific fidelity. Extensive experiments on three compositional image generation benchmarks demonstrate that OSPO significantly improves fine-grained alignment and reduces object hallucination, outperforming prior self-improving methods and even specialized diffusion-based text-to-image models.

We present Flowception, a novel non-autoregressive and variable-length video generation framework. Flowception learns a probability path that interleaves discrete frame insertions with continuous frame denoising. Compared to autoregressive methods, Flowception alleviates error accumulation/drift as the frame insertion mechanism during sampling serves as an efficient compression mechanism to handle long-term context. Compared to full-sequence flows, our method reduces FLOPs for training three-fold, while also being more amenable to local attention variants, and allowing to learn the length of videos jointly with their content. Quantitative experimental results show improved FVD and VBench metrics over autoregressive and full-sequence baselines, which is further validated with qualitative results. Finally, by learning to insert and denoise frames in a sequence, Flowception seamlessly integrates different tasks such as image-to-video generation and video interpolation.

The performance of Latent Diffusion Models (LDMs) is critically dependent on the quality of their visual tokenizers. While recent works have explored incorporating Vision Foundation Models (VFMs) into the tokenizers training via distillation, we empirically find this approach inevitably weakens the robustness of learnt representation from original VFM. In this paper, we bypass the distillation by proposing a more direct approach by leveraging the frozen VFM for the LDMs tokenizer, named VFM Variational Autoencoder (VFM-VAE). To fully exploit the potential to leverage frozen VFM for the LDMs tokenizer, we design a new decoder to reconstruct realistic images from the semantic-rich representation of VFM. With the proposed VFM-VAE, we conduct a systematic study on how the representation from different tokenizers impact the representation learning process throughout diffusion training, enabling synergistic benefits of dual-side alignment on both tokenizers and diffusion models. Our effort in tokenizer design and training strategy lead to superior performance and efficiency: our system reaches a **gFID (w/o CFG) of 2.22 in merely 80 epochs** (a 10xspeedup over prior tokenizers). With continued training to **640 epochs**, it further attains a **gFID (w/o CFG) of 1.62**. These results offer solid evidence for the substantial potential of VFMs to serve as visual tokenizers to accelerate the LDM training progress.

Scaling multimodal alignment between video and audio is challenging, particularly due to limited data and the mismatch between text descriptions and frame-level video information. In this work, we tackle the scaling challenge in multimodal-to-audio generation, examining whether models trained on short instances can generalize to longer ones during testing. To tackle this challenge, we present multimodal hierarchical networks so-called MMHNet, an enhanced extension of state-of-the-art video-to-audio models. Our approach integrates a hierarchical method and non-causal Mamba to support long-form audio generation. Our proposed method significantly improves long audio generation up to more than 5 minutes. We also prove that training short and testing long is possible in the video-to-audio generation tasks without training on the longer durations. We show in our experiments that our proposed method could achieve remarkable results on long-video to audio benchmarks, beating prior works in video-to-audio tasks. Moreover, we showcase our model capability in generating more than 5 minutes, while prior video-to-audio methods fall short in generating with long durations. Our project page: https://echoesovertime.github.io

Recent advances in fMRI pretraining have significantly improved visual decoding accuracy by leveraging cross-subject neuroimaging datasets. A prevailing strategy aligns individual fMRI signals into a shared feature space using subject-specific adapters, followed by a shared decoder. However, this unstructured feature space overlooks the redundancy and functional correlations among voxels and fails to incorporate the brain's intrinsic functional architecture centered on regions of interest (ROIs). To address these limitations, we propose ROITok, an ROI-guided fMRI pretraining framework. Our method introduces Sparse ROI Context Fusion to learn ROI-level visual representations and captures functional synergy between ROIs from cross-subject data. Inspired by Matryoshka Representation Learning (MRL), we design an embedding compression scheme that prioritizes the most informative visual components first, with later tokens adding progressively finer but still useful details. ROITok achieves strong transfer learning performance on the NSD and GOD datasets and shows strong resilience against high-level additive noises, while offering better interpretability. It allows for quantitative assessment of each brain region's contribution to decoding tasks. Our analysis shows that ROI-based pretraining can automatically learn the brain's visual hierarchy. Different ROIs can provide complementary contexts for decoding tasks; combining them improves decoding robustness.

Reinforcement learning (RL), earlier proven to be effective in large language and multi-modal models, has been successfully extended to enhance 2D image generation recently. However, applying RL to 3D generation remains largely unexplored due to the higher spatial complexity of 3D objects, which require globally consistent geometry and fine-grained local textures. This makes 3D generation significantly sensitive to reward designs and RL algorithms. To address these challenges, we conduct the first systematic study of RL for text-to-3D autoregressive generation across several dimensions. (1) Reward designs: We evaluate reward dimensions and model choices, showing that alignment with human preference is crucial, and that general multi-modal models provide robust signal for 3D attributes. (2) RL algorithms: We study GRPO variants, highlighting the effectiveness of token-level optimization, and further investigate the scaling of training data and iterations. (3) Text-to-3D Benchmarks: Since existing benchmarks fail to measure implicit reasoning abilities in 3D generation models, we introduce MME-3DR. (4) Advanced RL paradigms: Motivated by the natural hierarchy of 3D generation, we propose Hi-GRPO, which optimizes the global-to-local hierarchical 3D generation through dedicated reward ensembles. Based on these insights, we develop AR3D-R1, the first RL-enhanced text-to-3D model, expert from coarse shape to texture refinement. We hope this study provides insights into RL-driven reasoning for 3D generation. The code is released at https://github.com/Ivan-Tang-3D/3DGen-R1.

Existing ViT backdoor attacks based on backbone-overwriting full-tuning are computationally expensive and inflict performance degradation. This has forced adversaries towards the Visual Parameter-Efficient Fine-Tuning (PEFT) paradigm, dominated by adapter-based (e.g., LoRA) and prompt-based (e.g., VPT) approaches. While adapter security has seen initial study, the risks of the burgeoning prompt-based ecosystem remain critically unexplored. We fill this critical gap, exposing how the evolution of VPT towards dynamic and context-aware architectures can facilitate a far more dangerous and emergent threat.This vulnerability arises even though these dynamic modules unlock superior benign performance. We propose VIPER, an attack framework built on a lightweight, dynamic Visual Prompt Generator (VPG) that demonstrates this vulnerability. Critically, this dynamic architecture enables Functional Fusion: an emergent phenomenon where malicious logic and benign task utility are tightly fused into the same sparse, high-magnitude parameter core. This fusion creates a formidable "hostage" dilemma, as pruning the attack necessarily destroys the benign performance. Comprehensive evaluations show VIPER effectively addresses the attacker's trilemma: VIPER not only achieves state-of-the-art performance on clean data, but also maintains near-100% ASR even under 90% VPG-module pruning (where LoRA attacks collapse), while adding only an imperceptible 0.06ms (1.16%) of inference latency. VIPER's results, driven by Functional Fusion, expose a new, paradigm-level risk in dynamic prompt architectures.

Understanding dynamic 3D environments is essential for safe autonomous driving, particularly when reasoning about human-centric, nonrigid agents. However, existing weakly supervised occupancy prediction frameworks predominantly assume rigid-body motion and rely on simple frame-to-frame offsets, limiting their ability to capture fine-grained deformations and maintain temporal coherence. To address this issue, we propose DeGO, a deformable Gaussian occupancy framework that unifies decoupled Gaussian deformation with factorized 4D foundation-model distillation. DeGO disentangles rigid and nonrigid motion, enabling each Gaussian primitive to evolve through both deformation and offset-based updates. In parallel, a factorized 4D distillation strategy transfers cross-camera and cross-frame knowledge from the VGGT foundation model, producing foundation-aligned features that enhance temporal consistency. Experiments on the Occ3D-NuScenes benchmark demonstrate that our method achieves state-of-the-art performance under weak supervision, delivering 13.5% gains on human-centric instances and 10.9% overall improvements. These results highlight the effectiveness of deformation-aware and foundation-guided occupancy modeling for dynamic scene understanding.

Discovering physical laws directly from high-dimensional visual data is a long-standing human pursuit but remains a formidable challenge for machines, representing a fundamental goal of scientific intelligence. This task is inherently difficult because physical knowledge is low-dimensional and structured, whereas raw video observations are high-dimensional and redundant, with most pixels carrying little or no physical meaning. Extracting concise, physically relevant variables from such noisy data remains a key obstacle. To address this, we propose Pixel2Phys, a collaborative multi-agent framework adaptable to any Multimodal Large Language Model (MLLM). It emulates human scientific reasoning by employing a structured workflow to extract formalized physical knowledge through iterative hypothesis generation, validation, and refinement. By repeatedly formulating, and refining candidate equations on high-dimensional data, it identifies the most concise representations that best capture the underlying physical evolution. This automated exploration mimics the iterative workflow of human scientists, enabling AI to reveal interpretable governing equations directly from raw observations. Across diverse simulated and real-world physics videos, Pixel2Phys discovers accurate, interpretable governing equations and maintaining stable long-term extrapolation where baselines rapidly diverge.

Large multimodal language models have made rapid progress on vision-language tasks, yet their potential for zero-/few-shot object detection (ZSOD/FSOD) under a closed set of target classes has yet to be fully explored.. ZSOD/FSOD is hampered by data scarcity and catastrophic forgetting. Although vision-language models (VLMs) demonstrated excellent performance on multiple benchmarks, they typically rely on large-scale visual pre-training, which is inconsistent with the goal of FSOD to test generalization to new categories under limited supervision. We introduce AgentDet, a shared-blackboard multi-agent framework that unifies ZSOD and FSOD via pseudo-incremental learning. AgentDet decouples detection into four cooperating roles--Agent-Scout, Agent-Pinner, Agent-Curator, and Agent-Judge--which collaboratively maintain a Shared Blackboard and a Knowledge Base. For efficiency, we only train Agent-Judge by updating its image encoder and LLM-based detection head, which is a lightweight recipe that encourages generalization to previously unseen categories. On PASCAL VOC and MS COCO ZSOD/FSOD datasets, AgentDet achieves strongly competitive performance with state-of-the-art results in several settings.

End-to-end autonomous driving (E2E-AD) demands effective processing of multi-view sensor data and robust handling of diverse and complex driving scenarios, particularly rare maneuvers such as aggressive turns. The recent success of the Mixture-of-Experts (MoE) architecture in Large Language Models (LLMs) demonstrates that expert specialization enables strong scalability. In this work, we propose DriveMoE, a novel MoE-based E2E-AD framework, with a Scene-Specialized Vision MoE and a Skill-Specialized Action MoE. First, we introduce Drive-p0, a Vision-Language-Action (VLA) baseline adapted from Embodied AI for autonomous driving, which serves as the foundation model for DriveMoE. Building on this, we strengthen perception through a carefully designed Vision MoE, where a router adaptively selects context-relevant camera views. This mechanism is inspired by human driving cognition, in which attention is directed to key visual cues rather than to all sensory inputs simultaneously. Beyond perception, we introduce an Action MoE that augments the framework by training a router to activate specialized expert modules tailored to distinct driving behaviors. Within the Action MoE, we implement two distinct styles(Token-level Router and Trajectory-level Router) and extensively explore their applicability in autonomous driving. In Bench2Drive closed-loop evaluations, DriveMoE demonstrates robust performance across diverse driving scenarios, alleviates the mode-averaging effect that limits existing models, and achieves state-of-the-art results with significant improvements over Drive-p0. We will release our code and models of DriveMoE and Drive-p0.

Test-time training (TTT) enhances the robustness of pretrained models to out-of-distribution (OOD) data through auxiliary self-supervised tasks, without requiring labeled samples. However, existing TTT methods predominantly rely on decoder-based auxiliary objectives, which suffer from inefficient adaptation and weak coupling with the primary task. To solve these limitations, we revisit the mechanism of test-time training by analyzing masking-based pretrained models to uncover the fundamental source of their OOD robustness. Our investigation reveals that their generalization capability stems from a latent feature-level structural invariance, the consistency of encoded representations under masked perturbations. Building on this insight, we introduce LoTT-PC, a lightweight LoRA-based framework that operationalizes this invariance-preserving principle for 3D point cloud classification. LoTT-PC consists of two main components: (1) low-rank modulation units for parameter-efficient adaptation, and (2) a permutation-invariant alignment mechanism that enforces representation consistency through masked feature alignment. Extensive experiments on multiple benchmarks demonstrate that this unified design enables pretrained point cloud models to self-tune rapidly and reliably across diverse OOD scenarios, outperforming state-of-the-art methods by an average of 2.7% in accuracy under various corruption types.

The growing adoption of XR devices has fueled strong demand for high-quality stereo video, yet its production remains costly and artifact-prone.To address this challenge, we present **StereoWorld**, an **end-to-end framework** that repurposes a pretrained video generator for high-fidelity monocular-to-stereo video generation. Our framework jointly conditions the model on the monocular video input while explicitly supervising the generation with a **geometry-aware regularization** to ensure 3D structural fidelity.A spatio-temporal tiling scheme is further integrated to enable efficient, high-resolution synthesis.To enable large-scale training and evaluation, we curate a **high-definition stereo video dataset** containing over 11M frames aligned to natural human interpupillary distance (IPD).Extensive experiments demonstrate that StereoWorld substantially outperforms prior methods, generating stereo videos with superior visual fidelity and geometric consistency.

Camera-based Semantic Scene Completion (SSC) is able to comprehensively understand the entire scene, but it suffers from ambiguous predictions due to occlusions and incomplete information. Temporal SSC alleviates this issue, but existing models simply stack multi-frame temporal features, which can lead to inconsistencies between geometry and semantics over time. In this paper, we present ConSSC, a novel SSC method that learns Spatial-Temporal Consistency. It works by lifting historical frames into a 3D scene-level occupancy framework, aggregating 2D and 3D historical features from current voxels, and learning from 2D visibility and similarity cues in a temporal buffer. Specifically, our framework introduces two key components: the Hierarchical Voxel Refinement module, which extracts a coarse occupancy from depth and refines it through voxel-level representations, recovering missing information. The Temporal Semantic Aggregation module effectively integrates semantic features from different viewpoints and time points, enabling the reconstruction of occluded regions in the current frame using historical context, aggregating them into corresponding voxel features. Without additional sensors or data, ConSSC improves both geometric and semantic consistency. Extensive experiments on the SemanticKITTI and SSCBench-KITTI-360 datasets show that ConSSC outperforms state-of-the-art camera-based and temporal SSC baselines by a significant margin in terms of IoU and mIoU.

Recent advancements in text-to-image synthesis have been largely propelled by diffusion-based models, yet achieving precise alignment between text prompts and generated images remains a persistent challenge. We find that this difficulty arises primarily from the limitations of conventional diffusion loss, which provides only implicit supervision for modeling fine-grained text-image correspondence. In this paper, we introduce Cross-Timestep Self-Calibration (CTCal), founded on the supporting observation that establishing accurate text-image alignment within diffusion models becomes progressively more difficult as the timestep increases. CTCal leverages the reliable text-image alignment (i.e., cross-attention maps) formed at smaller timesteps with less noise to calibrate the representation learning at larger timesteps with more noise, thereby providing explicit supervision during training. We further propose a timestep-aware adaptive weighting to achieve a harmonious integration of CTCal and diffusion loss. CTCal is model-agnostic and can be seamlessly integrated into existing text-to-image diffusion models, encompassing both diffusion-based (e.g., SD 2.1) and flow-based approaches (e.g., SD 3). Extensive experiments on T2I-Compbench++ and GenEval benchmarks demonstrate the effectiveness and generalizability of the proposed CTCal. Our code is available at https://github.com/xiefan-guo/ctcal.

Vision-and-Language Navigation in Continuous Environment (VLN-CE) requires an agent to follow language instructions to navigate the target destination. With the advancement of large language models (LLMs), recent efforts have explored adapting them for zero-shot VLN-CE, offering a promising solution in addressing the drawbacks of poor generalization in the training-based paradigm. However, existing LLM-based works primarily perform naive reasoning for decision-making and lack feedback, e.g., reviewing historical errors and predicting future potentials. Consequently, it may suffer from continuous failure for those initial error tasks. In this paper, we rethink LLM-based zero-shot VLN-CE and propose a new paradigm, named EvoNav, to improve the agent's decision-making with future thought and history experience via Future Chain-of-Thought (F-CoT) and History Chain-of-Experience (H-CoE). F-CoT predicts future actions and landmarks as thoughts to assist navigation progress estimation and direction selection, while H-CoE summarizes historical trajectories and scenes as experience to improve navigation decision reliability. Both F-CoT and H-CoE cooperatively evolve the agent's decision-making. Extensive experiments in both the simulator and real-world environments demonstrate the effectiveness of our EvoNav.