ECCV 2026 论文列表

Despite recent progress in point cloud semantic segmenta-tion, obtaining deployment-grade predictions on individual scenes re-mains challenging. Interactive refinement offers a practical solution, yetexisting methods commonly interpret user clicks as single-point super-vision, leading to fragmented updates and requiring redundant user in-teraction. We identify this limitation as the Local Correction Trap. Ourkey insight is that a correction click more naturally expresses an in-tent to revise a semantically coherent region than to relabel an isolatedpoint. Based on this view, we propose InSeg, an interactive refinementframework that translates sparse clicks into reliable region-guided virtualsupervision. InSeg propagates corrective intent through feature affinity,spatial connectivity, and object-level priors to construct denoised virtuallabels. To support stable multi-round refinement, we further introducea history-aware mechanism that combines uncertainty-guided updateswith memory replay, reducing error accumulation and preventing previ-ously corrected regions from being overwritten. Extensive experimentsacross multiple benchmarks show that InSeg consistently improves seg-mentation quality and interaction efficiency, achieving strong gains withonly a few clicks.

The existence of adversarial attacks is often attributed tothe presence of non-robust features in neural networks. While prior de-fenses reduce their impact via pruning, masking, or feature recalibra-tion, we instead propose to jointly learn to amplify and attenuate thesesignals through a simple activation scaling mechanism. To this end,we introduce Activation Amplification and Attenuation (A3), alightweight plug-in module that enhances adversarial robustness withminimal modifications of the activations. A3 dynamically rescales theactivations using a learnable mask and a scaling factor derived from theoriginal activation magnitudes. The influence of adversarial perturba-tions can be amplified or attenuated using the same learnable parametersby simply flipping the sign of the scaling operation. The amplified signalsserve as negative references to construct novel contrastive and rankingloss functions. Experimental analysis shows that learning to degrade thepredictions in amplification mode simultaneously improves adversarialrobustness in attenuation mode. Moreover, A3 relies on only a smallnumber of learnable parameters, with most of its behavior being deter-mined by the scaling mechanism rather than additional network capac-ity. Extensive experiments demonstrate that integrating A3 into differentbackbones, datasets, and training methods consistently improves adver-sarial robustness while introducing negligible computational and mem-ory overhead compared to existing plug-in modules. Code is available at:https://github.com/tgoncalv/A3.

3D reassembly is a fundamental geometric problem, and inrecent years it has increasingly been challenged by deep learning meth-ods rather than classical optimization. While learning approaches haveshown promising results, most still rely primarily on geometric features toassemble a whole from its parts. This reliance limits their effectiveness incases where geometric information is incomplete or degraded, such as forsmall or eroded fragments. Additionally, solutions do not impose physi-cal constraints that explicitly prevent overlapping assemblies. To addressthese limitations, we introduce E-M3RF, an equivariant multimodal 3Dreassembly framework that takes as input the point clouds, containingboth point positions and color of fractured fragments, and predicts thetransformations required to reassemble them, using SE(3) flow match-ing. Each fragment is represented by both geometric and color features:i) 3D point positions are encoded as rotation-consistent geometric fea-tures using a rotation-equivariant encoder, while ii) the colors at each3D point are encoded with a transformer. The two feature sets are thencombined to form a multimodal representation. We experimented on fivedatasets: two synthetic datasets, Breaking Bad and Fantastic Breaks,and three real-world cultural heritage datasets, RePAIR, Presious, andFRACTURA, demonstrating that E-M3RF on the RePAIR dataset (themost challenging one) reduces rotation error by 23.1% and translationerror by 13.2%, while Chamfer Distance decreases by 18.4% comparedto competing methods. Project page: https://em3rf.github.io/.

Controllable video generation for complex hand-object interactions is a critical step toward building visual world models. However, existing methods often struggle to achieve fine-grained, 3D-consistent hand articulation in generated videos. By relying on dense 2D trajectories or implicit pose representations, they collapse crucial geometric structures into spatially ambiguous signals, leading to severe motion inconsistencies and hallucinated artifacts under egocentric occlusions. To address this, we propose leveraging sparse 3D hand joints as explicit control signals with three key advantages: explicit geometry to resolve occlusions, an intuitive interface for interactive editing, and cross-embodiment generalization to robotic hands. Built upon this, our efficient control module extracts occlusion-aware features from the source reference frame by penalizing unreliable visual features from hidden joints, and employs a 3Dbased weighting mechanism to handle dynamically occluded target joints during motion propagation. Meanwhile, it directly injects 3D geometric embeddings into the latent space to enforce structural consistency. To facilitate robust training and evaluation, we develop an automated annotation pipeline, yielding 1M high-quality egocentric video clips paired with precise hand trajectories. Experiments demonstrate that our approach outperforms state-of-the-art baselines, generating high-fidelity egocentric videos with realistic hand-object interactions.

Monocular depth estimation has seen remarkable progressthrough foundation models achieving robust zero-shot generalization, yettheir computational demands place them far beyond the reach of em-bedded and mobile platforms. Lightweight alternatives exist, but havebeen developed almost exclusively within single-domain, self-supervisedparadigms, failing silently under domain shift. We present ZipDepth, acompact monocular depth network that bridges this gap by combining anefficient reparameterizable encoder-decoder with large-scale knowledgedistillation from a foundation model over a large multi-domain train-ing set. Comprising just 6.1M parameters, ZipDepth runs at real-timerates from server GPUs to power-constrained devices, achieving the besttrade-off between zero-shot accuracy and deployment efficiency amonglightweight models across five benchmarks, taking a significant step to-wards the accuracy of foundation models with 50× more parameters.

Large-scale Vision-Language-Action (VLA) models excel atmapping natural language instructions to robotic actions. However, aprimary bottleneck in generalizing these models is the scarcity of high-quality real-world data. While fine-tuning requires massive successfuldemonstrations, the low success rates of current base policies in dy-namic environments often lead to compounding errors and inefficient,low-quality data collection. To enhance both real-world data acquisitionand long-horizon task success, we propose a proactive, online interven-tion framework. Our core innovation is a lightweight, composite WorldModel (WM) acting as an event-triggered “corrector”. Instead of con-tinuous frame-by-frame rollouts across the entire task horizon, our WMintervenes only at critical decision points (e.g., prior to grasping) to eval-uate the proposed action’s feasibility. If a high failure risk is detected, thesystem proactively rolls back to a heal state. A generative model thensynthesizes a short video of a successful future trajectory, which guidesthe robot’s re-attempt via an action decoder. We validate our frameworkacross multiple robotic grasping tasks on both simulation and real-worldsystems. Extensive results demonstrate that deploying world models asactive, on-demand correctors significantly improves task success ratesand efficiently harvests high-quality execution data by rescuing poten-tial failures before they occur.

We present Wid3R, a feed-forward neural network for multi-view visual geometry reconstruction that supports wide field-of-viewcamera models. Unlike existing methods that assume rectified or pinholeinputs, Wid3R directly models wide-angle imagery without explicit cali-bration or undistortion. Our approach leverages a ray-based representa-tion with spherical harmonics and introduces a novel camera model tokento enable distortion-aware reconstruction. To the best of our knowledge,Wid3R is the first multi-frame feed-forward 3D reconstruction methodthat supports 360◦ imagery. Moreover, we show that conditioning on di-verse camera types improves generalization to 360◦ scenes and alleviatesdata sparsity issues. Wid3R achieves significant performance gains, im-proving AUC@30◦ by up to +33.67 on Zip-NeRF (fisheye) and +77.33 onStanford2D3D (360). Project Page: https://jdk9405.github.io/Wid3R/

As virtual try-on (VTON) advances, a growing number ofreal-world scenarios have emerged, pushing beyond the ability of the ex-isting specialized VTON models. Meanwhile, universal multi-referenceimage editing models have progressed rapidly and exhibit strong gen-eralization in visual editing, suggesting a promising route toward moreflexible VTON systems. However, despite their strong capabilities, thestrengths and limitations of universal editors for VTON remain insuffi-ciently explored due to the lack of systematic evaluation benchmarks. Toaddress this gap, we introduce VTEdit-Bench, a comprehensive bench-mark designed to evaluate universal multi-reference image editing modelsacross various realistic VTON scenarios. VTEdit-Bench contains 24,220test image pairs spanning five representative VTON tasks with progres-sively increasing complexity, enabling systematic analysis of robustnessand generalization. We further propose VTEdit-QA, a reference-awareVLM-based evaluator that assesses VTON performance from three keyaspects: model consistency, cloth consistency, and overall image quality.Through this framework, we systematically evaluate 8 universal editingmodels and compare them with 7 specialized VTON models. Resultsshow that top universal editors are competitive on conventional tasksand generalize more stably to harder scenarios, but remain challengedby complex reference configurations, especially multi-cloth conditioning.

We propose the first compression approach for image-to-shape Diffusion Transformers (DiTs) that substantially reduces modelsize while preserving geometric fidelity. Despite remarkable progress in3D shape generation, large DiT-based models remain computationallyprohibitive in resource-constrained settings. Furthermore, it is difficult todirectly transfer existing diffusion model compression strategies developedfor different domains to 3D generation, and prior 3D efficiency approachesfocus primarily on inference speed rather than backbone compression. Toaddress this limitation, we build a geometry-aware compression frameworktailored to image-to-shape DiTs. Guided by the observation that 3DDiT layers exhibit non-uniform importance for geometry synthesis, weintroduce a vitality-guided framework integrating structured pruning,adaptive quantization, and targeted fine-tuning. Our method achieves upto 66% model-size reduction across state-of-the-art image-to-3D modelswhile maintaining synthesis fidelity comparable to full-sized counterparts.This highlights the potential of our framework as a plug-and-play solutionfor efficient 3D shape generation across diverse models.

Existing 3D Gaussian Splatting (3DGS) frameworks rely oncamera-speci_x001C_c rasterization, su_x001B_ering from inconsistent solid-angle sam-pling and degraded performance across heterogeneous camera models(e.g., perspective, _x001C_sheye, omnidirectional). To address this limitation,we propose UniTriSplat, a uni_x001C_ed 3DGS framework for universal camerasthat reformulates Gaussian splatting on the unit sphere via HEALPixdiscretization. Leveraging the equal-area property of HEALPix, we con-struct a spherical sampling grid aligned with the angular resolution ofinput images. We derive the forward rendering and gradient propagationof Gaussians directly in the spherical radian domain, yielding uniform op-timization behavior from narrow-FoV images to full 360-degree panora-mas. To enhance perceptual reconstruction quality, we additionally intro-duce a HEALPix-aware SSIM loss that respects spherical neighborhoodstructure. Extensive experiments across diverse camera models demon-strate that UniTriSplat consistently improves cross-camera generaliza-tion while preserving geometric _x001C_delity and rendering quality. Projectpage: https://yipengzhu0809.github.io/UniTriSplat/

We present UniMotion, to our knowledge the first unifiedframework for simultaneous understanding and generation of humanmotion, natural language, and RGB images within a single architecture.Existing unified models handle only restricted modality subsets (e.g.,Motion–Text or static Pose–Image) and predominantly rely on discretetokenization, which introduces quantization errors and disrupts temporalcontinuity. UniMotion overcomes both limitations through a core princi-ple: treating motion as a first-class continuous modality on equal footingwith RGB. A novel Cross-Modal Aligned Motion VAE (CMA-VAE) andsymmetric dual-path embedders construct parallel continuous pathwaysfor Motion and RGB within a shared LLM backbone. To inject visual-semantic priors into motion representations without requiring images atinference, we propose Dual-Posterior KL Alignment (DPA), which distillsa vision-fused encoder’s richer posterior into the motion-only encoder. Toaddress the cold-start problem—where text supervision alone is too sparseto calibrate the newly introduced motion pathway—we further proposeLatent Reconstruction Alignment (LRA), a self-supervised pre-trainingstrategy that uses dense motion latents as unambiguous conditions toco-calibrate the embedder, backbone, and flow head, establishing a stablemotion-aware foundation for all downstream tasks. UniMotion achievesstate-of-the-art performance across seven tasks spanning any-to-any un-derstanding, generation, and editing among the three modalities, withespecially strong advantages on cross-modal compositional tasks.

Continuous variational autoencoders (VAEs) are widely usedas visual tokenizers for latent diffusion models. While CNN-based VAEsreliably capture local details, they model long-range dependencies ineffi-ciently. Vision Transformers (ViTs) provide global context, but existingViT-VAEs exhibit three practical limitations in this setting: (i) weakresolution extrapolation when trained at a single resolution, (ii) slowconvergence with suboptimal fine-grained color and texture fidelity, and(iii) inconsistent scaling gains when increasing model size. To addressthese challenges, we propose TransVAE, a hybrid CNN-ViT VAE thatcombines a shallow CNN front-end for local feature extraction with adeep Transformer backbone for global modeling. We apply minimal yetprincipled architectural modifications and systematically analyze their ef-fectiveness, including a pure Rotary Position Embedding (RoPE) strategy,a multi-stage design, and a convolutional feed-forward network. Acrossmodel sizes from 44M to 2.3B parameters, TransVAE shows learning-efficient training and predictable scaling improvements, and it supportssingle-resolution training (256×256) with robust inference at higher resolu-tions (512 × 512/1024 × 1024). Moreover, TransVAE achieves competitivereconstruction and downstream generation performance compared tostrong tokenizers, such as FLUX-VAE and VA-VAE, offering a favorabletrade-off between reconstruction fidelity and generation-friendly latents.

We present RefVFX, a new framework that transfers complextemporal effects from a reference video onto a target video or image ina feed-forward manner. While existing methods excel at prompt-basedor keyframe-conditioned editing, they struggle with dynamic temporaleffects such as dynamic lighting changes or character transformations,which are difficult to describe via text or static conditions. Transferringa video effect is challenging, as the model must integrate the new tem-poral dynamics with the input video’s existing motion and appearance.To address this, we introduce a large-scale dataset of triplets, where eachtriplet consists of a reference effect video, an input image or video, and acorresponding output video depicting the transferred effect. Creating thisdata is non-trivial, especially the video-to-video effect triplets, which donot exist naturally. To generate these, we propose a scalable automatedpipeline that creates high-quality paired videos designed to preserve theinput’s motion and structure while transforming it based on some fixed,repeatable effect. We then augment this data with image-to-video effectsderived from LoRA adapters and code-based temporal effects generatedthrough programmatic composition. Building on our new dataset, wetrain our reference-conditioned model using recent text-to-video back-bones. Experimental results demonstrate that RefVFX produces visuallyconsistent and temporally coherent edits, generalizes across unseen effectcategories, and outperforms prompt-only baselines in both quantitativemetrics and human preference. See our website at this URL

3D object classification is a crucial problem due to its significant practical relevance in many fields, including computer vision, robotics, and autonomous driving. Although deep learning methods applied to point clouds sampled on CAD models of objects and/or captured by LiDAR or RGBD cameras have achieved remarkable success in recent years, achieving high classification accuracy remains a challenging problem due to the unordered point clouds and their irregularity and noise. To this end, we propose a novel state-of-the-art (SOTA) 3D object classification technique that combines topological data analysis with various image filtration techniques to classify objects when they are represented using point clouds. We transform every point cloud into a voxelized binary 3D image to extract distinguishing topological features. Next, we train a lightweight one-dimensional Convolutional Neural Network (1D CNN) using the extracted feature set from the training dataset. Our framework, TACO-Net, sets a new state-of-the-art by achieving 99.05% and 99.52% accuracy on the widely used synthetic benchmarks ModelNet40 and ModelNet10. It also achieved one of the highest accuracies among all the from-scratch supervised learning techniques when tested on the hardest variant of the ScanObjectNN dataset. When tested with ten different kinds of corrupted ModelNet40 inputs, the proposed TACO-Net demonstrates strong resiliency overall. The code can be found in the supplementary material.

Table parsing aims to convert table images into structured,machine-readable representations, a task requiring the joint perceptionof complex spatial layouts and textual content. While recent vision-language models (VLMs) enable end-to-end parsing, they typically relyon direct supervision of the final output, thereby bypassing the explicitintermediate reasoning that is crucial for understanding complex tablestructures. Furthermore, attempts to optimize these models using re-inforcement learning (RL) are often hindered by unstable or ambigu-ous reward designs, limiting potential performance gains. To addressthese limitations, we propose StrucTab, a table parsing model learnedthrough intermediate structural supervision and reward decomposition.At the modeling level, by decomposing the parsing process into human-inspired subtasks, such as row-column counting and merged-cell analy-sis, StrucTab progressively unifies them through a sequential reasoningstrategy. At the optimization level, we introduce Uni-TabRL, a unifiedRL framework that leverages decomposed rewards (validity, structure,and content) to provide stable and informative optimization signals. Fi-nally, at the evaluation level, we present TableVerse-5K, a large-scale,challenging benchmark encompassing diverse, real-world table scenarios.Extensive experiments demonstrate the state-of-the-art performance ofStrucTab across all evaluated public benchmarks and significant improve-ments on TableVerse-5K, validating the effectiveness of explicit structuralmodeling and decomposed reward optimization. Code and benchmark arepublicly available at https://github.com/VirtualLUOUCAS/StrucTab.

Recently, remarkable progress has been made in Uni_x001C_ed Mul-timodal Models (UMMs), which integrate vision-language generation andunderstanding capabilities within a single framework. However, a model'sstrong visual understanding often fails to transfer to visual generation:it may correctly judge prompt-image alignment while failing to gener-ate a faithful image from the same prompt. This raises a compellingquestion: Can a model improve itself by using its understanding moduleto reward its generation module? We introduce SRUM, a self-rewardingpost-training framework directly applicable to existing UMMs of vari-ous designs. SRUM creates a feedback loop where the model's own un-derstanding module acts as an internal _x0010_evaluator_x0011_, providing correctivesignals to improve generation without additional human-labeled data orexternal reward models. To provide comprehensive feedback, SRUM usesa global-local dual reward system: a global reward ensures overall vi-sual semantics and layout, while a local reward re_x001C_nes _x001C_ne-grained,object-level _x001C_delity. SRUM shows strong generalization, boosting perfor-mance on T2I-CompBench from 82.18 to 88.37 and on T2I-ReasonBenchfrom 43.82 to 46.75. Overall, our work establishes a powerful paradigmfor enabling a UMM's understanding module to guide and enhance itsown generation via self-rewarding.

Driving scene manipulation using real-world sensor data hasemerged as a promising alternative to traditional driving simulators. De-spite advances in language control and neural scene representations, ex-isting methods treat grounding, editing, and simulation as loosely con-nected stages, relying on heuristic object localization, manual guidance,and single-agent validation—thereby constraining semantic expressive-ness and hindering scalable, reactive scenario generation. We introduceSIMSplat, a driving scene editor built on scene-graph-based 4D Gaus-sian Splatting augmented with language-aligned features. By embed-ding appearance, motion, and location semantics directly into Gaus-sian scene-graph nodes, SIMSplat makes reconstructed scenes queryablethrough free-form natural language, bridging language understanding toobject-level editing and multi-agent simulation within a single frame-work. Building on this language-grounded scene graph, SIMSplat sup-ports diverse edits including fine-grained pedestrian manipulation, whilea multi-agent path refinement module propagates changes across allagents to ensure reactive, physically plausible simulations. The pipelinefurther integrates with Vision-Language Models for automated scenariomining. Experiments show that SIMSplat more than doubles baselinegrounding accuracy, achieves the highest task completion rate, and pro-duces the lowest failure rates across diverse driving scenarios.

Normalizing Flows (NFs) learn invertible mappings betweenthe data and a Gaussian distribution. Prior works usually suffer fromtwo limitations. First, they add random noise to training samples orVAE latents as data augmentation, introducing complex pipelines in-cluding extra noising and denoising steps. Second, they use a pretrainedand frozen VAE encoder, resulting in suboptimal reconstruction and gen-eration quality. In this paper, we find that the two issues can be solvedin a very simple way: just fixing the variance (which would otherwise bepredicted by the VAE encoder) to a constant (e.g., 0.5). On the one hand,this method allows the encoder to output a broader distribution of tokensand the decoder to learn to reconstruct clean images from the augmentedtoken distribution, avoiding additional noise or denoising design. On theother hand, fixed variance simplifies the VAE evidence lower bound, mak-ing it stable to train an NF with a VAE jointly. On the ImageNet 256×256generation task, our model SimFlow obtains a gFID score of 2.15, outper-forming the state-of-the-art method STARFlow (gFID 2.40). Moreover,SimFlow can be seamlessly integrated with the end-to-end representa-tion alignment (REPA-E) method and achieves an improved gFID of1.91, setting a new state of the art among NFs.

Diffusion Large Language Models (dLLMs) have achievedremarkable success in pure language tasks and recently shown promis-ing results in multimodal understanding. However, existing post-trainingmethods for multimodal dLLMs have not yet fully exploited the uniquestructure and inference paradigm of diffusion models, leaving room formore tailored optimization strategies. We present Bi-VRL, a reinforce-ment learning framework specifically designed for multimodal dLLMs,featuring a bidimensional optimization strategy that enables com-prehensive optimization across both reasoning and diffusion timestepdimensions. For the diffusion timestep dimension, our trajectory-leveloptimization learns a value function over the entire diffusion path, en-abling holistic credit assignment across denoising timesteps. For the rea-soning dimension, our thought-level optimization decomposes outputsinto reasoning steps and delivers targeted rewards to enhance logicalsoundness. This core strategy is augmented by a fine-grained multimodalalignment mechanism that systematically evaluates both the logical cor-rectness of reasoning steps and their grounding in visual evidence. Ourframework is architecture-agnostic, successfully applied to models withboth discrete vision tokens (MMaDA) and continuous vision embed-dings (LLaDA-V). Extensive experiments across challenging multimodalreasoning benchmarks demonstrate that Bi-VRL significantly outper-forms strong supervised fine-tuning baselines, establishing effective post-training methodology for multimodal dLLMs. We will release the codeat https://github.com/Gen-Verse/dLLM-RL.