ECCV 2026 论文列表

Anomaly generation is often framed as few-shot fine-tuningwith anomalous samples, which contradicts the scarcity that motivatesgeneration and tends to overfit category priors. We tackle the settingwhere no real anomaly samples or training are available. We proposeDelta-Denoising (DeltaDeno), a training-free zero-shot anomaly gener-ation method that localizes and edits defects by contrasting two diffusionbranches driven by a minimal prompt pair under a shared schedule. Byaccumulating per-step denoising deltas into an image-specific localizationmap, we obtain a mask to guide the latent inpainting during later diffu-sion steps and preserve the surrounding context while generating realis-tic local defects. To improve stability and control, DeltaDeno performstoken-level prompt refinement that aligns shared content and strength-ens anomaly tokens, and applies a spatial attention bias restricted toanomaly tokens in the predicted region. Experiments on public datasetsshow that DeltaDeno achieves great generation, realism and consistentgains in downstream detection performance. Code will be made publiclyavailable at https://github.com/CROVO1026/DeltaDeno.

Arbitrary-scale image super-resolution (ASISR) aims to recon-struct high-resolution images from low-resolution inputs over a continuousrange of upscaling factors. While traditional pixel-regression approachesoften produce overly smooth results that lack realistic details, recent dif-fusion methods can produce sharper and more realistic textures. However,these diffusion techniques frequently introduce the risk of structural hal-lucinations. To address these issues, we propose Fidelity- and Perception-Aware Local Implicit Attention (FPLIA), a framework that effectivelyintegrates fidelity-oriented features into a diffusion pipeline to produce re-alistic and faithful reconstructions for ASISR. We introduce a Fidelity andPerception Attention Module (FPAM), which applies both self-attentionand cross-attention to fidelity-oriented and perceptual features to en-hance representational capacity. To further exploit their complements,we design a Fidelity and Perception Select Module (FPSM) that adap-tively selects the most representative features for RGB values prediction.We conduct extensive experiments to validate the effectiveness of thesecomponents. Both qualitative and quantitative results show that FPLIAdelivers superior perceptual realism while maintaining reconstructionaccuracy on standard ASISR benchmarks. The source code is accessibleat the following repository: https://github.com/XUSean0118/FPLIA.

Large Chunk Test-Time Training (LaCT) has shown strongperformance on long-context 3D reconstruction, but its fully plastic inference-time updates remain vulnerable to catastrophic forgetting and overfit-ting. As a result, LaCT is typically instantiated with a single large chunkspanning the full input sequence, falling short of the broader goal ofhandling arbitrarily long sequences in a single pass. We propose ElasticTest-Time Training inspired by elastic weight consolidation, that sta-bilizes LaCT fast-weight updates with a Fisher-weighted elastic prioraround a maintained anchor state. The anchor evolves as an exponentialmoving average of past fast weights to balance stability and plasticity.Based on this updated architecture, we introduce Fast Spatial Memory(FSM), an efficient and scalable model for 4D reconstruction that learnsspatiotemporal representations from long observation sequences and ren-ders novel view-time combinations. We pre-trained FSM on large-scalecurated 3D/4D data to capture the dynamics and semantics of complexspatial environments. Extensive experiments show that FSM supportsfast adaptation over long sequences and delivers high-quality 3D/4D re-construction with smaller chunks and mitigates the camera-interpolationshortcut. Overall, we hope to advance LaCT beyond the bounded single-chunk setting toward robust multi-chunk adaptation, a necessary stepfor generalization to genuinely longer sequences, while substantially al-leviating the activation-memory bottleneck.

Recovering a sharp image from a motion-blurred observation remains challenging due to the inherently ill-posed nature of blind deblurring, especially under real-world non-uniform blur caused by complex camera and scene motion. Event cameras, with their microsecond-level temporal resolution and inherent sensitivity to motion, provide rich motion cues that can serve as strong priors for resolving such ambiguities. However, existing event-guided deblurring methods either rely on simplified blur assumptions or treat event data merely as generic features, without fully exploiting their physical relationship with the image formation process. In this paper, we propose a complete event-driven nonuniform blind deblurring framework that explicitly models the physical formation of spatially varying blur kernels. Specifically, we first estimate dense pixel-wise motion trajectories from the event stream via a differentiable event-alignment objective. These trajectories are then used to construct per-pixel point spread functions (PSFs), which provide a physically grounded blur operator. Based on this operator, we formulate a joint optimization framework that enforces blur consistency, event consistency, and image regularization. Finally, we design an unrolled optimization network that alternates between data-consistency reconstruction and learned image priors. Experiments on both synthetic and real-world datasets demonstrate that the proposed method improves deblurring performance while maintaining low computational complexity.

Recent paradigms in Random Projection Layer (RPL)–basedcontinual representation learning have demonstrated superior performancewhen building upon a pre-trained model (PTM). These methods insert arandomly initialized RPL after a PTM to enhance feature representationin the initial stage. Subsequently, a linear classification head is used foranalytic updates in the continual learning stage. However, under severedomain gaps between pre-trained representations and target domains, arandomly initialized RPL exhibits limited expressivity under large do-main shifts. While largely scaling up the RPL dimension can improveexpressivity, it also induces an ill-conditioned feature matrix, therebydestabilizing the recursive analytic updates of the linear head. To this end,we propose the Stochastic Continual Learner with MemoryGuard Super-visory Mechanism (SCL-MGSM). Unlike random initialization, MGSMconstructs the projection layer via a principled, data-guided mechanismthat progressively selects target-aligned random bases to adapt the PTMrepresentation to downstream tasks. This facilitates the construction of acompact yet expressive RPL while improving the numerical stability ofanalytic updates. Extensive experiments on multiple exemplar-free ClassIncremental Learning (CIL) benchmarks demonstrate that SCL-MGSMachieves superior performance compared to state-of-the-art methods.Project Page: https://rlinl.github.io/SCL/

Vision-language alignment—the stage that bridges pretrainedvision encoders and large language models—is widely treated as a formof pretraining requiring full-parameter updates. We challenge this viewand investigate what happens when low-rank adaptation is applied to theLLM during this stage instead. We find that low-rank alignment not onlyreduces computational costs but also outperforms full-parameter align-ment on most benchmarks. To understand this phenomenon, we system-atically characterize the implicit biases introduced by low-rank adap-tation during alignment. Empirically, we find that low-rank alignmentshifts model behavior from hallucinatory to conservative and preservesper-token linear separability of visual features that full-parameter align-ment disrupts—a phenomenon we term LS-curse. Geometrically, low-rank aligned models exhibit more homogeneous and structurally stablevisual representations, maintaining modality-specific knowledge ratherthan prematurely fusing entity-level semantics. Theoretically, we estab-lish two theorems showing that low-rank alignment induces preferencesfor parameter subspaces with flat gradients and feature subspaces robustto perturbations, providing a principled explanation for the observedstructure-preserving behavior. Extensive experiments cover ablation over100 alignment configurations, three families of low-rank operators, andvarious rank, encoder, and other settings.

Recent advances in video generation models have signifi-cantly accelerated video generation and related downstream tasks. Amongthese, video stylization holds important research value in areas suchas immersive applications and artistic creation, attracting widespreadattention. However, existing diffusion-based video stylization methodsstruggle to maintain stability and consistency when processing long videos,and their high computational cost and multi-step denoising make themdifficult to apply in practical scenarios. In this work, we propose RTR-DiT (DiT as Real-Time Rerenderer), a streaming video stylization frame-work built upon a Diffusion Transformer. We first fine-tune a bidirec-tional teacher model on a curated video stylization dataset, supportingboth text-guided and reference-guided video stylization tasks, and sub-sequently distill it into a few-step autoregressive model via post-trainingwith Self Forcing and Distribution Matching Distillation. Furthermore,we propose a reference-preserving KV cache update strategy that notonly enables stable and consistent processing of long videos, but alsosupports real-time switching between text prompts and reference images.Experimental results show that RTR-DiT outperforms existing meth-ods in both text-guided and reference-guided video stylization tasks, interms of quantitative metrics and visual quality, and demonstrates ex-cellent performance in real-time long video stylization and interactivestyle-switching applications.

Existing latent diffusion models excel at visual generationand editing tasks, employing autoencoders to project RGB images andvideos into latent spaces. However, their low-compression latents pro-duce redundant and excessive visual tokens, limiting scalability to high-resolution content such as 4K images and 2160p videos. One promis-ing approach for acceleration is to leverage deeply compressed latentspaces produced by deep compression autoencoders. However, trainingindustrial-scale models from scratch remains expensive and often fails tomatch the performance of pre-trained models. Alternatively, directly fine-tuning existing pre-trained models on deeply compressed latent spacesis appealing, but it introduces a critical challenge: a representation gapbetween the original latent space and the target compressed space, whichcan result in suboptimal convergence and even training instability. DC-Gen offers the first solution to tackle this challenge. Specifically, we in-troduce a novel embedding alignment process to bridge the representa-tion gap. Once the embedding spaces are aligned, only lightweight LoRAfine-tuning is required to unlock the pre-trained model’s full generativecapability. We validate the effectiveness of DC-Gen across tasks rangingfrom image and video generation to image editing. The resulting DC-Genmodels achieve quality comparable to pre-trained models while deliveringsubstantial speedups. For example, DC-Gen-FLUX reduces the latencyof 4K image generation by 53.8× on a single NVIDIA H100 GPU. Codeis available at https://github.com/dc-ai-projects/DC-Gen.

Diffusion Transformers (DiTs) have advanced video gener-ation with high-quality, temporally coherent results. However, extend-ing them to motion transfer, which requires following reference motionwhile aligning with a target prompt, remains challenging due to lim-ited understanding of motion and structure representations within DiTs.We analyze video DiTs at the attention-head level and identify distinctheads specialized for motion and spatial structure. Based on this insight,we propose a head-aware controllable motion transfer framework thatrequires no parameter updates. Our method refines motion cues frommotion-specialized heads via semantic correspondence guidance and pre-serves structure through selective feature injection. This head-level con-trol not only enables accurate motion transfer but also provides an in-terpretable foundation for controllable video generation with DiTs.

While Multimodal Large Language Models (MLLMs) showremarkable capabilities, their safety alignments are susceptible to jail-break attacks. Existing attack methods typically focus on text-imageinterplay, treating the visual modality as a secondary prompt. This ap-proach underutilizes the unique potential of images to carry complex,contextual information. To address this gap, we propose a new image-centric attack method, Contextual Image Attack (CIA), which employsa multi-agent system to subtly embed harmful queries into seeminglybenign visual contexts using four distinct visualization strategies. Tofurther enhance the attack’s efficacy, the system incorporates contextualelement enhancement and automatic toxicity obfuscation techniques. Ex-perimental results on the MMSafetyBench-tiny dataset show that CIAachieves high toxicity scores of 4.62 and 4.83 against the GPT-4o andQwen2.5-VL-72B models, respectively, with Attack Success Rates (ASR)reaching 86.83% and 91.02%. Our method significantly outperforms priorwork, demonstrating that the visual modality itself is a potent vector forjailbreaking advanced MLLMs. WARNING: This paper may contain ex-amples of harmful content for research purposes.

Multimodal Large Language Models (MLLMs) remain vul-nerable to jailbreak attacks exploiting cross-modal interactions. Existingcross-modal attacks rely only on cross-entropy optimization, causingvisual and textual perturbations to act independently and push modelrepresentations toward misaligned vulnerability subregions, limiting theircombined effectiveness. To address this, we propose Co-Steer, a universalattack framework built on a key insight: visual and textual jailbreaks,despite appearing misaligned, are projections of a unified vulnerabilitydirection in the representation space. Co-Steer first extracts a universalsteering vector by aggregating representation shifts from successful vi-sual and textual jailbreaks via SVD. It then optimizes a single universalperturbation pair across queries using a steering alignment loss that coor-dinates both perturbations toward the extracted direction, transformingmisalignment into coordination. Analyzing such attack mechanisms helpsexpose gaps in current multimodal alignment and informs more robustdefense design. Experiments show our method achieves state-of-the-artattack success rates on MM-SafetyBench, HarmBench, and AdvBenchacross InternVL2-8B, Qwen2-VL-7B, and MiniGPT-4-13B, with notableimprovements in transferability to commercial models. Warning: Thispaper might contain harmful content.

Person Re-Identification (ReID) faces severe challenges frommodality discrepancy and clothing variation in long-term surveillancescenario. While existing studies have made significant progress in ei-ther Visible-Infrared ReID (VI-ReID) or Clothing-Change ReID (CC-ReID), real-world surveillance system often face both challenges simul-taneously. To address this overlooked yet realistic problem, we definea new task, termed Cross-Modality Clothing-Change Re-Identification(CMCC-ReID), which targets pedestrian matching across variations inboth modality and clothing. To advance research in this direction, weconstruct a new benchmark SYSU-CMCC, where each identity is cap-tured in both visible and infrared domains with distinct outfits, reflectingthe dual heterogeneity of long-term surveillance. To tackle CMCC-ReID,we propose a Progressive Identity Alignment Network (PIA) that pro-gressively mitigates the issues of clothing variation and modality dis-crepancy. Specifically, a Dual-Branch Disentangling Learning (DBDL)module separates identity-related cues from clothing-related factors toachieve clothing-agnostic representation, and a Bi-Directional PrototypeLearning (BPL) module performs intra-modality and inter-modality con-trast in the embedding space to bridge the modality gap while furthersuppressing clothing interference. Extensive experiments on the SYSU-CMCC dataset demonstrate that PIA establishes a strong baseline forthis new task and significantly outperforms existing methods.

Video understanding is a fundamental capability for multimodal intelligence, and recent Multimodal Large Language Models (MLLMs) have achieved remarkable performance on Video Question Answering (VideoQA) benchmarks. However, existing benchmarks primarily evaluate whether models can perceive shallow visual cues, while rarely examining whether MLLMs can learn deeper knowledge or procedural skills from video tutorials and generalize them to downstream long-horizon agentic tasks. To address this gap, we introduce VG-GUIBench (Video-Guided GUI Benchmark), a new benchmark designed to evaluate whether MLLM-based GUI agents can follow video tutorials to complete corresponding GUI interactive tasks. Furthermore, we observe that the performance of models on both VideoQA and video-guided agentic tasks critically depends on effective keyframe extraction. Based on this observation, we propose TASKER (Task-driven And Scene-aware Keyframe searchER), a keyframe extraction algorithm that jointly considers task relevance and scene dynamics to identify informative frames. Experimental results demonstrate that TASKER achieves significant performance improvements on both VideoQA and video-guided agentic task benchmarks, outperforming the best baseline by 2.0% on the EgoSchema fullset and 1.8% on the NExT-QA dataset, respectively. These results further highlight the potential of generalized keyframe extraction methods for video understanding tasks. Our code and data are available at https://github.com/VG-GUI-TASKER/VG-GUI-TASKER.

Although VLMs are making fast progress in recognizing non-Latin scripts, they still face feasibility problems for less common scriptswith extremely low resources. Current Open-set text recognition methodsshow some good qualitative results, although they still lack quantitativeevidence. Specifically, the majority of open-set and zero-shot text recog-nition methods are confined to the CJK family, and to our knowledge,none handle more than one script family at the same time. In this work,we fill this gap with a unified, script-agnostic open-set text recognitionframework that can handle not just multiple scripts, but multiple scriptfamilies - Indic, Latin, and CJK scripts - in one set of partially sharedmodules. In this framework, we introduce a robust network structurethat can handle the complex character layout across a diverse collectionof scripts of different families. Furthermore, we introduce a heteroge-neous co-training method and tasks to enhance the model performanceon unseen scripts, and discuss the impacts from different data sourcesand task formats (character and word recognition). Experiments demon-strate that the proposed framework shows zero-shot learning capabilityon natural scene word images of Bengali, Gujarati, Japanese, Korean,and synthetic Yi samples. Specifically on the rare, yet actively used, Yiscript, our framework can achieve a word accuracy (ACR) of 16.10%,while the GPT5.2 shows an ACR of only 0, indicating a fundamentalcapability gap.

The rapid progress of generative models has intensi_x001C_ed theneed for reliable and robust detection under real-world conditions. How-ever, existing detectors often over_x001C_t to generator-speci_x001C_c artifacts andremain highly sensitive to real-world degradations. As generative archi-tectures evolve and images undergo multi-round cross-platform shar-ing and post-processing (chain degradations), these artifact cues be-come obsolete and harder to detect. To address this, we propose Real-centric Envelope Modeling (REM), a new paradigm that shifts de-tection from learning generator artifacts to modeling the robust distribu-tion of real images. REM introduces feature-level perturbations in self-reconstruction to generate near-real samples, and employs an envelopeestimator with cross-domain consistency to learn a boundary enclosingthe real image manifold. We further build RealChain , a comprehensivebenchmark covering both open-source and commercial generators withsimulated real-world degradation. REM achieves an average improve-ment of 6.9% over state-of-the-art (SOTA) methods across 12 bench-mark evaluations. Notably, it maintains exceptional generalization on theseverely degraded RealChain benchmark, outperforming existing meth-ods by 18.4%. These results establish a solid foundation for reliable syn-thetic image detection under challenging real-world conditions. The codeand datasets are available at https://github.com/handsome-rich/REM.

The integration of Large Language Models (LLMs) into au-tonomous driving has attracted growing interest for their strong rea-soning and semantic understanding abilities, which are essential for han-dling complex decision-making and long-tail scenarios. However, existingmethods typically feed LLMs with tokens from multi-view and multi-frame images independently, leading to redundant computation and lim-ited spatial consistency. This separation in visual processing hinders ac-curate 3D spatial reasoning and fails to maintain geometric coherenceacross views. On the other hand, Bird’s-Eye View (BEV) representa-tions learned from geometrically annotated tasks (e.g., object detection)provide spatial structure but lack the semantic richness of foundationvision encoders. To bridge this gap, we propose BEVLM, a frameworkthat connects a spatially consistent and semantically distilled BEV rep-resentation with LLMs. Through extensive experiments, we show thatBEVLM enables LLMs to reason more effectively in cross-view driv-ing scenes, improving accuracy by 46.0 %, by leveraging BEV featuresas unified inputs. Furthermore, by distilling semantic knowledge fromLLMs into BEV representations, BEVLM significantly improves closed-loop end-to-end driving performance in safety-critical scenarios acrossUniAD and VAD, with gains of up to 28.2 %.

Microscopic imaging provides essential visual evidence forstudying plant biology and pathology at the cellular and subcellular lev-els. However, existing benchmarks for vision-language models primarilyfocus on macroscopic plant imagery, while the microscopic domain re-mains underexplored. To address this gap, we present PlantMicro, a com-prehensive benchmark for evaluating vision–language models (VLMs) inmicroscopic plant imagery. PlantMicro integrates more than 5,000 imagescollected from diverse hosts, biological domains, and imaging modali-ties. Building on this diversity, we design a set of complementary tasksthat capture different aspects of microscopic image understanding. Tosupport these tasks, we construct over 9,000 VQA pairs that system-atically evaluate the capabilities of VLMs. Experiments on PlantMicroshow that current VLMs struggle with fine-grained recognition and bio-logically grounded reasoning. For example, GPT-5 achieves 34.93% accu-racy on the pathogen classification task, which is only modestly above a24.95% random guessing baseline. The results highlight a significant gapin the ability of current VLMs to comprehend microscopic plant images.PlantMicro provides a standardized foundation for advancing VLMs to-ward reliable and comprehensive microscopy-level plant understanding.PlantMicro is available at https://github.com/tqwei05/PlantMicro.

How can we infer a 3D hand pose when large portions of the handare heavily occluded by itself or by objects? Humans often resolve such ambigui-ties by leveraging contextual knowledge—such as affordances, where an object’sshape and function suggest how the object is typically grasped. Inspired by thisobservation, we propose a generative prior for 3D hand pose modeling guidedby affordance-aware textual descriptions of hand-object interactions (HOI). Ourmethod employs a diffusion-based generative model that learns the distributionof plausible hand poses conditioned on contextual signals, such as affordancedescriptions and image features. The affordance descriptions are designed to rep-resent the semantic intent and geometric structure of HOI, using the reasoningfrom a vision-language model (VLM) and grasp classification. We leverage thediffusion prior to refine the 3D pose predictions in hand reconstruction into moreaccurate and functionally coherent estimation. Our experiments demonstrate thatour affordance-guided refinement significantly improves 3D hand pose estima-tion performance on 3D hand affordance datasets, HOGraspNet and HO3D, overstate-of-the-art methods such as foundation models for hand reconstruction andthe latest diffusion priors for 3D hands.

We introduce Multi-Object Generative Perception (MultiGP),a generative inverse rendering method for stochastic sampling of all radio-metric constituents—reflectance, texture, and illumination—underlyingobject appearance from a single image. Our key idea to solve this in-herently ambiguous radiometric disentanglement is to leverage the factthat while their texture and reflectance may differ, objects in the samescene are all lit by the same illumination. MultiGP exploits this consen-sus to produce samples of reflectance, texture, and illumination from asingle image of known shapes based on four key technical contributions: acascaded end-to-end architecture that combines image-space and angular-space disentanglement; Coordinated Scheduling for diffusion convergenceto a single consistent illumination estimate; Axial Attention applied tofacilitate “cross-talk” between objects of different reflectance; and a Tex-ture Extraction ControlNet to preserve high-frequency texture detailswhile ensuring decoupling from estimated lighting. Experimental resultsdemonstrate that MultiGP effectively leverages the complementary spatialand frequency characteristics of multiple object appearances to recoverindividual texture and reflectance as well as the common illumination.