ECCV 2026 Papers

Multimodal large language models have significantly advancedzero-shot industrial anomaly detection, yet they remain highly constrainedwhen identifying logical anomalies. Fundamentally, logical anomalies arisefrom discrepancies between inspected products and their predefined con-straints. Current detection paradigms typically operate without accessto these explicit standards, forcing models to rely on pretrained com-mon sense rather than strict logical rules. To address this limitation,we introduce SCAN, a Systematic Criteria-driven ANomaly inspec-tion framework optimized through a two-stage post-training pipeline.We first employ supervised finetuning based on rule-by-rule Chain-of-Thought (CoT) reasoning to systematically verify explicit inspection cri-teria. We then apply reinforcement learning equipped with a rule-levelcredit assignment mechanism to enforce precise anomaly attribution andovercome severe errors caused by sparse feedback. However, such criteria-driven training demands large-scale logical anomaly data paired with ex-plicit inspection rules, yet existing datasets remain largely skewed towardstructural defects. To bridge this gap, we propose FLAW, a Fine-grainedLogical Anomaly dataset With criteria. This dataset leverages imagesynthesis and logical data retrieval to endow images with rich object-levelmetadata, enabling the construction of diverse inspection standards andlogical violations. Extensive experiments on MVTec LOCO demonstratethat our 8B-parameter SCAN model improves upon its base model by7.8% and establishes a new state of the art, outperforming GPT5.2-Chatand Gemini-3.1-Pro by 4.8% and 1.0%, respectively. Source code anddata will be publicly available at https://github.com/tzjtatata/SCAN.

Accurate visible-to-thermal image generation holds signifi-cant application value in fields such as vehicle detection and autonomousdriving. However, existing methods often suffer from the problem thatthe generated results are inconsistent with environmental conditions andvehicle states (e.g., moving/static). Moreover, the scarcity of synchro-nized and calibrated visible-thermal image pairs has further hinderedthe development of these fields. We propose AV2T-Gen, an innovativeenvironment and vehicle guided visible-to-thermal generation frameworkbuilt upon the Instruct-Pix2Pix Diffusion Models for aerial imagery,which distinguishes itself from existing methods by explicitly incorpo-rating environment parameters and vehicle state information for thefirst time. The framework first maps images to the latent space usinga VAE to process image features efficiently, and then generates direc-tional features for vehicles in different environmental conditions usingmultimodal feature information. Meanwhile, we also proposed a visible-thermal paired dataset with environmental conditions and vehicle states.Experimental results have demonstrated that the generated thermal im-ages outperform existing approaches in terms of physical realism andvehicle detail discrimination significantly, with more accurate reflectionof environmental and vehicle characteristics. The code is available athttps://github.com/NPU-CVPG/AV2T.

Few-shot 3D Point Cloud Semantic Segmentation (FS-PCS)aims to predict per-point labels for an unlabeled point cloud, given onlya few labeled examples. To extract representations from the limited la-beled set, existing methods have constructed prototypes with FarthestPoint Sampling (FPS). However, we found that this convention resultsin performance instability due to its sensitivity to FPS-induced varia-tions, while the prototype generation process remains underexplored inthe field. This motivates us to investigate deterministic prototype gen-eration method based on attention mechanism. Despite its potential, wefound that vanilla attention module suffers from the distributional gapbetween prototypical tokens and support features. To overcome this, weprovide a simple approach, White Aggregation and Restoration Module(WARM), which resolves the misalignment by wrapping cross-attentionwith whitening and coloring transformations. Specifically, whitening alig-ns the features to tokens before the attention process, and coloring subse-quently restores the original distribution to the attended tokens. This de-sign enables robust attention, thereby generating prototypes that capturethe semantic relationships in support features. WARM achieves state-of-the-art performance with a significant margin on the S3DIS dataset, andcompetitive performance on the ScanNet dataset. Further experimentsdemonstrate its effectiveness in deterministic prototype generation. Codeis publicly available at: https://github.com/JiyunIm00/WARM.git

Images tell us what a scene looks like, but rarely what itwould feel like to be there. While recent datasets pair visual sceneswith electronic-nose measurements, aligning smell signals with imagesremains challenging because many olfactory cues arise from contextualenvironmental factors that are not directly visible in pixels. We intro-duce SCENT, a multimodal framework that uses language guidance asa semantic bridge between vision and olfaction. Our approach leveragesVision-Language Models (VLMs) to generate scene descriptors captur-ing objects, environmental context, and plausible ambient smell cuessuggested by the visual scene. These descriptors provide semantic guid-ance for learning olfactory representations. We train a smell encoderthat maps electronic-nose signals into a shared embedding space alignedwith both visual and textual representations, and introduce a language-guided latent decomposition that separates object-specific odors fromcontextual environmental contributions. Experiments on the New YorkSmells dataset demonstrate that SCENT significantly improves cross-modal retrieval compared to vision-only baselines, achieving state-of-the-art performance on smell-to-image and smell-to-text retrieval tasks. Inaddition, our framework produces interpretable olfactory representationsthat enable the disentanglement of complex smell mixtures. Our resultsreveal the importance of contextual semantic information for groundingolfactory perception in multimodal learning and pave the way for futureresearch in this area.

Vision–Language Models (VLMs) are increasingly deployedin real-world applications. To ensure efficient inference, these deploy-ments typically rely on specialized system kernels (e.g., CUDA or cuDNN)for acceleration. However, existing system-level research on VLMs has fo-cused primarily on improving kernel efficiency and performance, whilelargely overlooking their potential impact on model security and ro-bustness. In particular, the security implications of inconsistencies be-tween training and inference kernels remain underexplored. To addressthis gap, we present the first systematic study revealing a new classof vulnerabilities: system-level backdoor attacks that exploit floating-point inconsistencies across training and inference kernels. We introduceVLMSysTrojan, a red-teaming framework that constructs models behav-ing benignly under standard training kernels but exhibiting backdooredbehavior when executed on specific target inference kernels. Specifically,VLMSysTrojan produces models whose backdoor triggers remain inactiveunder standard kernels and successfully evade three state-of-the-art back-door detection methods. Yet, when deployed on a target kernel, kernel-specific floating-point behavior activates the backdoor. Empirically, thecompromised models achieve a 99% attack success rate on triggered in-puts while maintaining normal accuracy on clean inputs. Moreover, theattack generalizes across multiple kernel configurations, exposing a previ-ously unrecognized risk in VLM deployment and highlighting the urgentneed for kernel-aware robustness analysis techniques.

End-to-end autonomous driving in urban environments re-quires robust decision-making under partial observability and complexmulti-agent interactions. Severe occlusions and dense traffic at inter-sections limit the perception capability of single-agent systems, moti-vating recent efforts on Vehicle-to-Infrastructure (V2I) cooperation forperception and planning. However, existing evaluation protocols face afundamental trade-off: open-loop evaluation fails to capture error accu-mulation and recovery from deviations, while closed-loop evaluation iscostly, difficult to scale, and often relies on simulated environments thatmay suffer from domain gaps. To bridge this gap, we propose VIPS, abenchmark for cooperative autonomous driving in V2I settings based onpseudo-simulation. VIPS extends pseudo-simulation by integrating ve-hicle and infrastructure observations. This enables scalable yet realisticevaluation of robustness and error propagation without full simulation.We further present CoS-V2X, a cooperative planning framework basedon sparse representations. CoS-V2X models vehicle–infrastructure inter-actions using compact features for efficient communication and robustdecision-making under heterogeneous observations. Code and dataset areavailable at https://vips2026.github.io.

Humans can efficiently parse continuous sensory streams,from hours to years, scaffolding an internal world model that groundsspatial reasoning and prediction. To mimic this capacity, spatial super-sensing [59] challenges multimodal models to move beyond linguistic un-derstanding toward true world modeling. However, their benchmark re-lies on synthetic long videos—formed by concatenating random shortclips—and is mostly limited to household scenes, leaving real-world con-tinuity and diversity underexplored. To address the gap, we introduceVSI-Super-Wild, a large-scale spatial supersensing benchmark builtfrom genuinely long, in-the-wild videos across diverse scenarios. Notably,inspired by cognitive studies on how humans structure experience, wesystematically probe the full triad of world state: the agent (observer),objects (scene items), and the environment (places and global layout).* †Equal contribution. Corresponding author.In total, VSI-Super-Wild comprises 442 real-world long-form videosacross 8 scene categories and 6,980 human-verified question-answer pairs.Evaluating multimodal models on VSI-Super-Wild exposes a funda-mental disconnect: despite advances in static image understanding, mod-els consistently fail at tasks that require coherent world-state trackingover time. We characterize how performance degrades with world-statecomplexity and temporal horizon, and diagnose four failure modes: spa-tial collapse, semantic shortcuts, insufficient update, and instance con-fusion. This taxonomy reveals that models lack the mechanisms to bindobjects, agents, and environments into a unified spatial world model—afundamental gap that defines the path forward for spatial supersensing.

In this paper, we reveal an important yet overlooked problem in image denoising: under signal-dependent camera noise models, dark regions suffer from inherently low Signal-to-Noise Ratio (SNR), as signal intensity decays far faster than noise variance diminishes, making detail recovery in dark areas fundamentally challenging. Yet rather than compensating for this difficulty, MSE-trained denoisers exacerbate it—reconstructing dark pixels up to 6× worse relative to their perband noise floor. This bias stems from two compounding factors: signaldependent noise inflates bright-pixel residuals, and the network’s Jacobian norm increases monotonically with brightness. Together, these cause bright regions to chronically dominate gradient updates at the expense of dark ones. To this end, we propose Brightness Bias-Robust Denoising (BBRD), a drop-in replacement for MSE loss that partitions pixels into brightness bands, normalizes per-band error by empirical noise variance, and applies Group Distributionally Robust Optimization (Group-DRO) to dynamically upweight whichever band is currently worst, with zero additional parameters or inference cost. Across 8 architectures and 2 datasets in our experiments, BBRD is the only method among 13 tested alternatives that improves each brightness band simultaneously, achieving up to +0.45 dB on dark bands, +0.32 dB on bright bands, and +0.65 dB aggregate Peak Signal-to-Noise Ratio (PSNR) on SIDD, with the largest per-band gains in the darkest regions where detail recovery matters most. Code is available at https://github.com/xmedlab/BBRD.

Geolocation encoders, which map geographic coordinates tolearned representations, are emerging as an effective means of capturingvisual and non-visual characteristics from a latitude-longitude pair alone.However, existing approaches project coordinates onto fixed bases (e.g.,spherical harmonics), allocating representational capacity uniformly anddevoting equal resources to the open ocean and to a developing city. Weintroduce Tessellating the Earth (TTE), a location encoder built fromlearnable Spherical Voronoi partitions that concentrates representationalcapacity where it is needed in a fully differentiable, end-to-end manner.Each Voronoi site carries its own embedding and migrates during train-ing toward discriminative areas. To bridge the gap between local spatialstructure and global semantic understanding, we introduce global seman-tic tokens: a set of shared learnable concept tokens that distill semanticknowledge from the satellite imagery into a compact vocabulary the lo-cation encoder can reference at inference, enabling geographically distantsites covering similar environments to share semantics. TTE sets a newstate of the art for location encoders across a suite of geospatial classifi-cation and regression tasks, and achieves the strongest results when usedas a geographic prior for fine-grained species classification on iNaturalist-2018. Code, and weights are available at https://github.com/mvrl/TTE.

Efficiently merging several models fine-tuned for differenttasks, but stemming from the same pretrained base model, is of greatpractical interest. Despite extensive prior work, most evaluations of modelmerging in computer vision are restricted to image classification usingCLIP, where different classification datasets define different tasks. In thiswork, our goal is to make model merging more practical and show itsrelevance on challenging scenarios beyond this specific setting. In mostvision scenarios, different tasks rely on trainable and usually heteroge-neous decoders. Differently from previous studies with frozen decoders,where merged models can be evaluated right away, the non-trivial cost ofdecoder training renders hyperparameter selection based on downstreamperformance impractical. To address this, we introduce the task align-ment proxy, and show how it can be used to speed up hyperparameterselection by orders of magnitude while retaining performance. Equippedwith the task alignment proxy, we extend the applicability of modelmerging to multi-task vision models beyond CLIP-based classification.

Accurate monocular metric depth estimation remains inher-ently ill-posed due to scale ambiguity. Light field cameras mitigate thisambiguity by capturing micro-baseline disparities within a single lensto recover metric depth, but current learning-based solutions are lim-ited by scarce training data and often overfit Lambertian appearancepatterns, restricting real-world robustness. To address this, we proposeSynLF, which alleviates data scarcity via physics-grounded light field(PG-LF) synthesis and performs robust real-world metric depth esti-mation through VisDepth, a prior-initialized iterative estimator with ex-plicit multi-view occlusion reasoning. During training, we synthesize lightfield supervision on-the-fly from RGB-D via physics-grounded light fieldsynthesis, including geometry-consistent view synthesis, depth-dependentdefocus, and non-Lambertian perturbations. At inference, the model isdirectly evaluated on real light field captures without fine-tuning. Ex-tensive experiments demonstrate that SynLF generalizes zero-shot toour custom real-world light field dataset, which features highly chal-lenging conditions such as transparency, reflections, specular highlights,and textureless regions. It achieves 30 mm absolute and 1.4% relativeerrors across 0.5–2.5 m, demonstrating that physics-grounded LF syn-thesis from large-scale RGB-D provides a scalable and storage-efficientalternative to conventional light field acquisition for accurate single-lensmetric depth estimation.

Empowering Large Multimodal Models (LMMs) with imagegeneration often leads to catastrophic forgetting in understanding tasksdue to severe gradient conflicts. While existing paradigms like Mixture-of-Transformers (MoT) mitigate this conflict through structural isola-tion, they fundamentally sever cross-modal synergy and suffer from ca-pacity fragmentation. In this work, we present Symbiotic-MoE, a unifiedpre-training framework that resolves task interference within a nativemultimodal Mixture-of-Experts (MoE) Transformers architecture withzero-parameter overhead. We first identify that standard MoE tuningleads to routing collapse, where generative gradients dominate expertutilization. To address this, we introduce Modality-Aware Expert Dis-entanglement, which partitions experts into task-specific groups whileutilizing shared experts as a multimodal semantic bridge. Crucially, thisdesign allows shared experts to absorb fine-grained visual semantics fromgenerative tasks to enrich textual representations. To optimize this, wepropose a Progressive Training Strategy featuring differential learningrates and early-stage gradient shielding. This mechanism not only shieldspre-trained knowledge from early volatility but eventually transformsgenerative signals into constructive feedback for understanding. Exten-sive experiments demonstrate that Symbiotic-MoE achieves rapid gener-ative convergence while unlocking cross-modal synergy, boosting inherentunderstanding with remarkable gains on MMLU and OCRBench.

Superquadrics have proven to provide a compact, geometri-cally meaningful representation for 3D objects. However, existing meth-ods suffer from limited reconstruction accuracy, are restricted to rigidprimitives, and lack robustness to partial point clouds. In this work,we present SuperFlex, an enhanced framework that expands the expres-sive power and applicability of superquadric decompositions. First, weintroduce a novel loss formulation which significantly improves recon-struction accuracy. Second, we include bending and tapering deforma-tions, enabling high-fidelity representation of curved and asymmetricgeometries. Finally, we leverage these high-quality decompositions assupervision to train a model that is robust to partial real-world pointclouds. Experiments demonstrate substantial improvements in recon-struction accuracy over both optimization- and learning-based baselineswhile maintaining a highly compact primitive representation. Projectpage: https://superflex3d.github.io.Point Cloud

Analytic Continual Learning (ACL) offers a computationallyefficient alternative to gradient-based approaches. Recent ACL methodsare based on Recursive Least Squares (RLS) and have achieved the state-of-the-art results compared to other alternatives. However, they faltersignificantly in Class-Incremental Learning scenarios characterized byLong-Tailed distributions. While the ill-conditioning of the autocorrela-tion (Gram) matrix is a known limitation of RLS, we demonstrate thatclass imbalance exacerbates this issue into a distinct spectral pathology:"tail" classes suffer from severe spectral collapse, rendering their sub-spaces numerically indistinguishable from noise. Standard Ridge Regres-sion (L2 ) fails to address this effectively as it applies isotropic regular-ization - a uniform penalty that is insufficient to stabilize the tail with-out over-shrinking the head. To address this, we propose Geometry-Spectral Rectification (GSR), a theoretically grounded frameworkthat treats long-tailed learning as a spectral regularization problem. Un-like standard isotropic regularization (Ridge) which uniformly penalizesall eigenvalues, GSR acts as an anisotropic spectral filter, selectively in-flating the collapsed eigenvalues of tail classes. We construct a structured,data-dependent spectral perturbation matrix ∆ that selectively inflatescollapsed tail eigen-directions of the Gram matrix. Theoretical analysisproves that GSR guarantees an improved stable rank for the Gram ma-trix, ensuring numerical stability. Extensive experiments show that GSRestablishes a new state-of-the-art for analytic CIL, offering a superiortrade-off between computational efficiency and robust generalization inlong-tailed settings.

Gait patterns play a critical role in human identification andhealthcare analytics, yet current progress remains constrained by small,narrowly designed models that fail to scale or generalize. Building a uni-fied gait foundation model requires addressing two longstanding barriers:(a) Scalability – Why have gait models historically failed to follow em-pirical scaling trends? (b) Generalization – Can one model serve thediverse gait tasks that have traditionally been studied in isolation? Weintroduce FoundationGait, the first scalable, self-supervised pretrain-ing framework for vision-based gait understanding. Its largest versionhas nearly 0.13 billion parameters and is pretrained on 12 public gaitdatasets comprising over 2 million walking sequences. Extensive exper-iments demonstrate that FoundationGait, with or without fine-tuning,performs robustly across a wide spectrum of gait datasets, conditions,tasks (e.g., human identification, scoliosis screening, depression predic-tion, and attribute estimation), and even input modality. Notably, itachieves 48.0% self-supervised rank-1 accuracy on the challenging in-the-wild Gait3D dataset (1,000 test subjects) and 64.5% on the largestin-the-lab OU-MVLP dataset (5,000+ test subjects), setting a new mile-stone in robust gait recognition. These results establish FoundationGaitas a strong and versatile foundation for future gait research. All codeand models: https://github.com/ShiqiYu/OpenGait.

3D instance segmentation methods typically rely on high-quality point clouds or posed RGB-D scans, requiring complex multi-stage processing pipelines, and are highly sensitive to reconstructionnoise. While recent feed-forward transformers have revolutionized multi-view 3D reconstruction, they remain decoupled from high-level semanticunderstanding. In this work, we present SegVGGT, a unified end-to-endframework that simultaneously performs feed-forward 3D reconstructionand instance segmentation directly from multi-view RGB images. Byintroducing object queries that interact with multi-level geometric fea-tures, our method deeply integrates instance identification into the visualgeometry grounded transformer. To address the severe attention disper-sion problem caused by the massive number of global image tokens, wepropose the Frame-level Attention Distribution Alignment (FADA) strat-egy. FADA explicitly guides object queries to attend to instance-relevantframes during training, providing structured supervision without extrainference overhead. Extensive experiments demonstrate that SegVGGTachieves the state-of-the-art performance on ScanNetv2 and ScanNet200,outperforming both recent joint models and RGB-D-based approaches,while showing favorable generalization to ScanNet++. The code is avail-able at https://github.com/IDEA-Research/SegVGGT.

Robustness to natural corruptions remains a fundamentalchallenge for deep neural networks. In this paper, we identify a robustnessfading phenomenon where shallow layers spontaneously develop robustrepresentations and flat loss landscapes in early training, yet these prop-erties are not preserved during standard convergence. To address this,we propose a framework that performs strategic interventions on train-ing dynamics to stabilize the empirically identified early-emergent robustpriors. Our approach includes two parameter-free strategies: Early-PhaseStabilization (EPS) and Asymmetric Weight Reversion (AWR), whichstabilize or recover robust shallow configurations without modifying themodel architecture or introducing learnable parameters. Extensive exper-iments demonstrate the efficacy of our framework across various bench-marks and architectures, yielding significant gains in downstream trans-fer, dynamic adaptation, and diverse computer vision applications.

Latent diffusion models (LDMs) achieve state-of-the-art im-age synthesis, yet their reconstruction-style denoising objective providesonly indirect semantic supervision: high-level semantics emerge slowly,requiring longer training and limiting sample quality. Recent works injectsemantics from Vision Foundation Models (VFMs) either externally viarepresentation alignment or internally by jointly modeling only a nar-row slice of VFM features inside the diffusion process, under-utilizingthe rich, nonlinear, multi-layer spatial semantics available. We intro-duce REGLUE (Representation Entanglement with Global-Local UnifiedEncoding), a unified latent diffusion framework that jointly models (i)VAE image latents, (ii) compact local (patch-level) VFM semantics, and(iii) a global (image-level) [CLS] token within a single SiT backbone.A lightweight convolutional semantic compressor nonlinearly aggregatesmulti-layer VFM features into a low-dimensional, spatially structuredrepresentation, which is entangled with the VAE latents in the diffusionprocess. An external alignment loss further regularizes internal repre-sentations toward frozen VFM targets. On ImageNet 256×256, REGLUEconsistently improves FID and accelerates convergence over SiT-B/2 andSiT-XL/2 baselines, as well as over REPA, ReDi, and REG. Extensive exper-iments show that (a) spatial VFM semantics are crucial, (b) non-linearcompression is key to unlocking their full benefit, and (c) global tokensand external alignment act as complementary, lightweight enhancementswithin our global–local–latent joint modeling framework. The code isavailable at https://github.com/giorgospets/reglue.

We propose Ref-GeNVS, a training-free, reflection-awaremethod for generative novel view synthesis (NVS) in mirror scenes. Ex-isting multi-view diffusion models often fail to recognize the mirror in thescene and cannot exploit reflected content for scene generation. To fix thisissue without additional training, our key idea is to treat a mirror imageas two complementary views. From input images, we estimate the mirrorplane and reflect camera poses to form virtual views. Based on this vir-tual view setup, we propose a two-stage generation method consisting ofMirror-gated attention and Reflection injection, which enables reflection-consistent NVS by explicitly leveraging reflection relationships in a multi-view diffusion model. Ref-GeNVS inherits the strong generalizability ofthe multi-view diffusion backbone, while it does not require finetuning.On synthetic and real scenes including mirrors, Ref-GeNVS outperformsrecent generative NVS methods by generating reflection-consistent andcontextually coherent novel views, revealing scene structure visible onlythrough mirrors. Project page: https://kim-geonu.github.io/Ref-GeNVS/