ECCV 2026 论文列表

The rapid advancements of Text-to-Image (T2I) models haveushered in a new phase of AI-generated content, marked by their growingability to interpret and follow user instructions. However, existing T2Imodel evaluation benchmarks fall short in limited prompt diversity andcomplexity, as well as coarse evaluation metrics, making it difficult to eval-uate the fine-grained alignment performance between textual instructionsand generated images. In this paper, we present TIIF-Bench (Text-to-Image Instruction Following Benchmark), aiming to systematicallyassess T2I models’ ability in interpreting and following intricate tex-tual instructions. TIIF-Bench comprises 5,000 prompts organized alongmultiple dimensions and categorized into three levels of difficulty andcomplexity. To rigorously evaluate robustness to prompt length, eachprompt is provided in both short and long versions with identical coresemantics. We further propose a novel Global Normalized Edit Distance(GNED) metric for text rendering and provide aspect-ratio-diverse ref-erence images for each prompt to assess style control. In addition, wecollect 100 high-quality designer-level prompts covering diverse scenariosfor comprehensive evaluation. To enable scalable and fine-grained evalua-tion, we explore the best paradigm for leveraging the world knowledgeencoded in large Vision-Language Models (VLMs) as automated binaryevaluators. Through extensive ablations, we develop a fully reproducibleevaluator that provides interpretable reasoning and reliable verification,enabling our benchmark to discern subtle variations in T2I model out-puts. Through comprehensive benchmarking of mainstream T2I modelson TIIF-Bench, we analyze the strengths and weaknesses of current T2Isystems and reveal the limitations of existing evaluation benchmarks.

Visuotactile sensors are indispensable for contact-rich roboticmanipulation tasks. However, policy learning with tactile feedback insimulation, especially for online reinforcement learning (RL), remains acritical challenge, as it demands a delicate balance between physics fi-delity and computational efficiency. To address this challenge, we presentTac2Real, a lightweight visuotactile simulation framework designed toenable efficient online RL training. Tac2Real integrates the Precondi-tioned Nonlinear Conjugate Gradient Incremental Potential Contact (PNCG-IPC) method with a multi-node, multi-GPU high-throughput par-allel simulation architecture, which can generate marker displacementfields at interactive rates. Meanwhile, we propose a systematic approach,TacAlign, to narrow both structured and stochastic sources of domaingap, ensuring a reliable zero-shot sim-to-real transfer. We further evalu-ate Tac2Real on the contact-rich peg insertion task. The zero-shot trans-fer results achieve a high success rate in the real-world scenario, verifyingthe effectiveness and robustness of our framework. The project page is:https://ningyurichard.github.io/tac2real-project-page/

Continual anomaly detection (CAD) addresses the need forindustrial inspection systems to adapt to evolving production conditions,yet existing methods share three critical gaps: unrealistic evaluation, nosystematic comparison, and no consideration of edge deployment con-straints. We introduce a unified benchmark combining discrete-task eval-uation on structural and logical anomalies, a novel continuous drift proto-col, the first head-to-head comparison of all published CAD methods, andcomputational efficiency profiling on edge hardware. Our results revealthat existing CAD methods do not consistently outperform traditionalapproaches with simple experience replay. Thus motivated, we proposeDINOSaur, a training-free method combining a frozen DINOv3 back-bone with spatially-indexed coreset memory and neighborhood-restrictedanomaly scoring. DINOSaur achieves zero forgetting by construction,outperforms all evaluated methods on every protocol except geometricdrift (where all methods collapse to chance), and runs at sub-100 ms in-ference on an NVIDIA Jetson Orin Nano, with on-device adaptation tonew tasks in under 30 seconds.

Despite recent progress, vision-language encoders strugglewith two core limitations: (1) weak alignment between language anddense vision features, which hurts tasks like open-vocabulary semanticsegmentation; and (2) high token counts for fine-grained visual repre-sentations, which limits scalability to long videos. This work addressesboth limitations. We propose T-REN (Text-aligned Region Encoder Net-work), an efficient encoder that maps visual data to a compact setof text-aligned region-level representations (or region tokens). T-RENachieves this through a lightweight network added on top of a frozenvision backbone, trained to pool patch-level representations within eachsemantic region into region tokens and align them with region-level textannotations. With only 3.7% additional parameters compared to thevision-language backbone, this design yields substantially stronger densecross-modal understanding while reducing the token count by orders ofmagnitude. Specifically, T-REN delivers +5.9 mIoU on ADE20K open-vocabulary segmentation, +18.4% recall on COCO object-level text-image retrieval, +15.6% recall on Ego4D video object localization, and+17.6% mIoU on VSPW video scene parsing, all while reducing tokencounts by more than 24× for images and 187× for videos compared to thepatch-based vision-language backbone. The code and model are availableat https://github.com/savya08/T-REN.

Open-Vocabulary Change Detection (OVCD) faces a funda-mental granularity gap: foundation models prioritize high-level semanticabstraction while change detection requires pixel-level spatial fidelity.Traditional OVCD methods rely on instance extraction models, intro-ducing spatial semantic ambiguities and leading to over-segmentationor under-segmentation in remote sensing scenarios. To bridge this gap,we propose Free-CD, a training-free OVCD framework that reformu-lates the task by predicting a change probability distribution rather thanenforcing binary change masks via instance boundaries. We introduceRIFI-Up, an upsampler that enforces a resolution-invariant feature in-version constraint, enabling high-fidelity detail recovery without seman-tic artifacts. Additionally, we propose a Bayesian Probability Correctionmechanism that separates change priors from semantic categorization,introducing a correction term to rectify change area discrimination andenforce bi-temporal semantic consistency. Extensive experiments on fourremote sensing datasets demonstrate that Free-CD consistently outper-forms current methods, establishing new standards in accuracy and gen-eralization for open-vocabulary change detection. Code is available athttps://github.com/20374230/FreeCD.

While garments are essential for realistic digital humans,their topological variety makes them much harder to model than para-metric bodies. Traditional tailoring relies on 2D sewing patterns, yetbridging these patterns to 3D geometry currently requires physical simu-lations. We present Stitched Embeddings, the first simulation-free frame-work to unify 3D garment reconstruction and sewing pattern inferencewithin a single bidirectional latent space. By leveraging the geometricpriors of a pretrained 3D foundation model, our approach overcomesthe data scarcity typically associated with high-quality garment mod-eling. We propose to use the BoxMesh as a critical intermediate repre-sentation to align 2D panels into 3D configurations without the com-putational overhead of a simulator. This architecture achieves state-of-the-art accuracy in pattern reconstruction while significantly improvingefficiency. Furthermore, our differentiable pipeline enables novel applica-tions, including pattern recovery from meshes and 3D editing from 2Dpatterns. Finally, this work provides a scalable link between neural 3Dvision and the physical garment manufacturing pipeline. Project page:

Vision-Language Models (VLMs) deployed as situated agentsin high-resolution visual environments require active perception — theability to dynamically decide where to look through operations like zoom-ing, cropping, and panning. However, current training paradigms pro-duce models that mimic the surface form of such operations withoutfunctionally depending on their outputs, a phenomenon we term lazyperception. We trace this to a fundamental learning asymmetry: whencoarse global views combined with language priors suffice for moderateaccuracy, the model has no incentive to learn harder multi-step visualsearch. If a model can succeed without actively looking, it will neverlearn to look. This motivates Starve to Perceive, a training paradigm thatconstrains visual bandwidth — restricting each observation to a tight to-ken budget so that no single view suffices for task completion, therebystrongly incentivizing strategic and efficient active perception. Despiterequiring no auxiliary losses, reward shaping, or architectural changes— serving as a minimal, plug-in modification to standard post-trainingpipelines — models trained under perceptual starvation achieve substan-tial gains of 5% average relative improvement across diverse benchmarks.Our codes and data will be publicly available at https://github.com/WhuanY/Starve2Perceive.

Pretrained deep denoisers can be used to solve a wide rangeof model-based image reconstruction tasks via Plug-and-Play (PnP) andRegularization-by-Denoising (RED) algorithms, without retraining pertask. These denoisers are trained only for single-step denoising. Usingthem as Image Reconstruction (IR) regularizers in an iterative processcan destabilize reconstruction. A common failure mode is the peak-and-collapse behaviour: metrics such as PSNR improve for early iterations andthen abruptly degrade, making these algorithms unreliable in practice.We propose a data-driven stabilization framework that (i) formalizes thisinstability of any IR operator through a local quantity and (ii) preventscollapse by regularizing this quantity adaptively, requiring no retrainingor modification of the given pretrained network. Our key idea is tocontrol the potentially unstable IR operator with a contractive operatorwhose stable iterates act as an anchor and prevent collapse. We furtherintroduce an efficient family of trainable contractive operators that serve asstrong anchors while remaining lightweight. Extensive experiments acrossproximal algorithms, denoiser architectures, noise levels, and imagingtasks show consistent, collapse-free performance and improved reliabilityof PnP and RED reconstruction.

Reducing visual token redundancy is critical for acceleratingMultimodal Large Language Models (MLLMs) without degrading cross-modal reasoning performance. Existing token pruning methods typicallyrely on single-layer signals, such as attention scores or token similari-ties, which overlook the cross-layer transformation of visual representa-tions and may exhibit positional bias in multimodal token sequences.To address this limitation, we propose a training-free token pruningframework based on Cross-Layer Spectral Evolution (CLSE). Instead ofmeasuring token importance from single-layer feature magnitudes, CLSEquantifies how token representations evolve across Transformer layers inthe frequency domain. This evolution reflects the transition from high-frequency structural details to low-frequency semantic abstractions. Weobserve that tokens with stronger spectral redistribution across layers aremore likely to be semantically active and should therefore be preserved.By modeling cross-layer token dynamics, CLSE provides a stable im-portance criterion that mitigates positional bias. Extensive experimentson both image and video benchmarks demonstrate that CLSE achievesa superior trade-off between efficiency and accuracy under aggressivetoken reduction. Across multiple MLLMs, CLSE reduces FLOPs, KVcache memory, and latency while maintaining competitive or improvedperformance.

We introduce a new problem domain for human action recog-nition: the fine-grained analysis of children’s gait behaviors from stan-dard RGB video. We specifically target the ambulatory patterns of chil-dren aged 3–17 years. Such behaviors arise naturally in the diagnosis andtreatment of several critical developmental and neuromuscular disorders,such as cerebral palsy and hemiplegia. Despite their clinical value, cur-rent 3D sensor-based gait analysis systems are expensive, intrusive, andoften impractical for young subjects. To address this, we introduce anew dataset comprising over 1,100 high-frame-rate (60 FPS) video se-quences from 110 subjects, accompanied by synchronized, anonymizedpose sequences. In each session, the child performs a 5-second "walk-around" task, capturing the gait cycle from multiple viewpoints. Cru-cially, we demonstrate that current state-of-the-art approaches, includ-ing gait foundation models and Multimodal Large Language Models(MLLMs), fail to effectively resolve these clinical nuances. We identifythe key technical challenges in analyzing these erratic and subtle mo-tor patterns and describe a unified end-to-end framework for decodingfundamental components of pediatric gait. Through comprehensive ex-perimental results, we demonstrate the potential of this dataset to drivenovel research questions and establish a rigorous baseline for automatedchild gait assessment. Project page: pediamedai.com/ChildrenGait

Tree-based speculative decoding accelerates autoregressive generation by verifying a branching tree of draft tokens in a single targetmodel forward pass. However, existing methods prioritize maximizing token-level likelihood or the number of accepted tokens while ignoring a critical “efficiency paradox”: the computational overhead of drafting and verifying big trees can grow super-linearly, particularly at scale. This often leads to negative wall-clock speedup when batch sizes increase or hardware saturation limits are reached. To address this, we propose SMART, a System-aware Marginal Analysis framework for Runtime Tree construction. SMART reformulates tree expansion as a hardware-aware optimization problem that directly maximizes end-toend speedup. By applying a principled marginal benefit–cost rule at inference time, SMART expands a node only when its marginal benefit– cost ratio exceeds the tree-level speedup. SMART is training-free and serves as a plug-and-play controller for existing frameworks like MSD and EAGLE. Extensive evaluations across three MLLMs (e.g., LLaVA, Qwen2-VL) and four LLMs (e.g., Llama-3.1, DeepSeek-R1) demonstrate that SMART consistently outperforms state-of-the-art baselines. It delivers an average additional speedup of 20.0% for MLLMs and 15.4% for LLMs across compute-bound batching regimes and diverse GPU architectures without performance loss.

While Vision-Language Models (VLMs) have empowered em-bodied agents to execute complex household tasks, they struggle toproactively handle dynamically emerging hazards during closed-loop in-teractions. Existing safety approaches often rely on runtime guardrailsto block unsafe actions or induce excessive caution, which severely stallstask progress instead of actively resolving the underlying risks. To breakthis safety–progress trade-off, we introduce the Self-Evolving Just-In-Time Memory framework, which reframes embodied safety from progress-stalling guardrails to proactive hazard mitigation. The framework con-sists of a Risk-Sufficient Topological Belief Graph (RSG) for persistentsafety-relevant state tracking under partial observability, an Agency-Grounded Factual Memory for precise hazard anticipation, and an Expe-rience Memory that injects procedural Meta-Skills to guide executable,progress-preserving mitigation. Furthermore, we propose an automatedTest-Verify-Write loop, allowing agents to continually refine their mit-igation Meta-Skills from execution traces at test time. Experiments onIS-Bench demonstrate that our framework substantially boosts the Safe-Success rate across multiple VLM backbones (e.g., +30.3% on Qwen3-VL-8B), enabling agents to proactively mitigate hazards without stallingtask progress. Code is available at https://github.com/DyMessi/JIT-Memory.

Integrating macro-scale radiology and micro-scale pathologyprovides complementary insights crucial for precision oncology. How-ever, this integration is often hindered by modality discrepancy. The fea-ture embeddings from heterogeneous foundation models reside on poorlyaligned latent manifolds, and naive static fusion may therefore mix unre-lated semantic neighborhoods, leading to topological degeneracy and se-mantic entanglement. To bridge this gap, we propose a novel multimodallearning framework for Synthesizing radiology and pathology EvidencEvia Knowledge-guided FLOW matching. Specifically, we construct ashared feature manifold anchored by learnable clinical prototypes, en-abling the disentanglement and synthesis of shared and modality-specificfeature evidence. We then reformulate multimodal integration as a trans-port problem via knowledge-guided flow matching and seek continuousflows that smoothly transport features onto this semantic substrate. Thisparadigm preserves the intrinsic topology of the feature spaces and re-duces feature collapse. Furthermore, an uncertainty-aware evidential ag-gregator is incorporated to synthesize cross-modal feature evidence forreliable decision-making. Extensive experiments on three cancer datasets(i.e., TCGA Glioma, Gastric Cancer, and Chondrosarcoma) demonstratethe effectiveness of SeekFlow. Notably, SeekFlow consistently achievesstate-of-the-art performance across all benchmarks, outperforming ex-isting multimodal fusion methods in both diagnosis and prognosis tasks.The code is available at https://github.com/HKU-MedAI/SeekFlow.

We introduce ScenarioControl, the first vision-language con-trol mechanism for learned driving scenario generation. Given a textprompt or an input image, ScenarioControl synthesizes diverse, realistic3D scenario rollouts – including map, 3D boxes of reactive actors overtime, pedestrians, driving infrastructure, and ego camera observations.The method generates scenes in a vectorized latent space that representsroad structure and dynamic agents jointly. To connect multimodal con-trol with sparse vectorized scene elements, we propose a cross-global con-trol mechanism that integrates cross-attention with a lightweight global-context branch, enabling fine-grained control over road layout and traf-fic conditions while preserving realism. The method produces tempo-rally consistent scenario rollouts from the perspectives different actorsin the scene, supporting long-horizon continuation of driving scenarios.Extensive experiments validate that the control adherence and fidelityof ScenarioControl compare favorable to all tested methods across allexperiments. To facilitate training and evaluation, we release a datasetwith text annotations aligned to vectorized map structures at the projectwebpage: https://light.princeton.edu/ScenarioControl

Pre-trained Vision-Language Models (VLMs) like CLIP haveproven highly effective as foundation models for various downstream ap-plications. However, prompt learning in VLMs encounters a performance-generalization dilemma: while prompts can be tuned to achieve highaccuracy on seen distributions, this tuning process often undermines theirgeneralizability to unseen data. The limited set of learnable prompts,which contextualize and condition the input to steer it toward the taskwithin the pretrained VLM, tends to overfit the training data, leading toa trade-off between task-specific performance and preserving generaliza-tion. To address this dilemma, we introduce SAMPLe (Sharpness-AwareMinimization Prompt Learning), a plug-in sharpness-aware optimizerthat enhances prompt generalizability by accounting for loss landscapesharpness. Unlike conventional methods, SAMPLe balances explorationand exploitation by satisfying objective function constraints at each step,dynamically adapting to the current optimization state based on the localcurvature and gradient properties. This approach reduces overfitting onseen distributions and improves adaptability to unseen data, preservingthe generalization potential of pre-trained VLM models. We integrateSAMPLe into multiple prompt learning frameworks, including CoOp,CoCoOp, MaPLe, TCP, and Co-Prompt, demonstrating its effectivenessacross diverse methods. Experiments show that SAMPLe elevates promptlearning frameworks and consistently outperforms existing optimizersacross diverse settings, establishing itself as a robust, model-agnosticsolution for prompt learning.

Subject-driven image generation and manipulation are fun-damental for personalized content creation, such as identity-preservingportrait synthesis, multi-character storytelling, and controllable photoediting. However, existing models still struggle with multi-reference set-tings, often showing identity drift and unstable editing quality. A keybottleneck is data: current resources are often limited in scale, diversity,and consistency of subject-level details, especially for unified support ofboth generation and manipulation. To address this gap, we introduceOpenSubject, a video-derived large-scale corpus for subject-driven gen-eration and manipulation. Our pipeline leverages cross-frame identitypriors through four key stages: (i) Video Curation. We apply resolutionand aesthetic filtering to obtain high-quality clips. (ii) Cross-FrameSubject Mining and Pairing. We utilize vision language model (VLM)-based category consensus, local grounding, and diversity-aware pairing toselect image pairs. (iii) Identity-Preserving Reference Image Syn-thesis. We introduce segmentation map-guided outpainting to synthesizeinput images for subject-driven generation and box-guided inpaintingto generate input images for subject-driven manipulation, together withgeometry-aware augmentations and irregular boundary erosion. (iv) Ver-ification and Captioning. We utilize a VLM to validate synthesizedsamples, re-synthesize failed samples based on stage (iii), and then con-struct short and long captions. Furthermore, we introduce a benchmarkcovering both subject-driven generation and subject-driven manipulation,evaluated with a VLM-based judge on identity fidelity, prompt adherence,and consistency. Extensive experiments demonstrate that training onOpenSubject significantly enhances performance in complex scenes.

Current instruction-guided video editing models struggle tosimultaneously balance precise semantic modifications with faithful mo-tion preservation. While existing approaches rely on injecting explicitexternal priors (e.g., VLM features or structural conditions) to mitigatethese issues, this reliance severely bottlenecks model robustness and gen-eralization. To overcome this limitation, we present SAMA (factorizedSemantic Anchoring and Motion Alignment), a framework that factor-ize video editing into semantic anchoring and motion modeling. First, weintroduce Semantic Anchoring which establish a reliable visual anchorby jointly predicting semantic tokens and video latents at sparse an-chor frames, enabling purely instruction-aware structural planning. Sec-ond, Motion Alignment pre-trains the same backbone on motion-centricvideo restoration pretext tasks (cube inpainting, speed perturbation, andtube shuffle), enabling the model to internalize temporal dynamics di-rectly from raw videos. SAMA is optimized with a two-stage pipeline: afactorized pre-training stage that learns inherent semantic-motion rep-resentations without paired video-instruction editing data, followed bysupervised fine-tuning on paired editing data. Remarkably, the factorizedpre-training alone already yields strong zero-shot video editing ability,validating the proposed factorization. SAMA achieves state-of-the-artperformance among open-source models and is competitive with leadingcommercial systems (e.g. Kling-Omni). Code, models, and datasets arereleased in https://cynthiazxy123.github.io/SAMA.

Diffusion Transformers (DiTs) achieve superior image gener-ation quality but suffer from quadratic computational complexity rela-tive to token count. While various token reduction (TR) methods havebeen proposed to mitigate this cost, they overlook the primary objectiveof generative models: minimizing recovery error, which requires reflect-ing output token similarity. Specifically, they rely solely on input tokensimilarity inherited from reduction-only ViT paradigms, leading to a fun-damental misalignment with this objective.To bridge this gap, we propose DiTo, a novel TR paradigm that shiftsthe focus toward output-centric token reduction. Based on the observa-tion that output token similarity is consistently preserved across adjacenttimesteps, DiTo utilizes prior-step similarities as an effective proxy toestablish token correspondences at a Matching timestep, and reuses theresulting correspondences across multiple subsequent Reduction steps.To optimize this interleaved scheduling, we propose Pair Match Ra-tio (PMR)-guided Interval Scheduling to determine the optimal match-ing frequency. Furthermore, to mitigate localized approximation errorsand resulting blocking artifacts caused by repeated reuse, we proposeFrequency-aware Token Matching by incorporating a selection-frequencypenalty. Extensive experiments demonstrate that DiTo consistently out-performs existing TR methods with 1.6–3.9 dB higher PSNR at compa-rable speedups, achieving a superior Pareto frontier.

Region-level retrieval aims to align user-specified image regions with relevantregions or textual descriptions, playing a crucial role in realworld applications such as e-commerce product search and RAG. Although recent Large Multimodal Models (LMMs)have made significant strides in multimodal retrieval, they primarily focus on global-level tasksand struggle to capture effective region-level representations. To bridge this gap, we presentRegRet, an LMM-based Region-level Retrieval framework that enhances the regional rep-resentations without compromising overall global retrieval performance. At its core, RegRetintegrates a Region-Aware Encoder to capture detailed regional features while balancing themwith the global background context. To further enhance the fine-grained understanding anddiscriminability of representations, we design a multi-stage training pipeline that includesdetailed localized captioning and regional contrastive learning tasks. In addition, consideringthe absence of region-level contrastive training data and the limited diversity of evaluationtasks in current benchmarks, we introduce the REGMB benchmark. It comprises 225k con-trastive pairs, covering four multimodal retrieval tasks. Extensive experiments validate theeffectiveness of our approach. RegRet outperforms strong baselines in the zero-shot setting.Further training with contrastive learning leads to an average improvement of more than 20%on both REGMB and public benchmarks, while achieving comparable or better results onglobal-level retrieval tasks. The code and data will be released for future research.