ECCV 2026 论文列表

A common way to improve vision-language model outputsis Best-of-N (BoN) selection: generate N candidate answers, score eachone, return the best. When the candidate pool and compute budget areheld fixed, the selected answer should be reproducible. We show it often isnot. Slightly adjusting the softmax temperature, toggling a probability-truncation option, or enabling dropout noise during the scoring passis enough to change the winning answer on 41–47% of test items —even though the candidate pool is identical. This failure is specific tolikelihood-based selection, where candidates are ranked by policy likeli-hoods or policy/reference likelihood ratios; a deterministic reward-modelor rule-based (e.g. RLVR) reranker assigns identical scores to identicaltext and is stable on a fixed pool by construction.Standard scoring conflates answer quality with two confounding factors:when token-level scores are summed, low-value suffixes can change thetotal without improving answer quality, and score magnitudes shift when-ever the model is updated. We propose DSPA, which separates two con-cerns into a test-time selector and a training-time filter. At test time,we score each candidate by comparing it token-by-token against a fixedreference model and averaging the result, which removes the additivelength accumulation of summed scores and keeps all scores on a fixedreference scale; this is the component responsible for selection stability.At training time, we filter preference pairs to prevent shortcuts based onanswer length, verbatim prompt copying, or hallucinated objects; thiscomponent improves the factual quality of the candidate pool, not thestability of selection. We evaluate stability by replaying the scoring stepon the same frozen pool under controlled perturbations (a protocol wecall SRP), where the anchored selector reduces the answer-flip rate from44.3% to 24.9%. Manual inspection shows roughly 70% of remainingflips are near-ties that differ only in phrasing. On identical pools, the an-chored selector also reduces object hallucination (POPE) from 9.6% to6.2%; the stability gain holds at 13B scale, while the utility gain transfersto Qwen2.5-VL-7B-Instruct.

Physical adversarial camouflage poses a significant threatto object detectors in safety-critical applications. However, synthesiz-ing robust 3D textures remains non-trivial due to the structural mis-match between static texture parameters and dynamic physical obser-vations. In this paper, we identify two fundamental optimization hur-dles: spatially inconsistent gradients arising from sparse rendering vis-ibility; and cross-view gradient misalignment, where contradictory sig-nals from diverse viewpoints cause destructive interference. To overcomethese, we propose a unified framework that reconciles spatial continu-ity with optimization stability. First, the Deformable Texture Field pa-rameterizes textures as continuous fields via learnable flow grids, en-forcing intrinsic smoothness while enabling adaptive geometric defor-mation to enhance the texture’s representational flexibility. Second, aGradient-Aligned Meta-Optimization strategy leverages multi-step ex-ploration trajectories to implicitly maximize gradient alignment acrossconflicting viewpoints. Extensive digital and physical experiments demon-strate that our method achieves state-of-the-art robustness and transfer-ability against advanced detectors under diverse viewing conditions.

World models have become central to autonomous driving,where accurate scene understanding and future prediction are crucialfor safe control. Recent work has explored using vision–language models(VLMs) for planning, yet existing approaches typically treat perception,prediction, and planning as separate modules. We propose UniDrive-WM, a unified VLM-based world model that jointly performs driving-scene understanding, trajectory planning, and trajectory-conditioned fu-ture image generation within a single architecture. UniDrive-WM’s tra-jectory planner predicts a future trajectory, which conditions a VLM-based image generator to produce plausible future frames. These predic-tions provide additional supervisory signals that enhance scene under-standing and iteratively refine trajectory generation. We further comparediscrete and continuous output representations for future image predic-tion, analyzing their influence on downstream driving performance. Ex-periments on the challenging Bench2Drive benchmark show that UniDrive-WM produces high-fidelity future images and improves planning perfor-mance by 7.3% in L2 trajectory error and 10.4% in collision rate overthe previous best method. These results demonstrate the advantagesof tightly integrating VLM-driven reasoning, planning, and generativeworld modeling for autonomous driving. The project page is available athttps://unidrive-wm.github.io/UniDrive-WM.

While Omni-modal Large Language Models have made strides in joint sensory processing, they fundamentally struggle with a cornerstone of human interaction: deciphering complex, multi-person conversational dynamics to accurately answer “Who said what and when.” Current models suffer from an “illusion of competence”—they exploit visual biases in conventional benchmarks to bypass genuine cross-modal alignment, while relying on sparse, low-frame-rate visual sampling that destroys crucial high-frequency dynamics like lip movements. To address this limitation, we introduce Visual-Registered Speaker Diarization and Recognition (VR-SDR) and the HumanOmni-Speaker Benchmark. By strictly eliminating visual shortcuts, this rigorous paradigm demands true end-to-end spatio-temporal identity binding using only natural language queries. To overcome the underlying architectural perception gap, we propose HumanOmni-Speaker, powered by a Visual Delta Encoder. By sampling raw video at 25 fps and explicitly compressing inter-frame motion residuals into just 6 tokens per frame, it captures fine-grained visemes and speaker trajectories without triggering a catastrophic token explosion. Ultimately, HumanOmni-Speaker demonstrates strong multimodal synergy, natively enabling end-to-end lip-reading and high-precision spatial localization without intrusive cropping, and achieving superior performance across a wide spectrum of speaker-centric tasks.

Segmenting 3D assets into their constituent parts is crucialfor enhancing 3D understanding, facilitating model reuse, and supportingvarious applications such as part generation. However, current methodsface limitations such as poor robustness when dealing with complex ob-jects and cannot fully automate the process. In this paper, we propose anative 3D point-promptable part segmentation model termed P3 -SAM,designed to fully automate the segmentation of any 3D objects into com-ponents. Inspired by SAM, P3 -SAM consists of a feature extractor, mul-tiple segmentation heads, and an IoU predictor, enabling interactive seg-mentation for users. We also propose an algorithm to automatically selectand merge masks predicted by our model for part instance segmentation.Our model is trained on a newly built dataset containing nearly 3.7 mil-lion models with reasonable segmentation labels. Comparisons show thatour method achieves precise segmentation results and strong robustnesson any complex objects, attaining state-of-the-art performance. Our codewill be released soon.

Reasoning agents increasingly rely on external tools such asweb search to answer complex queries. Reinforcement learning (RL) fine-tuning algorithms such as GRPO have improved long-form reasoningin text-only language models, particularly for coding and mathematics.Reliable tool use in multimodal agents, however, remains challengingbecause models must interpret text and images while integrating noisyretrieved evidence, often under sparse outcome-level supervision withoutexplicit verification signals. We present Self-Verification via Reinforce-ment Learning (SVRL), an RL-only finetuning framework that trainsmultimodal agents to verify and filter retrieved evidence within theirown reasoning traces, reducing reliance on external verifiers at infer-ence time. SVRL also introduces a search-aware penalty that discour-ages unnecessary tool calls and a query-diversity reward that encouragesdiverse, well-formed search queries, providing fine-grained feedback onwhen and what to search. Finetuning Qwen-2.5-VL-7B with SVRL ononly 5,000 visual question answering examples yields consistent gainsin multi-hop VQA generalization and tool efficiency across benchmarks.Overall, SVRL narrows the gap between compact agents and much largerproprietary models while requiring substantially lower training and in-ference cost.

Recently, Large Language Models (LLMs) have emerged aspromising layout agents for 3D scene generation. Existing layout agentsstill suffer from implausible layout generation because most of them con-vert 3D assets and 3D layouts into textual descriptions as inputs andoutputs, which involves severe information loss due to the modality gapbetween texts and 3D assets and 3D layouts. We propose NaLA, a native3D LLM layout Agent for high-quality 3D scene generation by placing 3Dassets in the scene. For the inputs, NaLA encodes 3D scene boundariesand 3D assets directly into the LLM, preserving fine-grained geometryand enabling explicit reasoning over relationships like collisions, surfacesupporting, and containment. To accurately output the positions and ori-entations of assets, NaLA adopts a coarse-to-fine prediction mechanismthat first predicts discrete poses in an autoregressive manner and thenrefines the discrete poses with a continuous regression. Trained on diverselayout datasets, NaLA attains strong geometric perception and layout co-herence. Experiments demonstrate that NaLA outperforms prior layoutagents in both generation quality and inference efficiency, with compre-hensive ablation studies to verify each component’s effectiveness.

For image editing with Multimodal Diffusion Transformers(MM-DiT), training-free methods face a critical trade-off: precise back-ground preservation limits editability, while high prompt fidelity de-grades the original scene and object structure. We argue that this tensionstems from suboptimal Key-Value (KV) injection strategies. Global KV-injection rigidly over-preserves the source, whereas strictly masked injec-tion acts as localized inpainting, destroying the edited object’s structuralidentity. We introduce LayerVerse, an optimization framework that re-solves this fundamental dilemma of layer selection for KV-injection byassigning specific layers to two distinct roles: Masked Injection to pre-cisely protect the background, and Global Injection to anchor the editedobject’s structure. Formulating optimal layer allocation as a combina-torial problem, we model the editing error based on pairwise layer in-teractions and solve it globally via Mixed-Integer Linear Programming(MILP). Supported by lightweight, single-step automatic masking, Lay-erVerse seamlessly integrates into training-free, inversion-based editors,achieving highly competitive results on the PIE-Bench benchmark.

Exemplar-free class-incremental learning (EFCIL) requires stable decision boundaries within a shifting feature space. While maintaining class-conditional Gaussian statistics provides a principled classification strategy, these parametric summaries remain sensitive to anisotropic representation drift. Existing methods often transport these statistics across tasks using a decoupled, post-hoc paradigm: optimizing a backbone without explicit geometric constraints can distort the legacy manifold, limiting the precision of retroactive alignment. In this paper, we formulate feature transport as an endogenous training constraint rather than a separate post-task step, presenting the Geometry-Anchored Transport Framework. First, we derive an Analytic Geometric Anchor via Mahalanobis-aligned regression to mitigate macroscopic anisotropic drift. Second, we introduce a Topology-Aware Evolution objective that regularizes localized manifold degradation while calibrating a residual network against the analytic prior. By coupling manifold evolution with transport constraints during the primary training phase, our framework mitigates evaluation errors without requiring decoupled fine-tuning. Experiments across CIFAR-100, TinyImageNet, and ImageNet-100 demonstrate that the proposed framework consistently improves upon existing post-hoc alternatives under strict exemplar-free constraints. The code is available at https://github.com/HXuSz11/GATF_ECCV2026.

We present RegHead, a framework for constructing semanticblendshape sets for animatable non-humanoid head avatars. With a fixedexpression vocabulary, semantic blendshapes provide a low-dimensionaland interpretable animation interface and support cross-identity retar-geting. Building such blendshape sets remains expensive because (i)expression-consistent supervision is scarce, (ii) generated 4D assets typ-ically lack correspondence, and (iii) facial motion is highly localized.We propose (1) a large-scale dataset of non-humanoid identities pairedwith a shared expression vocabulary, obtained by expanding a smallartist-rigged library via fine-tuned image editing; (2) a dense stochasticanchor motion representation tailored to localized facial deformations;and (3) a fast feed-forward registration model that converts unregis-tered expression meshes into a corresponded blendshape basis by pre-dicting anchor-based deformations from the neutral shape. Experimentsshow that our approach produces higher-fidelity expression meshes thanbaselines, while running orders of magnitude faster than optimization.We further demonstrate real-time retargeting from human face trackingsignals to non-humanoid characters, capturing both head pose and lo-calized facial motions. Our project page is available at https://snap-research.github.io/RegHead/.

Reconstructing high-fidelity, relightable 3D avatars from asingle in-the-wild image is a challenging ill-posed problem, primarilyhindered by the scarcity of high-quality PBR data and the complex-ity of disentangling illumination from intrinsic materials. In this paper,we present a data-efficient framework that leverages the robust priorsof a unified pre-trained diffusion backbone to sequentially address tex-ture completion, delighting, and material decomposition. Unlike exist-ing methods that rely on fragmented pipelines or extensive proprietarydatasets, we utilize cascaded Low-Rank Adaptations (LoRAs) to adaptthe strong generative prior of the diffusion model for each sub-task inUV space. Specifically, we first employ an Inpainting LoRA to completemissing UV textures caused by occlusion, leveraging the model’s seman-tic understanding to generate semantically and photometrically coher-ent details. Subsequently, a Light-Homogenization LoRA and a novelCross-Intrinsic Attention mechanism are introduced to remove baked-inlighting and collaboratively synthesize pixel-aligned PBR maps (Albedo,Normal, Roughness, Specular, and Displacement). To ensure physicalplausibility, we impose a UV-space differentiable BRDF shading lossduring the decomposition stage, forcing the generative process to adhereto the rendering equation without the artifacts typical of rasterization-based supervision. Extensive experiments demonstrate that our method,trained on fewer than 100 real 3D scans, generates comprehensive, 4K-resolution PBR assets with superior realism and generalization comparedto state-of-the-art methods, and all training code and model weights willbe released upon acceptance.

Controllable video generation remains challenging due tothe di"culty of specifying precise multi-object interactions using textprompts or motion-control inputs that primarily constrain pixel move-ment. In practice, trajectory-based control often requires users to drawaccurate tracks for multiple objects, which scales poorly with scene com-plexity and becomes ambiguous under occlusion or overlap. To enableflexible yet precise multi-subject control, we introduce GraphVid, agraph-conditioned image-to-video generation model that enables inter-active control through structured interaction graphs. We further cu-rate GraphVid-Bench, a large-scale interaction-centric video datasetwith structured relational annotations to enable training of interaction-aware video generation models. Despite using substantially less trainingdata and fewer trainable parameters than prior motion-control methods,GraphVid delivers strong controllability and video quality. Comparedwith Motion-I2V, GraphVid reduces FID by up to 39.9% and FVD by37.6%, while improving PSNR (9.87→15.98) and SSIM (0.38→0.61). Ourresults highlight the potential of structured semantic interfaces as a pow-erful paradigm for controllable video generation.PLAN Lab https://plan-lab.github.io/graphvid

Extracting dynamic 4D object interactions from massive, inthe-wild monocular videos offers a highly efficient data collection pathway for scaling Embodied AI and training VLAs. However, existing monocular 4D reconstruction methods primarily focus on isolated objects, often failing under the severe occlusions and complex dynamics inherent in multi-object interactions. To bridge this gap, we propose HAT- 4D, the first agentic framework designed to reconstruct the 3D geometry, temporal dynamics, and physical interactions of multiple objects from a single video. By integrating VLMs with a multi-level human-in-the-loop feedback mechanism, HAT-4D efficiently resolves depth ambiguities and interaction-induced occlusions during 3D generation and 4D propagation, yielding physically plausible assets without relying on expensive multicamera rigs. As a scalable data engine, HAT-4D facilitates the creation of MVOIK-4D, an open-world benchmark for monocular 4D interaction reconstruction, accompanied by a novel multi-dimensional evaluation protocol focused on physical plausibility and temporal consistency. Extensive experiments demonstrate that HAT-4D achieves SOTA performance on most evaluation metrics, while maintaining competitive semantic alignment. Ablation studies show that introducing a small amount of human feedback improves interaction reconstruction. Moreover, the data produced by HAT-4D effectively improves baseline performance when used for finetuning. Our data and code are available at project webpage.

Diffusion-based text-to-motion models synthesize realistic hu-man motions but often exhibit semantic drift from the input text. Mo-tion is inherently temporal, especially in compositional and long-durationsequences that require semantic consistency across multiple action seg-ments and smooth kinematic transitions throughout the trajectory. Weposit that the initial noise is central to this consistency: within the Gaus-sian noise space, certain instances, i.e. winning noise tickets, carry latentstructure that biases denoising toward particular motion semantics, evenunder null prompts. We propose WInning Noise Retrieval and Optimiza-tion (WINRO), a training-free, model-agnostic framework that improvestext–motion alignment by selecting and refining such tickets before diffu-sion sampling. WINRO maps random noises to motion features generatedunder null prompts, retrieves the best-aligned noise for a given text, andrefines it via a KL-regularized objective that reduces the residual seman-tic gap while preserving the Gaussian prior. An optional LoRA-basedadapter amortizes this refinement into a single forward pass. WINROconsistently improves text–motion fidelity across different base models,MDM and MotionLCM, on HumanML3D without retraining, improvestemporal robustness on the MTT benchmark, and generalizes to appli-cations such as motion stylization and spatial constraint satisfaction.

Event cameras provide several unique advantages over standard frame-based sensors, including high temporal resolution, low latency, and robustness to extreme lighting. However, existing learningbased approaches for event processing are typically confined to narrow, task-specific silos and lack the ability to generalize across modalities. We address this gap with REALM, a cross-modal framework that learns an RGBand Event-Aligned Latent Manifold by projecting event representations into the pretrained latent space of RGB foundation models. Instead of task-specific training, we leverage low-rank adaptation (LoRA) to bridge the modality gap, effectively unlocking the geometric and semantic priors of frozen RGB backbones for asynchronous event streams. We demonstrate that REALM effectively maps events into the ViT-based foundation latent space. Our method performs downstream tasks, such as depth estimation and semantic segmentation, by simply transferring linear heads trained on the RGB teacher. Most significantly, REALM enables the direct, zero-shot application of complex, frozen image-trained decoders, such as MASt3R, to raw event data. We demonstrate stateof-the-art performance in wide-baseline feature matching, significantly outperforming specialized architectures. Code and models are available at https://papers.starslab.ca/realm/.

Vision language action models (VLAs) have advanced general purpose robotic grasping and manipulation by enabling robots to interpret visual observations and natural language instructions to generate executable action sequences. However, existing VLAs often implicitly assume gripper invariance, despite grasping strategies being inherently embodiment-dependent. Different gripper types, such as parallel-jaw and suction, usually require distinct interaction strategies to achieve the same grasping objective. Moreover, current datasets for VLAs predominantly rely on parallel-jaw grippers, limiting gripper-aware learning. To address this gap, we introduce MiGA, a multi-gripper-aware dataset spanning five distinct gripper types across multiple robots with 103,000 demonstrations, explicitly capturing strategy divergence under shared task objectives. We further propose GVLA, which combines a new multi-gripper tokenizer with adapter-based policy routing. Our new gripper encoding induces structured embedding information that balances parameter sharing and strategy differentiation, while layer-wise probing confirms meaningful gripper-conditioned representations for VLAs. Intensive experiments in both simulation and real-world robots show that our GVLA outperforms the current baselines across evaluated settings. Our method also improves zero-shot generalization or few-shot adaptation to new objects or unseen tasks, and enable more efficient gripper adaptation.

4D volumetric video via Gaussian Splatting (GS) enables free viewpoint control of real-world scenes but does not enable motion editing. One challenge is achieving part-based control in multi-object scenes; prior works focus on single objects, assume rigid parts, or require manual annotations. MotionSplicer is a template-free and annotation-free method that automatically discovers 3D parts in GS. To bridge the semanticto-motion domain gap, it extracts initial parts from temporal difference images and resolves multi-view inconsistencies via an adaptive 3D unifyand-split process. For robust tracking, it optimizes a time-invariant feature grid with temporal refinement to learn spatially and temporally coherent skinning weights. Experiments show that MotionSplicer can be applied more generally to objects and scenes than state-of-the-art methods, and enables creative motion edits like isolation, ghosting, and echo.

Video Question Grounding (VideoQG) requires models toanswer questions and localize supporting temporal evidence. Despiteprogress in video-language models, they remain susceptible to variousbiases. Prior debiasing methods primarily focus on dataset-level correla-tions, while easily overlooking intra-video temporal bias, where visuallysimilar segments differ substantially in their causal relevance to answer-ing the question. To address this challenge, we present Wavelet-basedIntra-video Causal Intervention (WICI), a framework designed to dis-entangle causal temporal evidence from redundant visual contexts. OurWICI comprises two key components: (1) an Intra-video Causal Interven-tion module, which utilizes counterfactual reasoning to suppress irrele-vant intra-video biases, and (2) Wavelet-based Dynamic Modeling, whichdecomposes contextual variations between visually similar frames to cap-ture fine-grained temporal cues. Extensive experiments on two VideoQGbenchmarks demonstrate that the proposed WICI significantly improvesquestion grounding accuracy and yields more robust, faithful questionreasoning.

Upgrading embedding models typically requires expensivedatabase re-indexing, as new query embeddings are incompatible withexisting database embeddings. While Backward Compatible Training(BCT) mitigates this by enforcing compatibility during training, ex-isting approaches often require updating the backbone model. This isimpractical because of significant training cost, the risk of performanceregression, and limited access to proprietary model weights. We intro-duce Multi-modal Knowledge Preserving Adapter (MKP-Adapter),the first adapter-only BCT approach for Multi-Modal Large LanguageModels (MLLMs) that requires no backbone updates. We identified thatthe primary challenge in adapter-only BCT is preserving the knowledgeof the new embeddings while enforcing backward compatibility. Hence,we propose a multi-level preservation loss that maintains the geometricstructure of the embedding spaces throughout BCT. Furthermore, a focalre-weighting strategy is integrated to prioritize learning from challeng-ing samples. Experiments demonstrate that our method achieves strongbackward compatibility across diverse multi-modal benchmarks (image,text, visual document, and video retrieval tasks) and model types. No-tably, MKP-Adapter is trained solely on pre-extracted embeddings andrequires only negligible additional latency relative to the original back-bone forward pass, highlighting its efficiency.