ECCV 2026 Papers

In recent years, super-resolution (SR) methods leveragingpre-trained text-to-image diffusion models have shown impressive po-tential. However, their deployment in real-world applications remainsconstrained by the computational overhead of numerous inference steps.Recent one-step approaches, such as OSEDiff and TSD-SR, mitigate thisissue via score distillation techniques. However, their discriminators areoptimized solely with the diffusion loss, resulting in weak and unstablegradients for the generator, making it difficult to preserve fine details andstructural fidelity. To tackle these limitations, we introduce ASDSR, anovel one-step distillation framework tailored for real-world image super-resolution. Our approach enhances Adversarial Score Distillation (ASD)by adopting a WGAN-based formulation with a stabilized penalty termthat utilizes the full discriminator loss to improve the distillation pro-cess. Furthermore, we incorporate CLIP-derived semantic priors to alignreconstructed images with their high-quality counterparts in the fea-ture space, effectively guiding the generator toward perceptual realism.Extensive experiments on multiple real-world benchmarks demonstratethat ASDSR achieves superior performance compared to state-of-the-artmethods, offering a practical and efficient solution for real-world imagerestoration.

Recent research in geospatial machine learning demonstratesthat models pretrained with self-supervised learning on Earth observa-tion data can perform well on downstream tasks with limited labeleddata. However, most benchmark datasets have few data modalities andpoor global representation, limiting the ability to evaluate multimodalpretrained models at global scales. In order to fill this gap, we introduceMMEarth-Bench, a collection of five new environmental tasks with 12modalities, globally distributed data, and both random and geographictest splits. We benchmark a diverse set of pretrained models and findthat while (multimodal) pretraining tends to improve model robustnessin limited data settings, geographic generalization abilities remain poor.Moreover, a simple randomly initialized multimodal model is competi-tive given enough labeled data. Although data is abundant, models cancurrently only make use of the modalities on which they were pretrained.To solve this problem, we propose using all the modalities available attest time as auxiliary tasks for test-time adaptation. Our model-agnosticmethod for test-time training with multimodal reconstruction (TTT-MMR) can improve performance across all models and tasks on bothtest splits. Furthermore, geographic batching leads to a good trade-off be-tween regularization and specialization during TTT, which is especiallybeneficial for long-tail distributions. Our dataset, code, and visualizationtool are linked on the project page: lgordon99.github.io/mmearth-bench.

AI-generated image (AIGI) attribution presents a pressingchallenge that goes beyond mere AIGI detection, aiming to identify thesource model or technique responsible for a synthetic image. However,most previous source attribution methods operate in a closed-set manner,which necessitates retraining to recognize any novel category, preventingadaptation to the rapid evolution of image generation. In this work, wepropose a new paradigm for synthetic image attribution, termed few-shotattribution. This paradigm targets the reliable identification of unseengenerators using only limited samples, making it highly suitable for real-world applications. To facilitate this work, we construct OmniFake, alarge-scale, well-categorized synthetic image dataset that contains 1.17million images from 45 distinct generators. We further introduce Om-niDFA (Omni Detector and Few-shot Attributor), a few-shot attributionbaseline that not only assesses the authenticity of images but also deter-mines their synthesis origins. Experiments demonstrate that OmniDFAexhibits excellent capability in few-shot attribution and achieves state-of-the-art generalization performance in AIGI detection. Our dataset andcode are available at https://github.com/teheperinko541/OmniDFA.

Background-oriented schlieren (BOS) tomography reconstructs3D flow density fields from refraction-induced distortions observed fromone or more viewpoints. Capturing high-speed flows with frame-basedBOS typically requires high-speed cameras and bright illumination tocompensate for short exposure times. In this paper, we propose an event-based BOS tomography approach that reconstructs time-varying 4D den-sity fields from event streams. Leveraging the high temporal resolutionand high dynamic range of event cameras, the proposed approach enableshigh-speed airflow reconstruction under ambient lighting. We representthe spatiotemporal density field as a neural implicit field and render BOSobservations via refractive ray tracing. We utilize physics-informed regu-larization to improve reconstruction under sparse views. Experiments onsimulated and real data in single-view and orthogonal dual-view setupsdemonstrate accurate reconstructions.

Training-free controllable motion generation has attracted growing interest for enabling flexible constraint enforcement without constraint-specific training. However, existing training-free methods require constraints to be continuous objective-based with differentiable losses, while many real-world requirements are criterion-based and provide only discontinuous, sparse, or even black-box feedback. In this paper, we propose Motion-Inference-as-Control (MIC), the first training-free motion generation framework that handles both continuous objective-based and criterion-based motion constraints under a shared mechanism. The key idea is to cast diffusion-based motion generation as a stochastic control problem. This perspective not only provides principled and practically effective step-wise control laws that support criterion-based constraints without requiring differentiability and naturally accommodate objectivebased constraints as a special case, but also motivates a control-oriented constraint coordination mechanism that adaptively balances and reconciles motion constraints during generation. Experiments across diverse constraint settings demonstrate the effectiveness of our framework.

Counting individual trees is a fundamental task for environ-mental monitoring, yet remains largely unexplored with satellite imagery.At these resolutions, isolated trees may still be identifiable, but crownboundaries become ambiguous in dense forests, making the notion of anindividual tree inherently ill-defined. Moreover, large-scale manual an-notations of individual trees are prohibitively expensive. While scalablesupervision can be derived from airborne LiDAR, the resulting anno-tations are noisy and difficult to exploit effectively. We address thesechallenges by formulating tree counting as a spatial density matchingproblem supervised through Unbalanced Optimal Transport. This formu-lation naturally accommodates both precise localization of isolate treesand robust density estimation in dense forests. We further introduce aself-correction mechanism that leverages transport residuals to progres-sively refine noisy supervision during training. We evaluate our approachon TinyTrees, a new benchmark spanning three continents and threesatellite sensors, comprising over 216 million tree annotations (including639k manually verified instances) across 25 890 km2 . Our method con-sistently outperforms detection-based, regression-based, and transport-based distribution-matching baselines, demonstrating the effectivenessof unbalanced transport and reliability-aware supervision for large-scaletree counting from satellite imagery. Code, data and models are availableat github.com/dgominski/treematch.

Conformal prediction provides distribution-free reliability guar-antees for vision systems, but these guarantees depend on how predictionerrors are measured in the output space. Many vision tasks produce out-puts on curved spaces (e.g. gaze directions on the sphere or 3D headrotations), yet intermediate prediction heads, residuals, uncertainty es-timates, or conformal scores are often defined in flat coordinate chartssuch as yaw–pitch or Euler angles. We show that this scoring choice intro-duces systematic geometric distortion near coordinate singularities (largepitch angles on the sphere and poses approaching gimbal lock in 3D rota-tions). Across four datasets (ETH-XGaze, Gaze360, BIWI, AFLW2000-3D), slice-conditional coverage at a nominal 90% target drops by 30–50percentage points in these regions, falling to 38.9% on ETH-XGaze and42.0% on Gaze360 at gaze pitch above 70◦ , and to 57.5% on BIWI and55.2% on AFLW2000-3D at head pose pitch above 60◦ near gimbal lock,despite marginal coverage remaining near 90%. We prove that this isstructural. Scalar thresholding changes the size of chart-coordinate pre-diction sets but leaves their distorted axis ratios unchanged. To diagnosethis hidden failure mode, we show that a simple geometric quantity,the Riemannian volume density, strongly correlates with where cover-age collapse occurs. Finally, we show that coordinate-free geodesic scor-ing removes this distortion. It requires no retraining and adds negligiblecomputational cost.

Performance in face and speaker verification is largely drivenby margin-penalty softmax losses such as CosFace and ArcFace. Recentlyintroduced α-divergence loss functions offer a compelling alternative, par-ticularly due to their ability to induce sparse solutions (when α > 1).However, standard geometric margins are designed for the softmax func-tion and do not naturally extend to this generalized probabilistic frame-work. In this paper we propose Q-Margin, a novel α-divergence lossthat introduces a principled probabilistic margin. Unlike conventionalmethods that apply geometric penalties to the logits (unnormalized log-likelihoods), Q-Margin encodes the margin penalty directly into the ref-erence measure (prior probabilities). This formulation naturally encour-ages discriminative embeddings while preserving the beneficial sparsityproperties of the α-divergence. We demonstrate that Q-Margin achievescompetitive or superior performance on the challenging IJB-B and IJB-C face verification benchmarks and similarly strong results in speakerverification on VoxCeleb. Crucially, against ArcFace and CosFace base-lines trained under an identical recipe, Q-Margin consistently improvesat low False Acceptance Rates (FARs), a capability critical for practicalhigh-security applications. Finally, the extreme sparsity of the Q-Marginposteriors enables exact and memory-efficient training, offering a scalablesolution for datasets with millions of identities.

Video Large Language Models (Video LLMs) incur high in-ference latency due to a large number of visual tokens provided to LLMs.To address this, training-free visual token pruning has emerged as asolution to reduce computational costs; however, existing methods areprimarily validated on Multiple-Choice Question Answering (MCQA)benchmarks, where coarse-grained cues often suffice. In this work, wereveal that these methods suffer a sharp performance collapse on fine-grained understanding tasks requiring precise visual grounding, such ashallucination evaluation. To explore this gap, we conduct a systematicanalysis and identify sink tokens–semantically uninformative tokens thatattract excessive attention–as a key obstacle to fine-grained video un-derstanding. When these sink tokens survive pruning, they distort themodel’s visual evidence and hinder fine-grained understanding. Moti-vated by these insights, we propose Sink-Token-aware Pruning (SToP),a simple yet effective plug-and-play method that introduces a sink scoreto quantify each token’s tendency to behave as a sink and applies thisscore to existing spatial and temporal pruning methods to suppress them,thereby enhancing video understanding. To validate the effectivenessof SToP, we apply it to state-of-the-art pruning methods (VisionZip,FastVid, and Holitom) and evaluate it across diverse benchmarks cover-ing hallucination, open-ended generation, compositional reasoning, andMCQA. Our results demonstrate that SToP significantly boosts perfor-mance, even when pruning up to 90% of visual tokens. Our code is avail-able at https://github.com/rlqja1107/SToP

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.