ECCV 2026 Papers

Vision–language–action (VLA) policies often become brit-tle under deployment distribution shift when only limited target-domaindemonstrations are available. Subgoal-image planning addresses this is-sue by providing an explicit visual target. However, the target imagespecifies what future scene to reach, not how the robot should transitiontoward it. We propose TDSR-VLA, a VLA framework that reuses condi-tioning states from a diffusion-based Vision Planner to guide action gen-eration. The Vision Planner predicts a subgoal image and provides tworepresentations: a Sequence Representation (SR) grounded in the currentobservation and an Updated Sequence Representation (USR) refined dur-ing denoising steps. A flow-matching Action Expert generates continuousaction chunks by injecting these representations through layer-wise pre-fix key–value conditioning. On real-world tasks, TDSR-VLA achieves anaverage OpenArm score of 0.83, outperforming GR00T N1.5 (0.74), π0(0.71), and SmolVLA (0.58), with the largest gains on object-cooperativetasks (0.88 vs. 0.70). On LIBERO, TDSR-VLA reaches 82.0% on Goaland 56.0% on Long without robot-demonstration pretraining, outper-forming OpenVLA on both suites and π0 on Long. Ablation studies showthat planner-side conditioning states provide action-relevant transitioncues, improving long-horizon manipulation under limited training data.

While traditional image restoration focuses on perceptualquality, Task-Driven Image Restoration (TDIR) aims to maximize theperformance of downstream high-level vision tasks. Recent approachesleveraging generative priors have shown promise for TDIR; however,they typically suffer from computational inefficiency and potential se-mantic alteration by indiscriminately updating all latent tokens. In thispaper, we posit that not all visual information is equally important formachine perception. Through an analysis of the latent token space, weobserve that task-relevant cues are unevenly distributed across the to-ken sequence, exhibiting index-wise specialization. This suggests thatselectively refining a subset of tokens can be sufficient for task-drivenobjectives. Leveraging this insight, we propose TaskTok, a novel frame-work that selectively restores only task-relevant tokens via a learnabletoken switch and a lightweight token refinement module. Extensive ex-periments across image classification, semantic segmentation, and objectdetection demonstrate that TaskTok significantly enhances task perfor-mance with high computational efficiency. The source code is availableat our project page https://github.com/jimmy9704/TaskTok.

3D head stylization enables expressive reimagining of hu-man faces for creative visual experiences in digital media. Existing 3D-aware methods often require computationally intensive optimization orper-style fine-tuning, limiting flexibility and user control. To overcomethese challenges, we introduce StyleFusion360, a diffusion-based frame-work for multi-view consistent, identity-preserving 3D head stylizationfrom a single style reference image, without per-style training. Our ap-proach enhances the Style Fusion Attention mechanism with a style-conditioned key modulation mechanism that aligns content and stylerepresentations for fine-grained and controllable stylization. We furtherprovide a user-controllable slider for adjusting stylization intensity. Inaddition, StyleFusion360 supports local multi-edit stylization, enablingtargeted edits such as modifying hair or eyes independently. Extensiveexperiments on FFHQ and RenderMe360 demonstrate that StyleFu-sion360 produces high-quality, controllable, and visually compelling styl-izations, outperforming state-of-the-art GAN- and diffusion-based meth-ods across diverse style domains. Code is available at: https://github.com/furkanguzelant/StyleFusion360

When labeled data are scarce, off-the-shelf diffusion modelscan augment training sets for few-shot medical image classification, butnot all generated samples are equally useful for the downstream task. Ex-isting approaches largely improve synthetic data by increasing realism,diversity, or domain adaptation, while overlooking a more fundamen-tal question: how should sample usefulness for classification be measuredand optimized? We address this with Class-Contrastive Influence (C2I),a criterion that quantifies a sample’s usefulness through its gradient-based influence on the classifier. We find that effective samples exhibita strong C2I gap: their loss gradients align with validation gradientsfrom the same class and oppose those from other classes. Our anal-ysis further suggests that such high-C2I samples are hard, boundary-proximal examples that help refine the decision boundary and improverobustness. Building on this insight, we fine-tune diffusion models withreinforcement learning using a C2I-based reward to steer generation to-ward class-informative samples. Across several few-shot medical imagingbenchmarks, C2I-guided generation improves downstream accuracy androbustness over diffusion-based augmentation baselines, showing thatsynthetic augmentation is most effective when guided by task usefulnessrather than image quality alone.

Tiny object detection is critical for applications ranging from drone-based scene analysis to remote sensing. However, existing DETR frameworks suffer from inefficient encoder architectures and a severe query-object misalignment for tiny targets, hindering their deployment in real-time scenarios. To overcome these limitations, we propose SlimDETR, a novel framework tailored for real-time tiny object detection. Central to our approach is an Efficient Interaction (EI) encoder, which employs a gather-and-inject mechanism to aggregate global semantics. Uniquely, the EI encoder restricts semantic attention interactions exclusively to extracted foreground tokens, thereby significantly alleviating the computational overhead of conventional Transformer-based encoders. Furthermore, we introduce a Gaussian Target Guided (GTG) query selection strategy that leverages the Gaussian spatial distribution of targets to dynamically concentrate high-quality queries on regions containing tiny objects. Extensive experiments demonstrate that Slim-DETR achieves state-of-the-art performance, yielding 33.9% AP on VisDrone and 33.6% AP on AI-TOD-V2. Notably, it maintains real-time inference speeds across platforms, achieving a latency of 24.4 ms on an NVIDIA RTX 3090 GPU and 25.1 ms (40 FPS) on the embedded Jetson AGX Orin.

Vision-language models (VLMs) can follow complex textualinstructions, yet they struggle to reason from purely visual context. Inparticular, current models fail to infer shared concepts from sets of exam-ple images and apply them to new inputs. We introduce Visual ConceptInference from Sets (VICIS), a task that evaluates this capability. Givena small context set of images sharing a concept and a query image, themodel must generate new images that preserve the context-defined conceptwhile remaining consistent with the query. We show that state-of-the-artVLMs perform poorly on this task, often ignoring the visual contextor defaulting to biased generations. To address this gap, we propose atraining framework and architecture that learn to infer visual conceptsfrom image sets and extract concept-specific embeddings from queries.Experiments on synthetic data and large-scale ImageNet/WordNet datashow that our model generates more accurate and diverse outputs andgeneralizes to unseen concepts and modalities such as sketches.

Adapting large pretrained models to new tasks e!ciently andcontinually is crucial for real-world deployment but remains challengingdue to catastrophic forgetting and the high cost of retraining. Whileparameter-e!cient tuning methods like low-rank adaptation (LoRA) re-duce computational demands, they lack mechanisms for strict contin-ual learning and knowledge integration, without data replay or multi-adapters or MoE. We propose Share, a novel approach to parameter-e!cient continual finetuning that learns and dynamically updates a sin-gle, shared low-rank subspace, enabling seamless adaptation across multi-ple tasks and modalities. Share constructs a foundational subspace thatextracts core knowledge from past tasks and incrementally integratesnew information by identifying essential subspace directions. Knowledgefrom each new task is incorporated into this evolving subspace, enablingforward knowledge transfer while minimizing catastrophic interference.This approach achieves up to 100× reduction in trainable parametersand 281× memory savings over traditional LoRA methods, maintain-ing performance comparable to jointly trained models. A single Sharemodel can replace hundreds of task-specific LoRA adapters, support-ing scalable, asynchronous continual learning. Experiments across imageclassification, natural language understanding, 3D pose estimation, andtext-to-image generation validate its e"ectiveness, making Share a prac-tical and scalable solution for lifelong learning in large-scale AI systems.

While personalized image generation has achieved remark-able progress, multi-reference image generation (MRIG) remains a chal-lenging task. Most existing benchmarks fail to adequately evaluate com-plex MRIG scenarios, hindering further progress in this area. To bet-ter assess model performance on complex MRIG tasks, we introduceOmniRef-Bench, a benchmark that covers complex combinations of ref-erence image types and a large number of reference images. Evaluationson OmniRef-Bench show that mainstream open-source models strugglein complex MRIG scenarios, and their performance deteriorates signifi-cantly as the number of mixed-type reference images increases. To ad-dress this issue, we propose DyRef, a two-stage training framework.In the first stage, supervised fine-tuning equips the model with the ba-sic capability to handle complex MRIG tasks. In the second stage, weintroduce Difficulty-aware Advantage Reweighting (DAR) and Discrim-inative Reward Scaling (DRS). DAR dynamically adjusts the optimiza-tion objective to improve performance when handling a large numberof mixed-type reference images. DRS enlarges intra-group reward differ-ences for more effective policy optimization. Experiments demonstratethat DyRef significantly improves the performance of open-source modelson OmniRef-Bench and single-image editing benchmarks, demonstratingthe effectiveness and generalization capability of our approach. Our codeis available at https://github.com/Weistrass/DyRef.

The deployment of long-form video-language models (Video-VLMs) is fundamentally constrained by the prohibitive scaling of atten-tion complexity and KV-cache overhead as temporal and spatial res-olutions increase. Existing pruning methods rely on local frame-levelheuristics, which fail to capture shifting information density and oftenresult in cumulative budget misallocation over long horizons. To addressthese challenges, we propose S3 -prune, a stability-aware token budgetingframework that jointly models spatio-temporal demand while ensuringconsistent allocation over time. S3 -prune characterizes token demand byintegrating spatial uncertainty (SU) from patch-level embedding shiftsand segment transition (ST) from inter-segment semantic variations. Tomitigate temporal instability, we introduce stability accumulation (SA)via a Kalman-filtered latent estimator, which smoothens stochastic spikesand provides uncertainty-aware margins for adaptive budgeting. Guidedby these stabilized budgets, S3 -prune executes a hierarchical two-stageselection process. Extensive evaluations across seven benchmarks andvarious backbones demonstrate that S3 -prune significantly advances theefficiency-accuracy frontier, reducing inference latency while remainingrobust even under aggressive token retention settings.

Dynamic Hand Gesture Recognition is essential for naturalhuman-computer interaction but faces challenges in continuous local ex-ecution on consumer-grade hardware due to strict computational andlatency constraints. Existing solutions fail to reconcile the high MemoryAccess Cost (MAC) and dynamic footprint of Video Transformers with thelimited temporal modeling capacity of lightweight CNNs. To address theselimitations, we propose the Split-Scan Gesture Network (S2Gest), whichrethinks temporal modeling by leveraging continuous state-space evolutionto represent discrete sequential inputs. We introduce the Split-Scan Block(S2-Block), employing a novel tube-wise scanning strategy that partitionsfeature channels to process opposing temporal directions simultaneously.By structurally partitioning feature channels to capture complementaryforward and backward dynamics, this design achieves global bidirectionalmodeling with linear-time complexity and without introducing additionalparameters, effectively overcoming the causal unidirectionality inherent instandard State Space Models (SSMs). Evaluated across three benchmarks,the scalable S2Gest model family establishes a new state-of-the-art forlightweight architectures, matching the performance of large-scale base-lines with a sub-3M parameter footprint while consistently exceeding 100clips/s on consumer-grade GPUs. Code and weights are publicly availableat https://github.com/Chen-Ke-Fan/S2Gest.

Few-shot classification performance of Contrastive Language-Image Pre-training (CLIP) varies widely across datasets, especially whenclass names provide weak or ambiguous semantic priors. This issue is alle-viated in vision-only fine-grained methods, as they identify discriminativelocal features but require training on the full dataset. We thus introduce atraining-free approach that integrates discriminative patches into CLIPto reduce reliance on textual prompts. Our method identifies patcheswith high intra-class consistency and low inter-class ambiguity, forminga Class-Discriminative Patch Set (CDPS) for each category. Using CDPS,we enhance the recognition ability of CLIP through a hybrid classifiercombining global image-text alignment with local patch-based similarity.Extensive experiments on diverse benchmarks show that CDPS injectsfine-grained discriminatory power into CLIP and yields more robust few-shot recognition performance against ambiguous class names.

Video-based remote physiological measurement (RPM) ishighly accessible but remains fragile under varying illumination, skintones, and motion. Radio frequency (RF) radar is largely invariant to il-lumination and appearance, providing complementary cardio-respiratorymicro-motion cues; however, requiring radar at inference is often imprac-tical due to its limited ubiquity and deployment overhead. We proposeRPM-Distill, a physiology-guided cross-modal distillation frameworkthat leverages synchronized radar only during training while retainingvideo-only inference. Our key observation is that although RGB and RFwaveforms differ in sensing physics and time-domain morphology, theyshare similar latent periodic rhythm in the frequency domain. We thusdistill physiology-structured spectral evidence to improve robustness, vialosses that (i) anchor the fundamental peak, (ii) match the off-peak back-ground distribution, and (iii) preserve spectral morphology and sharp-ness. To avoid negative transfer under sample-level teacher quality andalignment uncertainty, a spectral policy network predicts sample-leveldistillation gates and component weights from the student–teacher spec-tral relation map, learned with a meta bilevel objective on a small labeledvalidation split. Through extensive experiments in challenging conditionsand cross-dataset settings, RPM-Distill reduces MAE by up to 81% andimproves correlation by up to 21% over the best video-only baseline.Code is at https://github.com/WJULYW/RPM-Distill.

Magnetic resonance imaging (MRI) inherently suffers fromnoise, which limits downstream medical analyses. In MRI, noise-freeimages are unobtainable; therefore, existing denoising approaches for-mulate surrogate training objectives, compromising between preservingdetail and concealing noise, causing domain shifts or incomplete de-noising. To enable denoiser training directly on unmodified, noisy im-ages, we exploit repeated acquisitions. This naturally constitutes a phys-ical Noise2Noise (pN2N) setting. For unrepeated data, we introducea diffusion-based re-noiser that synthesizes noisy image pairs, extend-ing pN2N to Renoise2Noise (ReN2N). Furthermore, we demonstratethat ReN2N improves generalization to unseen datasets. Additionally,we propose to combine pN2N or ReN2N with optional guidance fromco-acquired contrast, yielding four versions of our novel denoising frame-work: YADO (You Accurately Denoise real Observations). Across 14 testconditions, YADO consistently outperforms 17 state-of-the-art baselines,matching the quality of physically noise-suppressed images obtained viabrute-force averaging of independent acquisitions. YADO thus estab-lishes practical denoising for real-world acquisition settings.

Existing 3D understanding models are typically built on afixed geometric representation and often fail when the representationformat changes. Unlike domain generalization, which studies distribu-tion shift under a fixed input structure, we focus on representation shift,where identical 3D semantics are expressed through heterogeneous geo-metric encodings such as points, voxels, meshes, and 3D Gaussian prim-itives. As 3D representations continue to evolve beyond a predefinedset, this challenge is becoming increasingly important. We formalize thisproblem as Representation Generalization (RG), a new setting for unified3D learning. RG requires a model to share knowledge across representa-tions during training, remain effective when only a single representationis available at test time, and extend efficiently to unseen or newly emerg-ing representations. To this end, we propose ReGen3D, a unified frame-work with three modules: Unified Representation Tokenization (URT),which maps heterogeneous representations into a shared token space;Shared Memory Interaction (SMI), which enables cross-representationlearning during training while supporting inference from any single avail-able representation; and Efficient Representation Adapter (ERA), whichsupports lightweight adaptation to unseen representations without modi-fying the shared backbone. We further establish RepShift-3D for system-atic evaluation under the RG setting. Extensive experiments on classi-fication, segmentation, and reconstruction show that ReGen3D achievesstrong and robust performance across diverse representation scenarios.

Detailed image captioning demands both factual ground-ing and fine-grained coverage, yet existing methods have struggled toachieve them simultaneously. We address this tension with ReflectiveNote-Guided Captioning (ReflectCAP), where a multi-agent pipeline an-alyzes what the target large vision-language model (LVLM) consistentlyhallucinates and what it systematically overlooks, distilling these pat-terns into reusable guidelines called Structured Reflection Notes. At in-ference time, these notes steer the captioning model along both axes—whatto avoid and what to attend to—yielding detailed captions that jointlyimprove factuality and coverage. Applying this method to 8 LVLMs span-ning the GPT-4.1 family, Qwen series, and InternVL variants, Reflect-CAP reaches the Pareto frontier of the trade-off between factuality andcoverage, and delivers substantial gains on CapArena-Auto, where gen-erated captions are judged head-to-head against strong reference models.Moreover, ReflectCAP offers a more favorable trade-off between captionquality and compute cost than model scaling or existing multi-agentpipelines, which incur 21–36% greater overhead. This makes high-qualitydetailed captioning viable under real-world cost and latency constraints.

Recent advancements in video generation has shifted from bidi-rectional models for short videos to autoregressive ones for ultra long videogeneration. Previous models, which usually use sliding window attention torestrict inference cost, lack effective memory compression and retrieval forlong-term generation beyond the window size, leading to issues of forgettingand spatiotemporal inconsistencies. To enhance the retention of historicalinformation with a fixed memory budget, we additionally incorporate tem-poral recurrent neural network (RNN) layers into the diffusion transformer(DiT) model. Specifically, we found that a LSTM layer after attention at eachDiT layer achieves comparable performance to other state-of-the-art RNNblocks, such as Test-Time Training (TTT) and Mamba2. Moreover, existingdiffusion-RNN approaches often suffer from performance degradation dueto training-inference gap or the lack of overlap across windows. To addressthese limitations, we propose a novel Recurrent Autoregressive Diffusion(RAD) framework, which leverages recurrent blocks for memory update andretrieval and preserves local details by full attention on overlapping slidingwindows, with no training and inference gap. Experiments on Memory Mazeand Minecraft datasets demonstrate the superiority for long video generationby our framework with global memory and local attention.Project page: https://yeyutaihan.github.io/recurrent-autoregressive-diffusion/

World models aim to capture environmental dynamics andpredict future trajectories, showing growing potential for embodied in-telligence. Physics-informed 4D generation integrates physical simulationto predict 3D object interactions, offering a promising pathway towardworld models. However, this paradigm relies on manually imposed dy-namical assumptions rather than internalizing world dynamics, and thusstill leaves a gap toward a true world model. To bridge this gap, we pro-pose NeuIDO, a novel world dynamics modeling framework that learnsa unified intrinsic dynamics representation from visual observations, ad-vancing physics-informed 4D generation toward a world model. Specifi-cally, we formulate world modeling as a neural operator learning problemand introduce a two-stage training strategy to learn a generalizable map-ping from the visual observation distribution to the intrinsic dynamicsdistribution. Building on this observation-dynamics mapping, NeuIDOenables zero-shot dynamics inference directly from videos and can be fur-ther aligned with complex real-world dynamics via few-shot adaptation.Extensive experiments demonstrate that NeuIDO effectively unifies theintrinsic dynamics underlying diverse visual observations into a sharedrepresentation and rapidly infers dynamics in novel scenes.

Deep neural networks often produce poorly calibrated confi-dence estimates, overstating their certainty even when predictions are in-correct. Temperature Scaling (TS) remains the most widely used post-hoccalibration method due to its simplicity and e!ectiveness, yet its global,uniform rescaling of logits fails to correct the highly heterogeneous struc-ture of miscalibration observed across the confidence spectrum. In par-ticular, the largest correctness–confidence discrepancies arise in di!er-ent quantile regions depending on the setting, which standard TS leaveslargely unaddressed. We introduce Quantile-Adaptive Temperature Scal-ing (QaTS), a simple and e"cient post-hoc calibration method thatadapts the temperature as a function of a prediction’s empirical con-fidence quantile. By mapping confidences into the quantile space, QaTSnormalizes the calibration problem, makes the structure of miscalibrationexplicit, and enables a monotone temperature function that adapts acrossquantiles while leaving well-calibrated high-confidence predictions largelyunchanged. This quantile-aware formulation aligns naturally with a repa-rameterized Expected Calibration Error (ECE) objective and yields asample-wise temperature that is robust across a variety of challengingscenarios, such as class imbalance and distributional shifts. Across abroad range of datasets, architectures, evaluation scenarios and diversetasks, QaTS consistently, and substantially, outperforms state-of-the-artpost-hoc calibration methods, delivering more reliable and trustworthyconfidence estimates without modifying model predictions.

Slot Attention, an approach that binds different objects in ascene to a set of "slots", has become a leading method in unsupervisedobject-centric learning. Most methods assume a fixed slot count K, andto better accommodate the dynamic nature of object cardinality, a fewworks have explored K-adaptive variants. However, existing K-adaptivemethods still suffer from two limitations. First, they do not explicitlyconstrain slot-binding quality, so low-quality slots lead to ambiguousfeature attribution. Second, adding a slot-count penalty to the recon-struction objective creates conflicting optimization goals between reduc-ing the number of active slots and maintaining reconstruction fidelity.As a result, they still lag significantly behind strong K-fixed baselines.To address these challenges, we propose Quality-Guided K-AdaptiveSlot Attention (QASA). First, we decouple slot selection from recon-struction, eliminating the mutual constraints between the two objectives.Then, we propose an unsupervised Slot-Quality metric to assess per-slotquality, providing a principled signal for fine-grained slot–object binding.Based on this metric, we design a Quality-Guided Slot Selection schemethat dynamically selects a subset of high-quality slots and feeds them intoour newly designed gated decoder for reconstruction during training. Atinference, token-wise competition on slot attention yields a K-adaptiveoutcome. We conduct experiments on both object discovery and objectproperty prediction. Results show that QASA substantially outperformsexisting K-adaptive methods on both real and synthetic datasets. More-over, on real-world datasets, QASA surpasses K-fixed methods. The codeis available at https://github.com/ouyangtianran/QASA-tianran.Fig. 1: mBOi vs. slot count K on COCO [25]. K-fixed methods, such as SPOT [20]and DINOSAUR [33], show large performance fluctuations as K varies. By contrast,K-adaptive methods do not rely on a carefully tuned K. Our method substantiallyoutperforms existing K-adaptive baselines, including MetaSlot [26] and AdaSlot [16],and even surpasses strong K-fixed baselines.