ECCV 2026 论文列表

On-device vision-language models (VLMs) deployed on edgedevices often encounter distribution shifts between pre-training data andtest data, which can significantly degrade their performance. To miti-gate this degradation, device-server hybrid inference allows an on-deviceVLM to delegate uncertain test inputs to a more powerful server-sideVLM. However, frequently delegating to the remote server at test timeinflates data-transfer costs and server-call latency, limiting the practi-cality of hybrid inference in real-world applications. Thus, to reduce thedelegation frequency without sacrificing the benefits of server assistance,we present CURE, a cumulative knowledge reuse framework for device-server hybrid inference, which efficiently enhances on-device VLMs byaccumulating and reusing the server-side VLM’s knowledge. As a result,CURE enables the on-device VLM to approximate server-level predic-tions with far fewer delegations, thereby amortizing both data-transfercosts and server-call latency across the test stream. Extensive experi-ments across 13 datasets show that CURE closely matches the accuracyof an always-delegate-to-server baseline, reduces delegation frequency by66.74% on average, and improves end-to-end runtime by up to 3.67×,demonstrating the practicality of CURE for real-world deployment.

Large Multimodal Models (LMMs) show strong few-shotgeneralization, but industrial anomaly detection remains difficult becausedefects are small, input resolution is limited, and textual standardsare not always grounded in visual evidence. Recent optimization-basedmethods improve alignment through fine-tuning, but they often requiremany defective samples, which are unavailable in early deployment. Wepresent Global Logic and Local Search (GLLS), a training-freeframework for reference-guided multimodal in-context verification. GLLSuses a Part-Aware Visual-Logical Atlas to organize normal references andstructured specifications in the inference context. It combines a Global& Logic Stream, where SAM 3 extracts partially checkable visual facts,with a Fine-Grained & Actions Stream, where MCTS selects localevidence crops under a fixed budget. Experiments on MMAD-QA andadditional anomaly detection datasets show consistent gains over matchedand general-purpose baselines, while keeping the final diagnostic decisiontraceable to explicit visual evidence throughout the inspection trace.

Diffusion models generate conditional samples by progres-sively denoising Gaussian noise, yet the denoising trajectory can stall atvisually plausible but low-quality outcomes with conditional misalign-ment or structural artifacts. We interpret this behavior as local optimain a surrogate quality landscape: Once early denoising commits to a sub-optimal global structure, later steps mainly sharpen details and seldomcorrect the underlying mistake. While existing inference-time approachesexplore alternative diffusion states via re-noising with fixed strength ordirection, they exhibit limited capacity to escape steep quality plateaus.We propose Controlled Random Zigzag Sampling (Ctrl-Z Sampling), ascalable sampling strategy that detects plateaus in quality landscapevia a surrogate score, and allocates exploration only when a plateau isdetected. Upon detection, Ctrl-Z Sampling rolls back to noisier states,samples a set of alternative continuations, and updates the trajectorywhen a candidate improves the score, otherwise escalating the explo-ration depth to escape the current plateau. The proposed method ismodel-agnostic and broadly compatible with existing diffusion frame-works. Experiments show that Ctrl-Z Sampling consistently improvesgeneration quality over other inference-time scaling samplers across dif-ferent NFE budgets, offering a scalable compute-quality trade-off. Codeavailable at: https://github.com/ShunqiM/Ctrl-Z-Sampling.

Movie dubbing aims to synthesize speech that preserves thevocal identity of a reference audio while synchronizing with the lip move-ments in a target video. Existing methods fail to achieve precise lip-syncand lack naturalness due to explicit alignment at the duration level.While implicit alignment solutions have emerged, they remain suscep-tible to interference from the reference audio, triggering pronunciationand lip-sync degradation in in-the-wild scenarios. In this paper, we pro-pose a novel flow matching-based movie dubbing framework driven bythe Cognitive Synchronous Diffusion Transformer (CoSyncDiT), in-spired by the cognitive process of professional actors. This architectureprogressively guides the noise-to-speech generative trajectory by execut-ing acoustic style adapting, fine-grained visual calibrating, and time-aware context aligning. Furthermore, we design the Joint Semantic andAlignment Regularization (JSAR) mechanism to simultaneously con-strain frame-level temporal consistency on the contextual outputs andsemantic consistency on the flow hidden states, ensuring robust align-ment. Extensive experiments on both standard benchmarks and chal-lenging in-the-wild dubbing benchmarks demonstrate that our methodachieves the state-of-the-art performance across multiple metrics. Thecode is available at https://github.com/GalaxyCong/CoSyncDiT.

Joint image–feature generative modeling has recently emergedas an effective strategy for improving diffusion training by coupling low-level VAE latents with high-level semantic features extracted from pre-trained visual encoders. However, existing approaches rely on a fixedrepresentation space, constructed independently of the generative objec-tive and kept unchanged during training. We argue that the represen-tation space guiding diffusion should itself adapt to the generative task.To this end, we propose Coevolving Representation Diffusion (CoReDi),a framework in which the semantic representation space evolves duringtraining by learning a lightweight linear projection jointly with the diffu-sion model. While naïvely optimizing this projection leads to degeneratesolutions, we show that stable coevolution can be achieved through acombination of stop-gradient targets, normalization, and targeted reg-ularization that prevents feature collapse. This formulation enables thesemantic space to progressively specialize to the needs of image synthesis,improving its complementarity with image latents. We apply CoReDi toboth VAE latent diffusion and pixel-space diffusion, demonstrating thatadaptive semantic representations improve generative modeling acrossboth settings. Experiments show that CoReDi achieves faster convergenceand higher sample quality compared to joint diffusion models operatingin fixed representation spaces.

Weakly Supervised Video Anomaly Detection aims to detect abnor-mal events in long, untrimmed videos using only video-level supervision, provid-ing a practical solution to the high cost of temporally localized annotations. How-ever, abnormal events in real-world surveillance are often determined by complexcontextual interactions among actions, objects, and environments, making it dif-ficult for existing approaches to identify which factors actually cause an anomaly.Most prior methods rely on holistic representations or explanations that lack ex-plicit attribution to the semantic factors driving abnormal events. In this paper,we present CLUE-VAD, a structured semantic decomposition framework for ex-plainable weakly supervised video anomaly detection. Rather than relying solelyon holistic video representations or free-form explanations, CLUE-VAD decom-poses each video segment into structured semantic clues—Action, Environment,and Object—and estimates their category-aware contributions to anomaly predic-tion. By linking anomaly scores to clue-specific weights and keyword-level evi-dence, we provides factor-level attribution and supports transparent explanationgeneration. Extensive experiments on VAD benchmarks demonstrate that CLUE-VAD achieves strong detection performance among X-VAD methods while pro-viding fine-grained, clue-grounded explanations for anomaly predictions. Theproject page is available at https://visualsciencelab-khu.github.io/CLUE-VAD_project/.

We propose MotionAtlas, a system for detailed captioningof motion-centric videos, comprising (1) a dedicated human-annotatedbenchmark, (2) a scalable, high-quality pipeline to construct trainingsamples, and (3) a family of powerful Video-MLLMs. Unlike conventionalglobal motion captioning datasets, we focus on region-aware motion cap-tioning: given a video and a spatiotemporal mask, the model generatesprecise descriptions of motion within the target region, thereby alleviatingvisual clutter and motion entanglement and enabling reliable, quantifi-able evaluation. Concretely, we first build MotionAtlas-Bench, a compre-hensive benchmark comprising 2,073 multiple-choice questions, meticu-lously annotated for a curated set of high-quality, motion-centric videos,to evaluate fine-grained motion understanding of the objects in question.Second, we design a rigorous and scalable data pipeline that leveragesself-bootstrap refinement to suppress fine-grained hallucinations, yield-ing 159k high-quality motion captioning data. Third, we design a tai-lored training data composition strategy, which achieves consistent andsubstantial performance gains across diverse baseline Video-MLLMs, in-cluding Molmo2 and Qwen3-VL. For instance, MotionAtlas-4B surpassesQwen3-VL-4B by an average of 5.2 percentage points across general mo-tion benchmarks. The benchmark, dataset, and code have been released.

Multimodal Large Language Models (MLLMs) excel in di-verse vision tasks, but full-parameter retraining is computationally ex-pensive as real-world knowledge evolves. Existing continual learning meth-ods often suffer from semantic entanglement in parameter spaces acrosstasks, impeding the continuous deployment of models. This challenge isespecially pronounced in Anomaly Detection (AD), which exhibits tripleheterogeneity across modalities, domains, and defect scale variability,significantly complicating multi-task knowledge transfer. In this paper,we propose CL-Anomaly, a parameter-efficient fine-tuning frameworkbased on an isolation-sharing collaboration to enable continual learningfor anomaly detection with MLLMs. We introduce the task-private ex-pert PrivLoRA, which physically isolates task-specific subspaces in theparameter space to prevent semantic entanglement of anomaly knowl-edge in diverse scenarios. The Layer-Adaptive Shared Experts maintaincross-task representations within a unified feature space, enabling knowl-edge sharing between previous and new tasks. Furthermore, we proposea Layer-Adaptive Knowledge Transfer strategy that automatically se-lects and dynamically updates the layer-wise key shared experts of eachtask via a momentum-based mechanism, promoting effective knowledgetransfer across related anomaly detection tasks. Extensive experimentsacross three continual learning scenarios for anomaly detection, includ-ing class-incremental, cross-domain, and cross-modal, demonstrate thatCL-Anomaly outperforms state-of-the-art methods. Code is available athttps://github.com/WenDongyp/CL-Anomaly.

Despite the advancements of Large Multimodal Models (LMMs)in RGB vision, their ability to generalize to unseen visual modalitiesremains a largely unexplored challenge. We argue that different visualmodalities are merely distinct samplings of the same physical world.Therefore, effective generalization requires models to possess both modality-agnostic perception of scene semantics and the adaptability to modality-specific characteristics. To achieve this, we propose a training frame-work, VVM-Tuning, to equip LMMs with these capabilities throughmodality synthesis and modality contexts. Specifically, we synthesize di-verse appearance-varied images from RGB scenes, training the model todisentangle invariant semantics from varying visual appearances, andalign these appearances with language for visual concepts decoupledfrom modalities. We then introduce modality contexts in the promptand use instruction tuning to assist the model in mapping these appear-ance variations back to modality-related attributes, enabling zero-shotadaptation to unseen modalities during inference. To facilitate researchin this direction, we introduce VVM-Bench, a comprehensive bench-mark featuring 6 real and synthetic modalities to evaluate semantic per-ception and modality understanding. Experiments demonstrate that, viaour training on synthetic modalities, 5 tested models exhibit consistentimprovements on both real-world and novel synthetic modalities withoutin-modality training. Source code and data will be publicly available athttps://github.com/Hunter-Will/VVM-Tuning.

Diffusion Transformers (DiTs) have become a leading back-bone for video generation, yet their quadratic attention cost remains amajor bottleneck. Sparse attention reduces this cost by computing only asubset of attention blocks. However, prior methods often either drop theremaining blocks which incurs information loss, or rely on learned predic-tors to approximate them, introducing training overhead and potentialoutput distribution shifting. In this paper, we show that the missingcontributions can be recovered without training: after semantic cluster-ing, keys and values within each block exhibit strong similarity and canbe well summarized by a small set of cluster centroids. Based on thisobservation, we introduce SVG-EAR, a parameter-free linear compensa-tion branch that uses the centroid to approximate skipped blocks andrecover their contributions. While centroid compensation is accurate formost blocks, it can fail on a small subset. Standard sparsification typi-cally selects blocks by attention scores, which indicate where the modelplaces its attention mass, but not where the approximation error wouldbe largest. SVG-EAR therefore performs error-aware routing: a lightweightprobe estimates the compensation error for each block, and we computeexactly the blocks with the highest error-to-cost ratio while compensatingfor skipped blocks. We provide theoretical guarantees that relate atten-tion reconstruction error to clustering quality, and empirically show thatSVG-EAR improves the quality-efficiency trade-off and increases through-put at the same generation fidelity on video diffusion tasks. Overall,SVG-EAR establishes a clear Pareto frontier over prior approaches, achiev-ing up to 1.77× and 1.93× speedups while maintaining PSNRs of up to29.759 and 31.043 on Wan2.2 and HunyuanVideo, respectively.

Concept Bottleneck Models (CBMs) promise transparent rea-soning by predicting through human-interpretable concepts, yet theireffectiveness fundamentally depends on how well visual and textual rep-resentations are aligned or matched. Existing vision–language CBMs of-ten rely on pre-aligned encoders or global cosine similarity, which ob-scures fine-grained concept localization and fails to reflect true seman-tic geometry. In this work, we rethink concept alignment as dynamiccross-modal transport process instead of static projection and proposeOptimal Transport Flow Concept Bottleneck Model (OTF-CBM). Itfirst learns a data-driven semantic cost via Inverse Optimal Transport tomeasure cross-modal distances, and then performs unbalanced optimal-transport-based flow matching to model semantic transitions betweenvisual patches and textual concepts. With velocity-based concept activa-tion, OTF-CBM captures interpretable geometric relations without ODEintegration. Experiments further show that OTF-CBM achieves greatclassification accuracy and concept faithfulness, offering a new geomet-ric and dynamical perspective for interpretable cross-modal reasoning.Our code can be found at github.com/ChenyangZhang00/OTF-CBM.

In this paper, we present BLOB-Q, a novel Mixed-PrecisionPost Training Quantization (MPQ) approach for Vision Transformers(ViTs). MPQ aims to assign different bit-widths across layers under amodel size budget. Most existing MPQ methods either rely on layerwisesensitivity heuristics or surrogate objectives. These approaches simplifyoptimization, but they no longer optimize the same model-level objec-tive that determines accuracy. Others solve the global MPQ problemby non-heuristic and non-analytical algorithms, such as ReinforcementLearning, genetic algorithms or gradient-based learning. However, theyare inefficient for modern large vision transformers. In this work, we revisitthe global MPQ problem and show that this problem can be solved byefficient analytical algorithms while still maintaining its global-optimality.This comes from two key insights. Firstly, through systematic signalanalysis on quantization errors, we discover that under 4„6-bit range,the quantization error on converged ViTs are indeed small perturbations.This encourages that the global MPQ objective is possible to be quadrat-ically decomposed for empirical ViT models. Secondly, motivated by thesmall perturbation observation, we further discover that the global MPQobjective satisfies an additivity property. Utilizing the additivity property,the NP-hard global MPQ optimization problem can be decomposed intosub-problems and solved practically, while still retain the global optimal-ity. Specifically, we solve the decomposed analytical optimization problemusing dynamic programming algorithm, which efficiently find the globallyoptimal solution with only linear time complexity. Extensive experimentson numerous ViT models demonstrate the effectiveness of our approach.Results show that BLOB-Q significantly improves state-of-the-art andcan further reduce the size of ViT models to 4 bits to 6 bits withouthurting ImageNet accuracy. Moreover, BLOB-Q is highly efficient, onlyrequiring on average sub-2 minutes for regular ViT-S to ViT-B models.

Precise medical image segmentation is crucial for clinical di-agnosis and treatment planning, yet relies heavily on expensive expertannotations. Semi-supervised medical image segmentation (SSMIS) of-fers a cost-effective solution but typically operates under the assump-tion of independent and identically distributed (i.i.d.) data, defaultingto random sampling. While statistically valid at scale, this strategy suf-fers from severe representation bias in low-data regimes, failing to cap-ture the heterogeneous medical data manifold. To address this, we pro-pose a highly data-efficient framework driven by distribution alignment.First, we introduce an offline Distribution-Aware Sample Selection strat-egy. By leveraging Vision Foundation Models (VFMs) and our designedDensity-K-Center algorithm, we explicitly identify representative struc-tural anchors, establishing a more representative labeled domain. Second,to bridge the remaining distribution gap, we propose the Memory-guidedCopy-Paste (MCP) module. Tailored for the inherent class imbalance inmedical scans, MCP leverages a semantic memory mechanism to retrievehistorically consistent priors for cross-domain alignment, encouragingsemantic consistency. Coupled with an easy-to-hard progressive sched-ule, this framework effectively mitigates early-stage pseudo-label noise.Extensive experiments on six diverse 2D and 3D datasets demonstratestrong segmentation performance, particularly in extremely low-labeledscenarios (e.g., 1/16 ratio).

Text-driven inversion of generative models is a core paradigmfor manipulating 2D or 3D content, unlocking numerous applicationssuch as text-based editing, style transfer, or inverse problems. How-ever, it relies on the assumption that generative models remain sen-sitive to natural language prompts. We demonstrate that for state-of-the-art native text-to-3D generative models, this assumption often col-lapses. We identify a critical failure mode where generation trajectoriesare drawn into latent “sink traps”: regions where the model becomesinsensitive to prompt modifications. In these regimes, changes to theinput text fail to alter internal representations in a way that altersthe output geometry. Crucially, we observe that this is not a limita-tion of the model’s geometric expressivity; the same generative mod-els possess the ability to produce a vast diversity of shapes but, as wedemonstrate, become insensitive to out-of-distribution text guidance. Weinvestigate this behavior by analyzing the sampling trajectories of thegenerative model, and find that complex geometries can still be repre-sented and produced by leveraging the model’s unconditional generativeprior. This leads to a more robust framework for text-based 3D shapeediting that bypasses latent sinks by decoupling a model’s geometricrepresentation power from its linguistic sensitivity. Our approach ad-dresses the limitations of current 3D pipelines and enables high-fidelitysemantic manipulation of out-of-distribution 3D shapes. Project web-page: https://daidedou.sorpi.fr/publication/beyondprompts

While Unified Multimodal Models (UMMs) show remarkablereasoning capabilities, their spatial intelligence remains limited to pas-sive 2D Question-Answering (QA). In this paper, we argue that true spa-tial intelligence demands active construction: not only recognizing a 3Dstructure in pixel space, but also building and modifying it. We introduceMindBlock, a benchmark that challenges models’ active generative con-struction in pixel space across two primary axes: Spatial Assembly,which evaluates step-by-step compositional and causal reasoning, andSpatial Structure, which probes spatial equivariance through localsub-component rotations and global viewpoint transformations. To movebeyond pixel-level metrics, we propose 3DGS-Eval, a novel validationprotocol using 3D Gaussian Splatting to reconstruct implicit scenes frommodel-generated multi-view images. This allows us to quantify structuralconsistency, verifying for the first time whether a model’s generative out-put admits a coherent internal 3D world model. Furthermore, we con-duct a deep-dive diagnostic analysis into the representational groundingof spatial logic, disentangling whether structural consistency relies ontextual Chain-of-Thought (CoT) as a symbolic scaffold, or emerges as anative spatial intuition within the generative latent space. Our findingsreveal a significant “perception-execution” gap: while current models cor-rectly identify the intended spatial state, they fail at active construction.They struggle to maintain spatial equivariance without explicit symbolicscaffolding. MindBlock provides a rigorous foundation for the next gen-eration of embodied, physically-grounded multimodal AI.

Visual reasoning may require models to interpret images/videosand respond to implicit text queries across diverse output formats, frompixel-level segmentation masks to natural language descriptions. Exist-ing approaches rely on supervised fine-tuning with task-specific architec-tures. For example, reasoning segmentation, grounding, summarization,and visual question answering each demand distinct model designs andtraining, preventing unified solutions and limiting cross-task/-modalitygeneralization. Hence, we propose DT-R1, a reinforcement learning frame-work that trains large language models (LLMs) to construct digital twin(DT) representations of complex multi-modal visual inputs and then rea-son over these high-level representations as a unified approach to visualreasoning. Specifically, we train DT-R1 using GRPO with a novel rewardthat validates both structural integrity and output accuracy. Evaluationsin six visual reasoning benchmarks, covering two modalities and four tasktypes, demonstrate that DT-R1 consistently achieves improvements overstate-of-the-art task-specific models. DT-R1 opens a new direction wherevisual reasoning emerges from reinforcement learning on with DT repre-sentations.

Continual visual question answering (VQA) requires learningnew question types while retaining previously acquired visual ground-ing and reasoning. Existing methods mainly reduce parameter interfer-ence, but do not explicitly control how sequential adaptation changesthe coupling between a question and its visual evidence. We proposeQ-STEER, a task-agnostic framework that separates plasticity from re-tention through a shared question-conditioned controller. For plasticity,probe-based MoE-LoRA self-expansion activates new expert slots onlywhen the current capacity is insufficient. For retention, a KL-based late-layer attention-drift signal conditions a low-rank correction to attentionlogits; the drift is diagnostic rather than an attention-matching loss. Oncontinual VQA v2, Q-STEER obtains 53.68 average performance and4.51 average forgetting, and reaches 51.00/51.48 on Novel/Seen compo-sitional splits. Experiments on a heterogeneous visual-instruction stream,together with seed, ablation, grounding, sensitivity, and overhead anal-yses, support the effectiveness of drift-aware steering and on-demandexpansion.

Proactive tamper localization embeds an imperceptible sig-nal into an image prior to distribution, enabling pixel-level manipulationdetection. Existing methods assume a spliced (SP) setting, where syn-thesized regions are composited onto the original background, leavingembedded signals intact. However, real-world diffusion-based inpaint-ing operates in a fully regenerated (FR) setting, where the entire im-age undergoes denoising, disrupting background signals and renderingexisting frameworks ineffective. We propose APT, a semi-fragile latent-space perturbation that embeds a dense, vector-wise localization signal.By aligning each spatial feature vector toward a fixed anchor direction,APT localizes tampering via the alignment disparity between synthesizedforeground and anchor-aligned background features after inpainting. Theproposed hard negative mining loss and noisy perturbation branch fur-ther enforce uniform alignment. Experiments on COCO demonstratethat APT achieves an FR IoU of 0.92, outperforming the strongest base-line (WAM, 0.84), while existing methods collapse to near-random per-formance (AUC ≈ 0.5), establishing APT as a practical forensic frame-work generalizable across tampering types unknown at test time.

Multimodal Large Language Models (MLLMs) achieve re-markable progress in widespread vision-language tasks, while achievingfaithful MLLM reasoning remains a critical challenge: generating con-tent grounded in comprehensive visual evidence without hallucination.This challenge often stems from the inherent perceptual limitations ofthe single visual encoder design in most MLLMs. While many workshave turned to integrating multiple vision experts, existing methods failto leverage the unique insights of diverse experts due to unspecializedmixture and alignment strategies. To resolve these challenges, we in-troduce MoSVE (Mixture of Specialized Vision Experts), a holisticframework built upon the core insight of inter-expert complementarity.To achieve fine-grained mixture, we introduce Query-guided Com-plementary Clustering, which selectively preserves text-critical andinformative visual perception without redundancy. To cultivate expertspecialization, we propose Complementary Rejection Fine-Tuning,which explicitly routes hard samples from the anchor MLLM to the mostvisually-disparate auxiliary expert, enhancing unique insights withouthomogenization. Extensive experiments demonstrate that MoSVE notonly mitigates hallucinations across the POPE, CHAIR and MMVP, butalso improves general multimodal reasoning on MMBench. Ultimately,MoSVE provides an advanced and efficient solution for faithful MLLM