ECCV 2026 论文列表

Video-to-video diffusion models achieve impressive singleturn editing performance, but practical editing workflows are inherently iterative. When edits are applied sequentially, existing models treat each turn independently, often causing previously generated regions to drift or be overwritten. We identify this failure mode as the problem of cross-turn consistency in multi-turn video editing. We introduce Memory-V2V, a memory-augmented framework that treats prior edits as structured constraints for subsequent generations. Memory-V2V maintains an external memory of previous outputs, retrieves task-relevant edits, and integrates them through relevance-aware tokenization and adaptive compression. These technical ingredients enable scalable conditioning without linear growth in computation. We demonstrate Memory-V2V on iterative video novel view synthesis and text-guided long video editing. Memory-V2V substantially enhances cross-turn consistency while maintaining visual quality, outperforming strong baselines with modest overhead.

Audio-driven avatar interaction demands real-time, stream-ing, and in_x001C_nite-length generation_x0016_capabilities fundamentally at oddswith the sequential denoising and long-horizon drift of current di_x001B_usionmodels. We present Live Avatar, an algorithm-system co-designed frame-work that addresses both challenges for a 14-billion-parameter di_x001B_usionmodel. On the algorithm side, a two-stage pipeline distills a pre-trained bidirectional model into a causal, few-step streaming one whoseKV cache stores noisy rather than clean history. This lossy condition-ing, counter-intuitively, suppresses error accumulation and enables mod-els trained on short clips to generalize stably beyond 10,000 seconds.On the system side, this noisy conditioning also breaks the sequen-tial sampling bottleneck, allowing Timestep-forcing Pipeline Parallelism(TPP) to assign each GPU a _x001C_xed denoising timestep, converting the se-quential di_x001B_usion chain into an asynchronous spatial pipeline that simul-taneously boosts throughput and improves temporal consistency. LiveAvatar achieves 45 FPS with a TTFF of 1.21 s on 5 H100 GPUs, andto our knowledge is the _x001C_rst to enable practical real-time streaming ofa 14B di_x001B_usion model for in_x001C_nite-length avatar generation. Our projectpage is at https://liveavatar.github.io/.

Large foundation models trained on extensive datasets demon-strate strong zero-shot capabilities in various domains. Knowledge dis-tillation has become an established tool for transferring knowledge fromfoundation models to small student networks when data and model sizeare constrained. However, the efficacy of distillation is often hampered bylimited training data coverage. This can result in a covariate shift betweentraining and test data, which in turn can lead the student to exploitspurious features or even shortcut learning. We address this problemby introducing a novel diffusion-based data augmentation strategy thatgenerates images by maximizing the disagreement between the teacherand the student, effectively creating challenging samples that the studentstruggles with, thus mitigating the problem of covariate shift. Experi-ments demonstrate that, compared to state-of-the-art diffusion-based dataaugmentation baselines, our approach is best or second-best in samplemean accuracy and improves the worst group and mean group accuracyon CelebA-HQ, SpuCo Birds and BAR as well as the spurious score onSpurious ImageNet under covariate shift.

The quadratic time and memory complexity of the attention mechanism in modern Transformer-based video generators makes endto-end training for ultra-high-resolution videos prohibitively expensive. Motivated by this limitation, we introduce a training-free approach that leverages video Diffusion Transformers pre-trained at their native scale to synthesize higher-resolution videos without any additional training or adaptation. At the core of our method lies an inward sliding-window attention mechanism, which originates from a key observation: maintaining each query token’s training-scale receptive field is crucial for preserving visual fidelity and detail. However, naive local window attention, unfortunately, often leads to repetitive content and exhibits a lack of global coherence in the generated results. To overcome this challenge, we devise a dual-path pipeline that backs up window attention with a novel cross-attention override strategy, enabling the semantic content produced by local attention to be guided by another branch with a full receptive field and, therefore, ensuring holistic consistency. Furthermore, to improve efficiency, we incorporate a cross-attention caching strategy for this branch to avoid the frequent computation of full 3D attention. Extensive experiments demonstrate that our method delivers ultra-highresolution videos with fine-grained visual details and high efficiency in a training-free paradigm. Meanwhile, it achieves superior performance on VBench, even compared to training-based alternatives, with competitive or improved efficiency.

State Space Models (SSMs) have surfaced as a promisingarchitecture in Video Frame Interpolation (VFI), as they can capturelong-range dependencies with linear computational complexity. However,their predefined scanning order limits their effectiveness in modelingthe dynamic motion trajectories inherent in VFI problems. To tacklethis challenge, we propose Motion-Guided Mamba for Video FrameInterpolation (MGMVFI), an adaptation of the selective state spacemodel tailored explicitly for VFI. MGMVFI introduces Motion-GuidedSerialization (MGS), which leverages optical flow to define a motion-adaptive 1D input order for the SSM. This aligns the causal state up-dates with semantically related tokens, enabling motion-consistent fea-ture propagation, particularly for large and dynamic motions. Addition-ally, to mitigate the unreliable feature representations caused by inaccu-rate optical flow estimates, we introduce contextual synthesis that uti-lizes the surrounding spatial context for robust inter-frame feature syn-thesis. These components are seamlessly integrated within our tailoredMamba architecture, which also employs a lightweight refinement blockto enhance local detail reconstruction at a reduced computational cost.Extensive experiments on standard VFI benchmarks demonstrate thatMGMVFI achieves state-of-the-art performance, particularly on complexand dynamic motions, thereby establishing a new direction for sequencemodeling in video interpolation.

3D dog pose estimation is hindered by two limitations in ex-isting benchmarks: distributional bias toward canonical viewpoints, sta-ble poses, and narrow morphologies, and alignment degradation betweenimages and 3D annotations. We instead synthesize a large-scale bal-anced benchmark (4.2M images with exact 3D ground truth) by decou-pling pose, shape, and texture into independent libraries (70K retargetedposes, 500 breed shapes, 1,000 appearance maps) and rendering directlyto guarantee pixel-perfect alignment and broad view coverage. However,this balanced distribution exposes a challenge previously masked by bi-ased data: non-canonical combinations of pose, shape, and viewpointcreate multimodal 3D ambiguities that deterministic regression cannotresolve. We therefore propose a flow-based probabilistic framework thatmodels the conditional distribution of 3D poses given an image, pro-ducing multiple plausible hypotheses under ambiguity while convergingto precise estimates for clear views. Experiments on StanfordExtra, An-imal3D, and our synthetic benchmark demonstrate that our method,trained solely on synthetic data, outperforms state-of-the-art approaches(79.9 vs. 33.2 PCK on the synthetic benchmark). Dataset available here.

Decomposing outgoing surface radiance into material and illumination during inverse rendering is essential for applications such as relighting and augmented reality, yet it is severely ill-posed since multiple combinations can result in the same observed colour. Capturing an object under multiple lighting conditions usually helps resolve this ambiguity as it constrains the optimization towards correct solutions. In this work, we explore the potential of reconstructing rigidly moving objects—which provides observations of diverse light-surface interactions—to resolve the material-lighting ambiguity in inverse rendering. For this purpose, we introduce a relightable approach that marries object tracking and reconstruction with inverse rendering for general rigidly moving objects. Our experimental analysis on synthetic data demonstrates that motion can be an advantage for disentangling material and lighting: the reconstructed material is significantly more accurate when the object is observed under rigid motion than when it is static. Moreover, results on RGB videos of real hand-held objects show that our pipeline preserves this advantage even under noisy real-world conditions.

In industrial environments, new product categories arrive se-quentially, requiring continual anomaly detection without access to pastdata. Normalizing Flows (NFs) provide exact density estimation but suf-fer from catastrophic forgetting as parameter updates across tasks dis-tort the density manifold. While parameter isolation can prevent inter-ference, it must preserve the strict invertibility and Jacobian validityof NFs. To satisfy these requirements, we exploit the inherent propertythat affine coupling layers maintain transformation validity regardless ofsubnet parameterization. Based on this, we propose DeCoFlow, which de-composes subnets into a frozen universal base and task-specific low-rankadapters to isolate updates. We further introduce Task-Specific Align-ment, Auxiliary Coupling Layers, and Tail-Aware Loss to compensatefor frozen-base rigidity. DeCoFlow achieves state-of-the-art image-levelAUROCs of 98.40% on MVTec-AD and 93.00% on VisA, while maintain-ing parameter-level zero forgetting (0.00% FM under correct routing)with only 2.27M parameters per task.

Generalized Category Discovery (GCD) is an intriguing open-world problem that has garnered increasing attention: given partially la-belled data, the goal is to correctly recognize known classes while discov-ering coherent novel categories from unlabelled samples. Recent GCDmethods typically adapt foundation models by jointly optimizing su-pervised classification and unsupervised discovery objectives on mixedlabelled–unlabelled data. While effective, this coupled training can en-tangle closed-set recognition and open-set discovery, leading to objectiveconflict and biased predictions, and may disturb the semantic geome-try of pretrained representations under limited labels and noisy pseudo-labels. We propose CloSeR, a simple plug-and-play framework that in-jects Closed-Set Relational knowledge into GCD training. CloSeR firstbuilds a domain-adapted closed-set teacher by tuning lightweight block-wise adapters on labelled known-class data while keeping the founda-tion model backbone frozen, thereby preserving pretrained priors at lowtraining cost. It then transfers the teacher’s knowledge to downstreamGCD via Unified Relational Distillation (URD), which distills comple-mentary global sample-to-prototype relations to anchor known-class se-mantics and local sample-to-sample relations to preserve neighborhoodstructure, using separate feature pathways to reduce optimization in-terference. CloSeR is head-agnostic and readily integrates with bothparametric and non-parametric GCD methods. Extensive experimentswith DINO and DINOv2 backbones on six benchmarks (CIFAR-10/100,ImageNet-100, CUB, Stanford-Cars, and FGVC-Aircraft) show consis-tent gains over GCD baselines, achieving state-of-the-art performance.Project page: https://visual-ai.github.io/closer/

Category-level 6D pose estimation of articulated objects froma single-view input is a fundamental yet challenging task. Its core dif-ficulty lies in occlusion-induced pose and shape ambiguities, which arefurther exacerbated by highly entangled pose-shape representations andreal-world depth noise. Existing generative methods often neglect theinherent correlations between pose and shape, resulting in geometricallyinconsistent or physically implausible predictions. To address this limi-tation, we propose a novel generation-and-verification framework for thistask: First, a Pose-Shape Hypothesis Generator jointly models the dis-tributions of pose and shape to generate intrinsically coupled hypothesispairs, providing explicitly paired candidates for subsequent verification.Second, a Part-Level Point Cloud Denoising module recovers clean pointclouds from noisy and partial depth observations, ensuring robust inputsfor the verification stage. Finally, a Cross-Space Consistency Verifica-tion module explicitly verifies the geometric consistency between the ob-served point clouds and canonical shapes using a continuous SDF-basedscoring mechanism, which effectively prunes geometrically implausiblehypotheses. Extensive experiments on challenging benchmarks demon-strate that our method significantly improves the robustness and accu-racy of pose estimation in complex, heavily occluded scenarios. Projectpage: https://github.com/KaifengT/ArtPS.

Deep generative models, while revolutionizing image gener-ation, largely operate as opaque “black boxes”, hindering human un-derstanding, control, and alignment. While methods like sparse autoen-coders (SAEs) show remarkable empirical success, they often lack theo-retical guarantees, risking subjective insights. Our primary objective is toestablish a principled foundation for interpretable generative models. Wedemonstrate that the principle of causal minimality – favoring the sim-plest causal explanation – can endow the latent representations of gener-ative models with clear causal interpretation and robust, component-wiseidentifiable control. We introduce a novel theoretical framework for hi-erarchical selection models, where higher-level concepts emerge from theconstrained composition of lower-level variables, better capturing thecomplex dependencies in data generation. Under theoretically derivedminimality conditions that manifest as sparsity constraints, we show thatlearned representations can be equivalent to the true latent variables ofthe data-generating process. Empirically, applying these constraints totext-to-image diffusion models allows us to extract their innate hierarchi-cal concept graphs, offering fresh insights into their internal knowledgeorganization. Furthermore, these causally grounded concepts serve aslevers for fine-grained model steering, paving the way for transparent,reliable systems.

The reliance on extensively annotated 3D data confines cur-rent autonomous driving systems to a closed-set detection paradigm, lim-iting their robustness in real-world open-world scenarios where novel ob-jects frequently emerge. To bridge this gap, we propose BEVOpen3D,a novel distillation framework that facilitates open-world 3D object de-tection by effectively transferring knowledge from 2D vision-languagemodels into the 3D domain via the Bird’s-Eye-View (BEV) space undera partial-label setting. Our approach first generates initial 3D proposalsfor both seen and unseen categories using an open-vocabulary 2D detec-tor combined with a greedy spatial search strategy. We then introducea Triple-Source Label Refinement mechanism, which fuses original, lo-cally rectified (via a Seen-Guided Local Query), and globally re-examinedproposals to produce high-quality pseudo labels for unseen-category. Fi-nally, we propose a Heatmap Proposal Distillation strategy, where animage-based BEV teacher transfers its vision-grounded semantic priorsto a LiDAR-based student through aligned BEV heatmaps, enabling thestudent to detect both seen and unseen objects without requiring addi-tional 3D annotations. Extensive experiments validate the effectivenessof BEVOpen3D in open-world 3D detection.

Diffusion models have significantly advanced video super-resolution (VSR) but remain largely constrained to fixed upsamplingscales. Conversely, while coordinate-based arbitrary-scale VSR methodsoffer scale flexibility, they inherently suffer from severe over-smoothingat large scaling factors. Integrating generative priors with continuous de-coding is promising but currently hindered by severe temporal flickeringcaused by the stochasticity of diffusion sampling. To address this, wepropose AVSR-Diff (Arbitrary-scale Video Super-Resolution with Dif-fusion), a novel decoupled framework that separates scale-agnostic la-tent denoising from continuous coordinate rendering, effectively avoid-ing computationally heavy resolution-specific sampling. Our approachintroduces a Temporally-Gated Feature Recurrence (TGFR) module toextract strictly aligned, temporally consistent latent priors. Furthermore,we design a continuous video VAE decoder incorporating a Scale-AwareFourier Refinement (SAFR) module to dynamically adapt frequencycomponents to any target scale. Extensive experiments demonstrate thatAVSR-Diff consistently preserves high-frequency details and strong tem-poral stability across various scales, surpassing state-of-the-art arbitrary-scale baselines. Remarkably, our framework outperforms recent fixed-scale generative models even on their native resolution.

WordArt (artistic text) features highly customized fonts,textures, and layouts, making WordArt-oriented scene TExt Recogni-tion (WATER) substantially more challenging than general Scene TextRecognition (STR). Existing STR datasets and methods, typically builtaround regular scene text and fixed-template inputs, struggle to scale toWATER. Thus, we aim to advance this task from both data and modelperspectives. On the data side, we construct a 2M synthetic dataset,WATER-S, with the scale improved by hundreds of times comparedto existing artistic text data. WATER-S consists of two complementarysubsets. One rendered by an upgraded rendering pipeline (SynthWor-dArt), which provides highly accurate and controllable synthetic Wor-dArt data. The other is generated by combining Qwen3-VL for promptmining and Z-Image for image synthesis, which improves the coverage ofrealistic and diverse data. On the model side, we propose WATERec.It adopts a visual encoder supporting arbitrary-shaped inputs and anautoregressive decoder to model complex layouts, structurally breakingthe bottleneck of fixed-template STR on WordArt. Experiments showthat this architecture outperforms prior STR methods, achieving state-of-the-art performance on irregular texts such as WordArt. Togetherwith WATER-R, carefully reorganized from existing real STR data, ourstrong baseline with the new synthetic data and model design reaches90.40% accuracy on WordArt-Bench, surpassing both general-purposeand OCR-specialized vision-language models by a large margin. Codeand data are available at https://github.com/YesianRohn/WATER.

Out-of-Distribution (OOD) detection is a critical safety requirement for deploying deep neural networks in open-world environments. While recent advances increasingly rely on more computationally intensive training methods involving synthetic outliers, contrastive objectives, or specialized loss functions, their gains often come with substantial computational overhead and implementation complexity. In this work, we revisit the fundamentals of OOD detection and uncover a key flaw in common distance-based detectors: sensitivity to feature magnitude. We show that low-norm OOD samples can appear closer to in-distribution (ID) class centroids than actual ID samples, evading detection. To this end, we introduce Angular Separation Learning (ASL), a simple and highly effective strategy that applies ℓ -normalization to features before 2 the final classification layer. This modification compels the network to optimize for angular separation, achieving robust feature learning without additional regularization mechanisms, synthetic samples, or costly negative mining. Through extensive experiments on diverse benchmarks, we demonstrate that ASL not only matches but often surpasses state-of-theart methods, especially in challenging near-OOD scenarios, while maintaining training efficiency. Our results indicate that a minimalist rethink of standard training can achieve superior OOD performance, prompting a re-evaluation of the complexity-to-performance trade-off in OOD detection. Codes are available at https://github.com/HAIV-Lab/ASL.

While Multimodal Large Language Models (MLLMs) have advanced video understanding, achieving precise temporal and crossmodal alignment in audiovisual video captioning remains a formidable challenge. Most existing approaches suffer from modality detachment and temporal incoherence, failing to accurately bind auditory events to visual entities or capture complex causal dynamics. To address these deficiencies, we propose TCA-Captioner, a framework specifically engineered to enhance Temporal and Cross-Modal Alignment for audiovisual video captioning. We first introduce the Observer-Checker-Corrector (OCC) framework, an iterative refinement strategy that generates high-fidelity, meticulously grounded training data. Leveraging a curated high-density human interaction dataset, TCA-Captioner is optimized to model sophisticated audiovisual interactions. Furthermore, we present TCA-Bench, a diagnostic benchmark utilizing a Decoupled Evaluation Protocol to isolate and quantify model proficiency in audiovisual binding and temporal relational reasoning. Extensive experiments demonstrate that TCACaptioner sets a new standard for temporally-coherent and synchronized audiovisual narratives.

Recent Gaussian Splatting SLAM (GS-SLAM) methods alignrendered and observed images through photometric and geometric resid-uals to achieve high-quality reconstruction and rendering. However, thesesignals weaken in low-texture and repetitive environments, which leadsto floaters, inconsistent geometry, and drift. Existing systems mitigatedrift through loop closure or global layout assumptions, but these mech-anisms are opportunistic or lack persistent structural entities that pro-vide reliable constraints throughout mapping. We introduce a structure-aware GS-SLAM framework that models planar structures as persistentPlanar Gaussian Instances (PGIs) within a 3D Gaussian map. EachPGI represents a planar surface associated with Gaussian primitives andmaintains a consistent identity across frames. This representation pro-motes planes from transient observations to map entities and convertsplanar structures into stable geometric constraints. Building on PGIs,we propose structure closure, which estimates relative camera pose frommultiple non-parallel planes without requiring trajectory revisits. Wefurther integrate PGIs into a joint optimization that combines photo-metric–depth supervision with plane-instance consistency to improvemulti-view geometric coherence and suppress floater growth. Experi-ments on public benchmarks show that the proposed system achievesstate-of-the-art performance in camera tracking, dense reconstruction,and novel-view rendering, with strong robustness in scenes with largetextureless surfaces and repetitive layouts. Our code is publicly availableat https://github.com/yanyan-li/StructureGS-SLAM.

Deep neural networks are highly vulnerable to adversarialperturbations, raising serious safety concerns in the real-world systems.While prior work mainly explores single-task attacks or jointly degradingall tasks in multi-task models, practical scenarios often demand moreselective and stealthy attack strategies. To address this challenge, wepropose Stealthy Multi-Task Adversarial Attack (SMTA2 ), a novelframework that selectively degrades a targeted task while strictly pre-serving the performance of non-targeted tasks. We formulate this objec-tive as a constrained multi-objective optimization problem and designtask-aware adversarial perturbations that maximize degradation on thetargeted task without causing collateral damage on non-targeted tasks.To enhance practicality, we further introduce an automated loss-weighttuning strategy that dynamically balances attack and preservation objec-tives. Experiments on two multi-task benchmarks NYUv2 and Cityscapesdemonstrate that SMTA2 achieves strong attack performance on tar-geted tasks while maintaining non-targeted tasks intact on both unde-fended and adversarially trained models, establishing the first systematicframework for stealthy and selective multi-task attack framework.

Deep models are successfully applied to various fields, whiledemanding large amount of annotated data for high-performance recog-nition. To remedy the data-hunger issue, self-supervised learning, espe-cially masked autoencoder (MAE), is a promising approach to effectivelypre-train the deep models. The MAE leverages random masking to con-struct pretext tasks where masked image patches are reconstructed byusing unmasked (visible) ones. In vision domain, however, inherent imageproperties of high redundancy and correlation in local neighbors couldinterfere with the mask-based pretext tasks. In this study, we analyze twomain processes of masking and reconstruction in MAE through the lensof difficulty of the pretext task. The analysis inspires us to propose simpleyet effective approaches based on filtering to improve random maskingas well as raw-pixel reconstruction by properly controlling difficulty ofMAE task with a negligible extra computation cost. In the experimentson image classification, the proposed method renders favorable perfor-mance improvement to MAE using ViTs.