CVPR 2026 Papers

For an unlabeled dataset containing both known and unknown categories, Generalized Category Discovery (GCD) aims to classify known categories while discovering unknown ones. Although existing methods achieve promising results on coarse-grained datasets, they struggle in fine-grained scenarios. We observe that attention artifacts where attention maps exhibit abnormally high responses on a few tokens--significantly interfere with fine-grained GCD by causing the model to overemphasize global semantics and overlook discriminative local cues. To address this issue, we propose the Token-Aware Refinement (TAR) framework, a plug-and-play module that mitigates attention artifacts and enhances local information. Unlike conventional approaches that rely solely on the first token for classification, TAR leverages the entire token sequence to better capture fine-grained details. Extensive experiments demonstrate that TAR consistently improves performance across multiple benchmarks. Our code is available at https://github.com/VectorYangYiStar/TAR.

Recent progress in Multimodal Large Language Models (MLLMs) has enabled mobile GUI agents capable of visual perception, cross-modal reasoning, and interactive control. However, existing benchmarks are largely English-centric and fail to capture the linguistic and interaction characteristics of the Chinese mobile ecosystem. They also focus on isolated skills such as GUI grounding or offline agent, lacking a unified and fine-grained framework to assess the full capability chain from perception to execution. To address this gap, we introduce GUI-CEval, the first comprehensive benchmark for Chinese mobile GUI agents, built entirely on physical device environments. GUI-CEval spans 201 mainstream apps across four device types and adopts a two-level structure that evaluates both atomic abilities and realistic application-level performance along five dimensions: perception, planning, reflection, execution, and evaluation. All data are collected and verified through multi-stage manual processes to ensure authenticity and reproducibility. Extensive experiments on 20 representative MLLMs and multi-agent systems show that while models such as Qwen2.5-VL and UI-TARS perform competitively, most MLLMs still exhibit clear weaknesses in reflective decision-making and post-action self-evaluation, limiting their reliability in real-world interactions. We hope GUI-CEval provides a comprehensive benchmark to guide capability diagnosis and advance the development of Chinese mobile GUI agents.

Machine unlearning aims to erase requested data from trained models without full retraining. For Reasoning Multimodal Large Language Models (RMLLMs), this is uniquely challenging: intermediate chain-of-thought steps can still leak sensitive information even when final answers are forgotten, and overly aggressive interventions easily damage general reasoning ability. Yet no benchmark jointly evaluates how well unlearning methods suppress reasoning-level leakage while preserving reasoning competence. We address this gap with RMLLMU-Bench, the first benchmark for RMLLM unlearning that extends standard forgetting metrics with dedicated measures of reasoning leakage and reasoning retention. A systematic evaluation on RMLLMU-Bench reveals that existing unlearning methods for MLLMs and Large (Language) Reasoning Models (LRMs) either leave substantial leakage in the reasoning process or severely degrade reasoning performance. To address these gaps, we propose R-MUSE (\underline R easoning-preserving \underline M LLM \underline U nlearning via \underline S ubspace guidance and Adaptive St \underline E ering), a training-free and inference-time intervention framework that steers internal representations to forget both answers and reasoning traces while explicitly preserving general reasoning. Experiments on RMLLMU-Bench demonstrate that R-MUSE achieves a substantially better balance between effective forgetting and reasoning retention.

High-quality segmentations are critical in vision tasks where boundary accuracy is important (e.g., medical diagnostics, quality control, etc.). Recently, promptable vision models have emerged as effective backbones for segmentation refinement frameworks. However, their performance not only hinges on prompt quality, they also must overcome noisy input masks and semantically ambiguous outputs from promptable models. Existing prompt-based refiners rely on fixed prompt rules, making them brittle to changing failure modes and new tasks or domains. We propose PromptMoE, a model-agnostic MoE-driven prompting refiner effective in segmentation refinement across tasks and domains. PromptMoE features three collaborative modules to refine an initial mask: our MoE-based Image-Informed Prompting framework (IIP) takes an image and coarse mask and produces a set of expert score maps to guide prompt generation, the Dynamic Expert Selector (DES) activates only the most relevant experts and fuses their maps to avoid dense evaluation and signal dilution, and the Prompt-Placement Explorer (PPE) explores the fused guidance map to place high-confidence spatially diverse point prompts. Across five benchmark datasets (BIG, VOC, DAVIS585, ECSSD, MSRA-B), PromptMoE achieves statistically significant gains over SOTA methods CascadePSP, SegRefiner, and SAMRefiner on semantic, instance, and salient tasks, with mean improvements of +6.24 IoU / +8.99 BIoU.

The recent surge in video generation has shown the growing demand for high-quality video synthesis using large vision models. Existing video generation models are predominantly based on the video diffusion transformer (vDiT), however, they suffer from substantial inference delay due to self-attention. While prior studies have focused on reducing redundant computations in self-attention, they often overlook the inherent spatio-temporal correlations in video streams and directly leverage sparsity patterns from large language models to reduce attention computations. In this work, we take a principled approach to accelerate self-attention in vDiTs by leveraging the spatio-temporal correlations in the latent space. We show that the attention patterns within vDiT are primarily due to the dominant spatial and temporal correlations at the token channel level. Based on this insight, we propose a lightweight and adaptive reuse strategy that approximates attention computations by reusing partial attention scores of spatially or temporally correlated tokens along individual channels. We demonstrate that our method achieves significantly higher computational savings (85%) compared to state-of-the-art techniques over 4 vDiTs, while preserving almost identical video quality (<0.06% loss on VBench).

Creativity is a complex phenomenon. When it comes to representing and assessing creativity, treating it as a single undifferentiated quantity would appear naive and under-whelming. In this work, we learn the first type-specific creativity reward model, coined CREward, which spans three creativity "axes," geometry, material, and texture, to allow us to view creativity through the lens of the image formation pipeline. To build our reward model, we first conduct a human benchmark evaluation to capture human perception of creativity for each type across various creative images. We then analyze the correlation between human judgments and predictions by large vision-language models (LVLMs), confirming that LVLMs exhibit strong alignment with human perception. Building on this observation, we collect LVLM-generated labels to train our CREward model that is applicable to both evaluation and generation of creative images. We explore three applications of CREward: creativity assessment, explainable creativity, and creative sample acquisition for both human design inspiration and guiding creative generation through low-rank adaptation.

Online reconstruction of dynamic scenes aims to learn from streaming multi-view inputs under low-latency constraints. The fast training and real-time rendering capabilities of 3D Gaussian Splatting have made on-the-fly reconstruction practically feasible, enabling online 4D reconstruction. However, existing online approaches, despite their efficiency and visual quality, fail to learn per-Gaussian motion that reflects true scene dynamics. Without explicit motion cues, appearance and motion are optimized solely under photometric loss, causing per-Gaussian motion to chase pixel residuals rather than true 3D motion. To address this, we propose MoRGS, an efficient online per-Gaussian motion reasoning framework that treats Gaussian movement as a core modeling object. Specifically, we efficiently leverage optical flow on a sparse set of key views as a lightweight motion cue to guide per-Gaussian motion toward the scene's true dynamics. To compensate for the sparsity and view-dependence of flow, we learn a per-Gaussian motion offset field that reconciles discrepancies between projected 3D motion and observed flow across views and time. In addition, we introduce a per-Gaussian motion confidence that separates dynamic from static Gaussians and weights Gaussian attribute residual updates, thereby suppressing redundant motion in static regions for better temporal consistency and accelerating the modeling of large motions. Extensive experiments demonstrate that MoRGS achieves state-of-the-art reconstruction quality and motion fidelity among online methods, while maintaining streamable performance.

Recent advances in image-to-video (I2V) generation have achieved remarkable progress in synthesizing high-quality, temporally coherent videos from static images. Among all the applications of I2V, human-centric video generation includes a large portion. However, existing I2V models encounter difficulties in maintaining identity consistency between the input human image and the generated video, especially when the person in the video exhibits significant expression changes and movements. This issue becomes critical when the human face occupies merely a small fraction of the image. Since humans are highly sensitive to identity variations, this poses a critical yet under-explored challenge in I2V generation. In this paper, we propose Identity-Preserving Reward-guided Optimization (IPRO), a novel video diffusion framework based on reinforcement learning to enhance identity preservation. Instead of introducing auxiliary modules or altering model architectures, our approach introduces a direct and effective tuning algorithm that optimizes diffusion models using a face identity scorer. To improve performance and accelerate convergence, our method backpropagates the reward signal through the last steps of the sampling chain, enabling richer gradient feedback. We also propose a novel facial scoring mechanism that treats faces in ground-truth videos as facial feature pools, providing multi-angle facial information to enhance generalization. A KL-divergence regularization is further incorporated to stabilize training and prevent overfitting to the reward signal. Extensive experiments on Wan 2.2 I2V model and our in-house I2V model demonstrate the effectiveness of our method. Our project is available at https://dof-gaussian.github.io/.

Direct Preference Optimization has achieved remarkable success in aligning diffusion models with human feedback. However, existing methods heavily rely on image-level preferences, which suffer from sparse rewards in the spatial dimension. This creates a fundamental misalignment: while an image may be globally preferred, it can contain locally inferior instances. Applying the same positive preference to these areas thus unfairly credits distracting regions while penalizing informative ones, leading to suboptimal performance and inefficient learning. To resolve this issue, we propose IAPO, an Instance-Aware Preference Optimization that introduces instance-level credit assignment to advance alignment from image-level to instance-level. We first construct a high-quality instance-level preference dataset by automatically identifying and relabeling corresponding instances in image pairs using vision-language models and object detection models. Leveraging this fine-grained dataset, we design a novel instance alignment loss with a dynamic reweighting mask that modulates instance-level loss within annotated bounding boxes, suppressing distractors to enforce fine-grained human preference alignment. Extensive experiments demonstrate that our method not only achieves state-of-the-art performance in multiple benchmarks but also attains higher training efficiency due to fine-grained instance-level preference labels.

In this work, we present a panoramic metric depth foundation model that generalizes across diverse scene distances. We explore a data-in-the-loop paradigm from the view of both data construction and framework design. We collect a large-scale dataset by combining public datasets, high-quality synthetic data from our UE5 simulator and text-to-image models, and real panoramic images from the web. To reduce domain gaps between indoor/outdoor and synthetic/real data, we introduce a three-stage pseudo-label curation pipeline to generate reliable ground truth for unlabeled images. For the model, we adopt DINOv3-Large as the backbone for its strong pre-trained generalization, and introduce a plug-and-play range mask head, sharpness-centric optimization, and geometry-centric optimization to improve robustness to varying distances and enforce geometric consistency across views. Experiments on multiple benchmarks (e.g., Stanford2D3D, Matterport3D, and Deep360) demonstrate strong performance and zero-shot generalization, with particularly robust and stable metric predictions in diverse real-world scenes. The project page can be found at: https://insta360-research-team.github.io/DAP_website/

Multimodal large language models (MLLMs) have achieved remarkable progress on various vision-language tasks, yet their visual perception remains limited. Humans, in comparison, perceive complex scenes efficiently by dynamically scanning and focusing on salient regions in a sequential "blink-like" process. Motivated by this strategy, we first investigate whether MLLMs exhibit similar behavior. Our pilot analysis reveals that MLLMs naturally attend to different visual regions across layers and that selectively allocating more computation to salient tokens can enhance visual perception. Building on this insight, we propose Blink, a dynamic visual token resolution framework that emulates the human-inspired process within a single forward pass. Specifically, Blink includes two modules: saliency-guided scanning and dynamic token resolution. It first estimates the saliency of visual tokens in each layer based on the attention map, and extends important tokens through a plug-and-play token super-resolution (TokenSR) module. In the next layer, it drops the extended tokens when they lose focus. This dynamic mechanism balances broad exploration and fine-grained focus, thereby enhancing visual perception adaptively and efficiently. Extensive experiments validate Blink, demonstrating its effectiveness in enhancing visual perception and multimodal understanding.

Reasoning segmentation enables open-set object segmentation via implicit text queries, therefore serving as a foundation for embodied agents that should operate autonomously in real-world environments. However, existing methods for reasoning segmentation require multimodal large language models with billions of parameters that exceed the computational capabilities of edge devices that typically deploy the embodied AI systems. Distillation offers a pathway to compress these models while preserving their capabilities. Yet, existing distillation approaches fail to transfer the multi-step reasoning capabilities that reasoning segmentation demands, as they focus on matching output predictions and intermediate features rather than preserving reasoning chains. The emerging paradigm of reasoning over digital twin representations presents an opportunity for more effective distillation by re-framing the problem. Consequently, we propose FastReasonSeg, which employs digital twin representations that decouple perception from reasoning to enable more effective distillation. Our distillation scheme first relies on supervised fine-tuning on teacher-generated reasoning chains. Then it is followed by reinforcement fine-tuning with joint rewards evaluating both segmentation accuracy and reasoning quality alignment. Experiments on two video (JiTBench, RVTBench) and two image benchmarks (ReasonSeg, LLM-Seg40K) demonstrate that our FastReasonSeg achieves state-of-the-art reasoning segmentation performance. Moreover, the distilled 0.6B variant outperforms models with 20 times more parameters while achieving 7.79 FPS throughput with only 2.1GB memory consumption. This efficiency enables deployment in resource-constrained environments to enable real-time reasoning segmentation.

While iterative stereo matching achieves high accuracy, its dependence on Recurrent Neural Networks (RNN) hinders edge deployment, a challenge underexplored in existing researches. We analyze iterative refinement and reveal that disparity updates are spatially sparse and temporally redundant. First, we introduce a progressive iteration pruning strategy that suppresses redundant update steps, effectively collapsing the recursive computation into a near-single-pass inference. Second, we propose a collaborative monocular prior transfer framework that implicitly embeds depth priors without requiring a dedicated monocular encoder, thereby eliminating its associated computational burden. Third, we develop FlashGRU, a hardware-aware RNN operator leveraging structured sparsity and I/O-conscious design, achieving a 7.28xspeedup, 76.6% memory peak reduction and 80.9% global memory requests reduction over natvie ConvGRUs under 2K resolution. Our PipStereo enables real-time, high-fidelity stereo matching on edge hardware: it processes 320x640 frames in just 75ms on an NVIDIA Jetson Orin NX (FP16) and 19ms on RTX 4090, matching the accuracy of large iterative based models, and our generalization ability and accuracy far exceeds that of existing real-time methods.

Cross-domain few-shot image interpretation (CD-FSII) has been significantly advanced by fine-tuning pre-trained visual feature models using limited labeled samples in target domains. However, profound cross-domain distribution discrepancies, along with inherent conflicts between extensive object visual appearance variations and limited annotations, trap those existing pure visual feature representations into some non-transferable short-cut patterns, thus degrading their cross-domain generalization capacity. To mitigate this problem, we present a simple yet effective cross-modal visual feature enhancement framework which primarily contributes in the following three aspects. 1) We make the first attempt to introduce linguistic descriptions of image attributes to regulate the pre-trained visual feature model for specific target image adaptation. Specifically, image-level attributes (e.g., object appearance in individual images) and domain-level attributes (e.g., overall style and background characteristics of the dataset) are extracted using a pre-trained image captioning model and a large language model (LLM), respectively, to construct comprehensive linguistic characterizations. 2) A lightweight residual cross-attention scheme is developed to seamlessly embed linguistic descriptions of image attributes into visual feature representations, thereby compensating for the limitations of purely visual cues in capturing cross-domain transferable high-level semantic characteristics. 3) The proposed framework is task-agnostic and can be seamlessly integrated with off-the-shelf pre-trained visual feature models. It demonstrates superior generalization performance compared to several state-of-the-art methods across multiple CD-FSII benchmarks, including image classification, semantic segmentation, and object detection. We will release all code and data to facilitate further research.

Multimodal large language models (MLLMs) that "think with images" can interactively use tools to reason about visual inputs, but current approaches often rely on a narrow set of tools with limited real-world necessity and scalability. In this work, we first reveal a critical and previously overlooked weakness: even state-of-the-art MLLMs are surprisingly brittle, showing significant performance degradation on images with simple orientation changes, underscoring the need for more robust tool-based reasoning. To address this, we propose CodeVision, a flexible and scalable "code-as-tool" framework where the model generates code as a universal interface to invoke any image operation, moving beyond fixed tool registries. We train our model using a two-stage methodology, beginning with Supervised Fine-Tuning (SFT) on a high-quality dataset curated for complex, multi-turn tool composition and error recovery, followed by Reinforcement Learning (RL) with a dense process reward function to encourage strategic and efficient tool use. To facilitate this research, we construct new SFT and RL datasets and introduce a challenging new benchmark suite designed to rigorously evaluate robustness to orientation changes and multi-tool reasoning. Experiments on Qwen2.5-VL and Qwen3-VL series show that our approach significantly improves model performance and fosters emergent capabilities such as flexible tool composition, efficient chained execution, and robust error recovery from runtime feedback. Code is available at https://github.com/ByteDance-BandAI/CodeVision.

Topology reasoning is crucial for autonomous driving. Current methods primarily focus on instance-level learning for centerline detection, followed by a sequential module for topology reasoning that relies on simplified MLP layers. Moreover, these approaches often neglect the importance of point-to-instance (P2I) relationships in topology reasoning. To address these limitations, we present TopoHR (Topological Hierarchical Representation), a novel end-to-end framework that establishes cyclic interaction between centerline detection and topology reasoning, allowing them to iteratively enhance each other. Specifically, we introduce a hierarchical centerline representation including point queries, instance queries, and semantic representations. These multi-level features are seamlessly integrated and fused within a hierarchical centerline decoder. Furthermore, we design a hierarchical topology reasoning module that captures both fine-grained P2I relationships and global instance-to-instance (I2I) connections within a unified architecture. With these novel components, TopoHR ensures accurate and robust topology reasoning. On the OpenLane-V2 benchmark, TopoHR refreshes state-of-the-art performance with significant improvements. Notably, compared with previous best results, TopoHR achieves +3.8 in \mathrm DET _ \text l , +5.4 in \mathrm TOP _ \text ll on subset_A and +11.0 in \mathrm DET _ \text l , +7.9 in \mathrm TOP _ \text ll on subset_B, validating the effectiveness of the proposed components. The code will be shared publicly upon paper acceptance.

With diffusion models demonstrating superior capabilities in image editing, more users now rely on online servers for content manipulation via textual prompts rather than traditional offline tools. Despite servers attempting to prevent the proliferation of maliciously edited content via active defense like watermarking, this approach is not conducive to passive detection by third-party forensics. To address this limitation, we propose a plug-and-play controllable denoising termed Forensic-Friendly Image Manipulation (FFIM), which simultaneously satisfies user editing requirements while facilitating forensic analysis. Specifically, FFIM comprises three phases: Controllable Projection, Implicit Detection, and Explicit Guidance. Phase I enforces orthogonality between the variance of random noise and image features to ensure clear demarcation between the edited and unedited regions. Phase II implicitly evaluates whether this demarcation meets detection requirements; if not, Phase III explicitly introduces a surrogate detection model and adversarially adjusts the random noise to maximize the feature discrepancy between these regions. Experiments across four datasets demonstrate the superiority of FFIM over baseline methods, achieving up to +6.6% F1 in pixel-level localization and +27.3% AUC in image-level detection. Importantly, these forensic gains are attained without compromising visual quality, as evidenced by comparable manipulation in both subjective user studies and objective quality assessments. We envision that the proposed method will be widely adopted by generative AI service providers, enabling more comprehensive information authenticity from a passive defensive standpoint.

Classifying fine-grained visual concepts under open-world settings, i.e., without a predefined label set, demands models to be both accurate and specific. Recent reasoning Large Multimodal Models (LMMs) exhibit strong visual understanding capability but tend to produce overly generic predictions when performing fine-grained image classification. Our preliminary analysis reveals that models do possess the intrinsic fine-grained domain knowledge. However, promoting more specific predictions (specificity) without compromising correct ones (correctness) remains a non-trivial and understudied challenge. In this work, we investigate how to steer reasoning LMMs toward predictions that are both correct and specific. We propose a novel specificity-aware reinforcement learning framework, SpeciaRL, to fine-tune reasoning LMMs on fine-grained image classification under the open-world setting. SpeciaRL introduces a dynamic, verifier-based reward signal anchored to the best predictions within online rollouts, promoting specificity while respecting the model's capabilities to prevent incorrect predictions. Our out-of-domain experiments show that SpeciaRL delivers the best trade-off between correctness and specificity across extensive fine-grained benchmarks, surpassing existing methods and advancing open-world fine-grained image classification. Code and model are publicly available at https://github.com/s-angheben/SpeciaRL.

Maritime multimodal vision faces significant challenges due to the complexity and variability of oceanic weather and environmental conditions. While modern vessels are commonly equipped with visible and infrared imaging systems, the complementary nature of these modalities fundamentally depends on accurate cross-modal registration. However, the absence of paired visible-infrared datasets that realistically capture diverse maritime scenarios has severely hindered progress in this field. To overcome this limitation, we present MMVIP, the first large-scale visible-infrared maritime vision dataset covering a wide spectrum of weather conditions and sea states. The dataset contains 128,100 images and 50 video sequences with precise spatial-temporal alignment. Comprehensive evaluations across image registration, fusion, maritime object detection, and cross-modal image translation tasks demonstrate the dataset's effectiveness and challenge. Furthermore, MMVIP establishes a new benchmark for advancing multimodal maritime perception. The dataset and corresponding benchmarks are publicly available at: https://github.com/yyppptjr/MMVIP.