ECCV 2026 Papers

Image Quality Assessment (IQA) in open-world environmentsremains challenging due to limited generalization and interpretability.Recent approaches based on multimodal large language models (MLLMs)introduce textual reasoning for quality prediction, yet their judgmentsrely heavily on semantically biased internal representations, making theminsensitive to low-level perceptual degradations. We propose IQA-T1, atool-based visual evidence reasoning framework that augments MLLMreasoning with explicit perceptual observations. During inference, themodel autonomously invokes specialized analysis tools to generate struc-tured visual evidence, such as noise residual maps, gradient statistics,and frequency spectra, which are progressively integrated into the rea-soning process. To support this paradigm, we construct Q-Tool, a datasetcontaining 11k multimodal reasoning chains grounded in tool-generatedevidence. Extensive experiments on seven IQA benchmarks show thatIQA-T1 achieves the best overall performance across datasets while pro-ducing interpretable and evidence-grounded quality assessments. Codeand dataset are available at https://github.com/zibuyu-02/IQA-T1.

Dataset condensation for action segmentation synthesizescompact, informative representations of long, untrimmed video datasets.The existing approach relies on Variational Autoencoders and an iter-ative latent optimization; it is computationally expensive and suffersfrom over-smoothed reconstructions and rigid temporal constraints. Thispaper proposes to shift the condensation paradigm from optimization-based inversion to deterministic latent mapping. By leveraging Denois-ing Diffusion Implicit Models, we represent action segments as continu-ous trajectories anchored by sparse latent points in the noise manifold.To maximize representational efficiency, we introduce an adaptive allo-cation mechanism that dynamically redistributes the anchoring budgetbased on segment-wise reconstruction difficulty. Extensive experimentsdemonstrate that our framework significantly outperforms state-of-the-art methods in segmentation performance across common datasets. No-tably, our approach achieves performance parity with real data trainingwhile maintaining a condensation ratio of 2.4% on Breakfast dataset.

Semi-supervised medical image segmentation (SSMIS) aimsto achieve accurate segmentation with limited labeled data. However,existing methods still struggle in multi-organ scenarios. Limited anno-tations exacerbate class imbalance and feature entanglement among co-occurring organs, leading to false activations and performance degrada-tion. To overcome these challenges, we propose a dual-space patch-drivencomplementary learning framework for semi-supervised multi-organ seg-mentation (DPCL). Our framework addresses these problems from bothimage and feature space. In the image space, we combine uncertaintyand class frequency to form a probabilistic sampling strategy for im-age decomposition, which generates informative patches to enhance localfeature perception and promote balanced learning across organs. In thefeature space, we design two complementary objectives to achieve ro-bust feature disentanglement. The patch-based contrastive loss focuseson decoupling features of co-occurring neighboring organs within lo-cal patches, mitigating semantic confusion among adjacent structures.Meanwhile, the global-based contrastive loss minimizes the gap betweenglobal and local semantics, improving model sensitivity to fine-graineddetails and enhancing consistency across different regions. These com-plementary losses collaboratively enhance inter-organ feature discrimi-nation while mitigating false activations. Extensive evaluations on fourpublic multi-organ medical segmentation datasets (Synapse, FeTA2021,SegTHOR, and ACDC) demonstrate that DPCL outperforms state-of-the-art methods, achieving substantial improvements in segmentationperformance.

Current generative models, including GANs and diffusionmodels, have reached an outstanding level of photorealism, posing sig-nificant risks to privacy and security. To ensure real-world applicability,deepfake detectors must generalise effectively to unseen generators. How-ever, most existing approaches rely on supervised training with both realand fake images, which limits their generalisation especially across gener-ators categories (e.g. GANs vs DMs). In this work, we introduce µFlow ,a one-class deepfake detector trained only on real images without relyingon pseudo-deepfakes or synthetic artifacts. Our approach builds on theobservation that averaging multiple images amplifies consistent genera-tive traces, producing highly discriminative feature representations. Weleverage this property by modelling the distribution of features extractedfrom averaged images and training a normalizing flow to align the featurespace of individual images with this distribution. This alignment yieldsa likelihood-based criterion that separates real and fake samples whilepromoting strong generalisation. We evaluate µFlow on a fully out-of-distribution setting, where both real and fake datasets are unseen duringtraining. Experimental results show that our method significantly out-performs SOTA detectors. Project page: opontorno.github.io/MuFlow.

Vision-Language-Action (VLA) models aim to unify per-ception, language understanding, and action generation, offering strongcross-task and cross-scene generalization with broad impact on embodiedAI. However, current VLA models often lack explicit step-by-step rea-soning, instead emitting final actions without considering affordance con-straints or geometric relations. Their post-training pipelines also rarelyreinforce reasoning quality, relying primarily on supervised fine-tuningwith weak reward design. To address these challenges, we present VLA-R1, a reasoning-enhanced VLA that integrates Reinforcement Learningfrom Verifiable Rewards (RLVR) with Group Relative Policy Optimiza-tion (GRPO) to systematically optimize both reasoning and execution.Specifically, we design an RLVR-based post-training strategy with ver-ifiable rewards for region alignment, trajectory consistency, and outputformatting, thereby strengthening reasoning robustness and executionaccuracy. Moreover, we develop VLA-CoT-13K, a high-quality datasetthat provides chain-of-thought supervision explicitly aligned with affor-dance and trajectory annotations. Furthermore, extensive evaluations onin-domain, out-of-domain, simulation, and real-robot platforms demon-strate that VLA-R1 achieves superior generalization and real-world per-formance compared to prior VLA methods. We plan to release the model,code, and dataset following the publication of this work.

World models have demonstrated significant promise for data synthesis in autonomous driving. However, existing methods predominantly concentrate on single-modality generation, typically focusing on either multi-camera video or LiDAR sequence synthesis. In this paper, we propose UniDriveDreamer, a single-stage unified multimodal world model for autonomous driving, which directly generates multimodal future observations without relying on intermediate representations or cascaded modules. Our framework introduces a LiDAR-specific variational autoencoder (VAE) designed to encode input LiDAR sequences, alongside a video VAE for multi-camera images. To ensure cross-modal compatibility and training stability, we propose Unified Latent Anchoring (ULA), which explicitly aligns the latent distributions of the two modalities. The aligned features are fused and processed by a diffusion transformer that jointly models their geometric correspondence and temporal evolution. Additionally, structured scene layout information is projected per modality as a conditioning signal to guide the synthesis. Extensive experiments demonstrate that UniDriveDreamer outperforms previous state-of-the-art methods in both video and LiDAR generation, while also yielding measurable improvements in downstream driving tasks.

We introduce a controlled subspace intervention framework to investigate how self-supervised Vision Transformers (ViTs) encode dense geometric information. While linear probing is widely used to assess geometric representations, it treats features as a black box, failing to disentangle the underlying topology. To address this issue, we decompose the weights of converged linear probes to isolate the low-rank subspaces containing explicit geometric signals using Singular Value Decomposition (SVD). Our perspective yields three key insights: (1) Pre-training objectives determine how features are encoded. DINOv2 aligns spatial features for efficient linear extraction, while Masked Autoencoders (MAE) tend to disperse these signals, requiring a broader spatial context. (2) Explicit geometric representations are highly compressible, suggesting dense predictive heads could potentially be constrained to low-rank subspaces with minimal performance loss. (3) The layer-wise task affinity suggests that geometric precision peaks at intermediate layers before yielding to semantic abstraction in the final layers. By connecting internal encoding mechanics with downstream performance, these findings provide a basis for effective feature selection and lightweight decoder design. The source code is available at https://github.com/Zhou-Weichen/Geosubprobe.

In-context segmentation (ICS) requires a model to segmenttarget regions in a query image using only a few reference images andtheir corresponding masks, without updating any parameters. Despiterecent progress, prior ICS studies have largely overlooked a critical as-pect: system robustness, i.e., whether the model can produce stable seg-mentation results for the same query under different references. In thiswork, we revisit ICS from the robustness perspective and introduce anovel paradigm, Concept-Guided In-Context Segmentation (CG-ICS),which performs segmentation by extracting high-level semantic conceptsfrom references rather than relying solely on low-level visual matching.Specifically, CG-ICS introduces a concept reasoning module that usesan MLLM to propose candidates and a SAM3-driven scoring functionwith tree-search refinement to select reliable textual concepts, togetherwith a parallel visual exemplar route that provides query-side spatialgrounding via a simple context construction. Both the textual conceptand the visual exemplar are then used to activate the segmentation ca-pability of a frozen SAM3 backbone. Extensive experiments on standardICS benchmarks demonstrate that CG-ICS not only achieves state-of-the-art accuracy but also substantially improves robustness, yielding amore reliable ICS system with significantly reduced variance across di-verse reference choices.

Understanding person-level bi-manual interactions requiresnot only detecting hands, but also identifying which two hands belongto the same person and what each hand interacts with. Existing hand–object interaction methods are mostly hand-centric: they treat each handas an independent instance, which can lead to ambiguous ownership inmulti-person scenes.We propose a person-centric formulation in which a single query pre-dicts a structured output for one person, including the human box,body pose, hand boxes and states, and interaction targets. We intro-duce part-aware deformable attention to allocate attention acrosshuman, hand, and pose-specific reference regions, enabling one query tocapture the full person structure. We further unify detection and inter-action reasoning with a hand-to-query relationship matrix, whereeach hand selects its interaction target from the detected query set plusa learnable off token, directly recovering the target’s box and class with-out separate object regression.We build a COCO-based dataset with person-centric bi-manual interac-tion annotations and define structured metrics for evaluating hand statesand complete hand–object tuples. Experiments with a transformer-baseddetector show that our formulation improves person-level bi-manual in-teraction parsing and provides an effective unified framework for jointdetection, pose estimation, and hand reasoning.Project page: https://lgecto-ail-vil.github.io/SingleQuery-BHOI/

Distilling video generation models to extremely low infer-ence budgets (e.g., 2–4 NFEs) is crucial for real-time deployment, yetremains challenging. Trajectory-style consistency distillation often be-comes conservative under complex video dynamics, yielding an over-smoothed appearance and weak motion. Distribution matching distilla-tion (DMD) can recover sharp, mode-seeking samples, but its local train-ing signals do not explicitly regularize how denoising updates composeacross timesteps, making composed rollouts prone to drift. To overcomethis challenge, we propose Self-Consistent Distribution Matching Distil-lation (SC-DMD), which explicitly regularizes the endpoint-consistentcomposition of consecutive denoising updates. For real-time autoregres-sive video generation, we further treat the KV cache as a quality param-eterized condition and propose Cache-Distribution-Aware training. Thistraining scheme applies SC-DMD over multi-step rollouts and introducesa cache-conditioned feature alignment objective that steers low-qualityoutputs toward high-quality references. Across extensive experiments onboth non-autoregressive backbones (e.g., Wan 2.1) and autoregressivereal-time paradigms (e.g., Self Forcing), our method, dubbed Salt, con-sistently improves low-NFE video generation quality while remainingcompatible with diverse KV-cache memory mechanisms. Project page:https://xingtongge.github.io/Salt.

The impact of real-world noise on Open Vocabulary ObjectDetectors (OV-ODs) remains poorly understood due to their architec-tural complexity. We present our comprehensive analysis, Robust Onion,an empirical study that uses controlled synthetic visual degradations topeel OV-ODs layer-by-layer, revealing how, why, and where robustnessdegrades, systematically analyzing feature collapse. Our findings revealthat models with similar vision backbones exhibit comparable robust-ness, driven by similar feature collapse at similar layers, while factorssuch as pretraining strategy, architectural nuances, and caption supervi-sion contribute little. Robustness is primarily governed by the image do-main rather than annotations, explaining the similar robustness impacton COCO and LVIS, and why datasets like ODinW-13 can give an im-pression of inflated robustness due to large, isolated objects. Finally, wevalidate our insights by improving robustness on real-world BDD-100K,WiderFace, and VisDRONE via our lightweight plug-and-play NN &TK0 approach, using 96× fewer trainable parameters than end-to-endtraining. We also explain the prior works’ robustness observations.

Weakly-supervised video scene graph generation (WS-VSGG)aims to parse video content into structured relational triplets withoutbounding box annotations and with only sparse temporal labeling, sig-nificantly reducing annotation costs. Without ground-truth boundingboxes, these methods rely on off-the-shelf detectors to generate objectproposals, yet largely overlook a fundamental discrepancy from fully-supervised pipelines. Fully-supervised detectors implicitly filter out non-interactive objects, while off-the-shelf detectors indiscriminately detectall visible objects, overwhelming relation models with noisy pairs. We ad-dress this by introducing a learnable pair affinity that estimates the like-lihood of interaction between subject–object pairs. Through Pair AffinityLearning and Scoring (PALS), pair affinity is incorporated into inference-time ranking and further integrated into contextual reasoning throughPair Affinity Modulation (PAM), enabling the model to suppress non-interactive pairs and focus on relationally meaningful ones. To providecleaner supervision for pair affinity learning, we further propose Relation-Aware Matching (RAM), which leverages vision-language grounding toresolve class-level ambiguity in pseudo-label generation. Extensive exper-iments on Action Genome demonstrate that our approach consistentlyyields substantial improvements across different baselines and backbones,achieving state-of-the-art WS-VSGG performance.

While existing feed-forward Gaussian splatting models of-fer computational efficiency and can generalize to sparse view settings,their performance is fundamentally constrained by relying on a singleforward pass for inference. We propose ReSplat, a feed-forward recurrentGaussian splatting model that iteratively refines 3D Gaussians withoutexplicitly computing gradients. Our key insight is that the Gaussiansplatting rendering error serves as a rich feedback signal, guiding therecurrent network to learn effective Gaussian updates. This feedbacksignal naturally adapts to unseen data distributions at test time, enablingrobust generalization across datasets, view counts, and image resolutions.To initialize the recurrent process, we introduce a compact reconstructionmodel that operates in a 16× subsampled space, producing 16× fewerGaussians than previous per-pixel Gaussian models. This substantiallyreduces computational overhead and allows for efficient Gaussian updates.Extensive experiments across varying number of input views (2, 8, 16, 32),resolutions (256×256 to 540×960), and datasets (DL3DV, RealEstate10K,and ACID) demonstrate that our method achieves state-of-the-art perfor-mance while significantly reducing the number of Gaussians and improvingthe rendering speed. Our project page is at haofeixu.github.io/resplat.

Personalization models generate images guided by a few subject references, while style transfer methods aim to produce images aligned with a global style derived from a reference image. Recent approaches perform well when the reference image contains a single object, effectively capturing a global style that encompasses all implicit attributes. However, when applied to complex real-world scenes containing multiple objects with distinct attributes, existing methods struggle to selectively control generation using attributes from specific local regions of the reference image. Due to their primarily global conditioning strategies, these methods cannot reliably guide generation using attributes from specific local elements in the reference image. Moreover, existing methods typically employ a single identifier token to capture all details from the reference, resulting in a lack of individual, attribute-level control. Motivated by these limitations, we propose RefDiT, a novel framework for reference-guided image generation. RefDiT takes as input a reference image, a text prompt, and an optional user-provided guidance control. RefDiT employs local region guidance using the attributes of local elements. It constructs an attribute-aware conditioning signal from the reference image by performing attribute-level decomposition of the identifier tokens to train low-rank adapter (LoRA) blocks of a diffusion transformer (DiT)-based generative model. RefDiT leverages multi-modal joint attention to learn the correspondence between identifier tokens and local regions in the reference image, enabling more effective local guidance. RefDiT achieves a local attribute matching score of 0.88, outperforming state-of-the-art methods such as UnZipLoRA (0.54), K-LoRA (0.58), and B-LoRA (0.42). Compared to commercial models like Gemini-Banana (0.74) and GPT-5 (0.84), user studies indicate that participants find RefDiT’s outputs to be of similar quality.

Although Contrastive Language-Image Pretraining (CLIP)has shown promising transfer ability for ophthalmic image analysis, itsopaque decision-making process hinders interpretability, especially inophthalmic settings where probability-only outputs may be difficult forclinicians and patients to trust. Providing visual evidence for predictionscan therefore make models more inspectable and interpretable. How-ever, existing explainability methods often struggle to produce spatiallyconsistent and semantically aligned heatmaps that highlight prediction-relevant anatomical or pathological evidence. To bridge this gap, wepropose OTSMDL, a post-hoc framework for interpretable region selec-tion that couples submodular-inspired scoring with unbalanced optimaltransport. We formulate the identification of salient evidence regions as asubset selection problem. Specifically, submodular-inspired utilities firstconstruct an initial candidate subset and a score-guided source distri-bution, which are then used to initialize the Optimal Transport (OT)process. Subsequently, OT performs global transport-based re-rankingover all candidates, constructing a unified cost matrix to encourage cross-region competition and align candidate superpixels with text prompts.We evaluate OTSMDL on seven ophthalmic datasets spanning color fun-dus photography, OCT, and glaucoma assessment, with additional Ima-geNet experiments for out-of-domain analysis. Under standard post-hocDeletion/Insertion protocols and ground-truth mask localization met-rics, OTSMDL generally outperforms baselines in region-level evidencegrounding, achieving top-1/2 performance in most evaluation settings.

Multimodal Large Language Models (MLLMs) have achievedremarkable progress but still struggle with complex visual reasoningtasks requiring multi-step perception and logical deduction. While ex-plicit visual generation incurs prohibitive computational costs, existinglatent approaches often rely on external experts or lack rigorous cogni-tive logic. In this paper, we introduce ProLaViT (Progressive LatentVisual Thought), a framework empowering MLLMs to perform struc-tured visual derivation in the continuous latent space. Unlike worksdependent on heterogeneous external models, ProLaViT leverages anendogenous self-distillation mechanism, utilizing the model’s own vi-sual encoder to supervise latent thoughts. To facilitate this, we con-struct a scalable programmatic synthesis pipeline enabling the modelto internalize algorithmic precision without inference-time tools. We de-sign two reasoning paradigms: (1) Coarse-to-Fine Causal Chain for spa-tial tasks, guiding attention from global context to local targets. (2)Dialectical Reasoning Chain for logical tasks, incorporating counter-factual thinking for verification. Furthermore, we propose a Distance-Weighted Diversity Loss to impose topology-aware constraints, prevent-ing feature degeneration by enforcing semantic distinctiveness. Exten-sive experiments demonstrate that ProLaViT outperforms baselines onvision-centric benchmarks, achieving superior accuracy and interpretabil-ity with high efficiency.

Sparse Mixture-of-Experts (MoE) offers an attractive pathto scaling discrete diffusion vision–language models (dVLMs), but wefind that naïvely combining MoE with multimodal diffusion training canfail catastrophically. Specifically, MoE-dVLMs suffer from expert selec-tion collapse on visual tokens, where most image tokens are routed toonly a few experts and overall multimodal performance degrades sharply.We trace this failure to an upstream geometric mismatch at the modalityboundary rather than router optimization. Visual embeddings producedby deep Pre-Norm vision encoders have much larger norms than text em-beddings; when injected into a Pre-Norm language tower, the residualdynamics cause attention drowning and representational inertia, yield-ing highly concentrated, low-rank visual representations that are difficultfor the router to discriminate. To resolve this issue, we propose Tower-Align, a lightweight post-projector modality-wise norm alignment thatglobally rescales visual tokens to match text norm statistics while pre-serving their relative structure. TowerAlign stabilizes MoE routing andimproves multimodal accuracy with negligible overhead. On a practicalMoE-dVLM built from LLaDA-MoE-7B-A1B with a SigLIP2-SO400M-Patch14-384 vision encoder, TowerAlign delivers consistent gains acrossdiverse benchmarks and largely closes the gap to dense diffusion baselinesunder matched data budgets.

Slide design requires personalizing both deck themes andpage layouts. Yet, current AI agent-based methods struggle with fine-grained, page-level design. Solely relying on prespecified templates oruser verbose instructions, they fail to capture latent design intents, leav-ing Page-level Slide Personalization (PSP) unresolved. To close this gap,this work formulates PSP as an inverse planning problem. We proposeto learn a design intent without assuming any knowledge of the spe-cific executing tools (e.g., PowerPoint, Beamer) being used. However,relinquishing control over these tools makes the problem intractable tooptimize end-to-end. To overcome this, we propose Spire, a principledframework to solve PSP approximately. By intentionally corrupting thevisual structures of clean slides, Spire creates a verifiable task to de-noise the corruption, whereby two agents learn to collaboratively refineexecutable designs via reinforcement learning (RL). We present a proofthat structural denoising is a consistent surrogate for PSP, and that themulti-agent formulation strictly reduces policy gradient variance in RL.Extensive experiments demonstrate the superiority of Spire.

Real-world object removal is challenging due to two key dif-ficulties: the target object’s non-local effects (e.g., shadows and reflec-tions) that are hard to model, and the fact that user-provided masksare often inaccurate or incomplete. With billions of parameters and tensof denoising steps, diffusion-based models achieve this goal at the ex-pense of massive computational cost, limiting their use in interactiveapplications and edge devices. To solve this problem, we present OSOR(One-Step Object Removal), which achieves efficient, effect-aware, andmask-robust object removal at the same time. Concretely, OSOR in-troduces (1) an occupancy-guided discriminator for precise boundarysupervision, enabling stable single-step diffusion training, (2) an alphahead, which leverages the knowledge in diffusion models to predict cor-rect removal regions with minimal overhead, handling imperfect masks,and (3) a semantic-anchored verification pipeline (SAVP) that filtersnoisy instruction-based triplets to produce effect-aware supervision atscale. Using SAVP, we curate CORNE with 280K verified removal pairsand further annotate AnimeEraseBench and TextEraseBench to evalu-ate model performance in more complex tasks. Experiments show thatOSOR surpasses strong multi-step diffusion baselines in perceptual qual-ity while achieving 4× to 30× faster inference. Code and resources areavailable at https://github.com/Zhouqm-Git/osor.