ECCV 2026 论文列表

Di(cid:27)usion Transformers (DiTs) achieve high-quality generation but are costly due to iterative sampling. Dynamic-resolution sampling reduces early-stage cost by denoising at low resolution; however, uniformly upsampling all latent tokens at resolution transitions incurs redundant computation and may degrade (cid:28)ne-detail consistency. Existing partial upsampling strategies typically rely on local latent structure cues or single-step statistics, making it di(cid:30)cult to jointly capture token(cid:21)text semantic relevance and token-wise representation dynamics across diffusion steps. We propose AViTS, an adaptive spatiotemporal token selection framework for dynamic-resolution DiTs. AViTS models spatial importance via latent(cid:21)text attention and temporal importance via token-level feature variation across di(cid:27)usion timesteps, and fuses them to enable spatiotemporal importance-aware selective upsampling: it prioritizes resolution re(cid:28)nement for critical tokens while deferring less important ones, thereby reducing redundant high-resolution computation and improving the quality(cid:21)e(cid:30)ciency trade-o(cid:27). AViTS achieves up to 6.34× on FLUX and nearly 9× FLOPs reduction on Qwen-ImageEdit and FLUX.1-Kontext-dev, orthogonal to distillation, quantization, and feature caching, and reaching 14.76× with distilled models. Code: https://github.com/QHR69/AViTS.

Human decision-making interprets the world through highlevel concepts, such as recognizing a bird by its belly color. To bridge the gap between opaque deep learning representations and human understanding, Post-Hoc Concept Bottleneck Models (post-hoc CBMs) project latent features onto interpretable concept spaces using auxiliary datasets or vision-language models. However, relying on target task accuracy as the primary measure of post-hoc CBM success obscures whether the learned concepts are semantically meaningful or merely predictive artifacts. For example, random concept projections can achieve competitive accuracy despite being semantically meaningless. In this work, we analyze the learned projections directly and identify two failure cases: First, for concept projections learned from auxiliary data, covariate shifts can lead to unfaithful concept representations for the target task. In particular, we provide an upper bound on the error introduced by this shift. Second, systematic label noise in surrogate concept labels generated by visionlanguage models leads to unfaithful projections. After formalizing these failure modes, we introduce novel metrics that decouple concept faithfulness from predictive accuracy. Our empirical results across real-world and synthetic benchmarks confirm that these metrics identify unfaithful behaviors that standard accuracy-based evaluation fails to detect4.

Implicit Neural Representations (INRs) achieve state-of-theart results across image representation, 3D modeling, and novel-view synthesis, yet their performance hinges on tightly coupled design decisions— activation family, initialization scheme, and per-layer hyperparameters— whose interdependencies render manual tuning unreliable and greedy search provably suboptimal. We introduce OptiINR, the first framework that recasts INR configuration as a global optimization problem over a mixed-variable space of discrete activation choices (e.g., SIREN, WIRE, FINER) and their continuous parameters. OptiINR consists of a structured layer-wise INR configuration space and an adaptive Bayesian optimization procedure for navigating it. The search space spans activation families, activation-specific continuous parameters, initialization/scaling choices, per-layer learning rates, and positional-encoding usage. Across five tasks—1D audio, 2D image, 3D occupancy, spatiotemporal PDE, and neural radiance field synthesis—OptiINR consistently outperforms handtuned baselines and prior automated methods, revealing task-dependent activation patterns that challenge prevailing one-size-fits-all design conventions. Code is available at https://github.com/Sippengg/OptiINR.

Scientific research follows multi-turn, multi-step workflows that require proactively searching the literature, consulting figures and tables, and integrating evidence across papers to align experimental settings and support reproducible conclusions. This joint capability is not systematically assessed in existing benchmarks, which largely underevaluate proactive search, multi-evidence integration and sustained evidence use over time. In this work, we introduce EpiBench, an episodic multi-turn multimodal benchmark that instantiates short research workflows. Given a research task, agents must navigate across papers over multiple turns, align evidence from figures and tables, and use the accumulated evidence in the memory to answer objective questions that require cross-paper comparisons and multi-figure integration. EpiBench introduces a process-level evaluation framework for fine-grained testing and diagnosis of research agents. Our experiments show that even the leading model achieves an accuracy of only 29.23% on the hard split, indicating substantial room for improvement in multi-turn, multievidence research workflows. Code and data are available at https: //github.com/RetroDnix/EPIBench.

Realistic virtual try-on (VTON) concerns not only faithful rendering of garment details but also coordination of the style. Prior art typically pursues the former, but neglects a key factor that shapes the holistic style—garment fit. Garment fit delineates how a garment aligns with the body of a wearer and is a fundamental element in fashion design. In this work, we introduce fit-aware VTON and present FitControler, a learnable plug-in that can seamlessly integrate into modern VTON models to enable customized fit control. To achieve this, we highlight two challenges: i) how to delineate layouts of different fits and ii) how to render the garment that matches the layout. FitControler first features a fitaware layout generator to redraw the body-garment layout conditioned on a set of delicately processed garment-agnostic representations, and a multi-scale fit injector is then used to deliver layout cues to enable layoutdriven VTON. In particular, we build a fit-aware VTON dataset termed Fit4Men, including 13,000 body-garment pairs of different fits, covering both tops and bottoms, and featuring varying camera distances and body poses. Two fit consistency metrics are also introduced to assess the fitness of generations. Extensive experiments show that FitControler can work with various VTON models and achieve accurate fit control. Code and data are released at https://github.com/tiny-smart/FitControler.

Aerial-Ground Re-Identi(cid:28)cation (AG-ReID) is constrained by the viewpoint-domain gap, as drastic viewpoint disparities occlude or distort discriminative features, making cross-viewpoint image retrieval challenging. While existing methods rely on paired cross-view annotations, real-world deployments, such as wilderness search-and-rescue (SAR), often lack target-domain data, requiring retrieval from groundlevel references alone. To our knowledge, we are the (cid:28)rst to address this challenge by formalizing the Single-View AG-ReID (SV AGReID) setting, where models trained on a single real viewpoint must generalize to an unseen viewpoint. We propose 3D Lifting-based Elevated Novel-view Synthesis (3D-LENS), a uni(cid:28)ed framework combining geometrically-consistent novel view synthesis that leverages large-scale 3D mesh reconstruction, with a robust representation learning scheme to mitigate synthetic-to-real bias. Unlike 2D generative baselines that su(cid:27)er from geometric inconsistencies or prior 3D methods that are restricted to class-speci(cid:28)c templates, our approach ensures view-consistent synthesis across diverse categories without prede- (cid:28)ned templates that fail to capture (cid:28)ne-grained details, such as carried objects. Extensive experiments demonstrate that our method achieves state-of-the-art performance on SV AG-ReID scenarios. Code and data will be released at https://github.com/TurtleSmoke/3D-LENS. • •

Red-teaming Text-to-Image (T2I) models is essential for safedeployment, yet it remains particularly challenging against implicit ad-versarial prompts. Unlike explicit adversarial prompts that can be read-ily identified and blocked, implicit ones are much harder to detect: theprompts appear benign on the text surface yet still lead to inappropri-ate visual content. To address this, we propose Adversarial Probing forImplicit VulnErabilities (AdvPIE), a multimodal agentic framework toexpose implicit vulnerabilities without requiring access to the parametersof target models. AdvPIE adopts a policy agent to generate and refineimplicit adversarial prompts based on the feedback from a judge agent.To construct informative feedback, the judge agent provides modality-specific safety evaluation at both global and relative levels across itera-tions. To effectively leverage the feedback, we propose a novel CumulativeAdversarial Decoding strategy for the policy agent, which dynamicallyreweights token distributions to favor tokens that lead to more harmfulimages while preserving sampling diversity. Extensive experiments onstandard and safety-aligned T2I models show that AdvPIE1 effectivelyuncovers implicit vulnerabilities, outperforming various baseline meth-ods. Warning: This paper includes content that may be disturbing.

3D point-cloud observations are inherently ambiguous in complex, cluttered manipulation scenes, where target objects may be partially occluded or tightly intermingled with visually similar distractors. As a result, standard 3D diffusion policies often struggle to localize and exploit task-relevant geometry as scene complexity grows. We propose Attention-DP3, a spatially object-aware 3D diffusion policy that injects object-level geometric cues via attention while keeping the DP3 diffusion backbone unchanged. Our pipeline performs open-vocabulary 2D segmentation on RGB images, then lifts predicted target masks into 3D using calibrated camera geometry to obtain object-centric geometric priors. We incorporate these cues through Tri-field Attentional Conditioning, which constructs three complementary fields: (i) a targetness field to anchor the target object, (ii) an intra-target saliency field to emphasize task-relevant geometry within the target, and (iii) a backgroundness field to suppress distractors and clutter. Experiments on Adroit, DexArt, MetaWorld, and the real-world SO101 platform show consistent improvements over DP3, achieving state-of-the-art performance across benchmarks. Notably, as distractor objects increase, DP3 drops sharply, whereas Attention-DP3 remains stable and outperforms DP3 by up to 31% under heavy clutter. The code is publicly available at https:// github.com/zhangzhongbo2213/Attention-DP3.

Self-supervised learning from large-scale video data has emerged as a dominant paradigm for visual representation learning. Since audio and visual streams naturally co-occur in video data, extending this success to jointly learn from both modalities is a natural next step, yet it remains challenging. Existing audio-visual self-supervised methods rely on modality-specific encoders and complex combinations of contrastive or reconstruction objectives, limiting cross-modal synergy and scalability. Joint Embedding Predictive Architectures (JEPAs) offer a simple, modality-agnostic alternative, but have to date been applied primarily to individual modalities. We introduce MJEPA, a jointembedding predictive architecture for audio-visual learning that uses a single, unified encoder for both modalities. Our approach uses only a single predictive objective, applied both within and across modalities. We show that cross-modal prediction is critical: without it, a shared encoder degrades below unimodal baselines; with it, each modality’s representation benefits from the other. Our frozen ViT-g model outperforms the best prior frozen baseline by over 6.8 mAP on AudioSet-20K, surpasses fully finetuned models on ESC-50 and FSD50K, and is competitive on video benchmarks despite using 10x less video data.

LiDAR point clouds are fundamental to various applications, yet the extreme sparsity of high-precision geometric details hinders efficient context modeling, thereby limiting the compression speed and performance of existing methods. To address this challenge, we propose a compact representation for efficient predictive lossless coding. Our framework comprises two lightweight modules. First, the Geometry Re-Densification Module iteratively densifies encoded sparse geometry, extracts features at a dense scale, and then sparsifies the features for predictive coding. This module avoids costly computation on highly sparse details while maintaining a lightweight prediction head. Second, the Cross-scale Feature Propagation Module leverages occupancy cues from multiple resolution levels to guide hierarchical feature propagation, enabling information sharing across scales and reducing redundant feature extraction. Additionally, we introduce an integer-only inference pipeline to enable bit-exact cross-platform consistency, which avoids the entropy-coding collapse observed in existing neural compression methods and further accelerates coding. Experiments demonstrate competitive compression performance at real-time speed. Code is available at https://github.com/pengpeng-yu/FastPCC.

Text-to-image diffusion models can emit copyrighted, unsafe, or private content. Safety alignment aims to suppress specific concepts, yet evaluations seldom test whether safety persists under benign downstream fine-tuning routinely applied after deployment (e.g., LoRA personalization, style/domain adapters). We study the stability of current safety methods under benign fine-tuning and observe frequent breakdowns. As true safety alignment must withstand even benign post-deployment adaptations, we introduce the SPQR benchmark (Safety–Prompt adherence–Quality–Robustness). SPQR is a single-scored metric that provides a unified, reproducible framework to evaluate how well safety-aligned diffusion models preserve safety, utility, and robustness under benign fine-tuning, by reporting a single leaderboard score to facilitate comparisons. We conduct multilingual, domain-specific, and out-of-distribution analyses, along with category-wise breakdowns, to identify when safety alignment fails after benign fine-tuning, ultimately showcasing SPQR as a concise yet comprehensive benchmark for T2I safety alignment techniques for T2I models. Warning: This paper features illustrative examples that may involve explicit, sexual, or violent imagery and language, which could be sensitive for certain readers.

Vision Transformers (ViTs) and their hierarchical variants have achieved strong performance in Computational Pathology (CPath). However, most are pre-trained on single-resolution Whole Slide Images (WSIs), limiting their generalization across arbitrary resolutions. Gigapixel WSIs inherently contain diagnostic patterns at multiple scales, including cellular morphologies, tissue architectures, and global context, mirroring how expert pathologists examine WSIs. We introduce MultiResolution Pyramid Transformer (MRPT), a model that hierarchically aggregates multi-resolution information from cellular to tissue and WSI levels. MRPT employs a biologically meaningful Consecutive Cross-Resolution Attention (CCRA) mechanism to capture scale-independent interactions and enforces multi-resolution semantic consistency by aligning embeddings across resolutions, yielding robust and generalizable WSI representations. Pre-trained in a multi-resolution self-supervised manner on 624M patches, 2.4M regions, and 36K WSIs, MRPT learns rich coarse-to-fine histopathology features. Extensive experiments on 34 diverse datasets show that MRPT surpasses recent foundation models and Multimodal Large Language Models (MLLMs) in cancer subtype classification, tissue phenotyping, and Visual Question Answering (VQA) for WSI understanding. Our code and models available on link.

Capturing spatial relationships from visual inputs is a cornerstone of human-like general intelligence, yet current Vision-Language Models (VLMs) still struggle to perform spatial understanding from everyday visual observations. To enhance the spatial ability within general architectures, we introduce Visual Spatial Tuning (VST), a comprehensive framework to cultivate VLMs with human-like visuospatial competence spanning spatial perception and reasoning. VST dataset comprises VST-P, a large-scale perception dataset with 4.1 million samples spanning 19 skills across single views, multiple images, and videos; and VSTR, a reasoning dataset with 135K samples that explicitly instruct models to reason in space. For spatial reasoning, we design Reconstruction with Text in Chain-of-Thought (RT-CoT), which represents scene layout as a textual intermediate and performs reasoning grounded on this reconstruction. We further propose a strategy termed prompting with BEV annotations to improve the faithfulness and coherence of spatial reasoning traces. Then, we adopt a progressive training pipeline: supervised fine-tuning to build foundational spatial perception abilities, followed by reinforcement learning to develop spatial reasoning, where a 3D-aware layout reward is proposed. Without the side-effect on general capabilities, the proposed VST consistently achieves state-of-the-art results on several spatial benchmarks, including 34.8% on MMSI-Bench and 61.2% on VSIBench. Moreover, VST substantially improves Vision-LanguageAction (VLA) models, paving the way for more physically grounded AI.

Adversarial training is the most effective defense against adversarial attacks, but it often suffers from robust overfitting, whose causes remain unclear. In this paper, we observe a counterintuitive phenomenon: the model robust stability consistently deteriorates during training. To quantify this behavior, we introduce the Robust Stability Rate (RSR) and define robust instability. Our analysis demonstrates that improving the Robust Stability Rate can effectively reduce robust instability, thereby mitigating robust overfitting. Building on this insight, we propose a novel framework called Relearning Adversarial Training (RAT). Our min-max optimization establishes a new adversarial dynamic where generated adversarial examples induce robust instability, while model training enforces robust stability, aiming to enhance the model Robust Stability Rate of the current learning model by leveraging knowledge from historical model. Experimental results demonstrate that RAT can be integrated with various methods to significantly enhance their performance. It effectively reduces robust instability and mitigates robust overfitting.

Identity preserving face synthesis (IPFS) aims to generate large-scale virtual face datasets to train robust face recognition (FR) models while mitigating privacy and bias issues of web-crawled data. While diffusion-based IPFS methods have made huge progress, FR models trained on synthetic data still underperform those trained on real images. This gap stems from limited style diversity and insufficient intraidentity variation. Through analysis, we find that identity embeddings from real images primarily encode core facial features and inherently contain intra-identity variation. To generate more realistic virtual images, we propose TextFace, a diffusion-based IPFS framework that combines compositional text control for overall style diversity and a principled score-space identity blending strategy to enhance intra-identity variation. The identity-irrelevant condition factors are modeled by compositional language attributes, while identity-inherent factors are diversified by identity blending, which injects small, stable structural variations during denoising without sacrificing identity consistency. Extensive experiments show that FR models trained on TextFace synthetic data consistently outperform prior IPFS pipelines and substantially narrow the gap to real-data training.

Humans achieve stable and adaptive grasps by seamlessly integrating visual perception and tactile feedback, a capability that remains challenging to replicate in robotic systems. Existing robotic grasping approaches predominantly rely on visual inputs and lack mechanisms for tactile-guided adaptation after contact, limiting robustness and generalization. To address this challenge, we propose a uni(cid:28)ed visuo-tactilefusion grasping framework that integrates grasp generation, feasibility prediction, and adaptive re(cid:28)nement. At its core, our method introduces an e(cid:30)cient visuo-tactile representation that tightly fuses object geometry with tactile feedback by associating tactile signals with (cid:28)nger identities. This uni(cid:28)ed representation supports contact-aware grasp pose generation during planning and tactile-guided re(cid:28)nement after contact, enabling the system to reason about (cid:28)ne-grained (cid:28)nger-object interactions and adjust grasps dynamically. Comprehensive experiments in both simulation and real-world environments demonstrate that our approach signi(cid:28)cantly enhances grasp success rates and generalization across diverse objects. • •

Reconstructing physically plausible 3D human-scene interactions (HSI) from a single image currently presents a trade-off: optimization based methods offer accurate contact but are slow (∼20s), while feed-forward approaches are fast yet lack explicit interaction reasoning, producing floating and interpenetration artifacts. Our key insight is that geometry-based human–scene fitting can be amortized into fast feed-forward inference. We present GRAFT (Geometric Refinement And Fitting Transformer), a learned HSI prior that predicts Interaction Gradients: corrective parameter updates that iteratively refine human meshes by reasoning about their 3D relationship to the surrounding scene. GRAFT encodes the interaction state into compact body-anchored tokens, each grounded in the scene geometry via Geometric Probes that capture spatial relationships with nearby surfaces. A lightweight transformer recurrently updates human meshes and re-probes the scene, ensuring the final pose aligns with both learned priors and observed geometry. GRAFT operates either as an end-to-end reconstructor using image features, or with geometry alone as a transferable plug-and-play HSI prior that improves feed-forward methods without retraining. Experiments show GRAFT improves interaction quality by up to 122% over state-of-the-art feed-forward methods and matches optimization-based interaction quality at ∼100× lower runtime, while generalizing seamlessly to in-the-wild multi-person scenes and being preferred in 64.8% of threeway user study. Project page: https://pradyumnaym.github.io/graft.

The core challenge of Aerial-Ground Person Re-Identification (AG-ReID) is the drastic viewpoint discrepancy, which induces a significant cross-view feature gap that hinders aerial-ground matching. However, existing AG-ReID methods overlook this gap and attempt to learn direct global semantic correspondences through auxiliary components (e.g., attributes, tokens, or prompts), which limits training effectiveness and increases computational overhead. To address these issues, we propose a novel AG-ReID framework, MCVL, which effectively bridges the cross-view gap through multi-space cross-view learning. Our method aligns both spatial and global feature spaces, learning view-invariant yet identity-discriminative representations to effectively handle viewpoint discrepancies. Specifically, we introduce a Deformable Homography Transformation (DHT) module that projects features across aerial and ground view planes, creating an intermediate cross-view spatial feature space. Based on this, we propose a Cross-View Learning (CVL) strategy that leverages both original and projected features to reduce the spatial feature gap and enhance identity discrimination while preserving perspective invariance. Additionally, we design a View Decorrelation Contrastive Loss (VDCL) to mitigate view-specific biases by pushing global features away from their corresponding view-specific prototypes while pulling together global features of the same identity, thereby suppressing view-related cues and strengthening identity-related representations. Extensive experiments on five AG-ReID datasets, including CARGO, AG-ReIDv1, AG-ReIDv2, LAGPeR, and G2APS-ReID, demonstrate the effectiveness of our proposed MCVL method.

Pixel-space reasoning enables Vision-Language Models to perform active perception: instead of answering from a single global view, the model can zoom into high-resolution regions to gather finegrained evidence. This mirrors human visual problem solving, which involves broad exploration (finding relevant cues) and local verification (confirming details). However, most pixel-space reasoning systems are trained with Group Relative Policy Optimization (GRPO), whose groupwise advantage normalization directly compares sampled rollouts that follow different perception strategies. Rollouts with different zoom depths and different visual intents (explore vs. verify) are normalized together, making exploratory multi-zoom behavior appear disadvantageous: the multi-zoom strategies newly discovered during training typically yield lower mean rewards than already-mastered single-zoom or no-zoom strategies. This leads to visual laziness: the policy does not explore active multi-zoom perception strategies and collapses into a zoom-averse state. To address this, we propose Visual-Intent Stratified GRPO (VIS-GRPO), a drop-in replacement for GRPO that restores fair learning signals for active perception. VIS-GRPO computes advantages only among strategically comparable rollouts by stratifying trajectories along (i) zoom depth and (ii) visual intent, separating broad search over diverse regions from local verification over overlapping regions. This alignment prevents exploratory trajectories from being overshadowed and encourages the learning of active perception. Extensive experiments demonstrate that VIS-GRPO enables effective multi-zoom active perception strategies and consistently improves performance across challenging visual understanding benchmarks like HR-Bench and MME-RealWorld.