ECCV 2026 Papers

Prevailing 2D-centric visuomotor policies exhibit a pronounceddeficiency in novel view generalization, as their reliance on static observa-tions hinders consistent action mapping across unseen views. In response,we introduce GenSplat, a feed-forward 3D Gaussian Splatting frameworkthat facilitates view-generalized policy learning through novel view ren-dering. GenSplat employs a permutation-equivariant architecture to re-construct high-fidelity 3D scenes from sparse, uncalibrated inputs in asingle forward pass. To ensure structural integrity, we design a 3D-priordistillation strategy that regularizes the 3DGS optimization, preventingthe geometric collapse typical of purely photometric supervision. By ren-dering diverse synthetic views from these stable 3D representations, wesystematically augment the observational manifold during training. Thisaugmentation forces the policy to ground its decisions in underlying 3Dstructures, thereby ensuring robust execution under severe spatial per-turbations where baselines severely degrade.

3D visual grounding aims to locate objects based on naturallanguage descriptions in 3D scenes. Existing supervised methods are lim-ited by generalization and recent zero-shot methods typically rely on apredefined Object Lookup Table (OLT) to query Visual Language Mod-els (VLMs) for reasoning about object locations via a single step ground-ing, which limits the applications in scenarios with undefined targets andcomplex queries. To address these problems, we present OpenGround, anovel zero-shot framework for open-world 3D visual grounding that re-mains compatible with recent zero-shot methods. OpenGround integratesTask-Chain Planning to decompose a query into a plan of context-to-target sub-goals for progressive grounding, and Context-Guided Percep-tion to perceive novel objects online under context guidance from thetask chain. We also propose a new dataset named OpenTarget, whichcontains over 7000 object-description pairs to mimic open-world eval-uation. Extensive experiments demonstrate that OpenGround achievescompetitive performance on Nr3D, state-of-the-art on ScanRefer, anddelivers a substantial 17.6% improvement on OpenTarget.

Precisely manipulating objects in a single photograph (trans-lation, rotation, scaling) while obeying 3D physical constraints remainsunsolved for diffusion-based editors. Current 2D methods lack spatialawareness and produce perspective violations. Forcing structural proxiesinto the latent space also disrupts variance homogeneity, and the result-ing self-attention leakage leads to ghosting and background blur. Thecore difficulty is asymmetric: the relocated object must follow a rigidgeometry, yet the uncovered background needs freedom to synthesizeplausible content. We present GeoEdit, a training-free Lift-Manipulate-Render-Denoise pipeline that satisfies both constraints. We decouplescene and object in 3D, align them through point correspondence, andrender a geometry-aligned proxy with a structural depth map. A Dual-Branch Denoising stage then refines this proxy: a video diffusion back-bone preserves object identity, while 3D constraints are injected into theforeground within a narrow denoising window at matching noise vari-ance (variance-homogeneous injection). The background denoises freely.Because the injected signal matches the native latent statistics, self-attention stays undisturbed. We also introduce GeoEditBench, a pose-aware benchmark covering object translation, object rotation, and cam-era movement with pose-aware evaluation metrics. Experiments confirmconsistent gains in geometric accuracy, identity fidelity, and backgroundquality, validated by automatic metrics and human studies. Code anddata are publicly available at https://github.com/Heey731/GeoEdit.

Active 3D Gaussian Splatting methods select views usingfactorised per-pixel metrics, which cannot represent relationships be-tween surfaces. Incompleteness that arises from unverified geometric re-lationships therefore goes undetected. We present ActiveStructure,which lifts information gain from pixels to scene-level relational topol-ogy through three innovations: (1) a Plane Scene Graph whose topolog-ical health, captured through node connectivity, edge completeness, andangular consistency, quantifies a form of structural completeness thatno pixel-level criterion can reach; (2) a gain decomposition that splitsthe value of a viewpoint into a rendering-fidelity term and a structural-integrity term that can be optimised independently; (3) informationalpolarity reversal, which lets scaffold planes switch from exploration at-tractors to suppressors and provides state-driven phase control tied tostructural objectives. Experiments on Replica and Matterport3D demon-strate state-of-the-art reconstruction accuracy and rendering quality.

Concept erasure methods aim to remove specific unsafe tar-get concepts in diffusion models while preserving image generation utility.To address the vulnerability that erased concepts can be easily recoveredunder adversarial attacks, adversarial concept erasure methods integrateadversarial optimization into the concept erasure process. However, exist-ing adversarial concept erasure methods face a trade-off between robust-ness and computational cost. We attribute this to adversarial optimiza-tion techniques that use random samples to approximate the adversarialobjective function. Adversarial optimization that uses a small numberof samples fails to produce adversarial embeddings that accurately cap-ture the target concept space. To mitigate this limitation, we proposeSemantic-Guided Adversarial Optimization, which uses a single sampleto produce adversarial embeddings that better capture the target con-cept space. We also propose Semantic-Guided Concept Erasure, whichautomatically maps the target concept to a semantically similar surro-gate. Extensive experiments on not-safe-for-work content, artistic styles,and object-related concepts demonstrate that our method, S-GRACE(Semantic-Guided Robust Adversarial Concept Erasure) achieves state-of-the-art erasure robustness and superior image generation utility, withsignificantly lower computational cost than existing methods. Our codeis available at https://github.com/Qhong-522/S-GRACE.

Semi-supervised learning for LiDAR semantic segmentationoften suffers from error propagation and confirmation bias caused bynoisy pseudo-labels. To tackle this chronic issue, we introduce RePL, anovel framework that enhances pseudo-label quality by identifying andcorrecting potential errors in pseudo-labels through masked reconstruc-tion, along with a dedicated training strategy. We also provide a theoret-ical analysis demonstrating the condition under which the pseudo-labelrefinement is beneficial, and empirically confirm that the condition ismild and clearly met by RePL. Extensive evaluations on the nuScenes-lidarseg and SemanticKITTI datasets show that RePL improves pseudo-label quality substantially, and in consequence, achieves the state of theart in semi-supervised LiDAR semantic segmentation.

3D Gaussian Splatting (3DGS) enables real-time novel viewsynthesis for static scenes. Extending it to dynamic scenes via deforma-tion fields has recently attracted significant attention, particularly fordynamic scene reconstruction and distractor-free reconstruction. How-ever, existing deformation networks lack explicit motion awareness: theyneither capture long-term motion intensity nor exploit short-term tempo-ral coherence, leading to inaccurate foreground deformation and pseudo-static residuals in the background. We present MVFusion-GS, a methodthat enhances deformation networks with two complementary motion-aware mechanisms. The Motion-Variance Guided Refinement ag-gregates per-Gaussian deformation statistics across time to estimate mo-tion variance and uses it to guide dynamic-static separation during defor-mation prediction. The MotionFormer Temporal Attention moduleapplies Transformer self-attention over neighboring timesteps to modellocal motion dependencies and improve temporal consistency. Extensiveexperiments on both dynamic scene reconstruction and distractor-free re-construction benchmarks demonstrate state-of-the-art performance, show-ing that explicit motion awareness improves both foreground motionmodeling and static background reconstruction.

All-in-One medical image restoration (MedIR) aims to ad-dress diverse tasks across modalities and degradation types using a singleuniversal model. Existing methods typically prioritize modeling inter-task heterogeneity (e.g., distinct data distributions and degradation types).However, they largely neglect the inherent homogeneity present in med-ical images, such as widely shared anatomical structures within andacross modalities, which can be leveraged to ease model training andimprove generalization. To this end, we propose UniH3 , a novel frame-work that Unifies Hierarchical Homogeneity and Heterogeneity for all-in-one medical image restoration. Specifically, to comprehensively exploithomogeneity, we introduce a Hierarchical Homogeneity Memory (H2 M)module that progressively distills intra- and inter-task homogeneity pri-ors from high-quality images during training, and adaptively retrieves themost relevant priors tailored to the input for guided restoration. Theseretrieved priors are then injected into the restoration pipeline via an ef-ficient Homogeneity-Guided Attention (HGA) mechanism. Furthermore,to comprehensively address heterogeneity, we design a Hierarchical Het-erogeneity Balancer (H2 B) that mitigates both inter- and intra-task con-flicts during optimization, facilitating balanced and effective multi-tasklearning. Extensive experiments on two large-scale benchmarks—MedIR-2D-500K and MedIR-3D-3K—demonstrate that UniH3 achieves state-of-the-art performance on both all-in-one and single-task medical imagerestoration. We hope this work establishes a strong benchmark and ad-vances the development of general-purpose medical image restorationmodels. Code is available at https://github.com/Yaziwel/UniH3.

Reference-based inpainting utilizes the information in thereference image to inpaint the missing region in the target image, aimingto generate the fact object behind the mask rather than a reasonable one.The key challenge is how to accurately locate the region in the referenceimage and maximally utilize it. We proposed UNet-Twice, a simple yetnovel framework. UNet-Twice uses the same UNet two times, combinedwith a well-constructed data pair to build a close connection betweenthe target and reference image. The two passes fuse spatially comple-mentary features generated from the target and the reference image,which provides intuitively explicit guidance for image generation. Ourproposed framework is simple and easy to extend to a multi-referencesetting. Extensive experiments demonstrate that our proposed frame-work outperforms existing reference-based inpainting methods in bothsingle-reference and multi-reference settings.

We study open-vocabulary 3D indoor layout generation, whichsynthesizes diverse and physically plausible indoor scenes from unla-beled 3D assets given free-form language instructions. Recent text-guidedlayout generation methods leverage large language models (LLMs) andvision-language models (VLMs) to synthesize structured scenes directlyfrom text descriptions. However, most existing approaches model inter-asset relations implicitly or rely on local pairwise constraints during lo-cal optimization. Such formulations are misaligned with the global andhighly non-convex feasible layout space, often producing locally plausiblebut globally inconsistent or physically infeasible scenes. To address theselimitations, we introduce a graph-based intermediate representation thatdecouples semantic coherence and physical feasibility, and propose ahybrid search-and-refinement strategy to generate indoor layouts withglobal semantic consistency and physical feasibility. First, we proposeGlobal Semantic Verification (GSV), which represents scenes as struc-tured scene graphs and enforces semantic constraints through rule-basedgraph verification. This explicit structural validation prunes contradic-tory configurations and produces a globally consistent semantic scaffoldfor scene generation. Second, we introduce Global Physical FeasibilitySearch (GPFS), a hybrid optimization framework that combines evo-lutionary search for global exploration with gradient-based refinementfor local exploitation. GPFS reduces dependence on VLM-proposed ini-tialization and improves robustness in highly non-convex and discon-tinuous feasible spaces. Together, GSV and GPFS shift layout genera-tion from local relational modeling and initialization-sensitive optimiza-tion toward globally consistent reasoning and exploration. Experimentsdemonstrate that our method achieves state-of-the-art performance onopen-vocabulary 3D indoor layout generation, improving both semanticconsistency and physical plausibility.

Intrinsic image decomposition aims to estimate physicallybased rendering (PBR) parameters such as albedo, roughness, and metal-licity from images. While recent methods achieve strong single-view pre-dictions, applying them independently to multiple views of the samescene often yields inconsistent estimates, limiting their use in down-stream applications such as editable neural scenes and 3D reconstruc-tion. Video-based models can improve cross-frame consistency but re-quire dense, ordered sequences and substantial compute, limiting theirapplicability to sparse, unordered image collections. We propose Geo-ID, a novel test-time framework that repurposes pretrained single-viewintrinsic predictors to produce cross-view consistent decompositions bycoupling independent per-view predictions through sparse geometric cor-respondences that form uncertainty-aware consensus targets. Geo-ID ismodel-agnostic, requires no retraining or inverse rendering, and appliesdirectly to off-the-shelf intrinsic predictors. Experiments on syntheticbenchmarks and real-world scenes demonstrate substantial improvementsin cross-view intrinsic consistency as the number of views increases, whilemaintaining comparable single-view decomposition performance. We fur-ther show that the resulting consistent intrinsics enable coherent appear-ance editing and relighting in downstream neural scene representations.

Visual prompt tuning has emerged as a parameter-e!cientfine-tuning approach for adapting large-scale Vision Transformers (ViTs)to downstream tasks. As its learnable prompts are applied in input andfeature spaces, prior to jointly going through attention in transformerlayers, the most commonly used scheme for fusing image and prompttokens is concatenation or addition. In this paper, we aim to study afundamental yet essential problem in visual prompt tuning: whethera single fusion scheme tends to yield better results, and whether thatwould be beneficial to develop a hybrid fusion scheme. To this end,we formulate the task as a bi-level optimization problem, and solve itleveraging di"erentiable architecture search. In this context, the learnableprompts and their fusion schemes are jointly optimized. To enrich thesearch space in the architecture search, we propose two additional fusionschemes, namely, a!ne transformation and cross-attention, in additionto concatenation and addition. Extensive experiments on 34 datasetsspanning VTAB-1k, FGVC, and HTA show consistent gains over prompt-tuning baselines. With a frozen ViT backbone, our method delivers afavorable accuracy–latency–parameter trade-o" compared with VPT-Deep and recent variants. Our findings reveal that how prompts fusewith image tokens plays a significant role in visual prompt tuning, anda hybrid fusion fashion can more e"ectively leverage layer semantics ofViTs, contributing a novel perspective for visual prompt-tuning research.

In recent years, super-resolution (SR) methods leveragingpre-trained text-to-image diffusion models have shown impressive po-tential. However, their deployment in real-world applications remainsconstrained by the computational overhead of numerous inference steps.Recent one-step approaches, such as OSEDiff and TSD-SR, mitigate thisissue via score distillation techniques. However, their discriminators areoptimized solely with the diffusion loss, resulting in weak and unstablegradients for the generator, making it difficult to preserve fine details andstructural fidelity. To tackle these limitations, we introduce ASDSR, anovel one-step distillation framework tailored for real-world image super-resolution. Our approach enhances Adversarial Score Distillation (ASD)by adopting a WGAN-based formulation with a stabilized penalty termthat utilizes the full discriminator loss to improve the distillation pro-cess. Furthermore, we incorporate CLIP-derived semantic priors to alignreconstructed images with their high-quality counterparts in the fea-ture space, effectively guiding the generator toward perceptual realism.Extensive experiments on multiple real-world benchmarks demonstratethat ASDSR achieves superior performance compared to state-of-the-artmethods, offering a practical and efficient solution for real-world imagerestoration.

Recent research in geospatial machine learning demonstratesthat models pretrained with self-supervised learning on Earth observa-tion data can perform well on downstream tasks with limited labeleddata. However, most benchmark datasets have few data modalities andpoor global representation, limiting the ability to evaluate multimodalpretrained models at global scales. In order to fill this gap, we introduceMMEarth-Bench, a collection of five new environmental tasks with 12modalities, globally distributed data, and both random and geographictest splits. We benchmark a diverse set of pretrained models and findthat while (multimodal) pretraining tends to improve model robustnessin limited data settings, geographic generalization abilities remain poor.Moreover, a simple randomly initialized multimodal model is competi-tive given enough labeled data. Although data is abundant, models cancurrently only make use of the modalities on which they were pretrained.To solve this problem, we propose using all the modalities available attest time as auxiliary tasks for test-time adaptation. Our model-agnosticmethod for test-time training with multimodal reconstruction (TTT-MMR) can improve performance across all models and tasks on bothtest splits. Furthermore, geographic batching leads to a good trade-off be-tween regularization and specialization during TTT, which is especiallybeneficial for long-tail distributions. Our dataset, code, and visualizationtool are linked on the project page: lgordon99.github.io/mmearth-bench.

AI-generated image (AIGI) attribution presents a pressingchallenge that goes beyond mere AIGI detection, aiming to identify thesource model or technique responsible for a synthetic image. However,most previous source attribution methods operate in a closed-set manner,which necessitates retraining to recognize any novel category, preventingadaptation to the rapid evolution of image generation. In this work, wepropose a new paradigm for synthetic image attribution, termed few-shotattribution. This paradigm targets the reliable identification of unseengenerators using only limited samples, making it highly suitable for real-world applications. To facilitate this work, we construct OmniFake, alarge-scale, well-categorized synthetic image dataset that contains 1.17million images from 45 distinct generators. We further introduce Om-niDFA (Omni Detector and Few-shot Attributor), a few-shot attributionbaseline that not only assesses the authenticity of images but also deter-mines their synthesis origins. Experiments demonstrate that OmniDFAexhibits excellent capability in few-shot attribution and achieves state-of-the-art generalization performance in AIGI detection. Our dataset andcode are available at https://github.com/teheperinko541/OmniDFA.

Training-free controllable motion generation has attracted growing interest for enabling flexible constraint enforcement without constraint-specific training. However, existing training-free methods require constraints to be continuous objective-based with differentiable losses, while many real-world requirements are criterion-based and provide only discontinuous, sparse, or even black-box feedback. In this paper, we propose Motion-Inference-as-Control (MIC), the first training-free motion generation framework that handles both continuous objective-based and criterion-based motion constraints under a shared mechanism. The key idea is to cast diffusion-based motion generation as a stochastic control problem. This perspective not only provides principled and practically effective step-wise control laws that support criterion-based constraints without requiring differentiability and naturally accommodate objectivebased constraints as a special case, but also motivates a control-oriented constraint coordination mechanism that adaptively balances and reconciles motion constraints during generation. Experiments across diverse constraint settings demonstrate the effectiveness of our framework.

Counting individual trees is a fundamental task for environ-mental monitoring, yet remains largely unexplored with satellite imagery.At these resolutions, isolated trees may still be identifiable, but crownboundaries become ambiguous in dense forests, making the notion of anindividual tree inherently ill-defined. Moreover, large-scale manual an-notations of individual trees are prohibitively expensive. While scalablesupervision can be derived from airborne LiDAR, the resulting anno-tations are noisy and difficult to exploit effectively. We address thesechallenges by formulating tree counting as a spatial density matchingproblem supervised through Unbalanced Optimal Transport. This formu-lation naturally accommodates both precise localization of isolate treesand robust density estimation in dense forests. We further introduce aself-correction mechanism that leverages transport residuals to progres-sively refine noisy supervision during training. We evaluate our approachon TinyTrees, a new benchmark spanning three continents and threesatellite sensors, comprising over 216 million tree annotations (including639k manually verified instances) across 25 890 km2 . Our method con-sistently outperforms detection-based, regression-based, and transport-based distribution-matching baselines, demonstrating the effectivenessof unbalanced transport and reliability-aware supervision for large-scaletree counting from satellite imagery. Code, data and models are availableat github.com/dgominski/treematch.

Conformal prediction provides distribution-free reliability guar-antees for vision systems, but these guarantees depend on how predictionerrors are measured in the output space. Many vision tasks produce out-puts on curved spaces (e.g. gaze directions on the sphere or 3D headrotations), yet intermediate prediction heads, residuals, uncertainty es-timates, or conformal scores are often defined in flat coordinate chartssuch as yaw–pitch or Euler angles. We show that this scoring choice intro-duces systematic geometric distortion near coordinate singularities (largepitch angles on the sphere and poses approaching gimbal lock in 3D rota-tions). Across four datasets (ETH-XGaze, Gaze360, BIWI, AFLW2000-3D), slice-conditional coverage at a nominal 90% target drops by 30–50percentage points in these regions, falling to 38.9% on ETH-XGaze and42.0% on Gaze360 at gaze pitch above 70◦ , and to 57.5% on BIWI and55.2% on AFLW2000-3D at head pose pitch above 60◦ near gimbal lock,despite marginal coverage remaining near 90%. We prove that this isstructural. Scalar thresholding changes the size of chart-coordinate pre-diction sets but leaves their distorted axis ratios unchanged. To diagnosethis hidden failure mode, we show that a simple geometric quantity,the Riemannian volume density, strongly correlates with where cover-age collapse occurs. Finally, we show that coordinate-free geodesic scor-ing removes this distortion. It requires no retraining and adds negligiblecomputational cost.

Performance in face and speaker verification is largely drivenby margin-penalty softmax losses such as CosFace and ArcFace. Recentlyintroduced α-divergence loss functions offer a compelling alternative, par-ticularly due to their ability to induce sparse solutions (when α > 1).However, standard geometric margins are designed for the softmax func-tion and do not naturally extend to this generalized probabilistic frame-work. In this paper we propose Q-Margin, a novel α-divergence lossthat introduces a principled probabilistic margin. Unlike conventionalmethods that apply geometric penalties to the logits (unnormalized log-likelihoods), Q-Margin encodes the margin penalty directly into the ref-erence measure (prior probabilities). This formulation naturally encour-ages discriminative embeddings while preserving the beneficial sparsityproperties of the α-divergence. We demonstrate that Q-Margin achievescompetitive or superior performance on the challenging IJB-B and IJB-C face verification benchmarks and similarly strong results in speakerverification on VoxCeleb. Crucially, against ArcFace and CosFace base-lines trained under an identical recipe, Q-Margin consistently improvesat low False Acceptance Rates (FARs), a capability critical for practicalhigh-security applications. Finally, the extreme sparsity of the Q-Marginposteriors enables exact and memory-efficient training, offering a scalablesolution for datasets with millions of identities.