ECCV 2026 Papers

Although Multimodal Large Language Models (MLLMs) havedemonstrated impressive performance in Medical Visual Question An-swering (Med-VQA), their reliance on global image features often lacksprecise pixel-level grounding, thereby limiting clinical trustworthiness.To bridge the semantic gap between high-level clinical reasoning andspatial localization, we propose MedREAL (Medical REasoning-drivenAnswering and Localization), a unified framework that seamlessly alignslinguistic reasoning with spatial grounding. Specifically, MedREAL in-troduces Seg Anchored Reasoning Pooling (SARP) to distill task-relevantsemantic evidence directly from [SEG] tokens within the MLLM’s hid-den states. Furthermore, a Reasoning-to-Visual (R2V) fusion mecha-nism is proposed to effectively inject these reasoning-aware features intoa segmentation pipeline for accurate mask decoding. To facilitate thisparadigm, we construct MedRAVS-13K, a comprehensive dataset com-prising 13,824 expertly validated samples across four diverse imagingmodalities. Extensive experiments demonstrate that MedREAL signifi-cantly outperforms state-of-the-arts, achieving 68.49% gIoU and 70.47%cIoU on benchmark evaluations. By generating evidence masks that arestrictly consistent with textual diagnoses, MedREAL provides a robust,interpretable framework for reasoning-driven medical image analysis.

Computed tomography report generation (CTRG) automatesradiology reporting to reduce clinical workload and facilitate patient care.Recent efforts in applying the rapidly developing large language mod-els (LLMs) have advanced the field; yet, they still face a fundamentalchallenge due to the large volume of 3D data: effectively reducing highfeature redundancy and computational burden while simultaneously ex-tracting information-rich representations. To address this challenge, thiswork presents a novel CTRG framework that fully imitates the coarse-to-fine visual search pipeline practiced by radiologists. Our frameworkfirst employs a ViT-based image-text alignment architecture to extracta global token and local patch tokens for each anatomical structure, en-hanced by a mask-guided sparse negative-entropy loss. An abnormality-proposal LLM then processes the global token and clinical metadata topropose a shortlist of candidate abnormalities. This shortlist promptsan abnormality-prompted local token filter (AP-LTF) to select the mostinformative patch tokens, effectively reducing redundancy while preserv-ing critical information. Finally, a report-generation LLM takes in theglobal token, selected local tokens, and clinical metadata to composea full report. During training, ground-truth abnormalities are prefixedto the reference report to enhance awareness of clinic-relevant findings,and group relative policy optimization (GRPO) aligns the abnormality-proposal and report-generation LLMs for collaborative efficacy. Exper-imental results on two public CTRG datasets demonstrate the supe-rior performance of our framework compared to existing state-of-the-artFig. 1: Comparison of token compression strategies for CTRG. (a) 3D spatial pooling[2, 8] achieves high compression rate but incurs significant information loss. (b) Mask-guided volume cropping [12, 22] yields region-specific tokens but omits intra-regionredundancy. (c) [26] selects the top K tokens with the largest attention weights toregion-specific queries, yet may be subject to dominant “sink tokens” [14, 29]. (d) Ourmethod first proposes a shortlist of candidate abnormalities using only a single globaltoken per structure and clinical metadata. Then, an abnormality-prompted local tokenfilter (AP-LTF) combines the shortlist, a learned score, and the attention weight for acomprehensive token selection, thereby preventing sink tokens from dominating.methods in terms of clinical efficacy, RaTEScore, and GREEN scores.Ablation studies further validate the effectiveness of its novel design.

Recent advances integrate physically grounded Newtonian dynamics with neural rendering frameworks, narrowing the gap between photorealistic scene reconstruction and physics-based animation. However, existing approaches focus on mechanically driven dynamics while neglecting temperature, a fundamental yet invisible physical factor underlying phenomena such as melting, solidification, and other thermomechanical processes. In this paper, we propose MeGAS, a novel framework that incorporates thermomechanical phase-change dynamics into 3D Gaussian Splatting (3DGS). Specifically, we propose a new thermomechanical dynamic Gaussian Splatting representation that augments 3DGS with temperature attributes and employs a heat advection-diffusion solver with MPM dynamics incorporating phase transitions, enabling physically plausible and visually realistic synthesis of thermophysical phenomena. Furthermore, a new topology-adaptive Gaussian rendering strategy is proposed to mitigate cracking and floaters under extreme deformation. Extensive experiments demonstrate that MeGAS produces physically consistent thermomechanical behavior while maintaining highfidelity photorealistic rendering, advancing toward physics-integrated world models.

Despite notable progress in text-guided medical image seg-mentation nowadays, these methods are limited to single-round dia-logues and fail to support multi-round reasoning, which is importantfor medical education scenarios. In this work, we introduce Multi-Round Entity-Level Medical Reasoning Segmentation (MEMR-Seg), anew task that requires generating segmentation masks through multi-round queries with entity-level reasoning, helping learners progressivelydevelop their understanding of medical knowledge. To support this task,we construct MR-MedSeg, a large-scale dataset of 177K multi-roundmedical segmentation dialogues, featuring entity-based reasoning acrossrounds. Furthermore, we propose MediRound, an effective baselinemodel designed for multi-round medical reasoning segmentation. To mit-igate the inherent error propagation within the chain-like pipeline ofmulti-round segmentation, we introduce a lightweight yet effective Judg-ment & Correction Mechanism during model inference. Experi-mental results demonstrate that our method effectively addresses theMEMR-Seg task and outperforms conventional medical referring seg-mentation methods. The project is available at https: // github. com/Edisonhimself/ MediRound .

Knowledge-based Visual Question Answering (KB-VQA) re-quires models to combine image understanding with external knowledge.Most prior methods use a fixed retrieve-then-generate pipeline with a pre-selected retriever and a static top-k setting, which is not adaptive duringreasoning. We propose ProMSA, a progressive multimodal search agentfor KB-VQA. Given an image-question pair, the agent iteratively choosesimage search, text search, or stop, under explicit tool-call budgets andwith deduplication to avoid redundant retrieval. For training, we first userejection-sampling SFT to learn valid tool-use formats, then optimize theagent with TN-GSPO, a sequence-level RL objective that normalizesupdates by both generation length and tool-interaction depth. Experi-ments on E-VQA and InfoSeek show consistent gains over strong RAGand agent baselines, and improved retrieval and end-to-end accuracy.Thecode is available at https://github.com/DingWu1021/Promsa.

Video-LLMs often attend to irrelevant frames, which is espe-cially detrimental for sports coaching tasks requiring precise temporal lo-calization. Yet obtaining frame-level supervision is challenging: expensiveto collect from humans and unreliable from other models. We improvetemporal grounding without additional annotations during training byexploiting the observation that related tasks, such as generation andverification, must attend to the same frames. We enforce this via a self-consistency objective over select visual attention maps of tightly-relatedtasks. Using VidDiffBench, which provides ground-truth keyframe an-notations, we first validate that attention misallocation is a meaningfulbottleneck. We then show that training with our objective yields gains of+3.0%, +14.1% accuracy and +0.9 BERTScore over supervised finetun-ing across three sports coaching tasks: ExAct, FitnessQA, and ExpertAF.

In the real world, fashion is about layering: adding a jacketover a shirt, or a sequence of adding and removing layers, rather thanjust a single-layer swap. This fundamental real-world task remains achallenge in existing Virtual Try-On (VTON) methods, which excel atsingle-layer replacement but are not designed to layer or de-layer anexisting outfit. This paper proposes Layering Virtual Try-On (LVTON),a layering benchmark and method that preserves an existing outfit whileenabling sequential layering. We find that current VTON paradigms arefundamentally ill-equipped for LVTON, as their reliance on cloth-agnosticrepresentations and single-item datasets discards essential layering context.Our key insight is that the LVTON challenge must be disentangled intotwo distinct competencies: (1) General VTON Priors (e.g., deformation,identity preservation) and (2) Specific Layering Knowledge (e.g., layeringorder and occlusion reasoning). First, our model obtains general VTONpriors by being trained on data produced by an automatic data generationpipeline that synthesizes samples from fashion videos via segmentationand inpainting. Second, the model is fine-tuned on a small, dedicatedLVTON dataset to learn the layering logic. Our method achieves state-of-the-art results on our LVTON benchmark and demonstrates superiorgeneralizability on traditional VTON benchmarks, setting new state-of-the-art results when fine-tuned and exhibiting zero-shot capabilities.

Lookup table (LUT)-based image denoising methods haveattracted increasing attention due to their high efficiency and hardware-friendly properties. However, existing RGB-LUT approaches require threeidentical LUTs to process RGB channels in parallel, resulting in largeon-chip SRAM consumption. A simple alternative is to apply LUT pro-cessing only to the luminance (Y) channel in the YUV color space toreduce memory usage. However, this naive strategy leads to degradedrestoration quality, since ignoring the chrominance (UV) channels in-troduces color distortion and residual artifacts. In this work, we pro-pose Hybrid-LUT, a YUV-based asymmetric channel-processing frame-work that combines LUT and filtering in a unified design. Specifically, amulti-band LUT branch with pixel-level weight fusion is applied to theY channel to recover fine textures, while lightweight filtering is used forthe UV channels to maintain color consistency. This design reduces LUTstorage by two-thirds compared with RGB-LUT methods while main-taining the same runtime throughput. Extensive experiments show thatHybrid-LUT achieves state-of-the-art (SOTA) performance across mul-tiple benchmarks with only 421 KB of storage. In particular, our methodsurpasses existing LUT-based denoising approaches by at least 0.63 dBCPSNR on real-world datasets, demonstrating its effectiveness for imagedenoising on resource-constrained edge devices.

We present HUGE-Bench, a benchmark for High-LevelUAV Vision–Language–Action (HL-VLA) tasks that tests whether anagent can interpret concise language and execute complex, process-orientedtrajectories with safety awareness. HUGE-Bench comprises 4 real-worlddigital twin scenes, 8 high-level tasks, and 2.56M meters of trajecto-ries, and is built on an aligned 3D Gaussian Splatting (3DGS)–Mesh hy-brid representation that combines photorealistic rendering with collision-capable geometry for scalable generation and collision-aware evaluation.We introduce process-oriented and collision-aware metrics to assess pro-cess fidelity and flight safety. Experiments on representative state-of-the-art VLA models reveal significant gaps in high-level semantic completionand safe execution, highlighting HUGE-Bench as a diagnostic testbed forhigh-level UAV autonomy.

While Large Vision-Language Models (LVLMs) have signifi-cantly advanced GUI agents’ capabilities in parsing textual instructions,interpreting screen content, and executing tasks, a critical challenge per-sists: the irreversibility of agent operations—where a single erroneous ac-tion can trigger catastrophic deviations. To address this, we propose theGUI Action Critic’s Data Flywheel System (GAIA), a training frame-work that enables the models to have iterative critic capabilities, whichare used to improve the Test-Time Scaling (TTS) of basic GUI agents’performance. Specifically, we train an Intuitive Critic Model (ICM)using positive and negative action examples from a base agent first. Thiscritic evaluates the immediate correctness of the agent’s intended actions,thereby selecting operations with higher success probability. Then, theinitial critic guides agent actions to collect refined positive/negative sam-ples, initiating the self-improving cycle. The augmented data then trainsa second-round critic with enhanced discernment capability. We conductexperiments on various datasets and demonstrate that the proposed ICMcan improve the test-time performance of various closed-source and open-source models, and the performance can be gradually improved as thedata is recycled. The code, dataset, and accompanying datasheet will bepublicly released at https://github.com/SeerRay-Lab/GAIA.

Unsupervised Anomaly Detection (UAD) excels in real-worldapplications especially with rare anomalies due to the elimination of la-beled defective samples. Most existing UAD algorithms adopt a compar-ison process where the features of the test image are compared againsttheir normal “prototypes”. Such a paradigm relies primarily on raw fea-tures from pre-trained models and is inefficient due to the excessive num-ber of prototypes. To enhance UAD performance, recent algorithms aimto establish a mapping from the test features to their respective proto-types . However, the proposed mapping models, typically based on distil-lation or attention mechanisms, perform the “reconstruction” implicitly,making the process unnecessarily complex and difficult to explain. In thispaper, we propose a simple yet effective scheme, termed Fast DynamicPrototype (FDP), to rapidly generate high-quality reconstructed proto-types for Multi-class UAD. First, a raw prototype bank is obtained viasimply clustering the deep features of normal samples. Secondly, withinthe proposed Dynamic Bank Adapter module, the raw bank is dynam-ically condensed and adapted based on the information of the currenttest image. The reconstructed prototype of a given test feature is thendefined as the weighted average of its nearest bank members. Finally,an auxiliary FDP utilizing a different backbone model is introduced tosuccessfully mitigate the impact of artifacts present in the primary FDP.The study’s extensive experiments demonstrate the superiority of theproposed algorithm, which outperforms all compared SOTA methods onthree well-known UAD datasets (MVTec-AD, MVTec-3D, and ViSA),while attaining a speed up to 477 fps. Our implementation is publiclyavailable at https://github.com/lmlpy/FDP.git.

Offline-trained detectors and matchers often degrade in de-ployment when encountering previously unknown targets, since offlinetraining cannot anticipate the appearance and structural variations ofunknown targets. We propose EVEE, a feature adaptation frameworkthat leverages temporally accumulated event evidence as test-time proxysupervision for adapting detection and matching to previously unseentargets. Rather than performing fully asynchronous event-level match-ing or using events merely as auxiliary inputs, EVEE constructs event-derived proxy supervision to drive label-free online updates. EVEE con-tains three components. (i) Training Warehouse (TraW) accelerates con-vergence within a limited observation window via multi-threaded updatesand best-weight broadcasting. (ii) Event Voting Bank (EVB) accumu-lates bidirectional positive/negative votes to separate foreground frombackground, thereby reducing the impact of background noise on themodel. (iii) ReWeight improves correspondence reliability by rescoringcandidates using appearance similarity, geometric consistency, and EVBpriors. Experiments on public datasets and a real RGB–event platform,show that EVEE achieves 1.9% higher average pose ACC/AUC on un-known targets than strong baselines. Meanwhile, it runs at 51.3 FPSfor matching and 14.52 FPS for online adaptation. The code is publiclyavailable at https://github.com/ZeJZhao/EVEE.

Recent advances in text-to-video (T2V) diffusion models have demonstrated remarkable generative capabilities, yet their reliance on loosely curated training data raises pressing safety and copyright concerns. Concept erasure offers a principled remedy by removing unwanted semantics from pretrained models while preserving remaining concepts. However, existing approaches typically operate at a coarse granularity misaligned with the fine-grained, distributed nature of concept representations, leading to incomplete removal or degraded generation quality. We argue that surgical erasure fundamentally requires intervention at the level of monosemantic features, where each unit encodes a single interpretable concept. To this end, we propose EraseSAE, a novel framework that leverages sparse autoencoders to achieve surgical concept erasure in DiT-based T2V diffusion models via a principled decompose-attributeerase pipeline. We first introduce the Partitioned Convolutional Sparse Autoencoder, which decomposes dense spatiotemporal activations into disentangled, interpretable sparse features while preserving spatiotemporal coherence. A contrastive attribution mechanism then contrasts activations from paired prompts to isolate concept-specific feature kernels. At inference, timestep-resolved spatiotemporal masks derived from the identified kernels confine erasure to regions where the target concept is active, leaving unrelated content intact. Extensive experiments across diverse diffusion models and concept erasure tasks demonstrate that EraseSAE achieves precise and robust concept removal with minimal quality degradation, substantially outperforming state-of-the-art methods. The code is available at https://github.com/HiDream-ai/EraseSAE.

Object removal must prevent the masked target from reap-pearing and reconstruct the occluded background with structural andcontextual fidelity, rather than merely filling a hole plausibly. Recentdataset-free approaches manipulate the diffusion model’s internal self-attention to prevent it from referencing the masked region, yet they failin two critical ways: (i) they treat the masked region as the sole fore-ground, misinterpreting non-target objects as background and regener-ating them, and (ii) they apply uniform attention constraints withoutdistinguishing diverse background subtypes, leading to textural blurringand structural misalignment. Both failures stem from the absence of ex-plicit background-aware reasoning. We propose EraseLoRA, a dataset-free framework that replaces attention surgery with background-awarereasoning and test-time adaptation. The first stage, Background-awareForeground Exclusion (BFE), leverages a multimodal large-language modelto separate target foreground, non-target foregrounds, and clean back-ground from a single image-mask pair. The second stage, Background-aware Reconstruction with Subtype Aggregation (BRSA), performs test-time optimization that treats inferred background subtypes as comple-mentary pieces, enforcing their consistent integration through recon-struction and alignment objectives without explicit attention interven-tion. As a model-agnostic plug-in applicable to diverse diffusion back-bones, EraseLoRA reconstructs backgrounds at least 23% more faithfulto the original scene than previous dataset-free methods while nearlyhalving unwanted foreground re-generation, and surpasses all dataset-driven approaches in both aspects despite requiring no training data.Code is available at https://shjo-april.github.io/EraseLoRA.

We present E(3)-Pose, a novel fast pose estimation methodthat jointly and explicitly models rotation equivariance and object sym-metry. Our work is motivated by the challenging problem of accountingfor fetal head motion during a diagnostic MRI scan. We aim to enableautomatic adaptive prescription of diagnostic 2D MRI slices with 6-DoFhead pose estimation, supported by rapid low-resolution 3D MRI vol-umes acquired before each 2D slice. Existing pose estimation methodsstruggle to generalize to clinical volumes due to pose ambiguities inducedby inherent anatomical symmetries, as well as low resolution, noise, andartifacts. In contrast, E(3)-Pose captures anatomical symmetries andrigid pose equivariance by construction, and yields robust estimates ofthe fetal head pose. Our experiments on publicly available and represen-tative clinical fetal MRI datasets demonstrate the superior robustnessand generalization of our method across domains. Crucially, E(3)-Poseachieves state-of-the-art accuracy on clinical MRI volumes, supportingfuture clinical translation. Our implementation is publicly available1 .

Scene Change Detection (SCD) evaluates changes using pre-defined query-reference (i.e., present-past) image pairs. However, thisformulation overlooks a critical dependency: the corresponding query-reference pair is assumed to be prepared in advance. In real-world appli-cations, such as mobile robots, future query views cannot be known inadvance, and thus their corresponding reference images cannot be pre-defined. To remove this dependency and push change detection towardmore practical applications, we introduce Environmental Change Detec-tion (ECD). A key aspect of ECD is to avoid unrealistically predefinedand aligned query-reference pairs and instead retrieve environmental cuesfrom an uncurated image database of reference scenes. To tackle thisnew challenging task, we additionally introduce an initial solution thatenables change detection under unknown and imperfect query-referenceconditions. The main idea of our solution is to retrieve multiple referencecandidates and aggregate semantically rich representations for changedetection. We further construct ECD benchmark sets by reformulatingthree standard change detection datasets. Extensive experimental resultsdemonstrate the efficacy of our solution in both ECD and SCD.

Leveraging Multimodal Large Language Models (MLLMs)via contrastive learning has become a mainstream paradigm for improv-ing the performance of Universal Multimodal Retrieval (UMR). However,previous works have ignored the grain blindness when adapting thecontrastive paradigm into retrieval tasks. Grain blindness refers to thetendency of the model to overlook grain-level information contained inthe query, which is crucial for effectively handling complex queries. Thisstems from contrastive learning treating samples as a binary classifica-tion (positive/negative), while ignoring the different information carriedby each negative sample. To address this, we argue that negatives shouldbe treated differently according to their similarity to the positive sample,enabling the model to learn distinct grain information from each nega-tive. In this paper, we introduce a simple but effective framework, calledELVA, a novel rule-based RL framework that mitigates grain blindnessthrough ranking-driven MLLMs. 1) Instead of relying on reward mod-els, we extend Reinforcement Learning with Verifiable Rewards (RLVR)to retrieval tasks, allowing the model to explore new ranking behav-iors without explicit ranking labels. 2) By utilizing rule-based rewards,our approach jointly optimizes the ranking of negative samples whileenlarging the similarity gap between positive and negative. To more pre-cisely measure grain blindness, we further introduce MRBench, a newbenchmark specifically designed for multi-grain query scenarios. ELVAachieves state-of-the-art results across standard retrieval benchmarks,and its notable 13.1% improvement on MRBench further demonstratesits effectiveness in alleviating grain blindness. Our code is available athttps://github.com/SeerRay-Lab/ELVA.

Sports video analysis is a challenging domain for multimodalunderstanding because it involves complex and dynamic human activi-ties. Despite rapid progress in Multimodal Large Language Models (MLLMs),long-horizon reasoning in sports videos remains difficult, as answeringquestions requires both locating and integrating temporally sparse evi-dence into reasoning. We attribute this limitation to two closely relatedfactors: insufficient supervision over temporally dispersed evidence andthe lack of methods for explicit temporal evidence localization and justi-fication. To address these gaps, we introduce SportsTime, a large-scalebenchmark for long-form sports video understanding, comprising 14K+open-ended QA pairs and 50K+ step-wise temporal evidence annota-tions. Building on SportsTime, we propose Chain-of-Time Reasoning(CoTR), which treats reasoning as a process of temporally groundedevidence composition. Specifically, during training, CoTR introduces atemporal-reward GRPO to encourage temporally grounded reasoning.During inference, it employs an anchor-observe-infer evidence-seekingloop to iteratively localize, verify, and compose temporal evidence beforeproducing the final answer. Experiments show that SportsTime exposessubstantial gaps in current MLLMs, while CoTR yields consistent gainsover strong baselines, improving both temporal compositional reasoningperformance and step-wise grounding quality. The dataset and code areavailable at https://github.com/ustiniansy/SportsTime.

Pruning visual foundation models has attracted considerableattention. However, existing methods focus on rigid point-to-point tokenalignment on a single dataset for pruning, suffering from two limitations:i) robustness degradation, and ii) task-specificity deficiency. To addressthese limitations, we propose a task-specific pruning pipeline, namedCut-ViT. Specifically, we first construct gram anchoring matrices fromboth spatial and semantic perspectives, and perform the subspace de-composition to extract the corresponding subspace bases. Basis-agnosticand residual constraints are then adopted to align the gram subspaces be-tween the native and pruned DINOv3 models along spatial and channeldimensions, enabling subnetworks to inherit robust feature representa-tions of native DINOv3. Furthermore, we design spectral entropy adapta-tion, which quantifies the information density of feature manifolds alongspatial and channel dimensions, thereby adapting the pruning objectiveto specific downstream tasks. Experiments show that Cut-ViT requiresapproximately one minute on a single A100 GPU to obtain subnetworksat various sparsity levels, using only 20.9% of the time and 45.5% of theGPU memory compared with previous methods, while achieving SOTAperformance on six tasks across nine datasets.