ECCV 2026 Papers

Underwater computer vision tasks, such as detection, restora-tion, and segmentation, are limited by the scarcity of large-scale and di-verse training data. We introduce WaterGen, a method for generatinglarge-scale, realistic, and diverse underwater images that provides inde-pendent control of the scene and water medium conditions. Our approachtreats underwater image generation as the decoupled control of two fac-tors: realistic and diverse scene content (what is in the image), and accu-rate and controllable water medium effects (what the water does to theimage). Existing methods generally achieve only part of this objective:they either provide controllability with limited realism or diversity, orgenerate realistic scenes without accurately and independently modelingwater-medium effects. Our key insight, that allows us to avoid this com-promise, is that scene generation and medium modeling can be decoupledwithin a latent diffusion framework, enabling diverse scene generation to-gether with accurate and controllable underwater appearance. To do this,we decompose underwater image synthesis into two stages. First, we fine-tune the latent diffusion U-Net using degradation-free underwater imagesso that it learns to generate diverse and realistic latent embeddings ofunderwater scene content without medium-induced degradation. Second,we formulate the physically accurate medium degradation synthesis asa conditional decoding process applied to these latent embeddings. Thisdecoupled design allows our model to generate diverse scenes with fullcontrol of underwater appearance. We leverage WaterGen to build large-scale synthetic underwater datasets that are diverse in scene structuresand accurate in water effects and pseudo-labels. We demonstrate that oursynthetic data consistently improve downstream performance in under-water restoration and semantic segmentation. Code and model weightsare available at https://github.com/jiayi-wu-umd/WaterGen.

Extracting structured, parametric 3D representations fromraw images remains a fundamental challenge in computer vision andgraphics. While recent advancements in the 3D Gaussian Splatting (3DGS)pipeline integrate planar primitives to yield compact and editable geom-etry, these approaches typically treat planes as isolated, discrete sets.This lack of topological connectivity hinders robust geometric reasoning,leading to fragmented reconstructions and misaligned boundaries thatfall short of the precision for rigorous spatial analysis and professionaldesign workflows. To address this, we introduce TopoGS, the first 3DGSframework to explicitly integrate both planar and topological constraintsfor coherent 3D reconstruction. Specifically, we extract global 2D topo-logical relationships from multi-view image segmentations and anchorGaussian primitives to these structural elements. This formulation en-ables the joint optimization of plane parameters, rendering fidelity, andtopological adjacency. By enforcing strict multi-view consistency along-side these topological constraints, our method significantly mitigates ge-ometric misalignments and produces connected, structured 3D models.Extensive evaluations on the ScanNet++ dataset demonstrate that To-poGS achieves state-of-the-art performance, providing a highly robustsolution for generating accurate, topologically sound, and visually faith-ful scene representations.

Reliable 3D dynamic perception requires models that can anticipatemotion beyond predefined categories, yet progress is hindered by the scarcityof dense, high-quality motion annotations. While self-supervision on unlabeledreal data offers a path forward, empirical evidence suggests that scaling unla-beled data fails to close the performance gap due to noisy proxy signals. In thispaper, we propose learning robust real-world motion priors entirely from scal-able simulation. We introduce SynFlow, a data generation pipeline for large-scalesynthetic LiDAR scene flow. Unlike prior works that prioritize sensor-specificrealism, SynFlow employs a motion-oriented strategy to synthesize diverse kine-matic patterns across 4,000 sequences (∼940k frames), termed SynFlow-4k. Thisrepresents a 34× scale-up in annotated volume over existing real-world bench-marks. Our experiments demonstrate that SynFlow-4k provides a highly domain-invariant motion prior. In a zero-shot regime, models trained only on our syntheticdata generalize across multiple real-world benchmarks, comparable to in-domainsupervised baselines on nuScenes and outperforming state-of-the-art methodson TruckScenes by 31.8%. Furthermore, SynFlow-4k serves as a label-efficientfoundation: fine-tuning with only 5% of real-world labels surpasses models trainedfrom scratch on the full available budget. We open-source the pipeline and datasetto facilitate research in generalizable 3D motion estimation. More detail can befound at https://kin-zhang.github.io/SynFlow.

Imitation Learning (IL) offers an effective approach for robotmanipulation by learning a mapping from visual inputs to actions. How-ever, this paradigm suffers from poor robustness under visual domainshifts (e.g., lighting conditions, camera viewpoints, etc.), often failingto perform manipulation. Our key insight in tackling this problem is toconstruct a Structure Map encoding the fine-grained object structuresthat remain invariant across visual domains and vital for manipulation.Based on this insight, we propose StructPolicy, a method elaboratelydesigned to incorporate the domain-invariant Structure Map into the ILpolicy through two modules: StructCon and StructEncoder. StructCon isan automated Structure Map construction module that leverages struc-ture primitives to enable flexible transformation and composition to forma wide range of objects. StructEncoder is a hierarchical network thatefficiently captures structural relationships and affordance from Struc-ture Map. This provides the IL policy with both domain-invariant andstructurally-aware features, guiding it toward robust manipulation un-der visual domain shifts. We extensively evaluate StructPolicy across 49manipulation tasks in multiple benchmarks and diverse real-world tasksunder various visual changes. The results demonstrate consistent andsignificant performance improvements across all tasks, validating thatStructPolicy enhances the effectiveness and robustness against visual do-main shifts of the IL policy, improving manipulation accuracy.

Generating complete 3D scenes from sparse, unconstrainedviews is a fundamental challenge in 3D vision which requires reason-ing beyond observed content while remaining computationally tractable.Existing feed-forward reconstruction methods are inherently limited tocontent visible in the input images, while 3D generative modeling is hin-dered by the high computational cost of dense volumetric representationsand the scarcity of large-scale 3D supervision. We introduce SPAR3S, asparse voxel-aligned 3D latent generative model for conditional scenecompletion without requiring ground-truth 3D data for supervision. Ourkey insight is to formulate 3D scene generation in a structured, compact,voxel-aligned 3D latent space where only occupied voxels are represented.We learn this sparse latent space directly from multi-view images usingphotometric supervision via differentiable 3D Gaussian Splatting. Givena partial set of observed voxels encoded from sparse input views, scenecompletion reduces to predicting the missing latent tokens and theirspatial support within the voxel grid. To this end, we train a maskedautoregressive transformer that jointly models voxel occupancy and la-tent token values, enabling efficient and spatially consistent generation ofunseen regions. We demonstrate the effectiveness of our method on syn-thetic indoor scenes, achieving higher novel-view quality than prior work.We further validate its generalization on RealEstate10k, highlighting itsapplicability to real-world data.

Video Temporal Grounding (VTG) aims to localize momentsin untrimmed videos based on language queries. A fundamental chal-lenge lies in learning cross-modal representations that capture bothsemantic correspondence and temporal specificity. Existing approachespredominantly rely on token-level visual-textual cross-attention, whichmay not fully capture two essential properties: generalizability acrossvideos and distinguishability within videos. In practice, semanticallyrelated vocabularies may correspond to similar visual concepts, whiletemporally sparse yet crucial semantics can be easily diluted by fre-quently occurring visual patterns. To address these challenges, we pro-pose Semantic-Explicit Detection Transformer (SE-DETR), a unifiedframework that explicitly models semantic structures for video-languagegrounding. We introduce a Semantics-Proxied Alignment (SPA) mod-ule that learns semantic proxies dynamically optimized during training,and enables concept-level alignment between multimodal representations.Furthermore, we design a Temporal Sparsity Modulation (TSM) mod-ule that estimates the temporal sparsity of semantic components anddynamically reweights word-guided visual features to highlight informa-tive concepts for accurate moment localization. Extensive experimentson five benchmarks demonstrate that SE-DETR consistently achievescompetitive performance across Video Moment Retrieval, Highlight De-tection, and Video Summarization tasks. The source code is avaliable at:https://github.com/hu-cheng-yang/ECCV26-SE-DETR.

Real-time inference of vision–language–action (VLA) models is essential for robotic control. While visual token pruning has shown strong potential for accelerating inference, most existing methods mainly base pruning decisions on shallow-layer cues and risk discarding visual information required by deep layers. To address this issue, we propose SAFE-Pruner, a plug-and-play pruning framework that incorporates attention cues of future layers into pruning decisions. Specifically, we identify semantic attention consistency, the tendency that VLA models concentrate their attention probability mass on the same semantic entity across control timesteps. Based on this observation, we design a forwardlooking strategy to forecast the token saliency in deep layers, which prevents the premature removal of critical tokens and leads to more stable acceleration. We further introduce a reference timestep refresh strategy that triggers updates upon attention shifts, thereby improving forecasting accuracy and pruning reliability. Extensive experiments across diverse evaluation settings demonstrate that our method achieves up to 1.89× speedup with a minimal degradation in success rate of less than 1.5%, while outperforming state-of-the-art methods by up to 1.9%.

Multi-perturbation adversarial training (MAT) aims to achieverobustness against multiple ℓp perturbations but suffers from robustnesstrade-offs between different threats. To address this, we employ a mixtureof experts (MoE) to route different threats through distinct model path-ways. However, naïve application of MoE encounters two critical chal-lenges: experts tend to overlook threat-specific features and redundantlycapture features shared across threats, and gating networks suffer fromthreat-agnostic routing where they learn nearly identical routing pat-terns across threats, thus preventing the construction of threat-specificmodel pathways. To this end, we propose Robust Mixture of Low-RankExperts (RoME), where each expert is a low-rank additive update tothe shared backbone, allowing it to capture threat-common featureswhile experts focus on threat-specific information. To address threat-agnostic routing, RoME introduces (i) dual-scale gating that exploitsthreat-discriminative signals from local and global level features, and(ii) threat-guided gating diversification that enforces diverse expert uti-lization across threats. Extensive experiments demonstrate that RoMEoutperforms existing state-of-the-art MAT in union robustness and nat-ural accuracy and improves robustness against unseen threats. Codes areavailable at https://github.com/wkim97/RoME.

In this paper, we challenge the conventional notion in 3DGS-SLAM that rendering quality is the primary determinant of trackingaccuracy. We argue that, compared to solely pursuing a perfect scenerepresentation, it is more critical to enhance the robustness of the ras-terization process against parameter errors to ensure stable camera posetracking. To address this challenge, we propose a novel approach thatleverages a smooth kernel strategy to enhance the robustness of 3DGS-based SLAM. Unlike conventional methods that focus solely on minimiz-ing rendering error, our core insight is to make the rasterization processmore resilient to imperfections in the 3DGS parameters. We hypothesizethat by allowing each Gaussian to influence a smoother, wider distribu-tion of pixels during rendering, we can mitigate the detrimental effectsof parameter noise from outlier Gaussians. This approach intentionallyintroduces a controlled blur to the rendered image, which acts as a reg-ularization term, stabilizing the subsequent pose optimization. While acomplete redesign of the rasterization pipeline is an ideal solution, wepropose a practical and effective alternative that is readily integratedinto existing 3DGS frameworks. Our method, termed Corrective BlurryKNN (CB-KNN), adaptively modifies the RGB values and locations ofthe K-nearest neighboring Gaussians within a local region. This dynamicadjustment generates a smoother local rendering, reducing the impactof erroneous GS parameters on the overall image. Experimental resultsdemonstrate that our approach, while maintaining the overall quality ofthe scene reconstruction (mapping), significantly improves the robust-ness and accuracy of camera pose tracking.https://github.com/xju-zsh/Robust-3DGS-based-SLAM-via-Adaptive-Kernel-Smoothing.git

Physical principles are fundamental to realistic visual sim-ulation, but remain a significant oversight in video generation. Thisgap highlights a critical limitation in rendering rigid body motion, acore tenet of classical mechanics. While computer graphics and physics-based simulators can easily model such dynamics using Newton for-mulas, modern video generative models discard the concept of objectrigidity during pixel-level global denoising. Existing methods attemptto tackle this problem through physical data augmentation, dynamicspre-simulation, or reinforcement learning with VLM ratings, but noneof these approaches accurately reflect physical principles or enable themodel to internalize physical knowledge. Motivated by these considera-tions, we introduce reinforcement learning with physically verifiable re-wards. We design a quantitatively verifiable metric that combines Trajec-tory Offset and Collision Detection, which can accurately capture rigid-body motion states and assess the quality of generated samples. Subse-quently, we extend this paradigm to a unified post-training framework,termed Mimicry-Discovery Cycle, which enables stable training onout-of-distribution scenarios while improving overall model performance.To validate our approach, we construct new benchmark PhysRVGBenchand perform extensive qualitative and quantitative experiments to thor-oughly assess its effectiveness. The code and demo can be found athttps://lucaria-academy.github.io/PhysRVG/.

Garment draping aims to fit clothing onto 3D human mod-els. Existing physics-based methods are computationally expensive andstruggle to integrate with general differentiable systems. In contrast, deeplearning-based methods often require extensive annotations and lack ex-plicit physical constraints, limiting their ability to model accurate details.We innovate by bridging physical models with deep learning and proposePhysDrape. It integrates three modules using forces as an intermediary:the force-driven GNN predicts forces at each node, the stretching solvermodels physical deformation, and the collision handler penalizes inter-penetration. Each module includes learnable parameters to fit underly-ing force propagation processes and physical properties, enabling gener-alization to unseen templates and control over properties like stiffness.PhysDrape enables end-to-end optimization, leveraging intrinsic garmentenergy for physically plausible deformation via self-supervised learning.Following the previous training and evaluation protocol on CLOTH3Ddataset, PhysDrape achieves lower energy scores, negligible interpene-tration, more realistic visualizations, and comparable time consumption.

Applying Multimodal Large Language Models (MLLMs) tolong-form video is challenged by the substantial computational cost ofprocessing long token sequences. Current training-free reduction meth-ods present a difficult trade-off between efficient and content diversity.Query-focused approaches are efficient but sacrifice content diversity,which is critical for general-purpose understanding and open-ended dia-logue. Conversely, methods that preserve diversity by operating on high-dimensional features often incur significant computational overhead.We introduce Perceptual Projection Pruning (P 3 ), a training-free frame-work that tries to address this trade-off. Our approach is based on the hy-pothesis that a low-dimensional yet informative representation can serveas an efficient proxy for diversity-aware visual token pruning. We firstproject high-dimensional features into a 3D perceptual space by comput-ing three complementary saliency axes: dynamic (motion), regional (con-textual contrast), and focal (local complexity). This representation thenserves as a foundation for a subsequent diversity-aware pruning strat-egy. Experiments across SOTA MLLMs show our method retains nearly95% of the vanilla model’s performance while pruning up to 75% of to-kens. Our work demonstrates that a carefully designed low-dimensionalrepresentation is highly effective for guiding diversity-aware visual tokenpruning in MLLMs.

Representation Autoencoders (RAEs) enable diffusion mod-els to operate in the feature spaces of pretrained visual encoders. How-ever, many off-the-shelf encoders are not optimized for faithful recon-struction and often discard fine-grained visual details. Finetuning theseencoders for image reconstruction can recover such details, but we showthat it also reduces the effective dimensionality of the resulting represen-tation space. We analyze how this altered geometry affects generation inhigh-dimensional feature spaces. Under this geometry, standard veloc-ity prediction in flow matching can require the model to fit orthogonalnoise directions outside the low-dimensional signal manifold, making op-timization inefficient. This motivates the clean data parameterization(x0 -prediction), which focuses learning on the underlying signal mani-fold. Across experiments with multiple reconstruction-finetuned featuresets, we show that x0 -prediction consistently improves text-to-image gen-eration performance.

Text-to-image diffusion models excel at generating images from nat-ural language prompts, yet fail to interpret numerical colors such as hex codes(#FF5733) and RGB values (rgb(255,87,51)). This limitation stems fromsubword tokenization, which fragments color codes into semantically meaning-less tokens that text encoders cannot map to coherent color representations. Wepresent NumColor, that enables precise numerical color control across multi-ple diffusion architectures. NumColor comprises two components: a Color To-ken Aggregator that detects color specifications regardless of tokenization, anda ColorBook containing 6,707 learnable embeddings that map colors to embed-ding space of text encoder in perceptually uniform CIE Lab space. We introducetwo auxiliary losses, directional alignment and interpolation consistency, to en-force geometric correspondence between Lab and embedding spaces, enablingsmooth color interpolation. To train the ColorBook, we construct NumColor-Data, a synthetic dataset of 500K rendered images with unambiguous color-to-pixel correspondence, eliminating the annotation ambiguity inherent in photo-graphic datasets. Although trained solely on FLUX, NumColor transfers zero-shot to SD3, SD3.5, PixArt-α, and PixArt-Σ without model-specific adaptation.NumColor improves numerical color accuracy by 4–9× across five models, whilesimultaneously improving color harmony scores by 10–30× on GenColorBenchbenchmark. The code is available at https://github.com/moatifbutt/numcolor/.

We introduce VidPair-Halluc, a new benchmark for evaluating video hallucination in large video models (LVMs) under rigorous and controlled conditions. Unlike previous benchmarks that primarily rely on text-based perturbations or adversarial questions while neglecting the consistency of visual backgrounds, VidPair-Halluc features video pairs with highly similar backgrounds but distinctly different foreground semantics, enabling precise attribution of model errors to genuine hallucination rather than background variation. The benchmark is constructed through PairFlow, a pipeline that leverages recent advances in text-to-image and video generation to systematically compose stories, generate coherent video clips, and assemble them into adversarial pairs. Covering both spatial and temporal reasoning across ten semantic aspects, VidPair-Halluc comprises 1K high-quality adversarial video pairs and 11K spatio-temporal QA pairs with control over background and foreground variations. Evaluations on mainstream LVMs show persistent difficulty with robust fine-grained video understanding in adversarial settings, and code and data are available at the project page.

Di!erentiable rendering has advanced novel view synthesis(NVS), yet applying it to real-world driving remains di"cult due tosparse capture viewpoints, dynamic objects, and limited multi-trajectorydata. We introduce the Multi-View Multi-Vehicle (MV2 ) datasetand benchmark for evaluating NVS models under large viewpoint changesin dynamic urban scenes. MV2 features synchronized captures from a car,scooter, and drone, each following distinct yet synchronized trajectories.Training NVS methods on one vehicle’s camera stream and testing on an-other enables evaluation under substantially larger viewpoint variationsthan existing single-trajectory datasets. All sequences are registered viaStructure-from-Motion and camera poses verified using manual pixel-level correspondence annotations, yielding 50 high-quality scenes with12000 images. Benchmarking recent NVS and camera pose estimationmethods shows that NVS performance degrades with increasing view-point disparity, and that feed-forward pose estimators notably lag behindoptimization-based approaches, highlighting MV2 as a rigorous testbedfor NVS in driving. The dataset, benchmark protocol, and project re-sources are available at https://mv2-dataset.github.io/.

High-resolution 3D radar data is scarce. Commodity mmWavesensors use small antenna arrays that limit angular resolution to several de-grees, and existing datasets provide only 2D range–azimuth maps or sparsepoint clouds rather than raw analog-to-digital converter (ADC) signals.Hardware scaling is expensive, synthetic-aperture scanning is impracticalat fleet scale, and learned synthesis methods are bottlenecked by the verydata shortage they aim to address. We present mmIR, an open-sourcedifferentiable frequency-modulated continuous-wave (FMCW) radar in-verse renderer that fits a physics-based forward model to real capturesand re-renders from dense virtual apertures to synthesize high-resolution3D radar data. Because radar resolution is too coarse to recover geom-etry directly, mmIR performs LiDAR-assisted inverse rendering: usingLiDAR-derived meshes as a geometric scaffold, mmIR optimizes per-vertex International Telecommunication Union (ITU) physics materials,vertex normals, and antenna beam patterns through end-to-end auto-matic differentiation of a phase-coherent multiple-input multiple-output(MIMO) forward model with multi-bounce propagation, polarization,and free-space diffraction. On seven outdoor and six indoor ColoRadarscenes, mmIR achieves 0.914 mean Pearson correlation on range–azimuthmaps versus 0.307 for Sionna-RT. Scenes trained on a cascaded imag-ing radar transfer to a co-located single-chip radar without re-training(0.554 correlation), and dense virtual arrays (100×100 elements) pro-duce single-frame 3D occupancy validated against LiDAR. Project page:https://mmwave-inverse-rendering.github.io/.

Online Video Instance Segmentation (VIS) relies on per-frame instance queries for causal detection, segmentation, and tempo-ral association. However, in realistic videos with occlusion, blur, andclutter, these predictions often become noisy—exhibiting false positives,missed instances, or inaccurate masks. Such noisy queries not only desta-bilize cross-frame association but also contaminate contrastive supervi-sion and propagate drift through memory-based matching. Despite ad-vances in recent online VIS methods, none explicitly identify or addressnoisy query contamination, leaving a fundamental weakness unexam-ined. We introduce MiNQVIS, the first noise-aware framework for on-line VIS that stabilizes both training and inference. Our approach con-sists of three complementary components: (1) query-level gating that fil-ters unreliable predictions using classification-guided thresholding andembedding-norm stabilization; (2) video-level prototypical contrastivelearning that aggregates only reliable matches into stable, noise-robustprototypes; (3) dual-memory association that fuses momentum embed-dings with history-best retrieval to prevent drift from corrupted frames.Experiments on YouTubeVIS 2019/2021 and OVIS show consistent im-provements over strong online VIS baselines, with large gains in challeng-ing scenes. Project Page: https://nothing898.github.io/MiNQVIS.

Depth estimation in multi-illumination scenes with multi-ple, spatially varying light sources remains a crucial yet less-exploredproblem. Illumination changes introduce shadows, specular highlights,and exposure shifts that distort local appearance cues, causing severedepth inconsistency or even failure. Existing depth foundation models,trained predominantly on uniformly lit data, degrade sharply under suchconditions. However, direct adaptation is challenging because groundtruth depth is typically limited for multi-illumination datasets, whilesynthetic relighting often incurs geometric distortions. To address thesechallenges, we propose LumiDepth, a framework that learns from multi-illumination RGB images. First, a Disagreement-Calibrated ProbabilisticPseudo Supervision (DCPS) module constructs high-quality pseudo la-bels while preserving diversity. Second, a Frequency-aware Consistencyand Distillation (FaCD) module improves cross-illumination stabilitywithout over-smoothing by enforcing low-frequency geometric consis-tency and distilling high-frequency structural details bi-directionally. Toenable systematic evaluation, we introduce ReMID, a real-world multi-illumination RGB-D benchmark, together with stability metrics thatquantify average and worst-case depth variation. Experiments acrossdiverse datasets demonstrate that LumiDepth achieves state-of-the-artoverall performance, markedly improving both consistency and accuracyby reducing depth variation by 30.2% and absolute relative error by24.8%. We further show our target-domain label-free design remains ef-fective for depth under other appearance shifts such as weather and sen-⋆sor noise.