ECCV 2026 Papers

Diffusion models are primarily trained for image synthesis,yet their denoising trajectories encode rich, spatially aligned visual pri-ors. In this paper, we demonstrate that these priors can be utilized fortext-conditioned semantic and open-vocabulary segmentation, and thisapproach can be generalized to various downstream tasks to make ageneral-purpose diffusion segmentation framework. Concretely, we intro-duce DiGSeg (Diffusion Models as a Generalist Segmentation Learner),which repurposes a pretrained diffusion model into a unified segmenta-tion framework. Our approach encodes the input image and ground-truthmask into the latent space and concatenates them as conditioning sig-nals for the diffusion U-Net. A parallel CLIP-aligned text pathway injectslanguage features across multiple scales, enabling the model to align tex-tual queries with evolving visual representations. This design transformsan off-the-shelf diffusion backbone into a universal interface that pro-duces structured segmentation masks conditioned on both appearanceand arbitrary text prompts. Extensive experiments demonstrate state-of-the-art performance on standard semantic segmentation benchmarks, aswell as strong open-vocabulary generalization and cross-domaintransfer to medical, remote sensing, and agricultural scenarios—withoutdomain-specific architectural customization. These results indicate thatmodern diffusion backbones, can serve as generalist segmentation learn-ers rather than pure generators, narrowing the gap between visual gen-eration and visual understanding.

In the task of human mesh recovery (HMR), multi-personscenes are particularly difficult to handle due to the many entities thatappear and occlusions between them over time. In particular for videoinputs, there is a need to track each entity reliably and consistently.Existing methods rely on pretrained human detection modules, increas-ing their runtime and limiting the number of tracked entities. We presentDETRAM, a unified framework for multi-person HMR and tracking thatsimultaneously detects, reconstructs, and tracks humans across time,both automatically and via user prompts. DETRAM uses a single trans-former decoder with an identity-consistent set of learnable query embed-dings that persist across frames: detection queries discover new people,tracking queries maintain pose and shape for existing individuals, andprompt queries follow user-specified identities. Our approach achievesstate-of-the-art tracking results on PoseTrack21, 3DPW, BEDLAM, andMuPoTS-3D, and competitive reconstruction accuracy on BEDLAM and3DPW, while uniquely supporting prompt-based tracking of individualsin multi-person scenes. To our knowledge, this is the first method tounify promptability and multi-person HMR with tracking in an end-to-end trainable framework, enabling user-directed human analysis in

Existing latent di!usion models typically couple scale withcontent complexity, using more latent tokens to represent higher-resolutionimages or higher–frame rate videos. However, the latent capacity re-quired to represent visual data primarily depends on content complex-ity, with scale serving only as an upper bound. Motivated by this obser-vation, we propose DCS-LDM, a novel paradigm for visual generationthat decouples information complexity from scale. DCS-LDM constructsa hierarchical, scale-independent latent space that models sample com-plexity through multi-level tokens and supports decoding to arbitraryresolutions and frame rates within a fixed latent representation. This la-tent space enables DCS-LDM to achieve a flexible computation–qualitytradeo!. Furthermore, by decomposing structural and detailed informa-tion across levels, DCS-LDM supports a progressive coarse-to-fine gen-eration paradigm. Experimental results show that DCS-LDM deliversperformance comparable to state-of-the-art methods while o!ering flex-ible generation across diverse scales and visual qualities.

Stereo-based 3D obstacle perception for autonomous driv-ing is currently constrained by an imbalanced triplet: deployment cost,detection accuracy, and open-set adaptability. While existing methodsstruggle to balance these three competing objectives, there is an urgentdemand for high-precision, real-time algorithms capable of detecting ar-bitrary obstacles in the wild. In this paper, we present DDStereo, a novelDual-Decoder Stereo Transformer that achieves a synergistic integrationof 3D object detection and Out-of-Distribution (OoD) road anomaly de-tection. Leveraging the geometric priors of stereo disparity, our approacheffectively couples 3D attribute regression with open-set foreground de-tection within a streamlined dual-branch decoder architecture. Conven-tional methods rely on complex feature-level fusion; DDStereo main-tains execution efficiency by employing a decoupled decoding strategyand shared object-level queries to ensure cross-modal target alignment.Extensive evaluations of public benchmarks demonstrate that DDStereonot only achieves state-of-the-art accuracy under open-set and closed-set protocols. Our method delivers real-time performance comparable tomonocular 3D detection baselines, providing a cost-effective solution forthe perception of obstacles of the normal and OoD category. Code andmodels are available at https://github.com/shiyi-mu/DDStereo.

Image demoiréing remains a challenging task due to the com-plex interplay between texture corruption and color distortions causedby moiré patterns. Existing methods, especially those relying on directimage-to-image restoration, often fail to disentangle these intertwinedartifacts effectively. While frequency-aware approaches offer a promisingdirection, their potential is hindered by the discrete transform (e.g., Haarwavelet or block-based DCT), which may suffer from spatial discontinu-ity, channel redundancy, and further cause error accumulation duringtheir fixed inverse processes. In this paper, we present Freqformer, aTransformer-based framework specifically designed for image demoiréingthrough targeted frequency separation. Our method performs an effectivefrequency decomposition that splits moiré patterns into high-frequencyspatially-localized textures and low-frequency scale-robust color distor-tions, which are then handled by a dual-branch architecture and anasymmetric training scheme tailored to their distinct characteristics. Wefurther propose a learnable Frequency Composition Transform (FCT)module to adaptively fuse the frequency-specific outputs, enabling con-sistent and high-fidelity reconstruction. To better aggregate the spatialdependencies and the inter-channel complementary information, we in-troduce a Spatial-Aware Channel Attention (SA-CA) module that refinesmoiré-sensitive regions without incurring high computational cost. Ex-tensive experiments on various demoiréing benchmarks demonstrate thatFreqformer achieves state-of-the-art performance with a compact modelsize. The code will be made publicly available at https://github.com/xyLiu339/Freqformer.

Despite significant advances in image segmentation, evenstate-of-the-art models produce masks with imperfect boundaries, se-mantic inconsistencies, and structural errors. Mask refinement addressesthese limitations, yet current approaches rely on simplistic syntheticnoise that fails to capture the complex error patterns of real segmen-tation models. We introduce Phoenix, a novel framework that lever-ages adversarial learning to generate semantically meaningful noise pat-terns and contrastive learning to model refinement relationships. Ourapproach consists of two key innovations: (1) Adversarial Mask Per-turbation, which employs embedding attacks to create semantic-awarenoise that mimics real segmentation errors, and (2) Contrastive MaskRefinement Learning, which establishes a tri-directional framework thatensures feature consistency within semantic regions while maintainingseparation between classes. Experiments demonstrate that Phoenix sig-nificantly outperforms existing methods across diverse tasks, while con-sistently enhancing state-of-the-art segmentation models with substan-tial improvements. Our code and project page are publicly available atTargetMaskMorphologicalMask (b) Morphological NoiseLess-Semantic, Unrealistic, and Non-Contextual NoiseAdversarialMask(ours)Semantic, Realistic, and Contextual False-Negative & False-Positive Noise(a) Qualitative Samples of Morphological and Adversarial Noise Masks (c) Adversarial NoiseFig. 2: Qualitative Comparison of noise patterns (a) between morphological andour adversarial noise masks. (b, c) Distribution of the Pearson correlation betweennoise location and image edge/texture maps on LVIS val. (b) Morphological noise isnarrowly concentrated near 0 (semantically uncorrelated), while (c) our adversarialnoise spans [−0.6, 0.8], where positive values indicate alignment with semantic struc-tures and negative values indicate alignment with homogeneous regions, mirroring thediverse error patterns of real models.

Online skeleton-based action recognition requires per-frameclassification using only past and current observations. We identify twokey challenges in this setting: (1) Not all frames are equal: discrimina-tive motion concentrates within specific temporal regions whose loca-tion varies across action categories, yet existing temporal modeling relieson fixed sampling patterns and causal attention that treat all positionsequally. (2) Not all futures need to be predicted: prior methods attempt toexplicitly predict future features to compensate for the unobserved por-tion, but the inherent ambiguity of human motion causes such predictionto degenerate into trivial identity mapping of the input. To address thesechallenges, we propose DART (Deformable Adaptive Reasoning withTemporal Queries). DART employs a Deformable Multi-scale TemporalNetwork (DMTN) that introduces input-dependent offsets into causalconvolutions, enabling content-adaptive sampling of the most informa-tive historical moments. The learned sampling positions naturally forma compact temporal memory, upon which a Temporal Query Decoder(TQD) operates: learnable queries interact with the memory to distillhigh-level action evolution patterns, providing complementary temporalreasoning without explicit future forecasting. Extensive experiments onNTU RGB+D 60, NTU RGB+D 120, and NW-UCLA demonstrate theeffectiveness of DART. On NTU RGB+D 60, DART achieves 73.99%AUC, surpassing the SOTA by +3.04%. Code is available.

The emergence of large-scale text-to-image diffusion (T2ID)models has led to significant advancements in generating high-qualityvisual content from textual prompts. However, these powerful capabil-ities have also raised growing concerns about the generation of harm-ful and copyrighted material. While existing concept erasure techniquescan effectively block the production of specific unwanted concepts fromprompts, they often fall short when it comes to erasing an entire cate-gory (including subcategories) and are typically limited to handling onlya few concepts at a time. In this paper, we introduce SubcategoricalUnlearning via Regularized Erasure (SURE), a novel method for re-moving entire subcategories from text-to-image diffusion models usingonly a single parent category as the target. Unlike prior approaches,SURE does not rely on sets of synonyms. Instead, it employs conceptspace to discover and eliminate the target category while preserving themodel’s overall utility. To further enhance erasure, SURE incorporatesLipschitz regularization, which encourages smoother model responses toperturbations around the target category. Specifically, the regularizationpromotes consistent behavior in the model’s latent space when exposedto slight variations of the category to be forgotten. This smoothnessconstraint aids in erasure while maintaining the model’s ability to gen-erate unrelated content. Extensive experiments conducted across threetasks—object removal, suppression of explicit content, and elimination ofartistic styles demonstrate that SURE achieves balanced performance inboth effective category erasure and preservation of non-target concepts.

Aerial-ground person re-identification (AG-ReID) is signifi-cantly challenged by heterogeneous geometric regimes and severe foreground-background imbalance. Traditional shared-backbone architectures suf-fer from transformation conflicts when modeling disparate aerial andground views, while residual connections in transformers often lead toidentity embedding contamination from persistent background activa-tions. To address these issues, we propose View-conditioned and Identity-Guided Adaptation (VIGA), a unified framework that mitigates AG-ReID-specific aerial-ground discrepancies at both the parameter and ac-tivation levels. VIGA incorporates View-Conditioned Low-Rank Adap-tation to dynamically modulate feed-forward weights using view-awarelow-rank residuals, thereby alleviating view-dependent geometric map-ping conflicts without substantial parameter overhead. Simultaneously,we introduce Identity-Guided Sparsification, which employs a soft mask-ing mechanism guided by identity semantics to suppress irrelevant back-ground noise while preserving the 2D spatial topology. Extensive evalua-tions on the LAGPeR and AG-ReID.v2 datasets demonstrate that VIGAachieves state-of-the-art performance, particularly in challenging cross-platform retrieval scenarios, by effectively learning robust, view-invariantrepresentations.

Large-vocabulary instance segmentation is constrained bylong-tailed category distributions and fine-grained inter-class ambiguity.While data synthesis offers a promising alternative, current paradigmshave complementary limitations: text-to-image (T2I) methods inheritnoisy pseudo-labels and struggle on rare classes, whereas copy-pastemethods compromise contextual realism. To address these issues, we pro-pose a hybrid pipeline coupling T2I generation with context-aware image-to-image (I2I) editing. The T2I branch provides broad category and scenediversity, while a teacher-student scheme ensures label reliability by se-lectively retaining only prompt-specified categories. To strengthen su-pervision for rare classes, we introduce VRAIN (Verified Rare-classAugmentation via INstructed editing), a novel I2I editor. VRAIN insertshigh-confidence instances at semantically appropriate locations withinin-the-wild scenes, yielding semantically coherent and visually naturaledits that reduce domain gaps and enable targeted augmentation. Onthe LVIS benchmark, our method surpasses existing baselines, improv-ing overall AP by up to +4.0 points and rare-class AP by up to +9.5points, while scaling effectively with backbone capacity.

World models have driven remarkable progress in physicalperception and dynamic forecasting, yet current mainstream implemen-tations predominantly rely on visible-light (RGB) observations, whichlimits their all-weather reliability. Developing a thermal world model isessential, but it is hindered by the unique distribution of thermal imageryand the extreme scarcity of continuous video datasets. In this paper, weidentify a fundamental challenge in extending standard masked modelingto the thermal domain: intra-frame spatial homogenization. Due to ther-mal equilibrium, distinct physical entities often exhibit near-zero temper-ature gradients. Consequently, unconstrained self-attention blurs seman-tic boundaries and collapses object topologies, severely degrading down-stream dynamics forecasting. To overcome this, we propose Thermo-JEPA, a geometry-grounded pre-training framework based on Learn-ing Using Privileged Information (LUPI). We extract intra-frame spatialtopologies from aligned RGB data and inject them as a confidence-gatedstructural prior into the thermal teacher’s attention mechanism. Thisexplicitly enforces semantic segregation without corrupting fine-grainedintra-object thermal gradients. Concurrently, we release RGBT-World,the largest unified, spatiotemporally aligned RGB-Thermal video corpustailored for generative pre-training. Extensive experiments demonstratethat Thermo-JEPA effectively prevents representation collapse, achiev-ing state-of-the-art zero-shot dynamics alignment on pure thermal inputsand significantly outperforming massive generic video foundation mod-els.

We present a framework for document-centric backgroundgeneration with multi-page editing and thematic continuity. To ensuretext regions remain readable, we employ a latent masking formulationthat softly attenuates updates in the diffusion space, inspired by smoothbarrier functions in physics and numerical optimization. In addition, weintroduce Automated Readability Optimization (ARO), which automat-ically places semi-transparent, rounded backing shapes behind text re-gions. ARO determines the minimal opacity needed to satisfy percep-tual contrast standards (WCAG 2.2) relative to the underlying back-ground, ensuring readability while maintaining aesthetic harmony with-out human intervention. Multi-page consistency is maintained througha summarization-and-instruction process, where each page is distilledinto a compact representation that recursively guides subsequent gen-erations. This design reflects how humans build continuity by retain-ing prior context, ensuring that visual motifs evolve coherently acrossan entire document. Our method further treats a document as a struc-tured composition in which text, figures, and backgrounds are preservedor regenerated as separate layers, allowing targeted background editingwithout compromising readability. Finally, user-provided prompts allowstylistic adjustments in color and texture, balancing automated consis-tency with flexible customization. Our training-free framework producesvisually coherent, text-preserving, and thematically aligned documents,bridging generative modeling with natural design workflows.

Dataset distillation (DD) has proven to reduce training costwhile preserving accuracy. While promising, the factors that make onedistilled dataset more effective than another remain poorly understood.In this work, we investigate this question through the lens of discretevisual tokenizers. Whereas many prior DD efforts emphasize matchingglobal data distributions, we suggest that the effectiveness depends onwhich semantic concepts are captured and how they are composed. Dis-crete visual tokenizers provide a finite vocabulary that enables directstatistical analysis of such compositional structure. Through quantita-tive analysis of token-level statistics, we introduce the structural scoreto measure the adequacy of token compositions. We observe that dis-tilled datasets with balanced token composition yield higher validationperformance. On the other hand, divergence from the original data doesnot necessarily harm performance. We further show that samples withhigh structural scores in the discrete token space can effectively guidediffusion-based DD. Our findings highlight the importance of token com-position in dataset effectiveness, offering a principled complement to dis-tributional similarity considerations in DD.

Remote photoplethysmography (rPPG) enables non-contactphysiological monitoring by capturing subtle skin color variations inducedby cardiac cycles. Despite its promise, rPPG remains highly sensitive toenvironmental illumination. Existing illumination-aware methods sufferfrom two key limitations: (1) limited out-of-distribution (OOD) gen-eralization due to domain discrepancies between training and testingdata, and (2) reliance on the Lambertian assumption, which neglectsnon-Lambertian specular highlights that frequently corrupt facial skinsignals. In this paper, we propose SHINE-PPG (Specular-Highlight In-trinsic Network for rPPG Estimation), a novel framework that leveragesnon-Lambertian intrinsic decomposition to decouple facial videos intoillumination, reflectance, and specular components in a self-supervisedmanner. By isolating physiological information within the intrinsic re-flectance, our method effectively suppresses both ambient lighting vari-ations and surface highlights to recover high-fidelity rPPG signals. Tofurther enhance robustness, we introduce an adversarial illuminationenhancement strategy that dynamically synthesizes challenging unseenlighting conditions during training, significantly improving OOD general-ization. Extensive experiments on five benchmark datasets demonstratethat SHINE-PPG consistently outperforms previous methods, particularlyunder complex and dynamic illumination scenarios. The code is availableat https://github.com/Edmond-Yang/SHINE-PPG.

We present SD3.5-Fla.sh, an efficient few-step distillation framework that brings high-quality image generation to accessible consumer devices. Our approach distills computationally prohibitive rectified flow models through a reformulated distribution matching objective tailored specifically for few-step generation. We introduce two key innovations: “timestep sharing” to reduce gradient noise and “split-timestep fine-tuning” to improve prompt alignment. Combined with comprehensive pipeline optimizations like text encoder restructuring and specialized quantization, our system enables both rapid generation and memory-efficient deployment across different hardware configurations. This democratizes access across the full spectrum of devices, from mobile phones to desktop computers. Through extensive evaluation including large-scale user studies, we demonstrate that SD3.5-Flash consistently outperforms existing few-step methods, making advanced generative AI truly accessible for practical deployment.

Knowledge Distillation (KD) transfers knowledge from a largeteacher model to a smaller student by aligning their predictive distri-butions. However, conventional KD formulations - typically based onKullback–Leibler divergence - assume that the teacher provides reliabletargets. In practice, teacher predictions are often noisy or overconfident,and existing correction-based approaches rely on ad-hoc heuristics andextensive hyper-parameter tuning, which hinders generalization. We in-troduce REDistill (Robust Estimator Distillation), a simple yet prin-cipled framework grounded in robust statistics. REDistill replaces thestandard KD objective with a power divergence loss, a generalizationof KL divergence that adaptively downweights unreliable teacher out-put. This formulation provides a unified and interpretable treatmentof teacher noise, requires only logits, integrates seamlessly into existingKD pipelines, and incurs negligible computational overhead. Extensiveexperiments on CIFAR-100 and ImageNet-1k demonstrate that REDis-till consistently improves student accuracy in diverse teacher-student ar-chitectures. Remarkably, it achieves these gains without model-specifichyper-parameter tuning, underscoring its robustness and strong general-ization to unseen teacher–student pairs.

We study the problem of physically plausible shadow castingwhen animating 3D Gaussian Splatting (3DGS) avatars, either individu-ally or in multi-avatar and object-interaction scenarios, within existing3DGS scenes. In contrast to prior methods that rely on binary hit tests andmesh-based shadow casters, our method performs shadow computationentirely in Gaussian space, without requiring any mesh reconstruction. Weintroduce RAGA, a Ray-Traced Gaussian Shadow Casting formulationbased on exact ray–Gaussian line integrals. For each occluding Gaussian,we integrate the opacity profile along the shadow ray and normalize bythe theoretical maximum integral, producing a weight that captures howthe ray traverses the occluder rather than merely whether an intersectionoccurred. To reduce temporal variance from clothing deformations in ani-mated avatars, we further introduce an avatar proxy representation thatstabilizes shadow casting while preserving visual fidelity. We implement*Work partly done as intern at Snap Inc.RAGA using custom CUDA kernels integrated with the NVIDIA OptiXframework; as such, our shadow tracer runs at rates of about 50 FPS.We evaluate on single-avatar, multi-avatar, and avatar–object interactionscenarios across multiple datasets, demonstrating substantially improvedshadow realism, temporal stability, and scene coherence. Our project pageis available at https://miraymen.github.io/raga/.

Heterogeneity in data distributions and model architectures poses significant challenges in federated learning (FL). While various heterogeneous FL (HtFL) approaches have been proposed to address these challenges, prototype-based FL (PBFL) has emerged as a promising framework that exchanges prototypes—per-class mean activations from the penultimate layer—rather than model parameters. However, existing PBFL methods suffer from suboptimal prototype separation in the feature space, which limits their discriminative capacity. We propose Prototype Normalization (ProtoNorm), a novel PBFL framework that systematically addresses this limitation. Inspired by the Thomson problem in classical physics, ProtoNorm optimizes global prototype configurations on a unit hypersphere, maximizing separation between class prototypes. Extensive experimental evaluation demonstrates that ProtoNorm consistently outperforms existing HtFL methods across diverse heterogeneous settings. Importantly, ProtoNorm maintains communication efficiency while operating entirely on the server side with less computation than existing PBFL methods, making it well-suited for resource-constrained environments.

Polarimetric imaging enables advanced vision applicationssuch as normal estimation and de-reflection by capturing unique surface-material interactions. However, existing applications (alternatively calleddownstream tasks) rely on datasets constructed by naïvely regroupingraw measurements from division-of-focal-plane sensors—where pixels ofthe same polarization angle are extracted and aligned into sparse im-ages without proper demosaicking. This reconstruction strategy resultsin suboptimal, incomplete targets that limit downstream performance.Moreover, current demosaicking methods are task-agnostic, optimizingonly for photometric fidelity rather than utility in downstream tasks.Towards this end, we propose PolarAPP, the first framework to jointlyoptimize demosaicking and its downstream tasks. PolarAPP introducesa feature alignment mechanism that semantically aligns the representa-tions of demosaicking and downstream networks via meta-learning, guid-ing the reconstruction to be task-aware. It further employs an equivalentimaging constraint for demosaicking training, enabling direct regressionto physically meaningful outputs without relying on rearranged data.Finally, a task-refinement stage fine-tunes the task network using thestable demosaicking front-end to further enhance accuracy. Extensiveexperimental results demonstrate that PolarAPP outperforms existingmethods in both demosaicking quality and downstream performance.Code is available here.