ECCV 2026 Papers

Single image reflection removal (SIRR) is challenging in realscenes, where reflection strength varies spatially and reflection patternsare tightly entangled with transmission structures. This paper presentsa diffusion model with prior modulation framework (FUMO) that in-troduces explicit priors for spatially adaptive conditioning and struc-turally faithful restoration. Two priors are extracted directly from themixed image, an intensity prior that estimates spatial reflection sever-ity and a high-frequency prior that captures detail-sensitive responsesvia multi-scale residual aggregation. We propose a coarse-to-fine train-ing paradigm. In the first stage, these cues are combined to gate theconditional residual injections, focusing the conditioning on regions thatare both reflection-dominant and structure-sensitive. In the second stage,a fine-grained refinement network corrects local misalignment and sharp-ens fine details in the image space. Experiments conducted on both stan-dard benchmarks and challenging images in the wild demonstrate com-petitive quantitative results and consistently improved perceptual qual-ity. The code is released at https://github.com/Lucious-Desmon/FUMO.

“Thinking with images” has recently emerged as a promisingdirection for vision-language models, where models actively interact withimages via iterative zoom-ins to acquire visual evidence for answering.Yet, prevailing frameworks ignore the quality of intermediate zoom-ins,boosting metrics but leading to an action–outcome disconnect. This fail-ure mode creates the illusion of thinking with images, where modelsappear visually grounded while relying on context-agnostic or redun-dant crops. We address this problem by introducing Visual RationaleLearning (ViRL), an end-to-end reinforcement learning paradigm thatgrounds the evidence acquisition process. ViRL ensures action faithful-ness and efficiency through: (1) process supervision with ground-truthvisual rationales, (2) rationale fidelity rewards that incentivize evidence-grounded zoom-ins, and (3) fine-grained credit assignment to penalizeredundant or erroneous steps. Experimental results show that ViRLachieves state-of-the-art performance across benchmarks spanning per-ception, hallucination, and perception-grounded reasoning. We also dis-sect the illusion of thinking with images and position visual rationales asa verifiable foundation for building trustworthy vision-language models.

Real-time understanding of continuous video streams is es-sential for interactive assistants and multimodal agents operating in dy-namic environments. However, most existing video reasoning approachesfollow a batch paradigm that defers reasoning until the full video con-text is observed, resulting in high latency and growing computationalcost that are incompatible with streaming scenarios. In this paper, weintroduce ThinkStream, a framework for streaming video reasoningbased on a Watch–Think–Speak paradigm that enables models toincrementally update their understanding as new video observations ar-rive. At each step, the model performs a short reasoning update and de-cides whether sufficient evidence has accumulated to produce a response.To support long-horizon streaming, we propose Reasoning-CompressedStreaming Memory (RCSM), which treats intermediate reasoning tracesas compact semantic memory that replaces outdated visual tokens whilepreserving essential context. We further train the model using a Stream-ing Reinforcement Learning with Verifiable Rewards scheme that alignsincremental reasoning and response timing with the requirements ofstreaming interaction. Experiments on multiple streaming video bench-marks show that ThinkStream significantly outperforms existing onlinevideo models while maintaining low latency and memory usage. Code,models and data are available at https://github.com/CASIA- IVA-Lab/ThinkStream.

Video generation has advanced to produce visually compelling and temporally coherent results. Yet, whether these models can genuinely think with video—executing symbolic rules, respecting physical laws, and pursuing intentional goals—remains an open question. Existing benchmarks only partially address this, often conflating visual quality with cognitive correctness. We introduce VWG-Bench (Video World Generalist Benchmark), a comprehensive benchmark spanning 9 reasoning dimensions and 38 fine-grained tasks. To enable precise diagnosis, we design a three-level VLM-as-Judge protocol that independently assesses video-level fluency, task-level rule adherence, and samplelevel goal realization. Evaluations of leading models reveal a striking gap: while models achieve strong rendering scores, they consistently fail on logic-heavy and rule-constrained tasks. To address this, we propose Vid-PRE (Video Prompt Reasoner and Enhancer), a model-agnostic prompt rewriter that offloads the cognitive burden of reasoning to a dedicated VLM. Trained via reinforcement learning with purely textbased rewards, Vid-PRE produces concise, constraint-aware prompts without the instability of video-level reward signals. Experiments show that Vid-PRE yields substantial reasoning improvements across multiple generators without architectural modifications. Together, VWG-Bench and Vid-PRE offer a rigorous diagnostic lens and a scalable path toward true think-with-video capabilities. All data and code are publicly available at https://huggingface.co/datasets/KlingTeam/VWG-Bench.

Generalized Category Discovery aims to recognize knowncategories while identifying novel ones within unlabeled data. Existingmethods, typically based on self-supervision and contrastive learning,often struggle to capture fine-grained distinctions, relying on superficialvisual cues rather than the intrinsic attributes humans use for categoriza-tion. We introduce Fourier Self-Supervision, that leverages the Fouriertransform of images to enhance the discrimination of subtle differencesand support the discovery of new categories. Our method employs adual frequency filtering strategy: a low-pass filter first extracts broad,abstract attributes that capture high-level category information, whilea high-pass filter emphasizes fine details such as edges and texturesthat are essential for fine-grained recognition. Each operates on a dedi-cated latent space, and their overlapping representations together yielda richer, more complete feature space. This dual-frequency approachnot only refines feature extraction to identify novel categories, but alsostrengthens the model’s discriminative power in fine-grained categorydiscovery. Experiments on multiple fine-grained datasets show that incor-porating Fourier Self-Supervision outperforms state-of-the-art methods,even when the number of classes is unknown, demonstrating its effec-tiveness for Generalized Category Discovery. Our code is available at:https://github.com/SarahRastegar/FourEx.

Vision-and-Language Navigation in Continuous Environments(VLN-CE) requires agents to follow natural-language instructions in un-seen scenes. While Large Models (LMs) have advanced VLN-CE, theirperformance remains severely degraded by real-world visual corruptions,a critical yet underexplored domain constraint. We introduce Tempo-ral Conditional Flow Decorruptor (FlowDec), a novel image restorationframework tailored for LM-based VLN-CE. FlowDec integrates a hybridtemporal conditioning strategy to align the generative flow path withhistorical context and employs action-centroid guided filtering to dy-namically assess and integrate outputs. Extensive experiments demon-strate that FlowDec outperforms state-of-the-art decorruption methodsin both navigation accuracy and generation latency. Our approach es-tablishes a robust, efficient paradigm for resilient embodied navigationin unpredictable real-world conditions.

Dataset distillation (DD) aims to create a smaller datasetthat encapsulates the essential knowledge of a larger dataset, therebyreducing storage demands and accelerating downstream training. Forlarge-scale dataset distillation, state-of-the-art methods achieve satisfac-tory performance by using soft labels generated by well-trained teachermodels during downstream training. However, it will cause some issues:(1) a substantial amount of additional storage is required to retain theteacher models, often significantly exceeding the storage needed for thesynthetic images; (2) generating labels through these teacher modelsslows down the downstream training process, counteracting the efficiencygoals of dataset distillation; and (3) downstream training guided by theseteacher models, according to our studies, yields suboptimal performance.Focusing on these drawbacks, in this paper, we propose plug-and-playparameter-efficient label generation techniques for dataset distillation,which maximizes the benefits of limited model parameters and can begeneralized to different DD methods, datasets, and settings. Specifically,we propose a DD-oriented model parameter reduction method that au-tomatically determines the optimal capacity of teacher models and elim-inates redundant parameters for dataset distillation tasks. Furthermore,for additional parameter space, we turn to model ensemble strategiesand propose guidelines to optimize the utilization efficiency of the ad-ditional space. Compared to the state-of-the-art methods, Flash-DD re-quires only 0.03% of the additional storage and significantly acceleratesdownstream label generation by 843.81× while maintaining compara-ble performance. Alternatively, with a mere 1.8% storage budget, it canboost accuracy by up to 13.4% over previous leading methods.

In diffusion and flow-matching generative models, guidance techniques are widely used to improve sample quality and consistency. Classifier-free guidance (CFG) is the de facto choice in modern systems and achieves this by contrasting conditional and unconditional samples. Recent work explores contrasting negative samples at inference using a weaker model, via strong/weak model pairs, attention-based masking, stochastic block dropping, or perturbations to the self-attention energy landscape. While these strategies refine the generation quality, they still lack a reliable control over the granularity or difficulty of the negative samples, and target-layer selection is often fixed. We propose Exponential Moving Average Guidance (EMAG), a training-free mechanism that modifies attention at inference time in diffusion transformers, with a statistics-based, adaptive layer-selection rule. Unlike prior methods, EMAG produces harder, semantically faithful negatives (finegrained degradations), surfacing difficult failure modes, enabling the denoiser to refine subtle artifacts, boosting the quality and human preference score (HPS) by +0.54 over CFG. We further demonstrate that EMAG naturally composes with advanced orthogonal guidance techniques, such as APG and CADS, further improving HPS. Code is available at https://github.com/drkkgy/EMAG.

Geometry transformers such as VGGT achieve strong per-formance by jointly reasoning over multiple views with global attention.However, scaling them to large view collections remains challenging dueto the quadratic cost of attention. Moreover, our empirical analysis re-veals that the reconstruction quality in VGGT is sensitive to the distri-bution of viewpoints. Simply increasing the number of views without suf-ficient viewpoint diversity can even degrade performance, as redundantviews introduce highly similar tokens that dilute informative geometricsignals in the attention mechanism. Motivated by this observation, wepropose a training-free and plug-and-play VGGT inference frameworkthat organizes views into diversity-aware balanced chunks. The chunksare constructed through combinatorial graph partitioning over visualdissimilarity and spatial dispersion. This view organization allows thetransformer to focus attention on geometrically informative views whilereducing redundant attention interactions. To estimate spatial dispersionwithout full pose estimation, we approximate spatial relationships via asoft pose propagation strategy based on visual similarity from a smallset of seed frames. Extensive experiments demonstrate improved perfor-mance in camera pose estimation, multi-view depth prediction, and 3Dreconstruction while reducing memory usage and inference latency. Ourframework also complements existing VGGT variants, enabling scalablemulti-view reconstruction without sacrificing geometric fidelity.

Recent advances in generative medical models are often con-strained by modality-specific scenarios that hinder the integration ofcomplementary evidence, such as imaging, pathology, and clinical notes.This fragmentation limits their development to true foundation modelsthat empower medical AI agents to learn from and predict across thefull spectrum of biomedical knowledge. To address these challenges, wepropose MeDiM, the first medical discrete diffusion model that learnsshared distributions across different medical modalities without requiringmodality-specific components. MeDiM unifies multiple generative tasks:it flexibly translates between images and text or jointly produces im-age–report pairs across domains in response to user prompts. It builds ona discrete diffusion framework that unifies vision and language represen-tations by modeling their shared probabilistic distribution. To empowerthe diffusion process to support unified and versatile medical generation,we employ a multimodal large language model (MLLM) as the diffusionbackbone, leveraging its rich prior knowledge and cross-modal reason-ing abilities. Because MLLMs are trained with causal (autoregressive)masking while diffusion denoising benefits from bidirectional context,MeDiM introduces two adaptation designs: 1) removing the causal at-tention mask to enable a fully bidirectional information flow essentialfor mutual alignment, and 2) injecting continuous timestep embeddingsto make the MLLM aware of the diffusion steps. Extensive experimentsvalidate MeDiM as a unified foundation model capable of high-fidelitymedical generation across various modalities, including medical imagegeneration (16.60 FID on MIMIC-CXR; 24.19 FID on PathGen) and re-port generation (0.2650 METEOR on MIMIC-CXR; 0.2580 METEOR onPathGen). In addition, the jointly generated medical image-report pairsimprove the downstream task performance (+6.43% BLEU-1, +18.57%BLEU-2, +31.58% BLEU-3, and +4.80% METEOR in PathGen), en-abling the use of multimodal inputs and the production of coherent,clinically grounded outputs.

Class-conditional generative models have emerged as accu-rate and robust classifiers, with diffusion models demonstrating clear ad-vantages over other visual generative paradigms, including autoregressive(AR) models. In this work, we revisit visual AR-based generative clas-sifiers and identify an important limitation of prior approaches: theirreliance on a fixed token order, which imposes a restrictive inductivebias for image understanding. We observe that single-order predictionsrely more on partial discriminative cues, while averaging over multipletoken orders provides a more comprehensive signal. Based on this insight,we leverage recent any-order AR models to estimate order-marginalizedpredictions, unlocking the high classification potential of AR models.Our approach consistently outperforms diffusion-based classifiers acrossdiverse image classification benchmarks, while being up to 25× more ef-ficient. Compared to state-of-the-art self-supervised discriminative mod-els, our method delivers competitive classification performance – a no-table achievement for generative classifiers. The code and models areavailable at: https://github.com/yandex-research/ar-classifier.

Single-image 3D human reconstruction often suffers fromover-smoothed textures and geometric inconsistencies. While diffusionmodels improve generative quality, their reliance on multi-view synthe-sis prior to 3D reconstruction is computationally expensive and proneto view inconsistency. We propose DiGS-Avatar, which reformulates thistask as an efficient, diffusion-based UV-latent completion task, ensur-ing 3D consistency by design. To capture accurate spatial structure, weintroduce a teacher-student framework where a multi-view teacher pro-vides geometrically aligned pseudo-ground-truth latents to supervise asingle-view diffusion student. Treating this inferred latent as a robuststructural skeleton, our method injects high-level semantic features to ac-curately recover fine textural details without disrupting spatial integrity.The refined representation is then decoded into 3D Gaussian primitives.Extensive experiments demonstrate that DiGS-Avatar achieves state-of-the-art or highly competitive visual fidelity and zero-shot generalization,while reconstructing a fully animatable 3D avatar in just 0.71 seconds.Code is available at https://github.com/KLMAV-CUC/DiGS-Avatar.

Astronomical imaging remains noise-limited under practicalobserving conditions. Standard calibration pipelines remove structuredartifacts but largely leave stochastic noise unresolved. Although learning-based denoising has shown strong potential, progress is constrained byscarce paired training data and the requirement for physically inter-pretable models in scientific workflows. We propose a physics-based noisesynthesis framework tailored to CCD noise formation in the telescope.The pipeline models photon shot noise, photo-response non-uniformity,dark-current noise, readout effects, and localized outliers arising fromcosmic-ray hits and hot pixels. To obtain low-noise inputs for synthesis,we stack multiple unregistered exposures to produce high-SNR bases. Re-alistic noisy counterparts synthesized from these bases using our noisemodel enable the construction of abundant paired datasets for supervisedlearning. Extensive experiments on our real-world multi-band datasetcurated from two ground-based telescopes demonstrate the effectivenessof our framework in both photometric and scientific accuracy. Code isavailable at https://github.com/ShuhongLL/Denoising-Deep-Sky.

Robot navigation typically assumes an obstacle-free path exists between start and goal. In real environments, however, clutter may block all routes. We introduce Lifelong Interactive Navigation, where a mobile robot with manipulation capabilities must move objects to forge paths and complete sequential object-placement tasks. Because environment modifications persist, decisions impact future navigability and task difficulty. We propose CoReLIN, an LLM-driven constraintbased reasoning framework with active perception. CoReLIN reasons over a structured scene graph to decide which objects to relocate, where to place them, and where to explore next. A standard motion planner executes reliable navigation and manipulation primitives. To evaluate longhorizon behavior, we introduce 2 new metrics - Long-term Efficiency Score (LES), a unified metric capturing success, execution efficiency, environment optimality, captured by Price of Clutter. In ProcTHOR- 10k, CoReLIN outperforms best baseline by 16% under standard metrics and LES, and transfers to real-world hardware5.

Multimodal embedding spaces in models like CLIP enablepowerful capabilities such as semantic similarity retrieval and cross-modalzero-shot classification. These embeddings compress high-level seman-tics into a single vector, which comes at the cost of primarily express-ing a dominant semantics like main object while suppressing other im-portant attributes such as camera angle or color tone. We propose atext-conditioned transformation of visual embeddings that makes suchattributes explicitly accessible. Given a natural language description ofan attribute category (e.g., “color” or “art style”), a network generates anaffine transformation that emphasizes the specified attribute. Condition-ing on text enables it to learn many attributes simultaneously, accessingthem at inference time through an intuitive interface. The network istrained to align transformed embeddings with the frozen latent space,enabling retrieval using existing large-scale embeddings without any re-encoding. When applied to a full set, the same mechanism transforms thelatent space for attribute disentanglement tasks such as multi-clustering.By operating directly in latent space, our method provides a unifiedand efficient framework for controlling embedding spaces, demonstratingstate-of-the-art performance across both attribute-based retrieval andmulti-attribute organization tasks with near-zero inference cost.

Stereo image compression fundamentally relies on exploitingcross-view redundancies. We observe that the performance of conven-tional correspondence-driven methodologies degrades severely in out-of-view (OOV) regions or when reference features are heavily quantized,resulting in blurred textures and prediction failure. In this paper, we pro-pose a novel generative stereo image compression framework driven bya controllable generative reference. Instead of relying on finding explicitcorrespondences from a degraded reference, we leverage the powerful pri-ors of a pre-trained diffusion model to synthesize a high-quality target-view reference. To bridge the gap between stochastic generative processesand strict stereo consistency, we introduce Joint Semantic-Spatial Con-trol for reference generation. This mechanism synergistically steers thesynthesis using a robust geometric anchor from the compressed left viewand an ultra-compact semantic condition extracted from the target view.Furthermore, to minimize the impact of structural inconsistency duringreference generation on the quality of the reconstructed image, we de-sign a Reliability-Aware Gating module. By dynamically evaluating thesynthesized prior in the feature space, this module adaptively regulatesgenerative information flow within the entropy model to improve thereconstructed image quality. Extensive experiments on standard bench-marks demonstrate that our framework achieves state-of-the-art rate-distortion performance, delivering remarkable bitrate savings and highvisual quality at low bitrates.

Semantic navigation is a fundamental task for embodiedagents operating in unseen environments, requiring both semantic un-derstanding and long-term decision-making. Recent foundation modelshave empowered agents with rich semantic priors for this task. How-ever, without structured global representations, decision-making oftenfalls back on local observations and greedy strategies, resulting in in-efficient exploration and myopic behaviors, especially in long-distancenavigation. To address these challenges, we propose a zero-shot semanticnavigation framework. Our method incrementally maintains an onlineHierarchical 3D Scene Graph (HSG) to form a multi-granular semantictopology over objects, zones, and regions, serving as a compact stateabstraction for global planning. Building on this memory, we introducea hierarchical belief-based planning framework that fuses semantic pri-ors with exploration evidence on the HSG, and performs finite-horizonrollouts on an HSG-based simulator to explicitly estimate the long-termexpected returns of candidate macro-actions. This enables globally con-sistent decisions and reduces redundant backtracking. Extensive experi-ments in high-fidelity simulation environments across multiple tasks anddatasets demonstrate that our method outperforms existing state-of-the-art methods, particularly in long-distance scenarios, where our approachimproves SR and SPL by an average of 9.4% and 5.0%, respectively.

Universal medical image segmentation seeks to use a singlefoundational model to handle diverse tasks across multiple imaging modal-ities. However, existing approaches often rely heavily on manual visualprompts or retrieved reference images, which limits their automationand robustness. In addition, naive joint training across modalities oftenfails to address large domain shifts. To address these limitations, wepropose Concept-to-Pixel (C2P), a novel prompt-free universal seg-mentation framework. C2P explicitly separates anatomical knowledge intotwo components: Geometric and Semantic representations. It leveragesMultimodal Large Language Models (MLLMs) to distill abstract, high-level medical concepts into learnable Semantic Tokens and introducesexplicitly supervised Geometric Tokens to enforce universal physical andstructural constraints. These disentangled tokens interact deeply withimage features to generate input-specific dynamic kernels for precisemask prediction. Furthermore, we introduce a Geometry-Aware InferenceConsensus mechanism, which utilizes the model’s predicted geometricconstraints to assess prediction reliability and suppress outliers. Extensiveexperiments and analysis on a unified benchmark comprising eight diversedatasets across seven modalities demonstrate the significant superiorityof our jointly trained approach, compared to universe- or single-modelapproaches. Remarkably, our unified model demonstrates strong general-ization, achieving impressive results not only on zero-shot tasks involvingunseen cases but also in cross-modal transfers across similar tasks. Codeis available at: https://github.com/Yundi218/Concept-to-Pixel.

Diffusion Transformers (DiTs) have achieved state-of-the-art(SOTA) performance in visual generative modeling, yet their training re-mains computationally prohibitive. While the recently proposed Momen-tum Orthogonalization (Muon) optimizer offers a promising alternativeto AdamW, its direct application to DiTs yields suboptimal late-stageconvergence. In this paper, we identify the root cause of this bottleneck:standard DiT architectures fuse functionally distinct weights (e.g., withinAdaLN and QKV layers) into unified tensors for computational efficiency.Applying Muon to these fused tensors inadvertently induces implicit sub-space coupling, which distorts update directions and degrades global op-timization. To address this, we introduce Chunked Muon (CMuon), asimple yet highly effective strategy that partitions these matrices intoindependent sub-components prior to orthogonalization. Extensive ex-periments demonstrate that a 675M-parameter DiT trained with CMuonachieves a FID of 1.18 on ImageNet 256 in just 200 epochs. This rep-resents more than a 2x training speedup over AdamW, while effectivelyovercoming the late-stage convergence plateaus of vanilla Muon.