ECCV 2026 Papers

Synthetic data for autonomous driving is surging, poweredby diffusion models that promise scalable scene generation. Yet key ob-stacles remain, since multi-view and temporal consistency often requiresbackbone fine-tuning or added layers, which can erode pretrained knowl-edge and weaken text alignment. Models also stay close to the train-ing distribution, struggling under adverse weather and unseen configu-rations, while fidelity favors frequent over rare classes. We address thesegaps with FrozenDrive, a controllable generative framework that pre-serves pretrained diffusion knowledge while achieving strong consistency.FrozenDrive conditions on driving-stack signals and text prompts, andexpands the context of frozen self-attention across views and frames topromote cross-view alignment and temporal coherence in one pass, with-out trainable spatio-temporal modules in the diffusion backbone. Anobject-focused constraint further improves fidelity for rare categories.Without weather- or scene-specific fine-tuning, FrozenDrive synthesizesglobally coherent multi-view driving scenes from text and surpasses priorbaselines under adverse and rare conditions. On nuScenes, FrozenDrive-augmented data improves AD model performance, especially at night andin rain, demonstrating strong robustness with scenario-targeted data.

Optical flow methods typically rely on task-specific inductivebiases, such as correlation volumes, feature warping, and iterative refine-ment, among others, to reach high accuracy. While effective, such biasesconstrain the model to predefined heuristics, which can limit its expres-sivity and lead to more complex pipelines and additional computationalcost. We present FreeFlow, a hierarchical transformer built without anyflow-specific components, using instead a single feed-forward encoder–decoder. FreeFlow combines three attention variants: window attentionfor local processing, shifted-window attention for cross-window informa-tion exchange, and a global attention operating at a reduced resolu-tion. The resulting architecture scales naturally with model capacity,enabling a consistent accuracy gain from small to large variants. Despitethe absence of standard inductive biases, FreeFlow achieves state-of-the-art results on major benchmarks, including Sintel (0.68/1.48 EPE onClean/Final), KITTI-2015 (3.23 Fl-all), and Spring (3.192 1px), whileremaining memory efficient at 1080p inference.

Cross-Domain Few-Shot Object Detection (CDFSOD) aimsto transfer knowledge from data-rich upstream generic domains to down-stream expert domains using scarce training data, where the significantdomain gap and data scarcity make it an unsolved challenge. To ad-dress this problem, we revisit a natural yet underexplored approach inCDFSOD: data augmentation, by directly synthesizing data through dif-fusion models to supplement limited training samples. However, due tolarge domain gaps, we find that current diffusion methods cannot pro-duce good results, leading to performance even lower than using theoriginal images. To address these limitations, we divide the domain gapsinto visual gaps and semantic gaps for separate analysis. For the visualgap, we find that the diffusion model cannot distinguish noise from use-ful information on expert domains, which can be mitigated by addingweakened noise. For the semantic gap, we find that the background se-mantics shows much smaller gaps between domains than foreground se-mantics, and we can bridge this gap by background inpainting. Basedon the above analysis, we propose a method (Selective Inpainting withTailored Noise, SITN) to dynamically take different strategies for down-stream data synthesis based on their different gaps from the generaldomain, including a Generation Module for adding tailored noise and aSelection Module to dynamically select the inpainting regions. Extensiveexperiments on 6 datasets of CDFSOD and 4 datasets of cross-domainfew-shot segmentation (CDFSS) validate that we can synthesize helpfuldata, achieving new state-of-the-art performance. Our codes is availableat https://github.com/zzzzj311-droid/Free-Lunch-SITN.

005 Multimodal large language models (MLLMs) have recently 005006 shown promise for industrial anomaly understanding, where models are 006007 expected not only to determine whether an object is defective, but also to 007008 localize visual evidence and explain defect-related decisions. However, ex- 008009 isting MLLMs often rely on language priors or shortcut reasoning when 009010 facing fine-grained industrial defects, resulting in inconsistent outputs 010011 such as correct answers with invalid locations, hallucinated defect boxes 011012 on normal samples, or rationales that contradict the final prediction. To 012013 address these issues, we propose EGVLR, an evidence-grounded vision– 013014 language reinforcement framework for anomaly reasoning. EGVLR fol- 014015 lows a unified Evidence-Driven Diagnostic Protocol (EDDP), where each 015016 response explicitly contains visual evidence, diagnostic logic, spatial loca- 016017 tion, and the final answer. The framework consists of four stages: Progres- 017018 sive Visual-Evidential Fine-Tuning (PVE-FT), Knowledge-Grounded In- 018019 struction Tuning (KG-IT), Geometry-Semantic Decoupled Preference 019020 Optimization (GS-DPO), and Box-Guided Segmentation Rendering (BGSR). 020021 PVE-FT first aligns the model with localized visual evidence through 021022 synthetic anomalies, 3 × 3 grid grounding, local verification decoys, and 022023 null-hypothesis calibration. KG-IT then internalizes industrial-domain 023024 QA behavior while preserving the same output schema. GS-DPO fur- 024025 ther optimizes answer correctness, geometric localization, BGE-based 025026 rationale semantics, and answer–location–rationale consistency. Finally, 026027 BGSR converts predicted bounding boxes into dense masks using an off- 027028 the-shelf segmentation backend. Experiments on MMAD demonstrate 028029 that EGVLR improves evidence-grounded industrial anomaly reasoning, 029030 especially on defect localization, reasoning-dependent QA, and spatial 030031 false-positive control. 031

Scene flow represents the 3D cartesian motion of points inthe world and can be used for applications such as navigation in dy-namic environments, human-robot interaction or non-rigid motion anal-ysis, among others. Despite the potential benefits of event cameras formotion estimation tasks, due to their high temporal resolution and low-latency, not many works have yet addressed their use for scene flowestimation. Progress may be limited by the unconventional data addingcomplexity to established processing pipelines, but also due to the lackof event camera datasets with scene flow ground truth. With the aimof filling this gap, we present E-MOTION: a large and versatile datasetrecorded with high-resolution event cameras suitable for depth, opticalflow and scene flow estimation. E-MOTION features a number of inde-pendent moving objects for which ground truth poses and segmentationmasks are also available. We release a total of 43 sequences with denseground truth maps and poses at 200 Hz. The dataset is available at:https://emotion.hds.utc.fr/

The O(N 2 ) complexity of attention over N tokens remainsa computational bottleneck in transformer models. Vector-Quantized(VQ) attention reduces this to O(M N ) by representing keys with Mcodewords, but applies uniform codebook capacity regardless of whereattention mass concentrates: high-attention regions of key space maybe coarsely approximated while low-attention regions waste representa-tional capacity. We propose Adaptive Vector-Quantized (AVQ) Atten-tion, which adaptively allocates codebook capacity based on attentionimportance. Starting from a small set of codewords, our method identifiesthe most important codes during the forward pass and refines them withpre-learned child codewords, achieving fine-grained quantization whereit matters most while maintaining coarse quantization elsewhere. We de-velop an implementation using custom Triton kernels that enables the fulladaptive refinement process, including importance scoring, child code-word insertion, and parent contribution replacement, to be carried outwithin the tiled computation paradigm of Flash Attention with minimaloverhead. Our approach maintains O(M N ) complexity while achievingimproved accuracy-efficiency trade-offs compared to fixed-codebook VQ-attention.

Adversarial Robustness Distillation (ARD) is the most viableway for transferring defensive capabilities from over-parameterized teach-ers to lightweight student models. However, existing ARD techniques relyon the assumption that the teacher’s decision boundary is a perfect geo-metric oracle. In this paper, we identify a critical gap from the stabilityperspective that such exact matching forces blind inheritance of teacher’ssensitivity noise, punishing the student’s stability even if they are morerobust. To address this challenge, we propose the AEGIS (AdversarialError Gating and Instability Suppression) framework that replaces sim-ple logit matching with gated supervision. It leverages a pair of logit andsensitivity gates to rectify potential errors/instability on the categoricaland gradient-level, respectively. Extensive evaluations on CIFAR-10/100,and Tiny-ImageNet show that AEGIS significantly outperforms state-of-the-art methods, achieving a 3.12% gain under the strong AutoAttackand preserving nearly 50% accuracy under large perturbation intensities(ϵ = 16/255) where traditional baselines collapse. Code is available athttps://github.com/JingqiJi03/AEGIS.

3D shape assembly requires predicting 3D pose by estimatingrotation and translation separately to align objects or fractures. For 3Dpose prediction, prior works utilize the entangled features that capture3D pose information well, but this mixed representation impedes gen-eralization. To address this problem, we propose the SOT encoder thatdisentangles 3D pose into an SO(3)-equivariant (rotation) feature and aT(3)-equivariant (translation) feature. The SOT encoder consists of twobranches: 1) a rotation branch that suppresses translation by project-ing features onto the translation null space, and 2) a translation branchthat suppresses rotation by translation loss and directly subtracts theSO(3)-equivariant representation from its SE(3)-equivariant counter-part. Across a variety of assembly models and diverse datasets, the SOTencoder consistently demonstrates improved generalization performancein 3D shape assembly. Furthermore, in-depth analyses indicate that thedisentangled features show improved equivariant behavior with respectto the target factor while exhibiting invariant behavior with respect tothe other. These results indicate that explicitly encouraging factor dis-entanglement is a straightforward and effective approach for 3D shapeassembly. The code is available at https://github.com/maroo-sky/SOT.

The rapid scaling of large language models (LLMs) and vision-language models (VLMs) has sharply increased compute and storagecosts, motivating extensive research on model pruning. Yet most prun-ing pipelines underuse data scaling during pruning, i.e., increasing thepruning-phase token budget. We revisit L0-regularized structured chan-nel pruning, valued for differentiability, incremental scheduling, and globaldecision-making. We uncover a long-horizon instability: the Lagrangemultiplier grows during training, amplifies pruning gradients, and causeslate-stage oscillations in both gradients and loss. We derive sufficient L2-norm bounds on pruning-specific gradients and reformulate the loss toenforce these bounds per update, replacing dual ascent with a coefficientdetermined by the gradient constraint. Building on this foundation, weintroduce DIGS (Differentiable, Incremental, Global, Scalable pruning),which enables stable and efficient structured channel pruning withoutauxiliary stabilizers while supporting much larger pruning token bud-gets. Applied to the Qwen2.5VL LLM backbone, DIGS scales pruningbeyond 100B tokens and achieves competitive results on public multi-modal benchmarks. Within this setting, larger pruning token budgetsyield monotonic performance gains at fixed sparsity.

Physically based differentiable rendering has proven to be apowerful tool for inverse rendering problems (e.g., 3D reconstruction, re-flectance estimation, lighting estimation). However, most existing meth-ods operate solely on radiometric intensity, discarding valuable polariza-tion cues that constrain scene geometry and material properties. Whileforward simulation of polarized light is well-defined via Mueller-Stokescalculus, extending reverse-mode differentiation to this domain presentssignificant challenges. The rank-deficient nature of common polarimet-ric operators, such as linear polarizers and diffuse reflections, violatesthe invertibility assumptions of standard gradient estimators like pathreplay backpropagation and results in numerical instability. We addressthis by proposing a robust, polarization-aware differentiable path tracingmethod. Our approach estimates unbiased gradients through a combina-tion of path replay and local caching. This formulation enables efficientand stable optimization of material and lighting parameters in complexscenes, broadening the applicability of physically based inverse rendering.Project page: https://vcai.mpi-inf.mpg.de/projects/DPPT/

Latent diffusion models pair VAEs with diffusion backbones,and the structure of VAE latents strongly influences the difficulty of dif-fusion training. However, existing video VAEs typically focus on recon-struction fidelity, overlooking latent structure. We present a statisticalanalysis of video VAE latent spaces and identify two spectral propertiesessential for diffusion training: a channel-wise eigenspectrum dominatedby a few modes, and a spatio-temporal frequency spectrum biased towardlow frequencies. To induce these properties, we propose two lightweight,backbone-agnostic regularizers: Latent Masked Reconstruction and Lo-cal Correlation Regularization. Experiments show that our Spectral-Structured VAE (SSVAE) achieves a 3× speedup in text-to-video gen-eration convergence and a 10% gain in video reward, outperformingstrong open-source VAEs. Code is available at: https://github.com/zai-org/SSVAE.

We present D-Rex, a person-specific framework for photore-alistic, relightable, expressive, and animatable full-body human avatarswith free-viewpoint rendering. Existing methods for relightable full-bodyavatars rely on explicit 3D intrinsic decomposition with analytic re-flectance models, which require accurate geometry registration and care-ful optimization to capture realistic light transport effects. This tightcoupling of relighting with avatar modeling has hindered expressiveness:to our knowledge, no existing method demonstrates strong facial anima-tion alongside relighting, limiting applicability in telepresence, gaming,and virtual production. We propose to decouple relighting entirely fromavatar modeling by treating it as an image-space post-process: a learnedtranslation from flat-lit, albedo-like renderings to a target HDR illumina-tion. To this end, we leverage the strong generative prior of a pre-trainedvideo diffusion relighting model, fine-tuned via LoRA on paired flat-litand relit frames captured in a light stage. The flat-lit driving framesare produced by an independent expressive full-body avatar frameworktrained under white-light conditions, requiring no modification to sup-port relighting, making D-Rex directly applicable to any white-lightavatar system. We demonstrate that D-Rex enables view- and tempo-rally consistent relighting while faithfully preserving expressive motionand fine-grained facial detail, outperforming physically-based relightableavatar baselines. Project page: https://vcai.mpi-inf.mpg.de/projects/DRex/

Bundle adjustment (BA) remains a critical refinement mod-ule for image-based 3D reconstruction and continues to improve geomet-ric accuracy even in learning-based pipelines. However, in low-parallaxand near-rotational regimes, classical Schur-based Levenberg–Marquardt(LM) often becomes ill-conditioned and yields unreliable pose and cali-bration estimates. We propose Gate-Guided CSS-BA, a solver-side mod-ification of Schur-LM that preserves the classical BA objective and trust-region framework while constraining each update to a geometrically in-formed low-dimensional subspace. By integrating Column Space Search(CSS) with geometry-aware gating, the method stabilizes the Schur-LMupdate without altering the estimation problem. In contrast to keyframeor state-selection approaches, all camera and point parameters remainin the optimization problem; only the update direction is restricted. Themethod serves as a drop-in replacement for existing BA pipelines. Exper-iments on both generic and challenging weak-geometry scenarios showmore stable optimization, improved relative pose accuracy, and compet-itive calibration behavior while maintaining reprojection quality.

Referring Remote Sensing Image Segmentation (RRSIS) hasachieved significant progress through the integration of VLMs and theSegment Anything Model (SAM). However, this progress largely relieson strong pre-trained capabilities, while leaving two fundamental limi-tations insufficiently addressed: (1) Architectural Weak-Coupling, wherethe unidirectional flow forces reliance on coarse VLM prompts and wastesSAM’s pixel-level structural guidance, causing localization drift; and (2)Object-Centric Semantic Bias, where models overemphasize dominantobject semantics while remaining insensitive to spatial reasoning cru-cial for RRSIS. Motivated by these observations, we propose CROSS, atightly integrated paradigm for RRSIS. First, we introduce Linguistic-Guided Cascaded Distillation (LGCD) to bridge the architectural gap,which distills SAM’s geometric affinities as soft regularizers into VLM in-termediate layers, injecting dense structural priors to refine localization.Second, Perspective-Spatial Contrastive Learning (PSCL) imposes cross-anchored constraints by mining mask-filtered deceptive distractors andspatial-linguistic counterfactuals as hard negatives, explicitly shatter-ing semantic shortcuts to enforce genuine logical consistency. Extensiveexperiments on RRSIS benchmarks demonstrate that CROSS achievesstate-of-the-art performance and maintains precise localization even un-der severe spatial description perturbations, standing as a robust newparadigm for RRSIS. https://clarence-cv.github.io/CROSS/.

Realistic motion generation from a single example remainschallenging due to the scarcity of large-scale, high-quality motion datasets,particularly for animals and custom characters. Existing approaches mainlyoperate in temporal or spatial domains, where controllability is imposedthrough joint-level or time-based constraints. However, motion signalsexhibit a natural hierarchical structure across frequency scales: low-frequency components encode the fundamental rhythm and movementpattern, while high-frequency components capture individual details andvariations. This observation suggests an alternative control dimensionbased on frequency-selective motion synthesis. We present CMDer (pro-nounced Commander), a Controllable Mode Decomposition-based SingleMotion Synthesis framework that leverages diffusion models to generatediverse motions from a single motion example. CMDer employs Multi-variate Variational Mode Decomposition (MVMD) to represent motionsequences as intrinsic frequency modes, providing an interpretable andstructured representation for motion synthesis. A lightweight U-Net dif-fusion model then learns to generate these modes, while a Mode En-hancement Block (MEB) improves local motion modeling. By perform-ing diffusion directly in the decomposed motion modes, CMDer enablesfrequency-selective control that allows users to preserve global rhythmicpatterns while varying fine-grained motion details. Experiments demon-strate that CMDer achieves generation quality comparable to state-of-the-art temporal-domain approaches while introducing a new control-lable dimension for motion synthesis. We further showcase diverse appli-cations including mode-based synthesis, cross-modality motion genera-tion, and controllable crowd animation.

Continual learning for text-to-image retrieval is essential foradapting retrieval systems to evolving user interests and emerging con-cepts. While existing methods mitigate forgetting through image-textsimilarity distillation, they overlook a critical real-world scenario: newtextual queries often describe different semantic facets of images frompreviously seen visual domains, a phenomenon we term “unbalancedcross-modal shift”. This shift arises from the inherent polysemy of vi-sual content and is strongly associated with severe forgetting. To studythis phenomenon, we introduce two diagnostic metrics: image domainoverlap measured via principal subspace similarity, and textual querydivergence quantified through image-conditioned query probability. Ourempirical analysis shows that unbalanced cross-modal shift significantlyamplifies forgetting, with existing methods struggling to retain imagesemantics under such conditions. To address this challenge, we proposeCaPCL (Caption-Preserved Continual Learning), which regularizes mod-els to maintain consistent caption generation for previously learned con-tent, thereby preserving multi-faceted image semantics. By leveragingauxiliary captions as semantic anchors, CaPCL delivers more robusttraining signals than contrastive similarity alone. We construct a bench-mark integrating five heterogeneous datasets to enable comprehensiveevaluation. Extensive experiments demonstrate that CaPCL consistentlyoutperforms state-of-the-art methods in both knowledge retention andadaptation.

Establishing reliable correspondences across diverse objectinstances is fundamental to robust 3D understanding and generalizablemanipulation. However, existing methods are often inconsistent acrossgeometries or vulnerable to non-canonicalized poses, making correspon-dence transfer unreliable. To address this, we propose XSemanticFlow,which learns correspondences from pure semantic features. Given per-object semantic feature _x001C_elds, XSemanticFlow applies hierarchical intra-object self-attention and inter-object cross-attention to predict soft cor-respondence maps. By restricting cross-attention strictly to semantic to-kens, XSemanticFlow avoids geometric over_x001C_tting, enabling it to predictgeneralizable correspondences across di_x001B_erent topologies and producealigned feature _x001C_elds. This enables XSemanticFlow to learn e_x001B_ectivelythrough self-supervised cross-pose alignment on large-scale unlabeledshapes, followed by supervised cross-instance _x001C_netuning on a cosegmen-tation dataset containing labeled object pairs from PartNet. We eval-uate our method on both 3D understanding and manipulation tasks.For 3D understanding, XSemanticFlow improves self-segmentation withSE(3) augmentations by +19.1 mIoU and +20.6 accuracy, and cross-instance cosegmentation by +10.9 mIoU and +12.5 accuracy over base-lines, demonstrating stronger transform consistency and semantic align-ment. For manipulation, XSemanticFlow provides a reliable alternativeto recent video-generation-based pipelines for zero-shot manipulation.Rather than relying on synthesized visual plans, we directly transfer con-tact regions and end-e_x001B_ector trajectories from a single reference demon-stration to novel instances, bypassing the brittle video generation failurescommon in contact-critical tasks.

Vision transformers (ViTs) have achieved remarkable gen-eralization across visual domains, yet little is known about how theyinternally represent the structure of the world. To address this gap, weuse Cross-Layer Transcoders (CLTs) to read concept circuits from ViTs:directed graphs whose nodes correspond to sparse, interpretable con-cepts and edges capture concept interactions across layers. Our methodyields two complementary views of model behavior. The global conceptcircuit is input-invariant and can be recovered directly from learnedcross-layer weights, exposing the reusable “world knowledge” encodedin the model. The instance concept circuit is input-dependent and iden-tifies the concepts and pathways actually used for a specific prediction,enabling faithful example-level explanations. We demonstrate the util-ity of concept circuits in three ways: (1) Automatic spurious correla-tion discovery: leveraging the statistics of our global concept circuitsto identify shortcut dependencies within the model. (2) Spurious cor-relation removal: intervening on the instance concept circuit to steerthe model towards correct predictions. Empirical results show that ourmethod outperforms existing counterparts by 11.0% on the Waterbirddataset. (3) Model comparison: contrasting the global concept circuits ofdifferent foundation models (e.g., CLIP vs. DINO) to reveal how super-vision paradigms shape representational structure. Our code is availableat https://github.com/deep-real/VisionCLT

Vision-Language Large Models (VLLMs) are promising forAI-generated image (AIGI) detection because they can produce both aprediction and a natural-language output. However, most existing VLLM-based detectors primarily fine-tune the language side while giving lim-ited attention to low-level visual forensic cues. They also often dependon manually crafted prompts or human-annotated rationales, which lim-its scalability. We present UC-VLM, a unified multi-stage frameworkfor AIGI detection that relies solely on binary supervision. UC-VLMfirst identifies effective instruction variants automatically. It then reusesthe same binary label within a multi-stage training framework: (i) a vi-sual discrimination objective that strengthens sensitivity to non-semanticforensic cues, and (ii) a label-conditioned generation objective that usesthe binary label to supervise textual outputs. This design turns weakbinary supervision into a shared supervision signal for both the visualpathway and the language output. Our key novelty is a unified multi-stage binary-supervised framework that consistently reuses the same au-thenticity labels for visual adaptation and label-conditioned text gener-ation, while leveraging automatically optimized instructions to reduceprompt sensitivity without requiring human-written rationales or hand-crafted prompts. Experiments show that UC-VLM achieves 96.1% aver-age accuracy on GenImage, exceeding the strongest prior result by 4.6%,and obtains 69.6% / 77.9% accuracy on Chameleon under ProGAN /SDV1.4 training, surpassing the best baseline by 11.2% / 15.3%, re-spectively.