ECCV 2026 Papers

Feed-forward 3D reconstruction for autonomous driving hasdemonstrated strong performance and promising generalization capabil-ities. However, pixel-level methods often suffer from multi-view inconsis-tencies and content duplication. While voxel-level approaches alleviatethese issues by predicting 3D Gaussian primitives in voxel space, theyremain limited by resolution and computational overhead. We introducean efficient feed-forward framework for scene reconstruction from uncal-ibrated images. The key insight is the explicit decoupling of instanta-neous generation necessity and long-term geometric persistence duringGaussian generation. Specifically, we introduce Generation Confidencedriven by reprojection errors to enable demand-aware Gaussian expan-sion, and Reuse Confidence constructed from depth uncertainty and tem-poral decay to model the long-term stability. To balance fidelity andsparsity, we employ differentiable voxelization to aggregate neighboringfeatures and suppress spatial redundancy, followed by confidence-awaresoft-gating pruning. Gaussian distillation is further introduced to enablea compact set of backbone Gaussians to approximate the full scene repre-sentation. Extensive experiments on multiple datasets demonstrate thatour method achieves strong reconstruction quality and robustness whilemaintaining a highly compact scene representation.

A pivotal step in autonomous driving simulation involves in-serting foreground vehicles with predefined trajectories into simulatedscenes. This process enhances scene diversity and facilitates the creationof various corner cases for testing and improving autonomous drivingmodels. However, existing methods often rely on pre-reconstructed 3Dassets, which frequently lead to lighting inconsistencies between the in-serted foreground and the background. Moreover, the reliance on limited,manually-curated 3D assets hinders large-scale deployment. To addressthese challenges, we propose DriveWeaver, a novel framework for con-trollable vehicle insertion in autonomous driving simulation. Specifically,for a masked target insertion area, DriveWeaver performs video inpaint-ing conditioned on vehicle point clouds to generate high-quality, tempo-rally consistent vehicles. This video-inpainting-based approach ensuresseamless blending between the foreground and background, while thereadily available point cloud conditions enable superior generalization.To support long-term generation, we further design a global-to-local hier-archical inpainting strategy, ensuring the consistent identity and appear-ance of the inserted vehicles. Meanwhile, we extract explicit 3D Gaussianrepresentations of the inserted vehicles through an urban reconstructionpipeline to enable real-time rendering for autonomous driving simula-tion. Extensive experiments across diverse datasets demonstrate thatour method outperforms existing baselines in visual realism and geomet-ric consistency, providing a robust tool for scalable autonomous drivingscene augmentation.

We introduce the Noise-Tilted Reverse Kernel (NTRK), areward-guided diffusion sampler that injects reward gradients through thenoise term, leaving the pretrained reverse kernel unchanged and requiringonly a single sample per step. Reward-guided sampling at inference timehas greatly expanded the versatility of pretrained diffusion models. Yet ex-isting methods face a trade-off. Gradient-based guidance shifts the reversemean, steering generation but pushing intermediate states outside theregion that the model was trained on and degrading quality. Search-basedmethods preserve quality but gain no gradient signal. No prior methodachieves both. NTRK resolves this by keeping the reverse mean fixed andbiasing the noise term toward high reward. This is enabled by a whiteningoperator, the central mechanism behind NTRK, which converts rewardgradients into noise-compatible perturbations without losing their guidingsignal. Across various reward alignment tasks, NTRK outperforms recentstate-of-the-art baselines without losing sample quality. Remarkably, onaesthetic generation, NTRK surpasses the reward of the best baseline at500 NFEs using only 25 NFEs, a 20× reduction in compute.

Although recent robot perception research emphasizes train-ing on data from diverse environments to improve generalization, mostexisting methods still rely on centralized learning, which is inefficient anddifficult to scale across heterogeneous robot platforms. Federated learn-ing (FL) offers an alternative by enabling distributed training withoutraw data transfer, but it suffers from severe performance degradation un-der domain shifts caused by heterogeneity across clients. In real roboticdeployments, data distributions often overlap across platforms, environ-ments, and sensing conditions, making it difficult to partition clientsinto clearly separated domains. However, this characteristic breaks theassumption of clearly separable client domains commonly used in clus-tered FL. To address this gap in robot perception, particularly in depthestimation, we introduce two realistic and unexplored non-IID scenariosthat reflect heterogeneity in terms of platform, environment, and depthdistribution. We then propose FeDepth, a descriptor-based clusteredFL framework that models client relationships through soft clustering.Unlike hard clustering methods that assume clearly separated clusters,FeDepth allows clients to participate in multiple clusters, capturing con-tinuous and ambiguous domain transitions commonly observed in roboticenvironments. Extensive experiments demonstrate that FeDepth con-sistently improves robustness over standard FL and clustered FL base-lines across multiple depth estimation architectures, providing a practicaland effective solution for federated robot perception. Our project pageis available at https://vision3d-lab.github.io/fedepth/.

While 2D di_x001B_usion models have achieved remarkable successin identity-preserving personalization, extending this capability to 3Dassets remains a signi_x001C_cant challenge due to the complexities of multi-view consistency and spatial control. Inspired by these 2D advance-ments, we present a novel personalization method for text-guided 3Dediting that enables compositional, object-level control through naturallanguage. Given a 3D input, we render orthogonal views and extractobject-level segmentation masks to isolate semantic components. Wethen learn distinct token embeddings for each component through a tai-lored two-phase optimization strategy: multi-view textual inversion withattention alignment, followed by full _x001C_ne-tuning of multi-view di_x001B_usionmodel. During inference, these disentangled tokens seamlessly composewith editing prompts to generate multi-view consistent images, whichare subsequently lifted into high-_x001C_delity textured 3D meshes. Extensiveevaluations across diverse editing scenarios demonstrate that our methodsuccessfully transfers the _x001D_exibility of 2D personalization to 3D, achiev-ing state-of-the-art edit faithfulness and identity preservation comparedto existing baselines.

Vision-Language-Action (VLA) models often fail to performthe same learned tasks under environmental shifts, such as changes incamera pose and shifts to a different but similar robot (e.g., from Pandato UR5e). Adapting these models to the shifted environment (i.e., targetdomain) often requires training on multiple demonstrations for each task,which are costly to collect. To reduce the burden of data curation andtraining, we propose an analogy-based method that adapts VLA modelsunder environmental shifts through weight vector arithmetic with domain-specific information addition, named Domain ARiThmetic (DART). Un-like prior approaches, DART requires collecting only a single demonstra-tion, enabling efficient adaptation. To accurately isolate domain-specificinformation for addition, DART performs subspace alignment between sin-gular components in weight vectors to filter out noisy components. In bothsimulated and real-world experiments, DART outperforms existing VLAadaptation methods in one-shot scenarios across diverse visual and embod-iment shifts. Code is available at https://github.com/snumprlab/dart.

Zero-Shot Industrial Anomaly Detection (IAD) is criticalfor rapid deployment in agile manufacturing. We investigate a highlypractical Batch Zero-Shot IAD setting, strictly constrained by smallintra-batch references and common pose variations. Directly applyingexisting mutual-scoring methods fails because training-free features lackgeneralized anomaly semantics and spatial robustness, resulting in am-biguous scoring distances and an “anomaly-to-anomaly shortcut” duringmatching, which can lead to missed detections. To address this, we pro-pose DeMuS (Decoupled Mutual Scoring), a novel two-stage learningframework. By training two lightweight adapters via a two-stage aux-iliary process, our method extracts highly discriminative anomaly se-mantics for accurate scoring (how-to-score), while explicitly decouplingrobust part-level correspondence learning (where-to-match) to guaranteereliable physical alignment. This synergistic mechanism effectively pre-vents semantic confusion under unaligned inputs. Extensive experimentson MVTec AD, VisA, and Real-IAD demonstrate DeMuS achieves state-of-the-art performance and superior robustness against pose shifts, main-taining efficient inference for online inspection. Our code is available athttps://github.com/evoLonation/DeMuS.

In the 3D medical image domain, vision-language pre-training is used to create vision-language encoders (VLEs) that can support radiologists by retrieving patients with similar abnormalities, predicting likelihoods of abnormality, or, with downstream adaptation, generating radiological reports. While the methodology holds promise, three challenges limit the capabilities of current 3D VLEs: data scarcity due to privacy concerns, high computational costs resulting from the volumetric nature of the images, and a domain shift between the long reports used for training and the short prompts used during inference for, e.g., zeroshot classification. As a consequence, natural-image VLE recipes do not directly transfer to 3D medical imaging. In this paper, we overcome these challenges by injecting additional supervision via a report generation objective and combining vision-language with vision-only pre-training, allowing us to leverage both image-only and paired image-text 3D datasets. Further, we propose a novel loss that addresses the domain shift between long reports and short textual prompts. Through these additional objectives, paired with best practices of the 3D medical imaging domain, we develop the Comprehensive Language–Image Pre-training (COLIPRI) encoder family. Our COLIPRI encoders achieve state-of-the-art performance in report generation, semantic segmentation, classification probing, and zero-shot classification. The model weights and inference code are freely available at https: //huggingface.co/microsoft/colipri.

The connoisseurship of antique Chinese porcelain demands extensive historical expertise, material understanding, and aesthetic sensitivity, making it difficult for non-specialists to engage. To democratize cultural-heritage understanding and assist expert connoisseurship, we introduce CiQi-Agent—a domain-specific Porcelain Connoisseurship Agent for intelligent analysis of antique Chinese porcelain. CiQi-Agent supports multi-image porcelain inputs and enables vision tool invocation and multimodal retrieval-augmented generation, performing finegrained connoisseurship analysis across six attributes: dynasty, reign period, kiln site, glaze color, decorative motif, and vessel shape. Beyond attribute classification, it captures subtle visual details, retrieves relevant domain knowledge, and integrates visual and textual evidence to produce coherent, explainable connoisseurship descriptions. To achieve this capability, we construct a large-scale, expert-annotated dataset CiQi-VQA, comprising 29,596 porcelain specimens, 51,553 images, and 557,940 visual question–answering pairs, and further establish a comprehensive benchmark CiQi-Bench aligned with the previously mentioned six attributes. CiQi-Agent is trained through supervised fine-tuning, reinforcement learning, and a tool-augmented reasoning framework that integrates two categories of tools: a vision tool and multimodal retrieval tools. Experimental results show that CiQi-Agent (7B) outperforms all competitive openand closed-source models across all six attributes on CiQi-Bench, achieving on average 12.2% higher accuracy than GPT-5.

Recent advancements in deep learning have actively addressedcomplex challenges within the Computer-Aided Design (CAD) domain.However, most existing approaches rely on task-specific models requir-ing structural modifications for new tasks, and they predominantly focuson point clouds or images rather than the industry-standard BoundaryRepresentation (B-rep) format. To address these limitations, we proposeBrepCoder, a unified Multimodal Large Language Model (MLLM) thatperforms diverse CAD tasks from B-rep inputs. By leveraging the codegeneration capabilities of Large Language Models (LLMs), we convertCAD modeling sequences into Python-like code and align them withB-rep. We then adopt a two-stage training strategy: First, pre-trainingon reverse engineering to learn geometric features and design logic. Sec-ond, effectively extending the model to various downstream tasks such ascompletion, error correction, and CAD-QA. Consequently, by interpret-ing B-rep as structural code, BrepCoder achieves superior generalizationacross diverse tasks, demonstrating its potential as a general-purposeCAD agent.

Text-driven Referring Video Object Segmentation (RVOS)aims to locate and segment target objects in videos given natural lan-guage. However, existing models are typically trained on 2D image orvideo datasets with naive segmentation losses, which overlooks the geo-metric consistency across frames and leads to weak spatial understand-ing. In this paper, we propose Geometry-enhanced Language-guidedVideo segmentation (GeoLaV), a two-stage framework that distills 3Dgeometric knowledge from images to enhance text-driven video segmenta-tion. In the first stage, we perform monocular geometry pretraining withmonocular novel-view synthesis, enabling the model to acquire geometry-consistent visual representations via spatial alignment on large-scale single-image datasets. In the second stage, we introduce geometry-aware distil-lation and fine-tune the model on video segmentation datasets, transfer-ring 3D structural knowledge from a general 3D prior model. This processreinforces 3D awareness and improves both spatiotemporal coherence andlanguage grounding in segmentation. Extensive experiments show thatour method using only image segmentation data already provides notablezero-shot generalization in RVOS. When combined with geometry-awaredistillation for fine-tuning on videos, our method achieves state-of-the-artperformance across multiple RVOS benchmarks. The code is available athttps://github.com/Tony1882880/GeoLaV.

Self-supervised learning (SSL) has significantly advanced 3Dmedical image analysis. However, existing pretraining methods typicallytreat multi-phase scans as isolated samples, neglecting the rich anatom-ical correspondence inherent in paired acquisitions. Clinically, contrast-enhanced CT (CECT) offers a structure-amplified view of the exactanatomy captured in a paired non-contrast CT (NCCT), yet the drasticintensity shifts caused by contrast agents make direct voxel-wise align-ment challenging. To harness this naturally occurring supervisory signalwhile decoupling structural topology from phase-specific appearance, wepropose CAST (Cross-phase Alignment of Structure and Topology),a novel paired pretraining framework. CAST employs a 3D CNN archi-tecture to explicitly align NCCT representations with CECT targets.Moving beyond conventional reconstruction, we introduce two feature-level constraints: (1) a Spectral Consistency module that utilizes3D wavelet decomposition to align frequency-aware boundaries whilesuppressing contrast-induced noise; (2) a Geometry-Aware Topo-logical Consistency module that preserves local relational graphsamong salient anatomical keypoints via dynamic top-hat sampling. Tosupport this, we construct a large-scale dataset comprising 13,850 pairedvolumetric CT scans. Extensive evaluations on public benchmarks andreal-world clinical cohorts demonstrate that CAST achieves state-of-the-art performance. Notably, our method delivers substantial gains inboundary-sensitive segmentation tasks and low-data regimes, proving itsefficacy in learning robust, phase-invariant anatomical representationsfor both NCCT and CECT downstream applications. Code is availableat https://github.com/alibaba-damo-academy/CAST.

We address the inverse problem of blind 3D scene recon-struction from extremely motion-blurred images, a scenario where tradi-tional Structure-from-Motion (SfM) pipelines fail. Existing approachestypically circumvent this bottleneck by relying on impractical sharp-imagesupervision. In this work, we introduce PRISM3D, a unified frameworkenabling robust reconstruction directly from severely degraded inputs.To overcome the lack of a reliable starting point, we propose a RobustInitialization strategy utilizing deep dense tracking method (VGGSfM)to recover global topology where feature matching fails. To the best of ourknowledge, we are the first to effectively leverage this paradigm to boot-strap 3D Gaussian Splatting from extreme motion blur. However, whilerobust, this initialization yields sparse and noisy geometry that causes de-terministic optimization to diverge. To resolve this, we propose a coupledsolution driven by probability and physics: we adopt a probabilisticformulation for geometric densification via Markov Chain Monte Carlo(MCMC) to robustly populate the sparse priors, while simultaneouslymodeling physical image formation via continuous Bézier Trajectories.Furthermore, while PRISM3D establishes a highly robust standalonepipeline, the availability of complementary event streams offers an op-portunity to push the reconstruction fidelity further. To exploit this, weintroduce PRISM3D-E, a multi-modal (RGB + Events) extension thatseamlessly integrates high-temporal-resolution events as structural priorsto maximize geometric recovery. Because existing datasets lack pairedevent streams under such severe degradation, we concurrently contributethe PRISM3D-E Benchmark to facilitate rigorous evaluation. Exten-sive experiments demonstrate that both our standalone RGB frameworkand its multi-modal extension establish new state-of-the-art performance.To facilitate future research in extreme motion deblurring, our projectpage, featuring extensive qualitative comparisons, video results, and theupcoming releases of our codebase and benchmark dataset, is availableat https://gopirajumatta.github.io/PRISM3D/.

Vision–language models (VLMs) achieve strong performanceon general-domain tasks but can degrade when visual distributions, ter-minology, and reasoning requirements diverge from general-domain pre-training assumptions, including in scientific imaging. We study this chal-lenge in materials-science transmission electron microscopy (TEM), adata-scarce setting that requires nanoscale visual grounding, modality-specific interpretation, and specialized materials-science knowledge. Weintroduce ATOMIC (Assistant for TEM Oriented Multimodal Instruc-tion and Conversation), a TEM-oriented vision–language model adaptedusing peer-reviewed, literature-derived data without manual annotation.Our pipeline extracts figure–caption pairs, isolates TEM subfigures, andgenerates multimodal instruction data via GPT-based instruction gen-eration. The training strategy separates two complementary supervisionsignals: VisionGround, which enforces image-only grounding, and Do-mainContext, which injects caption-conditioned scientific context. Theircombination, Blend, yields consistent improvements across perception-and knowledge-intensive queries. ATOMIC-7B-Blend achieves 75.2% ac-curacy on TEM-MCQ and 37.5% recall-oriented Answer Word Cover-age (AWC) on TEM-VQA, narrowing the gap to GPT-4o on AWC(40.4%) while enabling local deployment. On the TEM subset of MatCha,an independent external benchmark, ATOMIC-7B-Blend+FT outper-forms LLaVA-v1.5-7B+FT by 18.1 percentage points under identicalfine-tuning conditions. We release TEM-MCQ/TEM-VQA annotations,article URLs, figure IDs, normalized crop coordinates for sub-figure re-construction, the data-curation pipeline, and model weights to supportreproducibility within copyright constraints. These results show thatstructured literature-derived supervision offers an effective data-centricstrategy for adapting VLMs to materials-science TEM and may provide areproducible foundation for future studies of VLM adaptation in relateddata-scarce scientific-imaging domains.

Accurate, dense depth estimation is crucial for robotic per-ception, but commodity sensors often yield sparse or incomplete measure-ments due to hardware limitations. Existing RGBD-fused depth comple-tion methods learn priors jointly conditioned on training RGB distri-bution and specific depth patterns, limiting domain generalization androbustness to various depth patterns. Recent efforts leverage monoculardepth estimation (MDE) models to introduce domain-general geomet-ric priors, but current two-stage integration strategies relying on explicitrelative-to-metric alignment incur additional computation and introducestructured distortions. To this end, we present Any2Full, a one-stage,domain-general, and pattern-agnostic framework that reformulates com-pletion as a scale-prompting adaptation of a pretrained MDE model.To address varying depth sparsity levels and irregular spatial distribu-tions, we design a Scale-Aware Prompt Encoder. It distills scale cuesfrom sparse inputs into unified scale prompts, guiding the MDE modeltoward globally scale-consistent predictions while preserving its geomet-ric priors. Extensive experiments demonstrate that Any2Full achievessuperior robustness and efficiency. It outperforms OMNI-DC by 32.2%in average AbsREL and delivers a 1.4× speedup over PriorDA with thesame MDE backbone, establishing a new paradigm for universal depthcompletion. Codes and checkpoints are available at https://github.com/zhiyuandaily/Any2Full.

Post-training via Supervised Fine-Tuning (SFT) and Rein-forcement Learning (RL) is crucial for enhancing reasoning in Multi-modal Large Language Models (MLLMs), yet existing paradigms of-ten reach a performance bottleneck due to the limitations of staticdata. While current methods leverage self-reflection or self-evolution topush these boundaries, they still suffer from cognitive drift and hallu-cinated reasoning paths caused by low-quality synthetic data. To ad-dress these challenges, we propose Anchor E volution (AnE), a newparadigm that integrates truth-anchored data curation and model evo-lution, achieving faithful and steady performance gains at the reason-ing frontier. Specifically, we propose Truth Anchor Expansion, whichpinpoints the model failing frontier via trajectory rollouts and lever-ages ground-truth databases to retrieve high-fidelity anchors for faithfuldata curation. Subsequently, we introduce the Scaffold-Stripping Mech-anism to internalize reasoning capabilities. This mechanism first an-chors reasoning paths via scaffold-augmented supervision to mitigatethe learning complexity and distribution drift of direct SFT on rawdata, then leverages RL to strip the scaffold template, thereby effec-tively transitioning the reasoning paths into intrinsic model capabilities.Experimental results on multimodal reasoning benchmarks show thatour method substantially advances the model performance frontier, im-proving the base model by 10.3% across eight multimodal benchmarksand achieving state-of-the-art results. The code will be publicly availableat https://github.com/wangzehao0704/AnE.

Adversarial attacks pose a challenge to the reliability of deeplearning models, motivating effective detection methods. Existing tech-niques often rely on attack-specific assumptions, access to adversarialsamples, or knowledge of the underlying classifier (white-box). We pro-pose A4 D (Attack- and Architecture-Agnostic Adversarial Detector), acompletely black-box, zero-shot adversarial attack detection frameworkthat utilizes prompt-based similarity scores derived from CLIP. To thebest of our knowledge this is the first attempt to utilize CLIP for sucha task. The method is based on two key observations: (i) CLIP is sen-sitive even to small imperceptible non-semantic perturbations; (ii) Theshift in CLIP embedding space is not arbitrary and can be used as arobust attack indicator. Experiments across multiple attacks, datasetsand classifiers validate that A4 D achieves SOTA detection results in theattack-agnostic and classifier-agnostic setting.

Intraoperative 2D/3D registration, which aligns live X-rayimages with preoperative volumes, is essential for image-guided interven-tions. Previous regression-based methods suffer from limited generaliza-tion, thus requiring time-consuming patient-specific retraining. Inspiredby recent geometry foundation models such as DUSt3R, we proposeXPos3R, a generalizable pose regression method that eliminates preop-erative preparation. Unlike existing geometry models designed for homo-geneous inputs, XPos3R extends this paradigm to multi-modal inputs,namely 2D X-rays and 3D volumes. Specifically, we introduce an asym-metric encoder–decoder architecture that improves cross-modal featurealignment while maintaining computational efficiency. To scale trainingunder limited medical data, we adopt an anatomy-specific data cura-tion strategy and construct million-scale synthetic datasets. Evaluatedon real-world benchmarks, a single pretrained XPos3R surpasses patient-specific methods in both accuracy and robustness. With test-time opti-mization completed in seconds, it further reduces the 3D error to <4 mmand the reprojection error to <1 mm. The strong generalization, accu-racy, and efficiency of XPos3R highlight its clinical potential, while itsasymmetric framework may inspire broader cross-modal vision geometrytasks.

Population-level morphometric measurements are central toecological and evolutionary studies, but traditionally require controlledimaging or physical specimen handling, limiting their scalability. Wepresent WildProp, a framework for estimating wildlife body-proportiondistributions directly from large-scale, unconstrained image repositories.We cast morphometric estimation as a retrieval-driven correspondenceproblem: given a single user-annotated canonical image, WildProp per-forms pose-aware retrieval using foundation model features, transferspart endpoints via dense patch-level matching, filters predictions us-ing geometric consistency, and aggregates measurements across retrievedimages to estimate length-ratio distributions. Unlike supervised keypointpipelines, our approach adapts to arbitrary species and user-defined bodyparts without per-species training. Evaluations on three large morpho-metric datasets spanning birds and amphibians show median relative er-rors of 10–20%. We further demonstrate broad applicability through casestudies measuring proportions across diverse taxa, including birds, frogs,insects, and flowers. Ablations show that pose-aware retrieval is criticalfor stable estimation, while robust aggregation mitigates keypoint andpose noise. Our results suggest that carefully filtered 2D correspondencesover web-scale imagery can provide scalable morphometric proxies forcomparative analyses across taxa, geography, and seasonality.