ECCV 2026 Papers

Virtual Try-On (VTON) synthesizes realistic images of aperson wearing a target garment, with broad applications in e-commerceand fashion. Diffusion-based dual-UNet methods achieve strong resultsbut double the parameters by dedicating a separate network to garmentconditioning. Spatial concatenation offers a simpler single-network alter-native, yet both UNet- and DiT-based instantiations report that full fine-tuning is ineffective, and the community has settled for attention-onlytraining. We ask: why does full fine-tuning fail, and can this be resolved?Through what is, to our knowledge, the first visualization study of dual-UNet reference network behavior, we identify a unifying insight: garmentconditioning must be decoupled from the denoising process. Spatial con-catenation violates this by embedding the garment within the denoisingtarget, causing three conflicts: guidance leakage, gradient competition,and train-test discrepancy. We derive three design principles to restorethis decoupling and implement them as a pure recipe atop a standardarchitecture with no modification. The resulting model, DeCo-VTON(860M params), achieves single-network state of the art, matching thedual-UNet state of the art at half the cost while being preferred in humanevaluation.

Feed-forward 3D Gaussian Splatting methods enable single-pass reconstruction and real-time rendering. However, they typicallyadopt rigid pixel-to-Gaussian or voxel-to-Gaussian pipelines that uni-formly allocate Gaussians, leading to redundant Gaussians across views.Moreover, they lack an effective mechanism to control the total numberof Gaussians while maintaining reconstruction fidelity. To address theselimitations, we present F4 Splat, which performs Feed-Forward predic-tive densification for Feed-Forward 3D Gaussian Splatting, introduc-ing a densification-score-guided allocation strategy that adaptively dis-tributes Gaussians according to spatial complexity and multi-view over-lap. Our model predicts per-region densification scores to estimate therequired Gaussian density and allows explicit control over the final Gaus-sian budget without retraining. This spatially adaptive allocation reducesredundancy in simple regions and minimizes duplicate Gaussians acrossoverlapping views, producing compact yet high-quality 3D representa-tions. Extensive experiments demonstrate that our model achieves su-perior novel-view synthesis performance compared to prior uncalibratedfeed-forward methods, while using significantly fewer Gaussians.

Bridging abstract semantics and precise physical control re-mains a fundamental challenge in open-world robotic manipulation. Whilerecent data-driven policies show promise, their reliance on isolated con-tact points or latent affordance embeddings lacks the rigorous kinematicconstraints necessary for complex articulated objects.To overcome thelimitation, we introduce RelAfford6D, a novel training-free frameworkcentered on a Relational 6D Affordance Graph. Given a free-form in-struction, our system deduces a semantic topology linking a primary in-teracting part to its physical anchor. By elevating these topological nodesinto precise metric SE(3) poses via vision foundation models, we analyti-cally formulate downstream execution as a kinematic constraint satisfac-tion problem. The robot synthesizes continuous trajectories by trackingstrictly defined physical manifolds (e.g., revolute or prismatic orbits).Coupled with a closed-loop tracking mechanism for dynamic replanningagainst disturbances, our physically grounded approach achieves supe-rior zero-shot success rates, cross-category generalization and executionrobustness in both simulation and the real world environments, outper-forming existing data-driven baselines.

Reconstructing non-rigid objects with physical plausibilityremains a significant challenge. Existing approaches leverage differen-tiable rendering for per-scene optimization, recovering geometry anddynamics but requiring expensive tuning or manual annotation, whichlimits practicality and generalizability. To address this, we propose Re-conPhys, the first feedforward framework that jointly learns physicalattribute estimation and 3D Gaussian Splatting reconstruction from asingle monocular video. Our method employs a dual-branch architecturetrained via a self-supervised strategy, eliminating the need for ground-truth physics labels. Given a video sequence, ReconPhys simultaneouslyinfers geometry, appearance, and physical attributes. Experiments ona large-scale synthetic dataset demonstrate superior performance: ourmethod achieves 21.64 PSNR in future prediction compared to 13.27by state-of-the-art optimization baselines, while reducing Chamfer Dis-tance from 0.349 to 0.004. Crucially, ReconPhys enables fast inference(<1 second) versus hours required by existing methods, facilitating rapidgeneration of simulation-ready assets for robotics and graphics. The codeis available at https://github.com/chuanshuogushi/ReconPhys.

We introduce LivingWorld, an interactive framework forgenerating 4D worlds with environmental dynamics from a single image.While recent advances in 3D scene generation enable large-scale environ-ment creation, most approaches focus primarily on reconstructing staticgeometry, leaving scene-scale environmental dynamics such as clouds,water, or smoke largely unexplored. Modeling such dynamics is challeng-ing because motion must remain coherent across an expanding scenewhile supporting low-latency user feedback. LivingWorld addresses thischallenge by progressively constructing a globally coherent motion field asthe scene expands. To maintain global consistency during expansion, weintroduce a geometry-aware alignment module that resolves directionaland scale ambiguities across views. We further represent motion using acompact hash-based motion field, enabling efficient querying and stablepropagation of dynamics throughout the scene. This representation alsosupports bidirectional motion propagation during rendering, producinglong and temporally coherent 4D sequences without relying on expen-sive video-based refinement. On a single RTX 5090 GPU, generatingeach new scene expansion step requires 9 seconds, followed by 3 secondsfor motion alignment and motion field updates, enabling interactive 4Dworld generation with globally coherent environmental dynamics. Videodemonstrations are available at paper.pnu-cvsp.com/LivingWorld.

While personalizing Image-to-Video (I2V) diffusion modelswith specific visual effects is increasingly demanded for high-end genera-tion, current practice requires training a separate Low-Rank Adaptation(LoRA) module for each effect, incurring substantial data curation anditerative optimization costs that hinder interactive control. We presentPrompt2Effect, a weight-driven hypernetwork that amortizes per-effecttraining by directly synthesizing effect-specific LoRA weights in a singleforward pass. Unlike prior hypernetworks that regress adapter weightspurely from semantics, Prompt2Effect is explicitly conditioned on thefrozen base model weights, grounding prediction in the structural geom-etry of each layer. Furthermore, instead of predicting raw LoRA matri-ces, we introduce an SVD-canonicalized parameterization that resolvesfactorization ambiguity and stabilizes large-scale synthesis. Extensive ex-periments demonstrate that Prompt2Effect achieves on-par or superiorvideo quality and effect alignment compared to conventional LoRA fine-tuning, while reducing the computational cost from 56 GPU traininghours to 3.3 seconds of hypernetwork inference. When used as initializa-tion for subsequent fine-tuning, our predicted weights further improvefinal performance and accelerate optimization by approximately 10×.

3D object detection from LiDAR point clouds faces a fun-damental dilemma: voxel-based methods achieve efficiency at the costof geometric quantization, while point-based methods preserve fidelitybut suffer from prohibitive computational bottlenecks. Specifically, point-based architectures are crippled by slow downsampling strategies (e.g.,FPS) and expensive dynamic neighbor queries (e.g., k-NN) coupled withcostly continuous interactions. To tackle these systemic inefficiencies, wepropose PointLAM, a highly efficient and powerful point-based architec-ture driven by two synergistic innovations. First, to resolve the downsam-pling bottleneck, we develop the Laplacian Point Sampler (LPS). LPSemploys an implicit discrete Laplacian high-pass filter and Doubly SortedSampling to achieve fast, structure-aware foreground preservation. Sec-ond, to overcome local modeling latency, we design the Local HadamardAggregator (LHA). LHA decouples spatial indexing from feature repre-sentation using transient grids, and replaces complex continuous interac-tions with a Hadamard Gating mechanism for topology-aware, attentivemodulation. By coupling this local gating with Bi-Directional Mamba(BDM) layers for global sequence modeling, we formulate the Local At-tentive Mamba (LAM) block. Powered by this architecture, PointLAMachieves competitive performance on nuScenes and Waymo for point-based detectors. It rivals highly optimized voxel competitors while re-quiring a fraction of the computational footprint, demonstrating markedsuperiority in detecting small instances and handling extreme sparsity.Project page: https://pointlam.github.io/.

The rise of 3D Gaussian Splatting has revolutionized pho-torealistic 3D asset creation, yet a critical gap remains for their inter-active appearance refinement and editing. Existing approaches based ondiffusion or optimization are ill-suited for this task, as they are oftenprohibitively slow, lack the precision for fine-grained control, or—mostimportantly—are fundamentally destructive to the original asset’s iden-tity. To address this, we introduce SplatPainter, a state-aware feedfor-ward model that enables continuous, high-fidelity appearance editing of3D Gaussian assets from user-provided 2D views. Our method directlypredicts updates to the attributes of a compact, feature-rich Gaussianrepresentation and leverages Test-Time Training to create a state-aware,iterative workflow. The versatility of our approach allows a single ar-chitecture to perform diverse tasks, including high-frequency local detailrefinement, local paint-over, and consistent global recoloring, all at inter-active speeds while strictly preserving the asset’s original structure. Ourproject website is at: https://y-zheng18.github.io/SplatPainter/.

Vision-Language-Action Models (VLAs), which leverage theadvanced reasoning capabilities of Vision-Language Models (VLMs), showpromising generalization in complex autonomous driving scenarios. Ex-isting VLAs typically predict and optimize 3D trajectories from 2D im-ages. While intuitive, this 2D-to-3D prediction is inherently entangledwith camera parameters, leading to limited data scalability across het-erogeneous driving datasets. Moreover, directly optimizing in 3D spaceinduces severe convergence to trivial solutions, where VLAs rely on ego-status rather than visual scene understanding. To address these issues,we propose PixelPilot, a novel VLA featuring a decoupled planning andlifting paradigm. In the planning phase, PixelPilot reformulates scene un-derstanding and trajectory prediction as sensor-agnostic 2D-to-2D tasksin the image plane, thereby facilitating scalable training across diversedatasets. The planned 2D trajectories are then deterministically lifted to3D only during inference, ensuring the full exploitation of visual cues andgeneralization across different vehicles. To realize this paradigm, we pro-pose a knowledge-instilled policy learning strategy that applies dense, in-termediate rewards via Group Relative Policy Optimization (GRPO) toenforce a rigorous causal chain from visual perception to spatial planning.Extensive experiments demonstrate that PixelPilot achieves state-of-the-art performance in both open-loop and closed-loop settings, validatingits superior scalability and visual reasoning capabilities.

Current video generation models produce highly realistic vi-suals but frequently fail to maintain physical and causal consistencywhen facing complex dynamic events, such as collisions or collapses.We attribute this limitation to the fact that existing models primar-ily fit visual data distributions, entangling dynamics with appearancewithout explicit physical priors to govern motion evolution. To addressthis, we introduce PhysAlign, a framework for Dynamical Event-DrivenPhysically Consistent Video Generation. PhysAlign decouples physicalpriors from visual appearance via a Physical Dynamics Extractor anda discrete Universal Physical Codebook. Recognizing the temporal lo-cality of dynamic events, an Event-Aware Temporal Gating module dy-namically controls the injection of physical priors, while Physics-GuidedAlignment distills continuous physical topologies from a video founda-tion model. Furthermore, we construct EventPhy, a 26K-video datasetwith structured, VLM-generated causal annotations to benchmark dy-namical event-driven video generation. Extensive evaluations show thatPhysAlign achieves state-of-the-art physical reasoning, obtaining leadingSemantic Adherence and Physical Commonsense scores on challengingbenchmarks like PhysicsIQ and VideoPhy, while preserving high visualfidelity. Project Page: https://github.com/FanQi-AI/PhysAlign

Latent generative models compress images into learned em-beddings prior to synthesis, and the generation quality critically dependson how faithfully these embeddings preserve visual detail. We observethat while such embeddings are effective at reconstructing low frequencystructure, they struggle to recover sharp high frequency details that areessential for perceptual realism. Conventional reconstruction objectivesimplicitly prioritize coarse structural information over high frequencycontent, which can lead to overly smoothed outputs and degraded vi-sual quality in textured regions. Motivated by this observation, we pro-pose DeBaT, a Decoupled frequency Band Tokenizer that explicitly sep-arates the learning of low and high frequency band embeddings. Thisdecoupling enables accurate reconstruction of fine details while preserv-ing global coherence. Integrated into a latent diffusion based generativemodel, DeBaT allows for sharper and more realistic samples than previ-ous latent tokenizers, confirming that the explicit decoupling of high andlow frequency bands eases the preservation of visual details in learnedembedding spaces.

Image matting is highly ill-posed, especially when both theforeground and background are richly textured. While single-image mat-ting methods learn strong priors from data, they often struggle on thesechallenging cases. Existing approaches improve results by requiring ad-ditional signals such as green screens, polarized lighting, or clean back-ground images, but these typically rely on specialized capture setups.We present Parallax Portrait Matting, a practical two-frame mattingmethod that uses a second image captured with slight viewpoint change.Such a setting arises naturally in burst photography, where small cameramotion induces foreground-background parallax and provides comple-mentary observations for matting. Our pipeline estimates trimaps andforeground/background motion, then constructs aligned views for pre-diction. To handle imperfect motion estimation, the network uses thebackground-aligned pair for direct fusion and the foreground-aligned cuethrough cross-attention for error compensation. Experiments show thatour method recovers finer details and more accurate foreground colorsthan strong single-image matting baselines on challenging portrait cases.

We present ViewSplat, a view-adaptive 3D Gaussian splat-ting network for novel view synthesis from unposed images. While recentfeed-forward 3D Gaussian splatting has significantly accelerated 3D scenereconstruction by bypassing per-scene optimization, a fundamental fidelitygap remains. We attribute this gap to the limited capacity of single-stepfeed-forward networks to regress static Gaussian primitives that sat-isfy all viewpoints. To address this limitation, we shift the paradigmfrom static primitive regression to view-adaptive splatting. Instead of arigid Gaussian representation, our pipeline learns a view-adaptive latentrepresentation. Specifically, ViewSplat initially predicts base Gaussianprimitives alongside the weights of scene-conditioned View MLPs. Duringrendering, these MLPs take target-view coordinates as input and predictview-dependent residual updates for each Gaussian attribute (i.e., 3Dposition, scale, rotation, opacity, and color). This mechanism, whichwe term view-adaptive splatting, allows each primitive to rectify initialestimation errors, effectively capturing high-fidelity appearances. Exten-sive experiments demonstrate that ViewSplat achieves state-of-the-artfidelity while maintaining fast inference and real-time rendering; our largebackbone variant runs at 15 FPS during inference and 90 FPS duringrendering.

Prototypical part-based models provide explainable predic-tions by comparing input regions to learned prototypes. However, currentapproaches are burdened by complex, multi-stage training pipelines andheavily rely on auxiliary regularization to prevent prototype collapse. Toovercome these limitations, we introduce Orthonormal Prototype Align-ment Learning (OPAL), a single-stage, end-to-end framework that sim-plifies interpretable classification. Our approach anchors the latent spaceusing predefined orthonormal bases, embedding each class within a dedi-cated subspace spanned by fixed part-prototypes. To achieve precise partlocalization, OPAL enforces spatial competition across feature maps.This mechanism isolates sparse, discriminative regions, directing eachprototype to consistently attend to the same semantic concept across dif-ferent images. By framing classification as a direct representation align-ment task, our method eliminates the need for auxiliary losses. Extensiveexperiments on fine-grained benchmarks demonstrate that OPAL outper-forms both its non-interpretable counterparts and state-of-the-art part-prototype methods, delivering granular visual explanations by explicitlyrevealing the specific image regions driving every prediction. Code isavailable at https://github.com/ilancarretero/OPAL.

Monocular dense prediction has recently seen remarkablesuccess by repurposing pre-trained diffusion models. This opens a promis-ing yet challenging avenue for more efficient multi-task learning paradigm.However, existing multi-task diffusion methods often introduce parameter-heavy adapters, experts, or learnable task tokens, leading to computa-tional redundancy. In this paper, we reveal an inherent mechanism withinone-step diffusion models: the native, fixed timestep positional embed-ding can be repurposed as an endogenous task steering signal. Based onthis discovery, we propose Multi-task Unified eStimation via timestepEmbedding (MUSE), a parameter-free, single-model multi-tasking ap-proach for dense prediction. We interpret this mechanism via Mani-fold Decoupling, where discrete, fixed timestep values deterministicallysteer the generation process towards decoupled, task-specific manifolds inthe latent space. Extensive experiments across 10 datasets demonstratethat MUSE achieves highly competitive performance on both monoculardepth and normal estimation, and its efficacy generalizes across U-Netand DiT architectures. Our work offers a concise and efficient path to-ward generalist vision models by simply unlocking the latent potentialof existing generation infrastructure.

Scaling the width of Vision Transformers (ViT) often triggerssevere instability issues. This work identifies that these instabilities stemfrom the “high-activation domination” phenomenon frequently encoun-tered during feature normalization in high-dimensional spaces. Specifi-cally, the norm of a high-dimensional vector can be easily dominated by afew outlier elements with disproportionately large values. Upon normal-ization, the remaining elements are suppressed toward zero, leading tosignificant gradient propagation issues. To address this, this work intro-duces Multi-Head RMSNorm (MH-RMSNorm), which partitions a high-dimensional feature vector into multiple chunks and applies RMSNormto each segment independently. This simple, drop-in replacement sig-nificantly enhances the training stability and predictive performance ofhigh-dimensional ViTs. For example, Our 1280-dimensional ViT achieves85.5% top-1 accuracy on ImageNet-1K, whereas baseline models usingstandard LayerNorm or RMSNorm suffer from complete training col-lapse; when applied to DiT on ImageNet-256, our method achieves a1.95 FID, significantly outperforming standard ViT counterparts.

Human body fitting, which aligns parametric body models,such as SMPL, to raw 3D point clouds of clothed humans, serves as acrucial first step for downstream tasks like animation and texturing. Aneffective fitting method should be both locally expressive – capturingfine details such as hands and facial features – and globally robust tohandle real-world challenges, including clothing dynamics, pose variations,and noisy or partial inputs. Existing approaches typically excel in onlyone aspect, lacking an all-in-one solution. We upgrade ETCH to ETCH-X,which leverages a tightness-aware fitting paradigm to filter out clothingdynamics (“undress”), extends expressiveness with SMPL-X, and replacesexplicit sparse markers (which are highly sensitive to partial data) with im-plicit dense correspondences (“dense fit”) for more robust and fine-grainedbody fitting. Our disentangled “undress” and “dense fit” modular stages en-able separate and scalable training on composable data sources, includingdiverse simulated garments (CLOTH3D), large-scale full-body motions(AMASS), and fine-grained hand gestures (InterHand2.6M), improvingoutfit generalization and pose robustness of both bodies and hands. Ourapproach achieves robust and expressive fitting across diverse clothing,poses, and levels of input completeness, delivering a substantial perfor-mance improvement over ETCH on both 1) seen data, such as 4D-Dress(MPJPE-All, 33.0% ↓) and CAPE (V2V-Hands, 35.8% ↓), and 2) unseendata, such as BEDLAM2.0 (MPJPE-All, 80.8% ↓; V2V-All, 80.5% ↓).Code and models will be released at xiaobenli00.github.io/ETCH-X.

Despite the impressive manipulation capabilities of Vision-Language-Action (VLA) models, their operational safety under strictconstraints remains largely unverified. To address this, we introducea parametric safety benchmark to procedurally generate safety-criticalscenarios with comprehensive stochasticity. To overcome the scalabilitybottlenecks of human teleoperation, we develop a novel keypose-drivendata generation pipeline. Leveraging this infrastructure, we curate alarge-scale dataset of 19,664 strictly collision-free demonstrations withextensive domain randomization. We then conduct a systematic cross-paradigm evaluation of eight VLA and two embodied foundation mod-els. Our analysis reveals a critical generalization-safety tension: althoughhigh-diversity training fosters safer trajectories, task success remains fun-damentally bottlenecked by sub-optimal trajectory synthesis and seman-tic misalignment. By providing a scalable pipeline, a robust dataset, andprofound failure-mode insights, LIBERO-Safety establishes a crucialfoundation for developing safe and reliable VLA models.

Despite rapid progress in Multi-modal Large Language Mod-els (MLLMs), understanding long-form videos is still bottlenecked bylimited context windows. While recent keyframe sampling methods at-tempt to mitigate this by distilling video inputs into a compact set ofquery-relevant frames, navigating the vast spatio-temporal search spaceremains challenging, as spatial detail and temporal coverage often conflict.To address this, we introduce LENS, a training-free keyframe samplingframework that dynamically decides when to zoom in for fine-graineddetails and when to zoom out for broader context based on the textquery. Concretely, LENS adaptively allocates a limited frame budgetbetween spatial zoom-ins, which highlight query-relevant regions withinindividual frames, and temporal zoom-outs, which expand the temporalscope through multi-frame aggregation, enabling the model to reasonacross multiple granularities while capturing both high-fidelity details andlong-range context. Across diverse long-form video benchmarks, LENSconsistently outperforms prior state-of-the-art keyframe sampling meth-ods and delivers substantial gains over uniform sampling, improvingVideo-MME accuracy from 53.3% to 60.7% with Qwen2.5-VL. Code isavailable at https://github.com/zhangce01/LENS.