ECCV 2026 Papers

Optimizing vertex deformations from video for mesh anima-tion is constrained by rendering-based reconstruction losses. While ex-isting approaches improve mesh representations, supervision signals, oranimation paradigms, their supervision remains confined to the 2D do-main. Such 2D supervision is limited: it provides no signal for occludedregions and only indirect cues for visible areas. Consequently, these meth-ods often suffer from severe shape and motion artifacts. To address thislimitation, we propose Shape Flow Guidance (SFG), a sequence of 3Dshapes derived from videos, which serves as explicit 3D supervision formesh animation. Specifically, SFG is elicited by intervening in the sam-pling process of a pretrained mesh generator in a training-free manner.We further tailor a skeletal animation model that separates local defor-mation from global transformation. This model enables SFG to guidecomplex local motion while reserving rendering-based losses for simpleglobal motion. Extensive experiments confirm that our method signifi-cantly outperforms prior work qualitatively, quantitatively, and in termsof processing speed. Qualitative results are available on our project page:https://sfgmesh.pages.dev/.

Most existing mesh-based methods for garment simulationsuffer from sensitivities tied to mesh discretisation and resolution. Re-cent approaches such as NeuralClothSim employ continuous neural fieldsas a promising alternative, but remain constrained by a 2D curvilinearparameterisation that models a single cloth panel, limiting their abil-ity to represent garments with multiple panels stitched together. Thispaper introduces NeuralGarSim, a quasistatic garment simulator for-mulated directly in 3D Euclidean space. Our method accepts diverseinput representations—including distance fields, meshes, point clouds,and Gaussians—as undeformed garment states, making it compatible withoff-the-shelf reconstruction frameworks without additional preprocessing.We then represent the garment deformation as a neural field and define anonlinear Kirchhoff–Love shell model directly in R3 by applying tangentialdifferential calculus on the garment surface. This ensures that the neuralgarment simulation is parameterisation-space-free and naturally extendsto arbitrary topologies. By minimising a potential energy functional, ourmethod learns a 3D neural deformation field that predicts physicallyconsistent deformations across garments with multiple panels, seams,and holes. It supports multiple material models and trains 2–4× fasterthan its predecessor, NeuralClothSim, while retaining the continuous,consistent, and memory-adaptive behaviour1 .

Achieving precise, versatile whole-body character control inphysics-based animation remains challenging. Recent diffusion-based poli-cies generate rich and expressive motions but typically rely on gradient-based test-time guidance to satisfy task objectives, which is slow andcan reduce robustness. We introduce NaP-Control (Navigating Diffu-sion Prior for Versatile and Fast Character Control), abbreviated asNaP. Our method uses reinforcement learning to manipulate the latentnoise of a task-agnostic diffusion policy prior, steering it toward task-specific behaviors for fast, robust control with high motion fidelity. Incontrast to methods that rely solely on offline training, NaP interactswith the environment during training to correct motions and optimizetask rewards, improving success rates and enabling adaptation to chal-lenging scenarios. By directly predicting task-optimized diffusion noise,NaP eliminates iterative guidance during denoising and enables efficientinference. Experiments show that NaP attains higher success rates andfaster inference while preserving natural motion across diverse tasks.

Recent Vision-Language Models (VLMs) have demonstrated significant progress in video understanding. However, we notice that VLMs perform coarse perception and reasoning when answering finegrained motion questions. We investigate the reasons for this phenomenon and discover that motions in video change rapidly, so fixed frame rate sampling easily misses inter-frame motion details. Moreover, when answering motion-centric questions, VLMs tend to introduce motion-irrelevant contextual information and lack a structured and precise representation of motions, making it difficult to distinguish what motion the subject performs and when. To address these problems, we propose MotionChain, a training-free method that decouples motion question answering into a sequence of ⟨time range, subject, motion, environment⟩ tuples. Concretely, MotionChain leverages optical flow variations to calibrate when motions occur and extracts motion-relevant key frames as supplementary input, effectively alleviating the problem of missing critical motion information. For fine-grained motion question answering (QA), each tuple element in MotionChain can be converted into a visually grounded claim, providing the model with visual evidence. Experiments on motion-centric video QA benchmarks demonstrate the superiority of our MotionChain.

Multi-exposure fusion (MEF) expands the luminance rangebeyond what a single exposure can capture. Combining images takenat different exposure levels requires handling geometric differences whilenaturally merging their complementary brightness information. It of-ten demands generative completion where details are missing. Diffusion-based generative methods address these challenges, however, they arecomputationally expensive and struggle to preserve fine structures in sat-urated regions. We propose LIIFusion, a coarse-to-fine framework thatbalances fusion quality and efficiency in generative MEF. The coarsestage performs low resolution generative fusion, enhanced by an adaptiveexposure correction that recovers structure lost in saturated over-exposedareas. The fine stage adapts a local implicit image function into a multi-exposure fusion function: conditioned on the HR OE/UE sources and thecoarse output, it queries arbitrary target coordinates and fuses sourceevidence regard- less of the HR input resolution. LIIFusion achieves upto 3.5× speed-up over existing generative methods while maintainingor improving structural fidelity and perceptual quality. We believe thisframework provides an effective pathway toward making generative MEFmore practical in real-world applications.

Multimodal large language models (MLLMs) have rapidlyadvanced video understanding, achieving strong zero-shot and few-shotrecognition across standard benchmarks. Yet their ability to deny anaction by recognizing when an activity is not happening despite strongcontextual cues remains largely unexplored. We introduce UCF101-AD, a large-scale benchmark consisting of paired Action-Presence andAction-Denial clips, designed to evaluate this capacity for denial. Eachnegative video in UCF101-AD preserves the same contextual and mo-tion cues (persons, objects, locations) as its positive counterpart, butthe defining action itself is explicitly absent. Evaluating 20 state-of-the-art MLLMs reveals a consistent failure: models that exceed 85% accu-racy on the positive action classes collapse below 50% on its action-denial counterpart, indicating a strong inclination to affirm plausibleactions rather than verify that they truly occur. This exposes a criti-cal blind spot in modern video understanding: the inability to reasoncausally about whether a motion actually happens. To probe this is-sue, we explore a causal graph formulation, CausalAct, which expressesscene structure through natural-language prompts linking context, in-teraction, and motion. Incorporating such causal cues substantially re-duces false positives, demonstrating that denial is a learnable reason-ing skill. UCF101-AD provides a new lens for diagnosing and improv-ing causal reasoning in multimodal models. Dataset and relevant code:https://github.com/raiyaan-abdullah/Learn-to-Deny.

Latent diffusion models have emerged as a dominant frame-work for high-fidelity image and video synthesis, operating in compactlatent spaces with variational autoencoders (VAEs) to enhance compu-tational efficiency without compromising visual quality. However, con-ventional VAEs are suboptimal for video data as they employ fixed com-pression ratios that cannot adapt to the varying complexity of spatio-temporal content. We present KATok (Keep-or-Drop? Adaptive Tok-enizer for Compact Video Representation), a transformer-based VAEthat incorporates an adaptive token selector which is jointly learned withlatent tokens. By evaluating each token’s content-richness as keep-or-dropprobability, the token selector effectively discards uninformative tokens,naturally allowing data-dependent compression. Applying adaptive tok-enization to diffusion models may cause spatial misalignment, as tokendropping can disturb the original spatio-temporal structure. To allevi-ate this issue, we propose two position-prediction strategies: cascadedand joint generation, to ensure spatial consistency. We empirically showthat our model achieves strong reconstruction and generation quality at astate-of-the-art compression ratio. Further analysis on video data revealsthat this improvement is primarily achieved by reducing spatio-temporalredundancy and removing uninformative tokens, as supported by bothquantitative and qualitative results.

Current diffusion models face a fundamental tension betweenthe computational efficiency of latent-space generation and the fine-detailfidelity of pixel-space generation. Latent Diffusion Models (LDMs) relyon VAE compression, which can discard high-frequency visual cues andlimit pixel-level quality, while pixel-space diffusion often suffers from diffi-cult optimization when global structure and high-frequency details mustbe learned jointly. In this paper, we propose the Hybrid Latent-PixelDiffusion Transformer (Hi-DiT), a unified Diffusion Transformer thatbridges these two regimes within a single architecture. Motivated by thetemporal heterogeneity of denoising—early timesteps primarily establishcoarse global structure, whereas late timesteps increasingly emphasizehigh-frequency detail—we introduce a dual-stream design with temporalspecialization. Hi-DiT performs semantic planning in a compact latentpathway at early stages, and activates a pixel pathway at later stagesto synthesize high-frequency details using time-consistent noisy pixelembeddings. A Time-Gated Injection mechanism schedules the partic-ipation of pixel tokens only in the low-noise regime, and a lightweightsub-pixel prediction head enables efficient dense detail generation. Ex-tensive experiments demonstrate that Hi-DiT achieves state-of-the-artperformance, obtaining an FID of 1.06 on ImageNet 256 × 256, 1.26 onImageNet 512 × 512, and 5.15 on MS-COCO text-to-image generation.Code is available at: https://github.com/HiDream-ai/Hi-DiT.

Natural-language queries about 3D environments become actionable when responses are verifiable and metric. Verifiability requires explicit grounding to the referred 3D region, while metric answers report physical measurements in real-world units (e.g., size, thickness, clearance, and distance). Existing 3D large multimodal models (LMMs) approaches remain limited: conversational systems typically respond without explicit 3D grounding, while 3D grounding models are not designed for interactive, metric-aware dialogue. In this paper, we present Ground3DLMM, a unified model that takes a point cloud and an optional RGB image as input and supports 3D spatial conversation with (i) pointgrounded responses and (ii) metric numeric outputs at both object and part granularity, including multi-object queries. To evaluate this intersection of grounding and measurement, we define the 3D Grounded Measurement task, which requires predicting the referred 3D region and the corresponding metric quantities in real-world units. We introduce a large-scale dataset built on ScanNet and ScanNet++ datasets with dense object and part annotations and roughly 2.5M question-answer pairs spanning eight tasks, along with a manually verified test set. Extensive experiments on multiple datasets and tasks show that our proposed Ground3D-LMM model provides a strong baseline for grounded, metric-aware 3D conversational understanding. Our dataset and model are publicly available.

Existing instruction-based video editing datasets commonlyfocus on single-task appearance editing, failing to meet the complex cre-ative demands of real-world scenarios. To bridge this gap, we presentGoku, a large-scale dataset featuring 2 million high-quality, instruction-aligned video editing pairs, which is the first to extend task boundariesfrom basic appearance editing to multi-task and structural manipulations(e.g., precise control of subject movement). To tackle the data synthesischallenges inherent in these complex tasks, we design an efficient datasynthesis pipeline that decomposes complex edits into controllable sub-problems and introduce a progressive filtering system for data reliabil-ity throughout the whole process. Furthermore, we explore the optimalnetwork structures on Goku, and propose Goku-Edit. To deeply com-prehend complex editing instructions, Goku-Edit leverages an MLLMas its text encoder and adopts a decoupled dual-branch design: a dedi-cated mask branch handles structural control, freeing the main branchfor appearance rendering. A comprehensive video editing benchmark,Goku-Bench, is also proposed with 1,000 human-verified test cases and7 novel editing-specific metrics. Evaluated on Goku-Bench, Goku-Editobtains up to +8% improvement on other open-source models in termsof instruction following.

Recent progress in text-to-video generation has achieved remarkable realism, yet fine-grained control over camera motion and orientation remains elusive, especially with extreme trajectories (e.g., a 180° turnaround, or looking directly up or down). Existing approaches typically encode camera trajectories using relative or ambiguous representations, limiting precise geometric control and offering limited support for large rotations. We introduce GimbalDiffusion, a framework that enables camera control grounded in physical-world coordinates, using gravity as a global reference. Instead of describing motion relative to previous frames, our method defines camera trajectories in an absolute coordinate system, allowing accurate, interpretable control over camera parameters. Using panoramic 360◦ videos for training, we cover the full sphere of possible viewpoints, including combinations of extreme pitch and roll that are out-of-distribution of conventional video data. To improve camera control, we introduce null-pitch conditioning, a strategy that prevents the model from overriding camera specifications in the presence of conflicting prompt content (e.g., generating grass while the camera points toward the sky). Finally, we propose new benchmarks to evaluate gravity-aware camera-controlled video generation, assessing models’ ability to generate extreme camera angles and quantify their input prompt entanglement.

Synthetic data for autonomous driving is surging, poweredby diffusion models that promise scalable scene generation. Yet key ob-stacles remain, since multi-view and temporal consistency often requiresbackbone fine-tuning or added layers, which can erode pretrained knowl-edge and weaken text alignment. Models also stay close to the train-ing distribution, struggling under adverse weather and unseen configu-rations, while fidelity favors frequent over rare classes. We address thesegaps with FrozenDrive, a controllable generative framework that pre-serves pretrained diffusion knowledge while achieving strong consistency.FrozenDrive conditions on driving-stack signals and text prompts, andexpands the context of frozen self-attention across views and frames topromote cross-view alignment and temporal coherence in one pass, with-out trainable spatio-temporal modules in the diffusion backbone. Anobject-focused constraint further improves fidelity for rare categories.Without weather- or scene-specific fine-tuning, FrozenDrive synthesizesglobally coherent multi-view driving scenes from text and surpasses priorbaselines under adverse and rare conditions. On nuScenes, FrozenDrive-augmented data improves AD model performance, especially at night andin rain, demonstrating strong robustness with scenario-targeted data.

Optical flow methods typically rely on task-specific inductivebiases, such as correlation volumes, feature warping, and iterative refine-ment, among others, to reach high accuracy. While effective, such biasesconstrain the model to predefined heuristics, which can limit its expres-sivity and lead to more complex pipelines and additional computationalcost. We present FreeFlow, a hierarchical transformer built without anyflow-specific components, using instead a single feed-forward encoder–decoder. FreeFlow combines three attention variants: window attentionfor local processing, shifted-window attention for cross-window informa-tion exchange, and a global attention operating at a reduced resolu-tion. The resulting architecture scales naturally with model capacity,enabling a consistent accuracy gain from small to large variants. Despitethe absence of standard inductive biases, FreeFlow achieves state-of-the-art results on major benchmarks, including Sintel (0.68/1.48 EPE onClean/Final), KITTI-2015 (3.23 Fl-all), and Spring (3.192 1px), whileremaining memory efficient at 1080p inference.

Cross-Domain Few-Shot Object Detection (CDFSOD) aimsto transfer knowledge from data-rich upstream generic domains to down-stream expert domains using scarce training data, where the significantdomain gap and data scarcity make it an unsolved challenge. To ad-dress this problem, we revisit a natural yet underexplored approach inCDFSOD: data augmentation, by directly synthesizing data through dif-fusion models to supplement limited training samples. However, due tolarge domain gaps, we find that current diffusion methods cannot pro-duce good results, leading to performance even lower than using theoriginal images. To address these limitations, we divide the domain gapsinto visual gaps and semantic gaps for separate analysis. For the visualgap, we find that the diffusion model cannot distinguish noise from use-ful information on expert domains, which can be mitigated by addingweakened noise. For the semantic gap, we find that the background se-mantics shows much smaller gaps between domains than foreground se-mantics, and we can bridge this gap by background inpainting. Basedon the above analysis, we propose a method (Selective Inpainting withTailored Noise, SITN) to dynamically take different strategies for down-stream data synthesis based on their different gaps from the generaldomain, including a Generation Module for adding tailored noise and aSelection Module to dynamically select the inpainting regions. Extensiveexperiments on 6 datasets of CDFSOD and 4 datasets of cross-domainfew-shot segmentation (CDFSS) validate that we can synthesize helpfuldata, achieving new state-of-the-art performance. Our codes is availableat https://github.com/zzzzj311-droid/Free-Lunch-SITN.

005 Multimodal large language models (MLLMs) have recently 005006 shown promise for industrial anomaly understanding, where models are 006007 expected not only to determine whether an object is defective, but also to 007008 localize visual evidence and explain defect-related decisions. However, ex- 008009 isting MLLMs often rely on language priors or shortcut reasoning when 009010 facing fine-grained industrial defects, resulting in inconsistent outputs 010011 such as correct answers with invalid locations, hallucinated defect boxes 011012 on normal samples, or rationales that contradict the final prediction. To 012013 address these issues, we propose EGVLR, an evidence-grounded vision– 013014 language reinforcement framework for anomaly reasoning. EGVLR fol- 014015 lows a unified Evidence-Driven Diagnostic Protocol (EDDP), where each 015016 response explicitly contains visual evidence, diagnostic logic, spatial loca- 016017 tion, and the final answer. The framework consists of four stages: Progres- 017018 sive Visual-Evidential Fine-Tuning (PVE-FT), Knowledge-Grounded In- 018019 struction Tuning (KG-IT), Geometry-Semantic Decoupled Preference 019020 Optimization (GS-DPO), and Box-Guided Segmentation Rendering (BGSR). 020021 PVE-FT first aligns the model with localized visual evidence through 021022 synthetic anomalies, 3 × 3 grid grounding, local verification decoys, and 022023 null-hypothesis calibration. KG-IT then internalizes industrial-domain 023024 QA behavior while preserving the same output schema. GS-DPO fur- 024025 ther optimizes answer correctness, geometric localization, BGE-based 025026 rationale semantics, and answer–location–rationale consistency. Finally, 026027 BGSR converts predicted bounding boxes into dense masks using an off- 027028 the-shelf segmentation backend. Experiments on MMAD demonstrate 028029 that EGVLR improves evidence-grounded industrial anomaly reasoning, 029030 especially on defect localization, reasoning-dependent QA, and spatial 030031 false-positive control. 031

The O(N 2 ) complexity of attention over N tokens remainsa computational bottleneck in transformer models. Vector-Quantized(VQ) attention reduces this to O(M N ) by representing keys with Mcodewords, but applies uniform codebook capacity regardless of whereattention mass concentrates: high-attention regions of key space maybe coarsely approximated while low-attention regions waste representa-tional capacity. We propose Adaptive Vector-Quantized (AVQ) Atten-tion, which adaptively allocates codebook capacity based on attentionimportance. Starting from a small set of codewords, our method identifiesthe most important codes during the forward pass and refines them withpre-learned child codewords, achieving fine-grained quantization whereit matters most while maintaining coarse quantization elsewhere. We de-velop an implementation using custom Triton kernels that enables the fulladaptive refinement process, including importance scoring, child code-word insertion, and parent contribution replacement, to be carried outwithin the tiled computation paradigm of Flash Attention with minimaloverhead. Our approach maintains O(M N ) complexity while achievingimproved accuracy-efficiency trade-offs compared to fixed-codebook VQ-attention.

Adversarial Robustness Distillation (ARD) is the most viableway for transferring defensive capabilities from over-parameterized teach-ers to lightweight student models. However, existing ARD techniques relyon the assumption that the teacher’s decision boundary is a perfect geo-metric oracle. In this paper, we identify a critical gap from the stabilityperspective that such exact matching forces blind inheritance of teacher’ssensitivity noise, punishing the student’s stability even if they are morerobust. To address this challenge, we propose the AEGIS (AdversarialError Gating and Instability Suppression) framework that replaces sim-ple logit matching with gated supervision. It leverages a pair of logit andsensitivity gates to rectify potential errors/instability on the categoricaland gradient-level, respectively. Extensive evaluations on CIFAR-10/100,and Tiny-ImageNet show that AEGIS significantly outperforms state-of-the-art methods, achieving a 3.12% gain under the strong AutoAttackand preserving nearly 50% accuracy under large perturbation intensities(ϵ = 16/255) where traditional baselines collapse. Code is available athttps://github.com/JingqiJi03/AEGIS.

3D shape assembly requires predicting 3D pose by estimatingrotation and translation separately to align objects or fractures. For 3Dpose prediction, prior works utilize the entangled features that capture3D pose information well, but this mixed representation impedes gen-eralization. To address this problem, we propose the SOT encoder thatdisentangles 3D pose into an SO(3)-equivariant (rotation) feature and aT(3)-equivariant (translation) feature. The SOT encoder consists of twobranches: 1) a rotation branch that suppresses translation by project-ing features onto the translation null space, and 2) a translation branchthat suppresses rotation by translation loss and directly subtracts theSO(3)-equivariant representation from its SE(3)-equivariant counter-part. Across a variety of assembly models and diverse datasets, the SOTencoder consistently demonstrates improved generalization performancein 3D shape assembly. Furthermore, in-depth analyses indicate that thedisentangled features show improved equivariant behavior with respectto the target factor while exhibiting invariant behavior with respect tothe other. These results indicate that explicitly encouraging factor dis-entanglement is a straightforward and effective approach for 3D shapeassembly. The code is available at https://github.com/maroo-sky/SOT.

The rapid scaling of large language models (LLMs) and vision-language models (VLMs) has sharply increased compute and storagecosts, motivating extensive research on model pruning. Yet most prun-ing pipelines underuse data scaling during pruning, i.e., increasing thepruning-phase token budget. We revisit L0-regularized structured chan-nel pruning, valued for differentiability, incremental scheduling, and globaldecision-making. We uncover a long-horizon instability: the Lagrangemultiplier grows during training, amplifies pruning gradients, and causeslate-stage oscillations in both gradients and loss. We derive sufficient L2-norm bounds on pruning-specific gradients and reformulate the loss toenforce these bounds per update, replacing dual ascent with a coefficientdetermined by the gradient constraint. Building on this foundation, weintroduce DIGS (Differentiable, Incremental, Global, Scalable pruning),which enables stable and efficient structured channel pruning withoutauxiliary stabilizers while supporting much larger pruning token bud-gets. Applied to the Qwen2.5VL LLM backbone, DIGS scales pruningbeyond 100B tokens and achieves competitive results on public multi-modal benchmarks. Within this setting, larger pruning token budgetsyield monotonic performance gains at fixed sparsity.