ECCV 2026 Papers

High-resolution Synthetic Aperture Radar (SAR) imageryis critical for precision analysis such as automatic target recognition,yet its acquisition is costly. Although generative image super-resolution(ISR) models offer a promising alternative, current smooth-approximation-based diffusion frameworks often struggle to preserve the coherent scat-tering statistics, causing stochastic structural distortions that are lessconsistent with real SAR physics. To address this, we propose Semantic-Prototype-Guided Super-Resolution (ProSR), reformulating SAR ISR asa semantically-guided discrete token prediction task within a quantizedlatent space. By mapping signal features to discrete scattering primitives,ProSR preserves the impulsive nature of SAR without over-smoothing.Furthermore, we integrate a Self-Supervised Learning backbone intoSAR ISR to extract label-free semantic priors, overcoming label scarcity.Guided by these priors, we introduce Semantic-Aligned Detail Encodingto decouple high-frequency signals into discrete scattering primitives. Inparallel, the Semantic Prototype Map Generator explicitly constructssemantic prototype maps, allowing Prototype-Map-Guided Attentionto route the information flows within identical categories and mitigateinter-class interference. To validate our approach, we present a large-scale0.25 m resolution benchmark from the Umbra Open Dataset. Experimen-tal results show ProSR achieves superior visual quality while preservingessential scattering characteristics required for practical SAR applications.

E!ective collaboration begins with knowing when to ask forhelp. For example, when trying to identify an occluded object, a humanwould ask someone to remove the obstruction. Can MLLMs exhibit asimilar “proactive” behavior by requesting simple user interventions?To investigate this, we introduce ProactiveBench, a benchmark builtfrom seven repurposed datasets that tests proactiveness across di!erenttasks such as recognizing occluded objects, enhancing image quality, andinterpreting coarse sketches. We evaluate 22 MLLMs on ProactiveBench,showing that (i) they generally lack proactiveness; (ii) proactiveness doesnot correlate with model capacity; (iii) “hinting” at proactiveness yieldsonly marginal gains. Surprisingly, we found that conversation historiesand in-context learning introduce negative biases, hindering performance.Finally, we explore a simple fine-tuning strategy based on reinforcementlearning: its results suggest that proactiveness can be learned, evengeneralizing to unseen scenarios. We publicly release ProactiveBench asa first step toward building proactive multimodal models.

While instruction-based image editing, enabled by multi-modal generative models, has advanced significantly, existing bench-marks lack comprehensive evaluation of physics-based reasoning—a criti-cal capability for handling real-world scenarios. To address this, we intro-duce PhyEditBench, a benchmark designed to assess the physical under-standing of editing models. Guided by a hierarchical taxonomy, we estab-lish 4 primary classes and 12 subclasses. It comprises 238 high-quality,high-resolution, real-world instances—meticulously extracted fromvideos to capture authentic physical dynamics, alongside 35 syntheticAnti-Physics instances. Our empirical analysis of current SOTA editingmethods exposes substantial limitations in their physics-based reasoning.We further propose a training-free baseline named PhyWorld that usestest-time scaling and a latent reduction strategy. PhyWorld outperformscomparable models and suggests that the video generation process caneffectively serve as a reasoning mechanism for image editing. The projectpage is available at https://github.com/Previsior/PhyEditBench.

Infrared-visible image fusion (IVIF) is pivotal for multimodalperception, yet reconciling the inherent information disparity betweenthermal and textural features remains a fundamental challenge. Exist-ing prior-guided methods often rely on static constraints that induceoptimization conflicts or utilize extrinsic semantic priors from large-scale foundation models (e.g., CLIP/DINO), which frequently fail toexploit the intrinsic modality characteristics essential for high-fidelityfusion. To address these issues, we propose P²Fusion, a prior-guideddistillation-based framework that reformulates IVIF via dual intrinsicprompts. Instead of imposing hard-coded penalties, we distill image-intrinsic priors, thermal saliency and spatial quality—into learnable,dynamic regulators. Specifically, a Teach-to-Fuse mechanism providesdual-granularity progressive guidance, coupled with a Gated DynamicExpert Recalibration (GDER) module for decoupled feature refinement.This design enables the network to adaptively mediate modal competi-tion through expert specialization. Extensive experiments demonstratethat P²Fusion achieves state-of-the-art performance across five main-stream datasets. Notably, our framework demonstrates consistent per-formance advantages in fusion quality, achieving state-of-the-art resultsin 14 out of 20 key evaluation metrics across 5 benchmarks. Further-more, it effectively contributes to the robustness of downstream percep-tion, such as +3.2% mAP on MSRS, +0.5% mAP on M3FD and +0.9%mAP on DroneVehicle for object detection. Our code will be available athttps://github.com/YiShi99/P2Fusion.

Text-to-image (T2I) models face significant safety risks fromadversarial induction, yet current concept erasure methods often causecollateral damage to benign attributes when suppressing selected neu-rons entirely. This occurs because sensitive and benign semantics ex-hibit non-orthogonal superposition, sharing activation subspaces wheretheir respective vectors are inherently entangled. To address this issue,we propose OrthoEraser, which leverages sparse autoencoders (SAE) toachieve high-resolution feature disentanglement and subsequently rede-fines erasure as an analytical orthogonalization projection that preservesthe benign manifold’s invariance. OrthoEraser first employs SAE to de-compose dense activations and segregate sensitive neurons. It then usescoupled neuron detection to identify non-sensitive features vulnerableto intervention. The key novelty lies in an analytical gradient orthog-onalization strategy that projects erasure vectors onto the null spaceof the coupled neurons. This orthogonally decouples the sensitive con-cepts from the identified critical benign subspace, effectively preservingnon-sensitive semantics. Experimental results on safety demonstrate thatOrthoEraser achieves high erasure precision, effectively removing harm-ful content while preserving the integrity of the generative manifold, andsignificantly outperforming SOTA baselines. WARNING: This papercontains results of unsafe models.

There are two prevalent ways for automatic 3D scene con-struction: procedural generation and 2D lifting. Among these, panorama-based 2D lifting has emerged as a promising technique, leveraging pow-erful 2D generative priors to produce immersive, realistic, and diverse3D environments. In this work, we advance this technique to generategraphics-ready 3D scenes suitable for physically based rendering (PBR),relighting, and simulation. Our key insight is to repurpose 2D generativemodels for panorama perception of geometry, textures, and PBR mate-rials. Unlike existing 2D lifting approaches that emphasize appearancegeneration and neglect the perception of intrinsic properties, we presentOmniX, a versatile and unified framework for panorama generation, per-ception, and completion. Built upon cross-modal adapter structure andcyclic spatial operators, OmniX effectively repurposes pre-trained 2Dflow matching priors for joint modeling of multimodal, seamless equirect-angular representations. Furthermore, we construct a large-scale syn-thetic panorama dataset comprising high-quality multimodal panoramasfrom diverse indoor and outdoor scenes. Extensive experiments demon-strate the effectiveness and generality of OmniX as a unified frameworkfor panorama generation and perception across geometry, lighting, andsemantics, enabling graphics-ready 3D scene generation and opening newpossibilities for immersive and physically realistic virtual world creation.Project page is available at https://yukun-huang.github.io/OmniX/.

Collaborative perception mitigates occlusions and extendsthe perception range, yet its accuracy remains susceptible to both poseand feature noise. Given these limitations, we propose NoiseGT, anoise-driven framework built upon the principle that “Noise is a GoodTeacher”. Instead of treating noise as a disturbance, NoiseGT leveragesit as a self-supervised signal to enhance robustness at both the poseand feature levels. Specifically, a pose calibration module first appliesmulti-level noise injection, followed by graph-based alignment refinement.This process enables the module to learn pose-invariant alignment de-spite localization uncertainties. Meanwhile, a feature denoising moduleadopts a “high-noise-over-low-noise” strategy to inject high-magnitudenoise into bird’s-eye-view (BEV) features, guiding the denoiser towardnoise-robust representations without relying on clean intermediate super-vision. Together, these components form a noise-driven learning processthat promotes consistent fusion across agents. At the most challengingnoise setting ((σt , σr ) = (1.2 m, 1.2◦ )), NoiseGT achieves state-of-the-artperformance over previous methods, improving AP@0.5/0.7 by a rela-tive 12.03%/14.56% on V2X-Sim, 2.67%/0.89% on DAIR-V2X, and6.12%/9.69% on OPV2V. These results highlight the effectiveness andgenerality of noise-driven learning for robust collaborative perception.

Test-Time Adaptation (TTA) seeks to improve model robust-ness under distribution shifts by adapting parameters using unlabeledtarget data. However, in the absence of supervision, entropy-based adap-tation is fundamentally underconstrained: multiple distinct parameterupdates can achieve similarly low entropy while inducing drasticallydifferent decision boundaries. This phenomenon, known as underspecifi-cation, renders standard TTA brittle and prone to collapse into spuriousmodes. In this work, we reinterpret TTA through a posterior-inspired lensinduced by entropy minimization, where low-entropy solutions define apseudo-likelihood over parameters. Instead of committing to a single pointestimate, we introduce a particle-based diversification framework thatexplores multiple plausible adaptation trajectories simultaneously. Ourmethod can be viewed as a structured exploration of multiple plausibleadaptation solutions, implemented through multi-level diversification atthe output, parameter, optimizer, and input levels. Crucially, the frame-work acts as a plug-and-play wrapper compatible with existing TTAmethods. Extensive experiments on challenging benchmarks demonstrateconsistent gains in stability and robustness, achieving improvements of3–4% under mixed shifts, 2–3% with batch size one, and 1–2.5% underlabel shifts, outperforming state-of-the-art baselines. Our results suggestthat treating TTA as a multi-hypothesis inference problem, rather thana single-point optimization task, is key to mitigating underspecificationand enabling reliable real-world deployment.

Cross-modal music generation from text and visual inputsis uniquely challenging due to its subjective, one-to-many nature. Cur-rent methods face three major limitations: they are bottlenecked byscarce paired data, lack explicit one-to-many modeling, and isolate gen-eration from retrieval. To address these issues, we propose Meric, aunified framework for multimodal music generation and retrieval. Ourcore innovation is a music semantic anchor that decouples multimodalunderstanding from acoustic synthesis. We employ a two-stage train-ing pipeline. First, a Flow Matching decoder learns to synthesize audiofrom anchor embeddings using large, unpaired music corpora. Second, alightweight generative diffusion module maps multimodal inputs into thiscontinuous anchor space using limited paired data. This decoupled de-sign explicitly models cross-modal ambiguity and exploits unpaired datato bypass the paired data bottleneck. During inference, the predicted an-chor embedding simultaneously acts as a condition for acoustic synthesisand a query for music retrieval. Extensive experiments demonstrate thatMeric achieves state-of-the-art performance across text-to-music andvision-to-music generation and retrieval benchmarks.

High-resolution volumetric medical imaging is critical for clinical diagnosis, yet acquisition is often limited by scanner hardware, scan time, and for CT, radiation dose. Medical 3D Super-Resolution (Med3DSR) offers a computational alternative, but existing methods commonly rely on per-subject optimization, pretrained priors, or coordinatebased implicit representations, which compromise anatomical fidelity and limit efficiency. To address these limitations, we present MedGSSR, a fully end-to-end feed-forward framework that represents volumes as an explicit 3D Gaussian field for Med3DSR. Unlike coordinate-based implicit functions, our explicit 3D Gaussian representation naturally enhances signal continuity and local high-frequency fidelity. Specifically, MedGSSR explicitly decouples the reconstruction process into coarsegrained structural preservation and fine-grained textural refinement through the proposed Pyramid Anatomical Encoder and a Hierarchical Gaussian Projector. To support arbitrary-scale super-resolution, we introduce sub-voxel Gaussian decomposition and a Differentiable Gaussian Voxelizer that directly queries the continuous 3D intensity field, reducing discretization artifacts. Extensive experiments on MRI and CT benchmarks demonstrate that MedGSSR significantly outperforms state-ofthe-art methods. Notably, our framework exhibits robust generalizability across unseen datasets without requiring per-subject optimization, enabling fast inference and high-fidelity volumetric super-resolution in practical clinical settings. Our project webpage, including code, is at https://william2ai.github.io/medgssr

Single-source remote sensing object detection using opticalor SAR images struggles in complex environments. Optical images offerrich textural details but are often affected by low-light, cloud-obscured,or low-resolution conditions, reducing the detection performance. SARimages are robust to weather, but suffer from speckle noise and limitedsemantic expressiveness. Optical and SAR images provide complemen-tary advantages, and fusing them can significantly improve the detectionaccuracy. However, progress in this field is hindered by the lack of large-scale, standardized datasets. To address these challenges, we proposea new comprehensive dataset for optical-SAR fusion object detection,named Multi-resolution, Multi-polarization, Multi-scene, Multi-sourceSAR dataset (M4-SAR). It contains 112,174 instance-level aligned im-age pairs and nearly one million labeled instances with arbitrary orienta-tions, spanning six key categories. To enable standardized evaluation, wedevelop a unified benchmarking toolkit that integrates six state-of-the-art multi-source fusion methods. Additionally, we propose E2E-OSDet, anovel end-to-end multi-source fusion detection framework that mitigatescross-domain discrepancies and establishes a robust baseline for futurestudies. Extensive experiments on M4-SAR demonstrate that fusing op-tical and SAR data can improve mAP by 5.7% over single-source inputs,with particularly significant gains in complex environments.

Video Temporal Grounding (VTG) localizes the temporalboundaries of query-relevant moments in long, untrimmed videos, mak-ing video-language-model prohibitively expensive. While recent training-free token pruning has shown success in video question answering, naivelyapplying these objectives to VTG causes drastic degradation, as VTGcrucially depends on boundary-sensitive evidence and cross-frame rea-soning chains. We therefore identify two VTG-specific pruning principles:evidence retention, which keeps query-critical patches especially aroundevent boundaries, and connectivity strength, which preserves cross-frameconnectivity for long-range evidence aggregation. Building on these in-sights, we propose SemVID, a training-free pruning framework that con-structs a compact yet coherent token subset with complementary seman-tic roles. SemVID first allocates per-frame budgets by balancing queryrelevance and inter-frame variation to avoid over-pruned segments, andthen selects three types of tokens: object tokens for diverse query-criticalevidence, motion tokens to capture meaningful transitions and serve ascross-frame relays, and context tokens for scene continuity. Extensive ex-periments show that SemVID achieves a strong accuracy-efficiency trade-off, retaining up to 95.4% mIoU with only 12.5% visual tokens and de-livering up to a 5.8× prefill speedup, consistently outperforming priormethods under the same budgets. Our code is available here.

Class-agnostic segmentation enables open-world object seg-mentation beyond predefined categories, and extending it to the eventdomain helps enhance perception in high-speed motion and extreme illu-mination scenarios. However, the sparsity and low redundancy of eventdata lead to weak intra-class consistency, making event-based segmen-tation typically underperform compared to RGB-based counterparts.To address this issue, we propose IC2 -ESeg, an Intra-class Consistencyguided, Class-agnostic Event Segmentation framework, including threemodules: intra-class consistency mining (ICM), lightweight informationinjection (LIJ), and cross-modal distillation (CMD). Grounded in theevent generation mechanism, the ICM module constructs noise, motion,and brightness representations by analyzing statistical noise frequency,spatio-temporal distribution, and brightness gradients, thereby enrichingthe informational content of event data. The LIJ module efficiently inte-grates these representations into the image encoder, while the CMD mod-ule transfers the general segmentation capability of Segment AnythingModel (SAM) to the event domain. Experiments on RGBE-SEG demon-strate that IC2 -ESeg achieves state-of-the-art (SOTA) performance inclass-agnostic segmentation. We also construct ComScene, a high reso-lution RGB-event segmentation dataset encompassing diverse scenarios.Results on MVSEC and ComScene demonstrate the strong generaliza-tion of IC2 -ESeg.

Vision–language models (VLMs) have become a paradigmfor multimodal learning, yet remain unstable due to object hallucination,weak visual grounding, and catastrophic forgetting after full-parameterinstruction tuning. We claim these failures result from a lack of explicitcontrol over visual representation learning during the standard next-token prediction objective. As a result, visual embeddings thus becomepassively optimized and prone to injecting redundant or spurious signals.To counter this, we introduce Information-Regularized Attention (IRA),a stochastic attention mechanism that explicitly regulates the amount ofvisual information injected into the hidden states of intermediate trans-former layers. This local reparameterization translates uncertainty aboutvisual representations into local noise that is independent across datapoints. Beyond evaluating model performance, we also quantify embed-ding properties, where IRA produces smoother curvature trajectoriesand suppresses attention-sink across all layers, indicating a more stabletransformation of the visual signal. Our results suggest that stochasticattention is not merely a regularizer but a key contributor to represen-tation learning in a generative architecture, offering a new direction forbuilding more reliable VLMs.

Cross-domain Few-shot Segmentation (CD-FSS) aims to learn generalizable segmentation capability from abundant annotated samples in the source domain, enabling accurate segmentation of novel classes in the target domain with only a few annotated samples. Existing CDFSS methods mainly focus on mitigating feature distribution shifts caused by style gaps while ignoring significant differences in class semantic granularity and discriminative attributes across domains, leading to two key degradations in support-query matching: semantic over-alignment and attribute over-alignment. To this end, we propose the Dual Hierarchical Aggregation Network (DHANet), which comprises three key modules. First, the Hierarchical Spatial Aggregation (HSA) module performs multi-scale region aggregation of pixel features along the spatial dimension, generating hierarchical semantic-enhanced features to alleviate semantic over-alignment. Additionally, the HCA module conducts multi-scale attribute aggregation along the channel dimension, generating hierarchical attribute-enhanced features to mitigate attribute overalignment. Finally, we propose the Online Probabilistic Semantic Bank (OPSB), which progressively constructs and updates class probability distributions from query predictions during inference, and samples multiple pseudo-prototypes as additional support information to mitigate insufficient support. Extensive experiments on four target-domain datasets demonstrate that our method achieves state-of-the-art performance.

Novel view synthesis requires strong 3D geometric consis-tency and the ability to generate visually coherent images across di-verse viewpoints. While recent camera-controlled video diffusion modelsshow promising results, they often suffer from geometric distortions andlimited camera controllability. To overcome these challenges, we intro-duce GeoNVS, a geometry-grounded novel-view synthesizer that en-hances both geometric fidelity and camera controllability through ex-plicit 3D geometric guidance. Our key innovation is the Gaussian Splat-ting Feature Adapter (GS-Adapter), which lifts input-view diffusion fea-tures into 3D Gaussian representations, renders geometry-constrainednovel-view features, and adaptively fuses them with diffusion features tocorrect geometrically inconsistent representations. Unlike prior methodsthat inject geometry at the input level, GS-Adapter operates in featurespace, avoiding view-dependent color noise that degrades structural con-sistency. Its plug-and-play design enables zero-shot compatibility withdiverse feed-forward geometry models without additional training, andcan be adapted to other video diffusion backbones. Experiments across 9scenes and 18 settings demonstrate state-of-the-art performance, achiev-ing 11.3% and 14.9% improvements over SEVA and CameraCtrl, withup to 2× reduction in translation error and 7× in Chamfer Distance.

Large 3D foundation models such as MASt3R achieve state-of-the-art stereo reconstruction but are computationally demanding fordeployment under strict hardware constraints—a critical limitation indomains such as planetary exploration, where onboard computing isseverely restricted. We study how far such models can be compressedthrough knowledge distillation, using lunar stereo reconstruction as achallenging and practically relevant case study. Starting from a 688M-parameter MASt3R teacher fine-tuned on lunar imagery, we distill itsdense geometric predictions into a family of lightweight students span-ning different encoder types (CNN vs. ViT), decoder widths and depths,and training strategies. To bridge the dimensional mismatch betweenteacher and student, we propose a structured SVD-based initializationthat projects the teacher’s decoder weights into the student’s smallerlatent space, yielding a warm start that significantly improves conver-gence and final performance. Based on our results on lunar data, wecan obtain a distilled student that retains most of teacher’s reconstruc-tion accuracy while reducing the model size up to 7×, and even out-performs a baseline trained directly with sparse ground-truth annota-tions. Beyond compression, our study highlights both principles andpractical insights for distilling geometric foundation models: a convo-lutional encoder underperforms transformer-based alternatives (thoughpretraining availability remains a confounding factor), preserving en-coder capacity is more critical than maintaining a large decoder, feature-level distillation consistently outperforms output-only supervision, andSVD-based initialization improves optimisation stability. These findingsprovide practical guidelines for deploying 3D reconstruction models inresource-constrained environments. Code is publicly available at https://clementinegrethen.github.io/publications/ECCV.html.

Images shared on social media often expose geographic cues.While early geolocation methods required expert effort and lacked gen-eralization, the rise of Large Vision Language Models (LVLMs) now en-ables accurate geolocation even for ordinary users. However, existingapproaches are not optimized for this task. To explore the full potentialand associated privacy risks, we present GEO-Detective, an agent thatmimics human reasoning and tool use for image geolocation inference.It follows a procedure with four steps that adaptively selects strategiesbased on image difficulty and is equipped with specialized tools suchas visual reverse search, which emulates how humans gather externalgeographic clues. Experimental results show that GEO-Detective out-performs baseline large vision language models (LVLMs) overall, par-ticularly on images lacking visible geographic features. In country levelgeolocation tasks, it achieves an improvement of over 11.1% compared tobaseline LLMs, and even at finer grained levels, it still provides around a5.2% performance gain. Meanwhile, when equipped with external clues,GEO-Detective becomes more likely to produce accurate predictions, re-ducing the “unknown” prediction rate by more than 50.6%. We furtherexplore multiple defense strategies and find that GEO-Detective exhibitsstronger robustness, highlighting the need for more effective privacy safe-guards.

Zero-shot video temporal grounding (VTG) localizes eventsin untrimmed videos from natural language queries without task-specifictraining. Existing methods rely on frame-query feature matching, whichsuffices for simple events but struggles with complex multi-stage queriesthat require understanding temporal ordering and causal structure—a disparity we call the reasoning gap. We propose DART (Difficulty-Adaptive Routing for Temporal Grounding), which bridges this gapby coupling difficulty-aware routing with structured reasoning in largevision-language models. A query-conditioned Determinantal Point Pro-cess (DPP) serves a dual role: selecting diverse, query-relevant keyframesas temporal evidence, and providing spectral entropy as a difficulty indi-cator. Simple queries are routed to a Fast path for direct prediction, whilecomplex queries follow a Slow path with Temporal Markup Prompting,which decomposes localization into global event analysis, per-frame tem-poral role annotation, and boundary extraction. On Charades-STA andActivityNet Captions, DART achieves state-of-the-art zero-shot perfor-mance across both identically distributed and multiple out-of-distributionsettings, improving mIoU by up to 3.5 points over the strongest baselinewhile using over 7× fewer frames. The project homepage is available athttps://dart-vtg.github.io/.