ECCV 2026 Papers

Camouflaged Object Detection (COD) remains challengingdue to the deliberate alignment of foreground statistics with surround-ing background patterns, which induces strong scene-dependent ambi-guity. Existing approaches typically adopt object-centric modeling orreference-based augmentation. They often treat object appearance andbackground statistics independently, without explicitly modeling theirconditional relationship. This work treats COD as a scene-conditionedpattern-deviation reasoning problem and captures how a camouflagedobject deviates from its background scene. In this direction, we proposea pattern-aware scene-conditioned reasoning (PASR) method with twostages. The first stage constructs a scene-aligned background anchor viabackground-first retrieval, while the second stage performs a conditionedpatch-level deviation reasoning. A reference prototype library is con-structed in an offline and annotation-free manner, where foreground pro-totypes are derived from self-inferred rough masks rather than manuallycurated annotations. Unlike retraining-based paradigms, PASR achievesgeneralization through prototype-level expansion rather than parameteradaptation, enabling scalable inference across diverse background dis-tributions. The resulting deviation maps can serve as priors for genericsegmentation models. Extensive experiments on challenging COD bench-marks demonstrate that PASR outperforms existing unsupervised andweakly supervised methods and significantly narrows the gap with fullysupervised models.

Customized image editing equips diffusion models with spe-cific visual transformations using limited paired data, typically via Low-Rank Adaptation (LoRA). However, combining acceleration and effectLoRAs during inference often introduces interference, causing conceptbleeding, semantic drift, and degraded stylistic fidelity. In this paper, wepropose a unified paradigm that distills numerous customized conceptsand fast-inference capabilities into a single LoRA, effectively resolvingcompositional conflicts while reducing deployment overhead. To addressthe challenges of multi-concept distillation, particularly gradient van-ishing and training instability, we introduce CollectionLoRA, a multi-teacher hybrid distillation framework. Our framework incorporates: (1)Probabilistic Dual-Stream Routing (PDSR) that leverages general-domaindata as regularization to enhance robustness. (2) Automated Asymmet-ric Conditioning (AAC) utilizing Vision-Language Models (VLMs) andorthogonal trigger words to ensure concept isolation, and (3) the De-coupled Hybrid Objective (DHO) that synergizes Trajectory-AnchoredFlow Matching (TA-FM) priors with target simulation and distribu-tion matching, ensuring robust optimization across heterogeneous multi-task distributions. Extensive experiments demonstrate that our unifiedarchitecture mitigates multi-LoRA interference and achieves enhancedconcept fidelity, effective feature isolation, and high-quality rapid syn-thesis. Code is available at https://github.com/Qwen-Applications/CollectionLoRA.

Sim-to-real transfer remains a major obstacle for reinforce-ment learning (RL), especially for vision-based control where image ob-servations exacerbate the state-distribution shift between simulation andthe real world. Domain adaptation (DA) is a promising remedy for thischallenge. Prior sim-to-real DA works have demonstrated encouragingresults, yet these approaches typically assume substantially more targetdata, which is not available in practice. Indeed, their performance de-grades significantly when the target data budget is reduced. To addressthis challenge, we propose AIDA (Adaptive Imagination for DomainAdaptation), a domain adaptation framework for visual reinforcementlearning that addresses sim-to-real transfer under scarce target datawithout requiring additional interaction with the target environment.Our key idea is adaptive imagination: generating reliable and semanticimagination rollouts to augment limited target data. Specifically, AIDAemploys a distribution-shift-aware discriminator that truncates rolloutswhen imagined transitions drift into low-confidence regions, so that onlyreliable transitions contribute to the augmentation. On these reliabletransitions, AIDA introduces a self-consistency loss that cycles throughstate → image observation → state, penalizing discrepancies betweenthe original and reconstructed states. This provides additional adap-tation signals beyond the scarce target data. Our experiments demon-strate that adaptive imagination effectively truncates unreliable rollouts.By enforcing a self-consistency loss on the resulting reliable transitions,AIDA learns semantically meaningful state representations and outper-forms baselines across five MuJoCo tasks and two Gymnasium-Roboticstasks.

Camera pose matters. The position and orientation of eachviewpoint define a shared spatial coordinate frame that relates observa-tions across video frames. Yet this signal is absent from multimodal LLMs(MLLMs) for video understanding, which process frames as isolated 2Dsnapshots, instead of the persistent scene humans perceive. We revisitpose as a lightweight supervisory signal and introduce Cambrian-P , avideo MLLM augmented with per-frame learnable camera tokens and apose regression head. With a carefully designed sampling scheme, themodel achieves substantial gains of 4.5–6.5% on spatial reasoning bench-marks such as VSI-Bench, generalizes across eight additional spatial andgeneral video QA benchmarks, and, as a byproduct, achieves state ofthe art streaming pose estimation on ScanNet. Surprisingly, training onpseudo-annotated poses from in-the-wild video further improves generalvideo QA benchmarks, showing pose helps beyond spatial reasoning. To-gether, these results position camera pose as a fundamental signal forvideo models that reason about the physical world.

Universal photometric stereo (UniPS) aims to recover sur-face normals from images captured under unconstrained illumination,which fundamentally differs from calibrated photometric stereo (CPS),where lighting conditions are known. This gap forces most existing UniPSmethods to adopt fully end-to-end learning that struggles to disentangleillumination and surface geometry due to the lack of explicit physical con-straints. In this paper, we propose Bridge-UniPS, a novel frameworkthat bridges CPS and UniPS through a lighting adapter trained withoutdirect intermediate supervision. The adapter transforms images capturedunder unconstrained illumination into a sequence of image–illuminationpairs directly compatible with CPS models, allowing CPS to act as aphysics-aligned intermediate representation and thereby improving theaccuracy of surface normal recovery. The lighting adapter receives nodirect supervision on Bridge Images or target lighting directions, butis optimized through the final normal-estimation objective with gra-dients propagated through a frozen CPS network. It can still produceCPS-compatible Bridge Images while preserving illumination-geometryconsistency. This behavior suggests that modern learning-based CPSmodels provide useful physical priors to guide the reconstruction. Ex-tensive experiments on multiple benchmarks demonstrate that Bridge-UniPS achieves state-of-the-art performance under unconstrained illu-mination, exhibiting strong generalization and significantly improvingfine-scale surface normal accuracy.

Interactive 3D segmentation aims to extract object masks inpoint clouds with minimal user clicks. Despite recent progress, most ex-isting approaches still struggle with (i) coarse voxel resolution that blursfine boundaries under limited clicks and (ii) hard false positives causedby confusing background structures. These issues are exacerbated bydensity and scale shifts across datasets (e.g., dense RGB-D reconstruc-tions vs. sparse LiDAR scans), where fixed refinement heuristics andpurely click-driven decoding generalize poorly. To address them, we pro-pose NegROI — a novel transformer-based interactive framework thatcouples click-centric multi-resolution refinement with scene-conditionednegative prompts. Given a coarse voxel prediction, it refines only a lo-cal Region Of Interest (ROI) around the current click on a finer gridand fuses refined logits back to the coarse mask. To improve robustnessand efficiency, we introduce uncertainty-driven selective refinement thatprioritizes ambiguous regions. Meanwhile, we model hard backgroundpatterns via a set of scene-conditioned negative prompts obtained bycross-attention over scene tokens. We further stabilize these prompts witha diversity regularizer. Finally, we propose boundary-aware hard nega-tive mining to supervise negative-prompt attention toward boundary-proximal, high-confidence false positives. Our experiments on commonbenchmark datasets (i.e., ScanNet, S3DIS, and KITTI) demonstrate im-proved click efficiency and reduced false positives, with stronger cross-dataset robustness than the state-of-the-art baselines.

Multi-site functional MRI (fMRI) studies are essential forrobust neuropsychiatric diagnosis yet suffer severe domain shifts fromscanner heterogeneity, demographics, and site-specific acquisition pro-tocols. Traditional domain adaptation requires concurrent source andtarget data access, violating clinical privacy regulations. Moreover, func-tional connectivity matrices lie on the Symmetric Positive Definite (SPD)manifold, where Euclidean operations cause geometric distortions cor-rupting diagnostic patterns. We propose BrainRiem, a source-free do-main adaptation framework learning compact Riemannian brain pro-totypes via manifold-aware bi-level optimization. It employs the Log-Euclidean Metric to ensure prototypes remain valid SPD matrices, whileDirichlet Energy spectral calibration aligns their frequency characteris-tics with real brain networks. Only anonymized prototypes are trans-mitted to target sites, serving as stable anchors for training local mod-els without source data access and reducing leakage under the evalu-ated attacks. Comprehensive experiments on ABIDE and REST-meta-MDD show BrainRiem consistently outperforms state-of-the-art source-free, traditional, and graph domain adaptation methods across diversescanners and demographics. Notably, learned prototypes exhibit biolog-ically interpretable connectivity patterns aligning with established neu-roscience findings, validating the necessity of Riemannian geometry forbrain network analysis.

Existing label-deficient concealed object segmentation (LDCOS) methods either rely on consistency constraints within a meanteacher framework or employ the Segment Anything Model (SAM) as a fixed pseudo-label generator. Interestingly, task-specific segmenters and foundation models exhibit complementary failure modes: segmenters lack generalization under scarce labels, while SAM suffers from domain gaps in concealed scenes. A natural solution is to employ co-training to mutually promote both models. However, we observe that standard homogeneous co-training cannot exploit this complementarity because architecturally identical networks tend to share similar error patterns, especially when targets are heavily concealed. To address this, we present SCALER (Segmenter-Coordinated Asymmetric LEaRning), a framework that jointly optimizes a mean-teacher segmenter and a learnable SAM through two alternating phases with model-specific optimization strategies. In Phase I, the segmenter is updated under fixed SAM supervision using entropy-based image-level and uncertainty-based pixellevel weighting to suppress unreliable pseudo-label regions. In Phase II, SAM is updated via an augmentation invariance loss and a noise resistance loss, which exploit SAM’s inherent perturbation robustness rather than treating it identically to the segmenter. This asymmetric design is the key distinction from prior co-training and one-way distillation methods: each model’s learning strategy is tailored to its own inductive bias, enabling mutual enhancement. Experiments across eight LDCOS tasks demonstrate consistent gains, and SCALER improves both the lightweight segmenter and the foundation model, serving as a general plug-and-play paradigm for label-scarce conditions. The code is available at https://github.com/ChunmingHe/SCALER.

Real-world dynamics are inherently compositional: multiple entities move simultaneously within a shared scene, each exhibiting distinct motion patterns. Yet most existing video representations encode motion globally, without explicitly capturing localized motion for individual entities. Crucially, motion is defined relative to a global reference frame, including camera motion and scene layout. However, localized embeddings are often computed from cropped images or obtained by masking features after encoding, discarding the context needed to interpret motion. To address this, we introduce a promptable localized motion representation that produces persistent embeddings for user-specified regions defined by spatial masks. Rather than cropping the input or masking features, our model processes the full video and conditions motion encoding directly on the queried region. This yields temporally consistent, region-addressable embeddings that isolate local dynamics while retaining the global context required for disambiguation. We demonstrate object-level motion transfer, enabling controlled composition of dynamic scenes. Beyond generative control, our embeddings support localized action classification in multi-actor videos. Across both tasks, our approach improves controllability and outperforms global representations localized through cropping or post-hoc masking. Project page: https://compvis.github.io/WhatMoves

Controlling continuous, physically valid 3D surface evolutionis a fundamental challenge in computer vision and graphics. Current neu-ral representations typically rely on unstructured latent spaces or discreteiterations, lacking explicit geometric interpretability and struggling toproduce smooth deformations. While classical scale-space theory offersclear physical interpretability, its numerical irregularity isolates it frommodern deep learning. In this paper, we bridge this gap by establishing alearnable scale-space foundation. Our framework comprises three syner-gistic components: (1) the Variational Neighborhood Curvature (VNC)operator, an efficient, parallelizable, and scale-stable geometric metric,(2) the spacetime-balanced VNC-Flow algorithm, which translates ir-regular physical smoothing into structured deformation trajectories, and(3) a Controllable Variational Autoencoder (C-VAE) that learns thesetrajectories conditioned on normalized evolution time. Departing fromtraditional discrete iterations or unstructured latent interpolations, ourcontinuous neural surrogate induces a structured radial latent organiza-tion. This enables precise control over continuous shape abstraction andestablishes a structural prior for downstream applications.

Chain-of-Thought (CoT) prompting has proven remarkablyeffective for eliciting complex reasoning in large language models (LLMs).Yet, its potential in multimodal large language models (MLLMs) re-mains largely untapped, hindered by the absence of large-scale datasetsthat capture the rich, spatially grounded reasoning intrinsic to visual un-derstanding. Existing visual-CoT resources are typically small, domain-specific, or lack the structured stepwise supervision necessary for com-positional visual reasoning. In this paper, we introduce VisReason, alarge-scale dataset designed to advance visual Chain-of-Thought rea-soning. VisReason comprises 489K annotated examples spanning fourdiverse domains, each featuring multi-round, RoI-grounded rationalesthat guide MLLMs through interpretable visual reasoning steps. Build-ing upon this, we curate VisReason-Pro, a 165K subset produced witha stronger GPT annotator, enriched with detailed reasoning traces anddepth-augmented spatial annotations derived from monocular depth andsegmentation cues. Fine-tuning strong MLLM backbones on VisRea-son and VisReason-Pro yields substantial improvements in step-by-stepvisual reasoning accuracy, RoI localization, interpretability, and fine-grained/spatial reasoning performance. These results demonstrate thatVisReason equips MLLMs with more systematic and verifiable visualreasoning capabilities. We envision VisReason as a cornerstone for cul-tivating human-like visual reasoning, paving the way toward the nextgeneration of multimodal intelligence.

Large-scale video diffusion models achieve impressive visualquality, yet often fail to preserve geometric consistency. Prior approachesimprove consistency either by augmenting the generator with additionalmodules or applying geometry-aware alignment. However, architecturalmodifications can compromise the generalization of internet-scale pre-trained models, while existing alignment methods are limited to staticscenes and rely on RGB-space rewards that require repeated VAE de-coding, incurring substantial compute overhead and failing to general-ize to highly dynamic real-world scenes. To preserve the pretrained ca-pacity while improving geometric consistency, we propose VGGRPO(Visual Geometry GRPO), a latent geometry-guided framework forgeometry-aware video post-training. VGGRPO introduces a Latent Ge-ometry Model (LGM) that stitches video diffusion latents to geometryfoundation models, enabling direct decoding of scene geometry from thelatent space. By constructing LGM from a geometry model with 4D re-construction capability, VGGRPO naturally extends to dynamic scenes,overcoming the static-scene limitations of prior methods. Building onthis, we perform latent-space Group Relative Policy Optimization withtwo complementary rewards: a camera motion smoothness reward thatpenalizes jittery trajectories, and a geometry reprojection consistencyreward that enforces cross-view geometric coherence. Experiments onboth static and dynamic benchmarks show that VGGRPO improvescamera stability, geometry consistency, and overall quality while elim-inating costly VAE decoding, making latent-space geometry-guided re-inforcement an efficient and flexible approach to world-consistent videogeneration. Our project page is here.

Language guided image editing has gained significant pop-ularity, yet existing methods struggle with intricate object intersectionsand fine grained spatial relationships due to the absence of an explicitvisual reasoning process. While Chain of Thought (CoT) has been ex-plored to enhance reasoning, purely textual CoT or coordinate basedprompts are fundamentally limited in representing intricate visual lay-outs and lack the pixel level cues necessary for precise editing. To addressthese challenges, we propose Unlocking Complex Image Editing viaMultimodal Reasoning Edit (MURE), a natively multimodal frame-work that shifts the editing process from purely verbal reasoning to asequence of native interleaved textual and visual rationales. As aunified monolithic model, MURE enables seamless end to end optimiza-tion by sharing a latent space and reusing KV cache, which provides amore flexible intermediate visual process and mitigates the error prop-agation typical of multi tool systems. To further ensure reasoning reli-ability, we introduce the Multimodal Deep Confidence (MMDC)paradigm. This paradigm explores a tree of visual reasoning paths ateach step, pruning low quality branches via a deep confidence score froma reward model to ensure a high fidelity trajectory toward the final re-sult. We define the formal representation for native interleaved chainsand release CoT-Edit 14K, the first dataset comprising 14K high qual-ity multimodal reasoning examples. Extensive experiments demonstratethat MURE establishes a new state of the art in complex editing, achiev-ing an average gain of 7.5% on the MagicBrush benchmark and show-casing superior spatial reasoning capabilitiesThe code and CoT-Edit-14Kdataset are released at https://github.com/zhentao-zou/MURE.

The modern-day surge in popularity of wearable devices posesa fundamentally unique motion capture problem: reconstructing full-body movement from any set of sensing hardware worn at a given mo-ment. Yet, most research efforts assume fixed sensor configurations (e.g.,IMU suits or HMD-centric rigs) and cannot generalize across them.In contrast, we argue that motion capture should prioritize unobtru-sive and lightweight devices such as smartphones, smartwatches, smartglasses, and smart insoles, and study the interplay between them. To thisend, we make three contributions. First, we present a large-scale multi-modal dataset synchronizing these consumer-grade sensors with ground-truth 3D motion, spanning 50 diverse activities including everyday tasks,sports, and social interactions. Second, we propose WHIP, a baseline gen-erative model that reconstructs motion from arbitrary subsets of avail-able sensors, robustly handling missing modalities and producing physi-cally plausible motions. Third, we conduct a systematic study of sensorcomplementarity, quantifying how different modalities complement oneanother. Code and dataset are available at this URL.

Vision-language models (VLMs) have recently been adoptedfor open-vocabulary segmentation by aligning visual features with textprototypes. However, this alignment often degrades under domain shift,as text prototypes encode language priors learned from natural imagesthat may not hold in the target domain. To address this challenge, wepropose VPTA, a framework that adapts text prototypes to the tar-get domain to mitigate language prior bias using only unlabeled target-domain images. VPTA gathers reliable visual evidence from confidentpixels, estimates per-class evidence reliability to drive class-dependentprototype updates, and explicitly preserves prototype separation to pre-vent collapse between similar categories. This lightweight adaptationproceeds through a small number of iterations with negligible computa-tional overhead. We conduct extensive evaluations under diverse domainshifts across three representative domains, including remote sensing, au-tonomous driving, and natural scene segmentation. VPTA achieves state-of-the-art performance on ten remote sensing datasets and consistentlyimproves segmentation performance on autonomous driving and naturalscene benchmarks. These results demonstrate the strong generalizationability of VPTA and highlight its effectiveness as a practical solution forlabel-free prototype adaptation in vision-language models.

Continual Generalized Category Discovery (C-GCD) aimsto incrementally identify novel categories from sequential unlabeled datawhile preserving recognition of known classes, which is an essential ca-pability for open-world visual learning. A major bottleneck lies in am-biguous unlabeled samples that cannot be confidently assigned to knownclasses nor reliably grouped as novel ones, making pseudo-labeling brit-tle and often biasing learning toward familiar categories. In this work,we introduce Virtual Category-Guided Continual Generalized CategoryDiscovery by adapting Virtual Category Learning (VCL) to the contin-ual setting. Our method identifies uncertain samples and assigns them totemporary virtual categories, enabling safe and informative learning fromunlabeled streams without injecting noisy labels, while improving unla-beled data utilization and mitigating prediction bias. To further stabilizediscovery across sessions and enhance class separation, we augment VCLwith Expanded Neighborhood Contrastive Learning (ENCL), which ex-ploits extended neighborhood relations and an adaptive margin to learnmore discriminative and well-separated representations for both old andemerging classes. Extensive experiments on CIFAR-100, Tiny ImageNet,and ImageNet-100 demonstrate that our approach consistently outper-forms state-of-the-art methods, establishing a scalable and effective so-lution for C-GCD. Code is on: https://github.com/Mrxjh105/VC-CGCD

Recent deepfake detection methods leveraging vision foun-dation models (VFMs) like CLIP have made significant progress. How-ever, the abundant semantic information in large training datasets hasmade VFMs highly dependent on semantics. As a result, VFM-basedmethods perform well on images from the same category as the train-ing set but struggle with others. This bias limits generalization in real-world scenarios. To address this issue, we propose the Basis CorrectionNetwork (BCNet), which consists of two modules: the attention-guidedsemantic erasure (ASE), which adaptively identifies and erases seman-tic regions of the image by capturing the model’s semantic attention,and the normalized-gradient perturbation enhancement (NPE) scalesthe gradients of fake samples to concentrated values and adds them asa small perturbation to the original sample, helping the model recog-nize more forgery patterns and improving its ability to distinguish fakefrom real samples. This design ensures the model focuses on the coredistinction between real and fake categories rather than semantic infor-mation. Extensive experiments on 51 AI-generated datasets show thatBCNet achieves 96.7% generalization accuracy on WildRF (collectedfrom real social media) and outperforms other competitors by 8.7% onAIGI-Bench, offering a fresh perspective on semantic generalization indeepfake detection. The code is open-sourced and publicly available athttps://github.com/rstao-bjtu/BCNet.

Fine-grained 3D point cloud segmentation is essential forrobot manipulation and CAD editing. A key unresolved challenge is thatdifferent parts of the same object often share similar local geometry, mak-ing part boundaries difficult to distinguish. Existing methods constructlocal patches via K-nearest search and rely on deeper and larger networksto learn discriminative features. However, these approaches mainly focuson local geometry and lack global structural awareness, which leads toambiguous predictions under sampling noise and partial observations.In this work, we propose Topological Graph Attention Network (To-poGAT), a lightweight, plug-and-play backbone refinement module thatintegrates topological data analysis with Graph Attention Networks tointroduce global structural information into point-wise feature learning.When combined with existing backbones, TopoGAT improves segmenta-tion accuracy with a slight parameter increase. Extensive experiments onfine-grained 3D part segmentation validate the effectiveness of the pro-posed TopoGAT and show up to 1.0% improvement on ShapeNetPartdataset and 3.3% improvement on S3DIS dataset.

Masked modeling has emerged as a robust self-supervisedlearning framework. However, most methods rely on random masking,which disregards the structural properties of different data modalities. Toalign with the spatiotemporal and spectral characteristics of video andaudio data, we introduce a structured noise-based masking approach.By filtering white noise into different color noise distributions, we gen-erate structured masks that capture modality-specific patterns withoutrequiring handcrafted heuristics or access to the data. Our approach en-hances masked video and audio modeling frameworks without any addi-tional computational cost. Experiments show that structured noise mask-ing consistently outperforms random masking, underscoring the value ofmodality-aware masking strategies for representation learning.