ECCV 2026 Papers

Prior studies have demonstrated that diffusion classifiers achieverobust zero-shot classification performance. However, their effectivenessis strongly tied to the pretraining data distribution: they perform well inmajority, high-density regions of the data manifold, but are significantlyless accurate in minority, low-density regions. Although prior works onminority sampling have focused on generating more minority-like im-ages, what minority sampling fundamentally enables beyond generationremains underexplored. In this paper, we reveal a direct relationshipbetween minority sampling in generation and the perception capabil-ity of diffusion classifiers. Specifically, we show that enhancing minoritysampling broadens the coverage of underrepresented regions on the datamanifold, thereby improving diffusion-based recognition. To exploit thisconnection, we propose Self-Improving Diffusion Classifiers with Minor-ity Preference Optimization (MiPO), which fine-tunes a pretrained dif-fusion model using minority preference rewards. Using only arbitrarycaption data, MiPO generates candidate samples, rewards those thatbetter cover minority regions, and optimizes the model with LoRA andGroup Relative Policy Optimization, without additional image data, ex-ternal foundation models, or external reward models. This enables stable,prompt-adaptive minority sampling and translates low-density genera-tive coverage into improved zero-shot diffusion classification. To sum up,we show that diffusion classifier perception is biased toward majorityregions, demonstrate that this bias can be alleviated through minoritypreference optimization, and evaluate MiPO on five standard datasets.

Existing visual privacy benchmarks largely treat privacy asa binary property, labeling images as private or non-private based onvisible sensitive content. We argue that privacy is fundamentally compo-sitional. Attributes that are benign in isolation may combine to producesevere privacy violations. We introduce the Compositional Privacy RiskTaxonomy (CPRT), a regulation-aware framework that organizes visualattributes according to standalone identifiability and compositional harmpotential. CPRT defines four graded severity levels and is paired with aninterpretable scoring function that assigns continuous privacy severityscores. We further construct a taxonomy-aligned dataset of 6.7K imagesand derive compositional risk scores. By evaluating frontier and open-weight VLMs we find that frontier models align well with compositionalseverity when provided structured guidance, but systematically underes-timate composition-driven risks. Smaller models struggle to internalizegraded privacy reasoning. To bridge this gap, we introduce a deploy-able 8B SFT model that closely matches frontier-level performance oncompositional privacy assessment. Our dataset and models are publiclyavailable at: https://huggingface.co/collections/timtsapras23/cprt.

Semantic segmentation models remain vulnerable to adversarial perturbations, which can degrade pixel-wise predictions despite being imperceptible. Existing adversarial training methods primarily rely on iterative prediction-driven attacks that push features across decision boundaries between in-distribution (ID) classes while remaining close to the ID manifold. This limits their ability to expose more severe representation failures. In this paper, we propose SegOOD, a novel adversarial attack framework that explicitly encourages out-of-distribution (OOD)- like deviations in latent feature space. SegOOD introduces a Feature Wasserstein Separation objective to increase distributional discrepancy between adversarial features and ID semantic prototypes, along with a Weighted K-Nearest-Neighbor Separation objective to enforce local feature-level deviation. These objectives are efficiently implemented using prototype aggregation and feature-aware superpixel clustering. Furthermore, we integrate SegOOD with conventional prediction-driven attacks during adversarial training, improving robustness against both existing ID-style and our OOD-like perturbations. Extensive experiments demonstrate state-of-the-art robustness while maintaining competitive clean performance and no additional inference cost.

Although existing 3D dance generation methods performwell in controlled scenarios, they often struggle to generalize in the wild.When conditioned on unseen music, existing methods often produce un-structured or physically implausible dance, largely due to limited music-to-dance data and restricted model capacity. This work aims to pushthe frontier of generalizable 3D dance generation by scaling up bothdata and model design. 1) On the data side, we develop a fully auto-mated pipeline that reconstructs high-fidelity 3D dance motions frommonocular videos. To eliminate the physical artifacts prevalent in exist-ing reconstruction methods, we introduce a Foot Restoration DiffusionModel (FRDM) guided by foot-contact and geometric constraints thatenforce physical plausibility while preserving kinematic smoothness andexpressiveness, resulting in a diverse, high-quality multimodal 3D dancedataset totaling 100.69 hours. 2) On model design, we propose Choreo-graphic LLaMA (ChoreoLLaMA), a scalable LLaMA-based architecture.To enhance robustness under unfamiliar music conditions, we integratea retrieval-augmented generation (RAG) module that injects referencedance as a prompt. Additionally, we design a slow/fast-cadence Mixture-of-Experts (MoE) module that enables ChoreoLLaMA to smoothly adaptmotion rhythms across varying music tempos. Extensive experimentsacross diverse dance genres show that our approach surpasses existingmethods in both qualitative and quantitative evaluations, marking a steptoward scalable, real-world 3D dance generation.

Although fully supervised methods have substantially ad-vanced segmentation in boundary areas of point clouds, effective weaklysupervised approaches remain scarce. This is primarily because limitedsupervision rarely reaches boundary regions, leaving them lacking reli-able supervision. We propose a novel 2D-assisted pseudo-label propaga-tion paradigm that does not rely on the model’s own predictions or anyexternal foundation models, yet is able to generate high-purity pseudo-labels. Compared with SAM-based 2D-3D projection, our pseudo-labelsare purer and more uniformly distributed. Even under 1 pt/obj settingon S3DIS, our initial offline propagation achieves >94.4% accuracy (≈93k pts per scene). We decomposed the pseudo-labels generation pro-cess from the main network, and applied a divide-and-conquer strategy:supervision from interior pseudo-labels serves to stabilize the representa-tion of core class regions, while boundary pseudo-labels are leveraged toenhance boundary robustness. This design reduces the confirmation biasinherent in classic online labeling and alleviates the lack of boundary su-pervision in existing weakly supervised models. Experiments show thatour method outperforms existing state-of-the-art methods. Our codeswill be released at: https://github.com/paul-swu/Bound3D.

Recent advances in diffusion models have shown impressiveperformance in controllable image generation and dense prediction tasks.However, existing approaches typically treat diffusion-based controllablegeneration and dense prediction as separate tasks, overlooking the po-tential benefits of jointly modeling the heterogeneous distributions. Inthis work, we introduce UniGP, a framework built upon MMDiT, whichunifies controllable generation and dense prediction through simple jointtraining, without the need for complex task-specific designs or losses,while preserving the backbone’s versatile priors. By learning controllablegeneration and prediction under different conditions, our model effectivelycaptures the joint distribution of image-geometry pairs. UniGP is capableof versatile controllable generation, dense prediction, and joint generation.Specifically, the proposed UniGP consists of DUGP and a unified datasettraining strategy. The former, following the principle of Occam’s razor,uses only a copied image branch of MMDiT to model dense distributionsbeyond RGB, while the latter integrates heterogeneous datasets into aunified training framework to jointly model generation and perceptiontasks. Extensive experiments demonstrate that our unified model sur-passes prior unified approaches and performs on par with specializedmethods. Furthermore, we demonstrate that multi-task joint training pro-vides complementary benefits: generative priors enrich perceptual details,while perceptual learning improves structural alignment in generation.Project page: guoqincode.github.io/UniGP.

RL (reinforcement learning) methods (e.g., GRPO) for mul-timodal LLM perception ability has attracted wide research interest ow-ing to its remarkable generalization ability. Nevertheless, existing rein-forcement learning methods still face the problem of low data quality,where data samples cannot elicit diverse responses from MLLMs, thus re-stricting the exploration scope for MLLM reinforcement learning. Somemethods attempt to mitigate this problem by imposing constraints onentropy, but none address it at its root. Therefore, to tackle this problem,this work proposes Syn-GRPO (Synthesis-GRPO), which employs anonline data generator to synthesize high-quality training data with di-verse responses in GRPO training. Specifically, Syn-GRPO consists oftwo components: (1) data server; (2) GRPO workflow. The data serversynthesizes new samples from existing ones using an image generationmodel, featuring a decoupled and asynchronous scheme to achieve highgeneration efficiency. The GRPO workflow provides the data server withthe new image descriptions, and it leverages a diversity reward to su-pervise the MLLM to predict image descriptions for synthesizing sam-ples with diverse responses. Experiment results across three visual per-ception tasks demonstrate that Syn-GRPO improves the data qualityby a large margin, achieving significant superior performance to exist-ing MLLM perception methods, and Syn-GRPO presents promising po-tential for scaling long-term self-evolving RL. Our code is available athttps: // github. com/ hqhQAQ/ Syn-GRPO .

Autonomous agents navigating human society must masterboth production activities and social interactions, yet existing bench-marks rarely evaluate these skills simultaneously. To bridge this gap,we introduce StarDojo, a novel benchmark based on Stardew Valley,designed to assess AI agents in open-ended production–living simulations.In StarDojo, agents are tasked to perform essential livelihood activitiessuch as farming and crafting, while simultaneously engaging in socialinteractions to establish relationships within a vibrant community. Star-Dojo features 1,000 meticulously curated tasks across five key domains:farming, crafting, exploration, combat, and social interactions. Addition-ally, we provide a compact subset of 100 representative tasks for efficientmodel evaluation. The benchmark offers a unified, user-friendly interfacethat eliminates the need for keyboard and mouse control, supports allmajor operating systems, and enables the parallel execution of multipleenvironment instances, making it particularly well-suited for evaluatingthe most capable foundation agents, powered by multimodal large lan-guage models (MLLMs). Extensive evaluations of state-of-the-art MLLMsagents demonstrate substantial limitations, with the best-performingmodel, GPT-4.1, achieving only a 12.7% success rate, primarily due tochallenges in visual understanding, multimodal reasoning and low-levelmanipulation. As a user-friendly environment and benchmark, StarDojoaims to facilitate further research towards robust, open-ended agents incomplex production-living environments.

Music-driven dance generation aims to produce human mo-tion that is both rhythmically synchronized and semantically consis-tent with music. While recent neural approaches have achieved impres-sive visual realism, they typically model motion as a continuous signaland neglect its compositional nature, making generated dances struc-turally incoherent and difficult to control. In this work, we introducea structure-aware framework that models choreography as a sequenceof atomic movements—semantically interpretable motion events thatserve as the building blocks of dance. To construct this atomic move-ment vocabulary, we first segment large-scale dance data and clusterthem into atomic movement groups. We then employ a large languagemodel to semantically relabel and refine the clusters, yielding a set ofinterpretable and reusable atomic movements. Based on these atomicmovement annotations, we design a two-stage generation framework thatmirrors the human choreography process. In the atomic movement plan-ning stage, the model predicts the type, duration, and timing of atomicmovements conditioned on the input music, forming a symbolic danceallocation. In the completion stage, a transition-aware generator syn-thesizes smooth and stylistically coherent motion conditioned on theplanned structure. Extensive experiments demonstrate that our methodproduces dances with significantly improved structural coherence, rhyth-mic alignment, and perceptual naturalness compared to existing base-lines, while providing enhanced interpretability and controllable edit-ing through explicit structural representation. The code is available at(A) Real music and choreography exhibits highly structural formation.Repeated patterns Repeated patternsSimilar or symmetrical atomic movements(B) Existing end-to-end works overlook such structural organization.Fig. 1: (A) There exists explicit repeated patterns in music and choreography. (B)Existing methods, usually in an end-to-end form, treat music-to-dance generation as asimple sequence-to-sequence task, neglecting the inner structure.

Video polyp segmentation (VPS) is important for automatedcolorectal cancer screening, yet it is a very challenging task due to severecamouflage in mucosal scenes, complex intestinal anatomy, and substan-tial frame-to-frame appearance variations. Although Segment AnythingModel 3 (SAM 3) is able to provide strong representations for segmen-tation tasks, directly applying it to endoscopic videos usually cannotachieve satisfactory performance owing to domain shifts, boundary am-biguities, and unstable temporal propagation. To address these chal-lenges, we propose FST-SAM3, a novel parameter-efficient adaptationframework built on SAM 3 for VPS. FST-SAM3 introduces two inno-vative yet complementary components. First, we propose a Structure-Aware Frequency Enhancement (SAFE) module to strengthen structure-and boundary-sensitive cues under endoscopic appearance. Second, wedevelop a Spatio-Temporal Token Refinement Framework (STRF) torefine temporal information at the token level for more stable cross-frame segmentation. We freeze the SAM 3 backbone and learn only asmall set of additional parameters, keeping the inference characteristicsclose to SAM 3 while improving segmentation performance. We con-ducted extensive experiments on two benchmarking datasets (SUN-SEGand CVC-612) under standardized fully automatic (0-point) and first-frame prompted (1-point at t=0) protocols. Results demonstrate that ourmodel achieves consistent improvements over state-of-the-art approaches,including both task-specific baselines and representative SAM-based ap-proaches. Code is available at https://github.com/GavonW/FST-SAM3.

Gaussian Splatting (GS) has emerged as a dominating tech-nique for image rendering and has quickly been adapted for the X-ray Computed Tomography (CT) reconstruction task. However, despiteits growing popularity, the benefits of GS are typically not substantialenough to motivate a transition from well-established reconstruction al-gorithms. This paper addresses the most significant remaining limita-tions of the GS-based approach by introducing FaCT-GS, a frameworkfor fast and flexible CT reconstruction. Enabled by an in-depth optimiza-tion of the voxelization and rasterization pipelines, our new method issignificantly faster than its predecessors and scales well with projectionand output volume size. Furthermore, the improved voxelization enablesrapid fitting of Gaussians to pre-existing volumes, which can serve as aprior for warm-starting the reconstruction, or simply as an alternative,compressed representation. FaCT-GS is over 4× faster than the State ofthe Art GS CT reconstruction on standard 5122 projections, and over13× faster on 2k projections. Implementation and data available through:https://papieta.github.io/fact-gs/.

Knowledge-based Visual Question Answering (KB-VQA) re-quires models to retrieve visual entities matching the query image fromlarge-scale encyclopedic knowledge bases and answer related questions.Existing multimodal Retrieval Augmented Generation (mRAG) meth-ods rely on global visual features to match candidate entities, yet whenthe knowledge base contains numerous visually similar entities, the re-triever struggles to distinguish them, populating the candidate set withvisually similar but factually mismatched distractors. Since subsequentprocessing steps such as noise filtering are also confined to this fixedcandidate set, errors from failed retrieval inevitably propagate to the fi-nal answer. To address these challenges, we propose MMAgent-R2 , anagentic mRAG framework that integrates visual reranking and active re-jection as its internal verification mechanism. Visual reranking directlycompares query and candidate images, capturing discriminative detailsbeyond textual descriptions to precisely identify the target entity amongsimilar candidates; active rejection discards unreliable results and re-trieves additional candidates when no confident match is found, movingbeyond the fixed candidate pool. We design a composite reward functionwith step-level verification rewards and achieve joint optimization of ex-ternal retrieval, internal verification, and answer generation via GRPOtraining. Experiments on InfoSeek, E-VQA, and MMhops demonstratethat MMAgent-R2 achieves state-of-the-art performance, with particu-larly notable advantages in challenging retrieval scenarios and complexmulti-image multi-hop reasoning tasks.

Concealed object segmentation aims to segment objects thatblend into their surroundings, presenting significant challenges due tothe high similarity between objects and backgrounds. Existing methodsrely on task-specific designs and independent training, leading to limitedgeneralization and structural redundancy. Although a unified parameter-shared model for diverse concealed scenarios is highly desirable andpromising, it is hindered by two challenges: large representation gapsacross scenarios and the intrinsic difficulty of concealed targets. From afrequency-domain perspective, we observe that different concealed sce-narios exhibit more discriminative and interpretable spectral character-istics compared to those in the RGB domain. These findings motivate usto approach unified concealed scene segmentation from the perspectiveof frequency modulation. To this end, we propose Frequency Director(FreqDirect), a dynamic frequency adaptation framework that directsspectral representations across diverse concealed scenes. It features amixture of frequency experts to perform adaptive routing over frequencycomponents, capturing heterogeneous task-specific patterns, along witha spatial commonality anchor to complement the spatial and structurecues. These components collectively enable effective and efficient adap-tation to diverse concealed segmentation tasks within a single model.Extensive experiments on eight concealed scenarios demonstrate thatFreqDirect surpasses existing unified models and achieves performancecomparable to, or better than, specialist models.

Change captioning is the task of generating natural languagedescriptions that explain the changes between a pair of images. Althoughdifferent change types (e.g., color shifts, object additions) exhibit distinctvisual cues and require specialized reasoning processes, existing methodsoften overlook these distinctions. To address this limitation, we proposeMulti-Expert Diagnosis for Image Change (MEDIC), a novel frameworkthat introduces change-type awareness by explicitly modeling changecategories. We build our MEDIC as a memory network to dynamicallyretrieve type-relevant visual patterns conditioned on the input. This de-sign allows each expert to flexibly capture diverse variations within eachchange type and focus on the most informative cues for its designatedchange type. By routing inputs through type-specialized experts andlearning dedicated representations for each change category, MEDICgenerates more precise and type-aware change descriptions. Extensiveexperiments demonstrate that proposed MEDIC consistently outper-forms across diverse and challenging datasets. The code is available athttps://github.com/VisualAIKHU/MEDIC.

Multi-view 3D Visual Question Answering (MV3D-VQA) re-quires integrating partial observations into a coherent 3D scene repre-sentation and selecting informative viewpoints for multi-step spatial rea-soning. However, current multimodal LLMs are typically trained withsparse, answer-level supervision, which often yields inconsistent cross-view reasoning and brittle view selection. We present DR-MV3D (DenseReward for MV3D-VQA), a map-grounded learning framework that pro-vides dense, verifiable rewards to supervise the reasoning process. Our ap-proach decomposes MV3D-VQA into (i) allocentric global map construc-tion, (ii) question-conditioned view-trajectory planning, and (iii) egocen-tric grounding for answer prediction. To make intermediate steps learn-able without manual annotations, we introduce two rewards: a global con-sistency reward that aligns the predicted map with geometry-consistentpseudo targets from frozen 3D vision foundation models (e.g., VGGT+ SAM3), and a local trajectory reward that supervises ordered view-point selection. We optimize the full pipeline with trajectory-level pol-icy optimization (GRPO). Experiments on MindCube, VSI-Bench, andBLINK (MV) show that DR-MV3D consistently improves over strongmulti-image baselines, supporting the effectiveness of process-level densesupervision for multi-view 3D reasoning.Code is available at: https://github.com/kaist-cvml/dr-mv3d

Linear probing on frozen features is the standard evalua-tion protocol for vision foundation models, yet classical theory for d-dimensional classifiers demands Ω(d) labeled samples, far more thanpractitioners actually need. We resolve this discrepancy with the first non-vacuous PAC-Bayes generalization bounds for linear probing on frozenvision foundation-model features. The key insight is that foundation-modelfeatures do not fill Rd : covariance spectra decay steeply, concentratingvariance in a subspace of effective dimension deff ≪ d. A data-dependentGaussian prior aligned to this spectral structure yields a certificate whosecomplexity is governed by deff and the classification margin γ rather thanthe ambient dimension d. We validate the theory on ImageNet across 12vision encoders spanning discriminative, language-supervised, generative,and predictive pretraining: the bound is non-vacuous for all discrimi-natively pretrained models and ranks among the tightest PAC-Bayescertificates reported for any vision model. Three empirical predictionsfollow: generative pretraining yields markedly less certifiable features thandiscriminative pretraining, deff alone ranks foundation models withoutany labels, and larger models tend to be more certifiable because scalesteepens spectral decay.

Understanding and forecasting audience reactions to videocontent are crucial for improving content creation, recommendation sys-tems, and media analysis. To enable audience reaction prediction andother content engagement applications, we introduce Video2Reaction,a multimodal dataset that maps short movie segments to a distributionof induced emotions of viewers in the wild, as expressed through socialmedia. Video2Reaction spans more than 10,000 videos and serves asa reliable benchmark as well as a training resource for audience reactionprediction. To enable cost-e!ective continuous annotations as reactionsmay change over time, we develop a two-stage multi-agent pipeline usingonly open-source LLMs, achieving 86% correctness under blind humanverification despite the inherently noisy and subjective nature of the task.We establish the first benchmark for video-to-reaction-distribution pre-diction in the wild and show that pretrained foundation video modelsfail in zero-shot settings, while finetuning transforms them into state-of-the-art predictors capable of modeling both full reaction distribu-tions and dominant responses from video alone. However, the task re-mains challenging: even the strongest methods achieve only 77% Top-3F1 in dominant reaction prediction (LLaVA-Next), highlighting a sub-stantial gap in modeling collective audience reaction. Dataset and codeare available at our project page: https://information-fusion-lab-umass.github.io/video2reaction-bench.github.io/.

Although diffusion-based real-world image restoration (Real-IR) has achieved remarkable progress, efficiently leveraging ultra-large-scale pre-trained text-to-image (T2I) models and fully exploiting theirpotential remain significant challenges. To address this issue, we proposeResFlow-Tuner, an image restoration framework based on the state-of-the-art flow matching model, FLUX.1-dev, which integrates unifiedmulti-modal fusion (UMMF) with test-time scaling (TTS) to achieveunprecedented restoration performance. Our approach fully leveragesthe advantages of the Multi-Modal Diffusion Transformer (MM-DiT) ar-chitecture by encoding multi-modal conditions into a unified sequencethat guides the synthesis of high-quality images. Furthermore, we in-troduce a training-free test-time scaling paradigm tailored for imagerestoration. During inference, this technique dynamically steers the de-noising direction through feedback from a reward model (RM), therebyachieving significant performance gains with controllable computationaloverhead. Extensive experiments demonstrate that our method achievesstate-of-the-art performance across multiple standard benchmarks. Thiswork not only validates the powerful capabilities of the flow match-ing model in low-level vision tasks but, more importantly, proposes anovel and efficient inference-time scaling paradigm suitable for largepre-trained models. The source code is publicly available at https://github.com/Rorschach-1010/ResFlow-Tuner.

We introduce PPTArena, a benchmark for PowerPoint edit-ing that evaluates how agents modify real slides from natural-languageinstructions. Unlike benchmarks that rely on image-PDF renderings ortext-to-slide generation, PPTArena features 100 decks with over 1,300human-curated edits across 2,125 slides, spanning text, charts, anima-tions, and professional master styles. Each edit pairs a ground-truthdeck with a target rubric and is scored by two Vision-Language Model(VLM) judges: one rates instruction following from structural diffs, theother visual quality from slide images. On top of this benchmark, wepresent PPTPilot, a structure-aware agent that plans semantic edit se-quences, routes between programmatic tools and deterministic XMLoperations, and verifies each result in an iterative plan-edit-check loop.PPTPilot outperforms strong VLM-based agents by more than 10 per-centage points on compound, layout-sensitive, and cross-slide edits, withlarge gains in visual fidelity and deck-wide consistency. Despite this, allagents still struggle on long-horizon, document-scale tasks, underscoringhow hard reliable PowerPoint editing remains. We publicly release ourcode: https://github.com/michaelofengend/PPTArena.