CVPR 2026 Papers

Video Question Answering (VideoQA) has made significant strides by leveraging multimodal learning to align visual and textual modalities. However, current benchmarks overwhelmingly focus on questions answerable through explicit visual content - actions, objects, and events - directly observable within individual frames or short clips. To truly understand videos as humans do, models must go beyond what is directly shown, inferring hidden relationships and contextual cues that are only implied across frames. Current benchmarks fail to capture this essential aspect of video understanding. To address this gap, we introduce VRR-QA, a benchmark for Visual Relational Reasoning Beyond Explicit Cues. We curate our benchmark from creative and cinematic videos such as movies, that deliberately employ storytelling techniques which omit direct depictions of certain events or relations, requiring viewers to infer them. VRR-QA comprises 1K meticulously expert-annotated QA pairs drawn from 1K creative video clips covering 15 genres across 7 decades of content, from both live-action and animated titles. Our extensive evaluations on 14 leading VideoQA models reveals consistent and significant performance degradation, underscoring their reliance on surface-level visual cues and highlighting the difficulty of implicit reasoning. Even the best model substantially underperforms human baselines with only 64% accuracy. Performance variations across models further illustrate the complexity and diversity of the challenges presented by VRR-QA. By releasing both dataset and data collection framework, VRR-QA establishes a rigorous, diverse, and reproducible testbed for advancing VideoQA.

Recent advancements in video generation highlight that realistic audio-visual synchronization is crucial for engaging content creation. However, existing video editing methods largely overlook audio-visual synchronization and lack the fine-grained spatial and temporal controllability required for precise instance-level edits. In this paper, we propose AVI-Edit, a framework for audio-sync video instance editing. We propose a granularity-aware mask refiner that iteratively refines coarse user-provided masks into precise instance-level regions. We further design a self-feedback audio agent to curate high-quality audio guidance, providing fine-grained temporal control. To facilitate this task, we additionally construct a large-scale dataset with instance-centric correspondence and comprehensive annotations. Extensive experiments demonstrate that AVI-Edit outperforms state-of-the-art methods in visual quality, condition following, and audio-visual synchronization. Project page: https://hjzheng.net/projects/AVI-Edit/.

Newborns perceive the world with low-acuity, color-degraded, and temporally continuous vision, which gradually sharpens as infants develop. To explore the ecological advantages of such staged "visual diets", we train self-supervised learning (SSL) models on object-centric videos under constraints that simulate infant vision: grayscale-to-color (C), blur-to-sharp (A), and preserved temporal continuity (T)--collectively termed CATDiet. For evaluation, we establish a comprehensive benchmark across ten datasets, covering clean and corrupted image recognition, texture-shape cue conflict tests, silhouette recognition, depth-order classification, and the visual cliff paradigm.All CATDiet variants demonstrate enhanced robustness in object recognition, despite being trained solely on object-centric videos. Remarkably, models also exhibit biologically aligned developmental patterns, including neural plasticity changes mirroring synaptic density in macaque V1 and behaviors resembling infants' visual cliff responses. Building on these insights, CombDiet initializes SSL with CATDiet before standard training while preserving temporal continuity. Trained on object-centric or head-mounted infant videos, CombDiet outperforms standard SSL on both in-domain and out-of-domain object recognition and depth perception. Together, these results suggest that the developmental progression of early infant visual experience offers a powerful reverse-engineering framework for understanding the emergence of robust visual intelligence in machines. All code, data, and models are available at Github.

Vision-Language Models (VLMs) have shown strong performance on User Interface (UI) grounding tasks, driven by their ability to process increasingly high-resolution screenshots. However, screenshots are tokenized into thousands of visual tokens (e.g., about 4,700 for 2K resolution), which incurs significant computational overhead and dilutes attention. In contrast, humans typically focus on regions of interest when interacting with UIs. In this work, we pioneer the task of efficient UI grounding. We propose FocusUI, an efficient UI grounding framework that selects patches most relevant to the instruction while preserving positional continuity for precise grounding. FocusUI addresses two key challenges: (1) eliminating redundant tokens in visual encoding by constructing patch-level supervision that fuses an instruction-conditioned score with a rule-based UI-graph score, down-weighting large homogeneous regions to select distinct and instruction-relevant visual tokens; and (2) preserving positional continuity during visual token selection. We find that general visual token pruning methods suffer from severe accuracy degradation on UI grounding tasks due to broken positional information. To address this, we introduce a PosPad strategy, which compresses each contiguous sequence of dropped visual tokens into a single special marker placed at the sequence's last index to preserve positional continuity. Comprehensive experiments on four grounding benchmarks demonstrate that FocusUI surpasses GUI-specific baselines. On the ScreenSpot-Pro benchmark, FocusUI-7B achieves a 3.7% performance improvement over GUI-Actor-7B. Even with only 30% visual token retention, the performance of FocusUI-7B drops by just 3.2%, while achieving up to 1.44x faster inference and 17% lower peak GPU memory.

Visual encoding and decoding models act as gateways to understanding the neural mechanisms underlying human visual perception. Typically, visual encoding models that predict brain activity from stimuli and decoding models that reproduce stimuli from brain activity are treated as distinct tasks, requiring separate models and training procedures. This separation is inefficient and fails to model the consistency between encoding and decoding processes. To address this limitation, we propose NeuroFlow, the first unified framework that jointly models visual encoding and decoding from neural activity within a single flow model. NeuroFlow introduces two key components: (i) NeuroVAE is designed as a variational backbone to model neural variability and establish a compact, semantically structured latent space for bidirectional modeling across visual and neural modalities. (ii) Cross-modal Flow Matching (XFM) bypasses the typical paradigm of noise-to-data diffusion guided by a specific modality condition, instead learning a reversibly consistent flow model between visual and neural latent distributions. For the first time, visual encoding and decoding are reformulated as a time-dependent, reversible process within a shared latent space for unified modeling. Empirical results demonstrate that NeuroFlow achieves superior overall performance in visual encoding and decoding tasks with higher computational efficiency compared to any isolated methods. We further analyze principal factors that steer the model toward encoding-decoding consistency and, through brain functional analyses, demonstrate that NeuroFlow captures consistent activation patterns underlying neural variability. NeuroFlow marks a major step toward unified visual encoding and decoding from neural activity, providing mechanistic insights that inform future bidirectional visual brain-computer interfaces. Code and project page are publicly available at https://michaelmaiii.github.io/NeuroFlow-S.

Visual autoregressive models achieve remarkable generation quality through next-scale predictions across multi-scale token pyramids. However, the conventional method uses uniform scale downsampling to build these pyramids, leading to aliasing artifacts that compromise fine details and introduce unwanted jaggies and moire patterns. To tackle this issue, we present FVAR, which reframes the paradigm from next-scale prediction to next-focus prediction, mimicking the natural process of camera focusing from blur to clarity. Our approach introduces three key innovations: 1) Next-Focus Prediction Paradigm that transforms multi-scale autoregression by progressively reducing blur rather than simply downsampling; 2) Progressive Refocusing Pyramid Construction that uses physics-consistent defocus kernels to build clean, alias-free multi-scale representations; and 3) High-Frequency Residual Learning that employs a specialized residual teacher network to effectively incorporate alias information during training while maintaining deployment simplicity. Specifically, we construct optical low-pass views using defocus point spread function (PSF) kernels with decreasing radius, creating smooth blur-to-clarity transitions that eliminate aliasing at its source. To further enhance detail generation, we introduce a High-Frequency Residual Teacher that learns from both clean structure and alias residuals, distilling this knowledge to a vanilla VAR deployment network for seamless inference. Extensive experiments on ImageNet demonstrate that FVAR substantially reduces aliasing artifacts, improves fine detail preservation, and enhances text readability, achieving superior performance with perfect compatibility to existing VAR frameworks.

Dataset Distillation (DD) aims to compress large-scale datasets into a small number of condensed Images Per Class (IPC), enabling efficient network training. Previous coreset selection and synthetic-based DD methods achieve reasonable performance. However, our in-depth investigation reveals that existing methods share a common issue: pattern imbalance. Specifically, they either overemphasize class-general patterns representing the majority of each class or focus on fewer marginal patterns critical for model generalization. To address this issue, we propose a novel framework, Balanced Patterns Selection (BPS). Unlike prior methods that assume each class forms a single cluster, BPS models the multiple visual pattern distribution within each class via a hierarchical semantic structure inherent to the dataset. It then selects two complementary subsets in a balanced manner from the center (class-general patterns) and the margins (marginal patterns) of each pattern, producing a pattern-balanced coreset. Theoretically, we prove that the BPS-selected coreset aligns with the original dataset in both information coverage and performance. Moreover, its model-agnostic selection nature ensures cross-architecture generalization, while the Optimize-Once-for-All-IPCs property guarantees efficiency. Extensive experiments on four benchmarks demonstrate that BPS significantly outperforms existing state-of-the-art methods. The source code is available at: https://github.com/BeCarefulOfYournaoke/BPS.

We present FlashLips, a two-stage, mask-free lip-sync system that decouples lips control from rendering and achieves real-time performance, with our U-Net variant running at over 100 FPS on a single GPU, while matching the visual quality of larger state-of-the-art models. Stage 1 is a compact, one-step latent-space editor that reconstructs an image using a reference identity, a masked target frame, and a low-dimensional lips-pose vector, trained purely with reconstruction losses - no GANs or diffusion. To remove explicit masks at inference, we use self-supervision via mouth-altered target variants as pseudo ground truth, teaching the network to localize lip edits while preserving the rest. Stage 2 is an audio-to-pose transformer trained with a flow-matching objective to predict lips-pose vectors from speech. Together, these stages form a simple and stable pipeline that combines deterministic reconstruction with robust audio control, delivering high perceptual quality and faster-than-real-time speed.

Tremendous progress in visual scene generation now turns a single image into an explorable 3D world, yet immersion remains incomplete without sound. We introduce Image2AVScene, the task of generating a 3D audio-visual scene from a single image, and present SonoWorld, the first framework to tackle this challenge. From one image, our pipeline outpaints a 360deg panorama, lifts it into a navigable 3D scene, places language-guided sound anchors, and renders ambisonics for point, areal, and ambient sources, yielding spatial audio aligned with scene geometry and semantics. Quantitative evaluations on a newly curated real-world dataset and a controlled user study confirm the effectiveness of our approach. Beyond free-viewpoint audio-visual rendering, we also demonstrate applications to one-shot acoustic learning and audio-visual spatial source separation. Project website: https://humathe.github.io/sonoworld/

Temporal information is crucial for visual tracking, but existing multi-frame trackers are vulnerable to model drift caused by naively aggregating noisy historical predictions. In this paper, we introduce DTPTrack, a lightweight and generalizable module designed to be seamlessly integrated into existing trackers to suppress drift. Our framework consists of two core components: (1) a Temporal Reliability Calibrator (TRC) mechanism that learns to assign a per-frame reliability score to historical states, filtering out noise while anchoring on the ground-truth template; and (2) a Temporal Guidance Synthesizer (TGS) module that synthesizes this calibrated history into a compact set of dynamic temporal priors to provide predictive guidance. To demonstrate its versatility, we integrate DTPTrack into three diverse tracking architectures--OSTrack, ODTrack, and LoRAT--and show consistent, significant performance gains across all baselines. Our best-performing model, built upon an extended LoRATv2 backbone, sets a new state-of-the-art on several benchmarks, achieving a 77.5% Success rate on LaSOT and an 80.3% AO on GOT-10k. The source code is available at https://github.com/NorahGreen/DTPTrack.

Large vision-language models (LVLMs) are highly vulnerable to visual corruptions, substantially compromising their reliability and limiting real-world deployment. Prior work has attributed this degradation primarily to insufficient visual grounding and overreliance on language priors. However, these explanations often overlook the heterogeneous nature of corruptions, which perturb model perception in fundamentally different ways. We revisit this problem from a corruption-centric perspective and show that diverse corruptions can be organized along two complementary perceptual dimensions--shape and texture--which induce distinct failure modes. To address them, we propose Shape-Texture Dual-Path Contrastive Decoding (ST-CD), a training-free inference framework that constructs complementary contrastive views to diagnose and correct shape- and texture-induced biases through adaptive fusion. Experiments across multiple LVLMs and robustness benchmarks demonstrate that ST-CD consistently improves robustness under heterogeneous corruptions, suggesting that leveraging the complementarity between shape and texture provides a general and effective principle for building robust multimodal models.

Contrastive Language-Image Pre-training (CLIP) offers a new paradigm for Weakly Supervised Semantic Segmentation (WSSS) by generating Class Activation Maps (CAMs) from text-image alignment. Existing methods primarily rely on hand-crafted templates or general attribute descriptions generated by a large language model to construct text prototypes for querying visual features. However, these strategies faces two major limitations: the inherent modality gap in CLIP prevents text prototypes achieving tight alignment with visual features; and their static text prototypes cannot adaptively respond to target instances that exhibit diverse visual attributes. To address these challenges, our key insight is to directly construct instance-specific visual description prototype as query, thereby bypassing the suboptimal static text description optimization. To this end, we propose the Visual Description Assembly (VDA) framework. It employs a probabilistic model to map complex CLIP visual features into a structured latent space. This latent space allows us to explicitly disentangle and aggregate varied visual attributes, and then dynamically assemble them into instance-specific visual prototypes. Furthermore, to enhance the robustness of this prototype, we adaptively incorporate the semantically stable text prototype into it as the final query for generating superior CAMs. Experimental results show our method outperforms existing baselines, achieving state-of-the-art performance on WSSS benchmarks.

Streaming dense video captioning requires real-time processing of continuous visual input while determining precisely when and what to caption. Current approaches primarily focus on designing complex external memory mechanisms, failing to leverage Large Multimodal Models' (LMMs) inherent long-context capabilities. Moreover, existing methods employing threshold-based caption triggering face a severe Threshold-Gated Discrepancy (TGD) problem, a training-inference mismatch arising from data imbalance, where models predominantly predict silence tokens, requiring thresholds that vary drastically across videos with extremely narrow effective ranges. We introduce Takusen, an asynchronous temporal modeling two-agent framework comprising a Small Multimodal Model (SMM) as an Oracle agent and an LMM as a Listener agent. The Oracle agent processes sparse video inputs at an accelerated rate to detect event boundaries, while the Listener agent processes dense inputs to generate accurate captions when prompted by the Oracle's signals. This architecture eliminates threshold dependencies by fundamentally changing how silence/generation decisions are made, resolving the TGD problem. To enhance robustness against boundary prediction instabilities, we integrate uniformly distributed fixed decoding points with Oracle-predicted boundaries. Experiments on ActivityNet Captions and YouCook2 datasets demonstrate that Takusen achieves state-of-the-art performance with a simpler and more efficient design that balances temporal sensitivity with descriptive accuracy.

We introduce the Visual Personalization Turing Test (VPTT), a new paradigm for evaluating contextual visual personalization based on perceptual indistinguishability, rather than identity replication. A model passes the VPTT if its output (image, video, 3D asset, etc.) is indistinguishable to a human or calibrated VLM judge from content a given person might plausibly create or share. To operationalize VPTT, we present the VPTT Framework, integrating a 10k-persona benchmark (VPTT-Bench), a visual retrieval-augmented generator (VPRAG), and the VPTT Score, a text-only metric calibrated against human and VLM judgments. We show high correlation across human, VLM, and VPTT evaluations, validating the VPTT Score as a reliable perceptual proxy. Experiments demonstrate that VPRAG achieves the best alignment-originality balance, offering a scalable and privacy-safe foundation for personalized generative AI.

Generalized Category Discovery (GCD) aims to identify both known and novel classes from unlabeled data with the aid of labeled examples. While promising, most existing GCD methods rely on simultaneous access to labeled and unlabeled datasets--an assumption often impractical in real-world deployments. Continual Category Discovery (CCD) relaxes this requirement by adapting a pre-trained model to streaming unlabeled data, yet it typically assumes domain-consistent data distributions. This places a strong limitation on its applicability. In this work, we study Open Continual Category Discovery (OCCD), where the model must robustly discover previously unseen concepts from real-world data streams that may originate from heterogeneous and shifting domains. To address this, we propose an adaptive framework built on three key ideas. First, we propose a weight-aware separation module, which leverages partial unbalanced optimal transport for instance probability modeling and employs binary response spectrum quantization to generate cues for distinguishing known and unknown categories, enabling automatic sample separation. Second, for known categories, we introduce a cross-domain semantic alignment module that incorporates adversarial learning to perform adaptive prototype matching, thereby enhancing robustness against domain shifts. Finally, for unknown categories, we design a category topology consistency constraint that preserves semantic relationships between known and novel classes during distribution shifts. Experiments show our approach excels at discovering new categories while maintaining strong performance on known ones in evolving domains.

The limited understanding capacity of the visual encoder in Contrastive Language-Image Pre-training (CLIP) has become a key bottleneck for downstream performance. This capacity includes both Discriminative Ability (D-Ability), which reflects class separability, and Detail Perceptual Ability (P-Ability), which focuses on fine-grained visual cues. Recent solutions use diffusion models to enhance representations by conditioning image reconstruction on CLIP visual tokens. We argue that such paradigms may compromise D-Ability and therefore fail to effectively address CLIP's representation limitations. To address this, we integrate contrastive signals into diffusion-based reconstruction to pursue more comprehensive visual representations. We begin with a straightforward design that augments the diffusion process with contrastive learning on input images. However, empirical results show that the naive combination suffers from gradient conflict and yields suboptimal performance. To balance the optimization, we introduce the Diffusion Contrastive Reconstruction (DCR), which unifies the learning objective. The key idea is to inject contrastive signals derived from each reconstructed image, rather than from the original input, into the diffusion process. Our theoretical analysis shows that the DCR loss can jointly optimize D-Ability and P-Ability. Extensive experiments across various benchmarks and multi-modal large language models validate the effectiveness of our method. The code is available at https://github.com/boyuh/DCR.

Vision-Language Models (VLMs) have achieved remarkable progress across tasks such as visual question answering and image captioning. Yet, the extent to which these models perform visual reasoning as opposed to relying on linguistic priors remains unclear. To address this, we introduce VisRes Bench, a benchmark designed to study visual reasoning in naturalistic settings without contextual language supervision. Analyzing model behavior across three levels of complexity, we uncover clear limitations in perceptual and relational visual reasoning capacities. VisRes isolates distinct reasoning abilities across its levels. Level 1 probes perceptual completion and global image matching under perturbations such as blur, texture changes, occlusion, and rotation; Level 2 tests rule-based inference over a single attribute (e.g., color, count, orientation); and Level 3 targets compositional reasoning that requires integrating multiple visual attributes. Across more than 19,000 controlled task images, we find that state-of-the-art VLMs perform near random under subtle perceptual perturbations, revealing limited abstraction beyond pattern recognition. We conclude by discussing how VisRes provides a unified framework for advancing abstract visual reasoning in multimodal research (https://visres-bench.github.io).

Data matters. In computer vision, data (or pixels) are the primary source of information containing signals that span from low-level attributes to high-level concepts. At scale, the success of modern vision systems has been closely tied to how data is curated for semantic understanding (e.g., ImageNet). Recent trends in spatial intelligence and physical world understanding further highlight the importance of real-world signals that preserve spatial structure, beyond purely semantic signals. This motivates a shift toward curating data that better captures spatially grounded information across diverse environments. In this work, we demonstrate that training on 2B web-crawled images with a self-curation strategy on masked autoencoder (MAE) can learn strong representations for dense prediction tasks, while remaining simple, stable, and efficient. Our model, codenamed "Pixio", is an enhanced masked autoencoder (MAE) with more challenging pre-training tasks and more capable architectures. Pixio yields dense representations achieving promising results across a wide range of dense prediction tasks in the wild, including monocular depth estimation (e.g., Depth Anything), feed-forward 3D reconstruction (i.e., MapAnything), and visual segmentation (e.g., SAM). Our results suggest that data curation can significantly contribute to dense representation learning.

Recent Vision-Language-Action (VLA) models map visual-textual inputs to robotic actions via end-to-end architectures, yet this approach entangles visual understanding with task-specific actions. This leads to an exhaustive collection of full operational sequences and parameter redundancy across tasks, while generic third-person camera setups require fine-tuning for different hardware due to implicit hand-eye assumptions. We argue that decoupling how robots see from how robots act is a missing primitive in VLA systems. We present EgoRoC, a plug-and-play egocentric alignment head that precedes any task policy and exposes only a thin 6-DoF pose interface. EgoRoC establishes task-agnostic viewpoint consistency from a wrist-mounted (first-person) camera and then alternates alignment with manipulation, while a diffusion-based online hand-eye module corrects the action in the end-effector frame for hardware-agnostic deployment. Trained once from static wrist-target image pairs with relative poses, rather than full manipulation trajectories, EgoRoC leaves downstream VLAs unchanged. By turning egocentric alignment into a reusable capability, EgoRoC reduces training redundancy, strengthens zero-shot cross-scene transfer, and scales across VLA backbones without manual calibration. Across simulation and real settings, attaching EgoRoC consistently boosts success rates, especially on long-horizon and out-of-distribution tasks, and improves data efficiency during fine-tuning.