ECCV 2026 Papers

Controllable traffic simulation is critical for testing autonomousdriving systems, yet existing approaches often require retraining largegenerative models with extensive annotated data. We introduce a lightweightcontrol adaptation framework that enables multi-modal controllability(sketch, latent behavior codes, and text) for pretrained state-of-the-artdiffusion and autoregressive traffic models. By modulating intermediatefeatures through identity-initialized FiLM layers, our method efficientlyadds new control modalities while preserving the base model’s generativeprior. Evaluated on Waymo Open Sim Agents Challenge, our approachdemonstrates strong controllability with less than 1% of the paired con-trol data. Through context-aware condition transfer, our framework en-ables counterfactual scenario generation and long-tail synthesis whilemaintaining stable closed-loop driving realism and safety. Our framework* Equal contribution.unlocks new possibilities for controllable traffic simulation, enabling tar-geted scenario generation through lightweight adaptation of pretrainedgenerative models. Project page: https://ecosim-web.github.io/

Generating large dynamic 3D worlds is important for appli-cations such as virtual reality, simulation, and robotics. However, the lackof large-scale 3D dynamic training data prevents training world-scale 3Dgenerative models directly. Recent progress in video generation providespowerful priors for generating 3D dynamic visual content. As a result,current approaches typically generate videos from multiple camera viewsusing image or video diffusion models and fuse them into a 3D represen-tation. However, these models generate each view independently, makingit difficult to maintain motion consistency across viewpoints. Existing so-lutions address this problem by sequentially generating each view whileupdating an explicit 3D representation, which introduces a significantcomputational bottleneck. In this work we propose D YN W ORLD, a frame-work that enables parallel generation of motion-consistent videos acrossmultiple views. Our key insight is that the latent tokens of transformer-based video generators implicitly encode spatial and dynamic informa-tion that can be geometrically related across cameras. We thus intro-duce a Spatial Token Memory that stores latent video tokens along withtheir estimated 3D positions. During generation, tokens from one vieware back-projected into 3D and re-projected into other views, allowingmultiple video generators to share consistent motion context through at-tention. This mechanism enforces cross-view motion consistency whileallowing all views to be generated simultaneously. Experiments showthat our method produces large, diverse, view-consistent dynamic worldswhile reducing generation time by an order of magnitude compared tosequential pipelines.

The softmax activation in multihead attention (MHA) is thede facto standard for attention-based models in visual perception tasks.However, standard softmax can produce noisy attention patterns thatdilute relevant features and degrade its performance. In this paper, wepropose Denoising Attention or DnA, in which, first, a positive queryidentifies which image features belong to the correct class, and a neg-ative query identifies closely associated but irrelevant image features.DnA then projects these interactions into two distinct subspaces withlarger principal angles, promoting subspace separation and improved dis-criminability. Using a ViT-B backbone, our proposed DnA achieves anabsolute gain of 0.8% on ImageNet-1K compared to the baseline. Wefurther show improvements across multiple visual understanding tasks,including video understanding with video transformers (1.8%) and videoLLMs (0.5%). Our extensive empirical analyses justify the design choicesinvolving two interacting subspaces and the denoising effect of DnA. Thecode is publicly available at https://github.com/rjccv/DnA.

Cervical cancer remains a leading cause of female morbidityand mortality, making early lesion detection critical. Although RCNN-and YOLO-based detectors have improved performance, they still relyon complex post-processing and are limited in global reasoning. DETRoffers a new paradigm, but its direct application is challenged due to(1) large intra-class scale variation, (2) subtle inter-class differences, and(3) high inter-annotator variability. To address these challenges, we pro-pose CerDETR, a DETR-based framework empowered by cell priors. Itintroduces a training-only Prior Corrector branch to capture intra-classscale and inter-class feature variations. It employs an Auto MultiscalePrior Generator (Auto MPG) to produce precise multiscale cell priorscovering nuclei, cytoplasm, and clusters, an IoU and Contain-Guided(ICG) matching strategy for robust prior assignment under annotationnoise, and a Prior Query Enhance module to leverage lesion character-istics and optimize prior embeddings for discriminative feature learning.Experiments on four public datasets show that CerDETR consistentlyoutperforms RCNN-, YOLO-, and DETR-based methods, demonstrat-ing its effectiveness, generalization, and clinical potential. Source code isavailable at https://github.com/imAzhou/CerDETR.

Mixture-of-Experts (MoE) has become a prevalent paradigmfor scaling Vision Transformers efficiently. To ensure computational scal-ability and prevent expert overload, Vision MoE architectures employ acapacity-bounded token dispatch mechanism, where each expert’s pro-cessing budget depends on the inference batch size. This work identifiesthis batch-dependent behavior as an overlooked attack surface, and pro-poses a stealthy supply-chain backdoor attack that exploits this prop-erty through a three-phase framework. First, we inject a backdoor intoan early MoE layer. Second, we train a neutralizer in a deeper MoElayer that suppresses the backdoor under normal capacity. Third, weconfigure a batch-adaptive capacity factor that preserves high capacityfor small batches while reducing it for large batches, naturally disablingthe neutralizer via token overflow at deployment-scale batch sizes. Theattack remains in dormant mode during small-batch security audits andenters activation mode during large-batch deployment. Experiments onV-MoE and Swin-MoE across ImageNet-100 and GTSRB demonstrateactivation-mode attack success rates of 76–87% with dormant-mode ASRbelow 9%, while evading Neural Cleanse, STRIP, Fine-Pruning, and Acti-vation Clustering. Our findings reveal a fundamental security risk arisingfrom batch-dependent execution in scalable Vision MoE architectures.

Face forgery detection (FFD) is essential in preventing themisuse of diverse generation methods like generative adversarial net-works (GANs) and diffusion models. Although advanced FFD methodshave started to explore the benefits of Vision-Language Model (VLM),they either rely solely on visual modality or optimize visual-linguisticmodalities independently, leaving researches on cross-modal interactionlargely unexplored. We find that visual-linguistic asymmetry in-herent in VLM tends to cause cross-modal misalignment, underminingits ability to discriminate and generalize on visually similar real-forgedfaces. This finding shows that this asymmetry is twofold: (1) cross-modal: visual-linguistic spaces affect VLM’s generalizability oppositely.(2) intra-modal: spaces with different-level semantics have inverse effectson VLM’s discriminability and generalizability. Building upon these in-sights, we propose Asy-Det, a parameter-efficient detector that follows anasymmetry-guided visual-linguistic interaction paradigm, to fully unlockthe cross-modal potential of VLM for reliable FFD. With only 4.41Mtrainable parameters, Asy-Det achieves an image-level AUC of 92.94%on unseen Celeb-DF-v2 dataset.

Multimodal large language models (MLLMs) are promisingfor forgery detection, but most methods treat them as fixed black boxes.Preliminary studies have shown that this passive use overlooks instabili-ties in the internal data flow, leading to representation drift. In this work,we introduce the Asymmetric Anchoring Paradigm (AAP), an open-boxapproach that reshapes the MLLM’s internal data flow (rather than ap-pending external components) by re-purposing its pre-trained visual en-coder as an active truth anchor. AAP has two key steps: (1) for realimages, we impose an anchor-alignment constraint that pulls represen-tations toward the truth anchor, yielding a highly stable, low-variancemanifold; (2) for tampered images, we measure their deviation from theanchor and use the resulting error map, which spatially quantifies depar-tures from real-world priors, as a precise cue for the localization decoder.Comprehensive evaluations on SID-Set and OpenSDID demonstrate thatAAP substantially improves detection accuracy and the localization IoUwithin manipulated regions. By opening the black box and anchoring tothe encoder’s real-world priors, AAP turns MLLMs from passive compo-nents into actively regularized detectors for practical forgery detection,offering a new insight for the field. The code is open-sourced and publiclyavailable at https://github.com/rstao-bjtu/AAP.

Large pretrained vision encoders are central to multimodalintelligence, powering applications from on-device vision processing tovision-language models. Since these applications often demand real-timeprocessing of massive visual data, reducing the inference cost of visionencoders is critical. Quantization offers a practical path, but it remainschallenging even at 8-bit precision due to so-called outliers. In this work,we propose RegCache, a training-free algorithm that mitigates outliers inlarge-scale pretrained vision encoders and serves as a plug-in module thatcan be applied on top of other quantization methods. RegCache intro-duces outlier-prone yet semantically meaningless prefix tokens into thevision encoder, which prevent other tokens from having outliers. Notably,we observe that outliers in vision encoders behave differently from thosein language models, motivating two technical innovations: middle-layerprefixing and token deletion. Experimental results show that our methodconsistently improves quantized model performance across various visionencoders, particularly in extremely low-bit regimes (e.g., 4-bit).

Depth from Defocus (DfD) is the task of estimating a densemetric depth map from a focus stack. Unlike previous works overfittingto a certain dataset, this paper focuses on the challenging and practi-cal setting of zero-shot generalization. We first propose a new real-worldDfD benchmark ZEDD, which contains 8.3× more scenes and signifi-cantly higher quality images and ground-truth depth maps comparedto previous benchmarks. We also design a novel network architecturenamed FOSSA. FOSSA is a Transformer-based architecture with noveldesigns tailored to the DfD task. The key contribution is a stack atten-tion layer with a focus distance embedding, allowing efficient informa-tion exchange across the focus stack. Finally, we develop a new trainingdata pipeline allowing us to utilize existing large-scale RGBD datasetsto generate synthetic focus stacks. Experiment results on ZEDD andother benchmarks show a significant improvement over the baselines,reducing errors by up to 55.7%. The ZEDD benchmark is released athttps://zedd.cs.princeton.edu. The code and checkpoints are re-leased at https://github.com/princeton-vl/FOSSA.Input: Focus stack and corresponding focus distancesFocus Distance = 0.8m Focus Distance = 1.6m Focus Distance = 3.1m Focus Distance = 8.1mDFF-FV 𝛿! = 0.71 MoGe-2 𝛿! = 0.45 𝛿! = 0.98Existing depth from defocus Monocular depth models suffer Ours Ground Truthmodels are not generalizable from scale ambiguity*Equal contribution.

Retargeting human–object interaction demonstrations tophysics-based simulation requires reproducing not only body motion butalso the object motion and contacts that make manipulation succeed.However, position-only hand trajectories do not specify the contact forcesneeded to manipulate objects, and directly tracking them can overcon-strain contact-rich finger behavior. We introduce WristMimic, a wrist-guided whole-body control framework that explicitly separates contact-free body motion from contact-rich hand manipulation. The contact-freebody and wrist are guided by kinematic pose targets, whereas the fin-gers are not directly supervised by human hand pose. Instead, they learngrasping and manipulation behaviors from object tracking and contactoutcomes. Our key insight is that the wrist is the natural gate betweenthese two regimes. It is largely free from contact and can be tracked kine-matically, yet it determines the global hand configuration and places thefingers within reachable grasp affordances. To ensure reliable wrist place-ment during interaction, we introduce wrist-specific reset constraintsand reward prioritization. Experiments show that WristMimic matchesor surpasses methods using full finger pose supervision while enablingfinger-agnostic retargeting across diverse hand embodiments.

Camera motion control is essential for directing viewpointchanges in generative systems. However, existing methods typically condi-tion the generation process on a single specific modality, such as explicitpose trajectories or reference videos, limiting their ability to support het-erogeneous user inputs. To address this limitation, we present TriMotion,a modality-agnostic framework for camera-controlled video generationthat maps video, pose, and text inputs, describing the same cameratrajectory into a shared motion embedding space. Learning such a spacerequires synchronized supervision across modalities. Therefore, we buildthe Motion Triplet Dataset by extending a Multi-Cam Video Dataset withgeometry-grounded motion descriptions derived from camera extrinsics.We further introduce a latent motion consistency objective that lever-ages the motion embedding space to encourage the generated video tofollow the target camera trajectory directly in latent space, avoiding thecost of pixel-space decoding. Extensive experiments show that TriMotiongenerates high-quality videos that accurately follow the target cameratrajectories across all three modalities. Beyond standard generation, theshared motion embedding space also enables flexible applications such assequential motion composition and cross-modal motion interpolation.

We introduce a novel mathematical framework for analyz-ing and generating complex tree-shaped 3D objects, such as botanicaltrees and plants, which deform both in their 3D geometry and branchingstructure. Unlike previous works, which either consider only the skeletalstructure of tree-like objects or approximate their 3D geometry usingbranch thickness, the proposed framework accurately models both the3D geometry of the tree branches and the way they are interconnected.In this paper, we first generalize the Square Root Normal Fields (SRNF)representation, originally proposed for the statistical analysis of genus-0surfaces, to tree-shaped 3D objects. We then treat tree-shaped 3D ob-jects as points on a novel Riemannian tree-shape space equipped with anovel Riemannian metric that measures the amount of surface bendingand stretching, and structural changes one needs to apply to one 3D tree-shape to align it with another. This way, deformations become trajecto-ries in this novel tree-shape space. We analyze the theoretical propertiesof this novel tree-shape space and the corresponding metric and developalgorithms for computing point-wise and branch-wise correspondencesand geodesic paths between complex 3D trees. We finally show how touse these building blocks for (1) computing statistical summaries, i.e.means and modes of variation, of collections of tree-shaped 3D objects,and (2) synthesizing novel tree-shaped 3D objects by sampling fromprobability distributions fitted to a population of tree-shaped 3D objects.We demonstrate the performance and utility of the proposed frameworkon real and synthetic plants and botanical trees and show that it sig-nificantly outperforms the state-of-the-art. Additional results and sourcecode are available at https://tahmina979.github.io/Tree_in_SRNF/

Although existing video editing methods are generally feasi-ble, they often require many costly iterations and still struggle to deliverhigh-quality yet satisfying editing results. We attribute this limitationto the prevalent data-to-data paradigm, which is less compatible withmodern generative models than noise-to-data generation. To address thisgap, we revisit video editing from a noise-to-data perspective and pro-pose Streaming-Generation-based Video Editing (StreamEdit), whichpreserves few-step sampling while seamlessly injecting source-video con-ditions. Built on pre-trained streaming generation models, StreamEditintroduces dual-branch fast sampling with a self-attention bridge andcross-attention grounding/boosting to satisfy both sampling and con-ditioning requirements. We further propose source-oriented guidance toimprove target-generation quality, and a visual prompting strategy to en-hance editing flexibility and practicality. The method is effective, robust,and generalizable across different models. Extensive experiments on di-verse video editing tasks show that StreamEdit consistently outperformsexisting approaches, even in few-step settings with minimal time cost.Code and results are available at: dsl-lab.github.io/StreamEdit/.

Benefiting from the strong ability to capture data distributions, diffusion models have become powerful tools for solving image inverse problems. The key is to controllably steer the sampling trajectory toward the measurements while respecting the diffusion prior. In this work, we introduce Stochastic Optimal Control Sampling (SOCS), which models the denoising process as a dynamical system and injects control signals via SOC. Previous SOC-based approach addresses inverse problems by optimizing over the entire trajectory, which is computationally expensive. In contrast, we derive a closed-form control update and apply it at each sampling step, pulling the measurement-consistent clean prediction back onto the denoising flow. In SOCS, we can readily modulate the control strength to align with the diffusion model’s native capabilities and thereby enhance perceptual quality. Our method is compatible with a variety of linear stochastic differential equation backbones. Extensive experiments across a broad spectrum of image inverse tasks demonstrate that SOCS achieves accurate measurement-aligned reconstructions with improved visual fidelity and stronger quantitative performance. Code is available at https://github.com/zjqwq01/SOCS-DIP.

As Text-to-Video (T2V) models progress towards higher vi-sual realism, the artifacts and errors they produce are highly nuanced,fine-grained, and spatio-temporally localized. Vision Language Models(VLMs) are actively being adopted as automatic evaluators for videogeneration, driven by the promise of their perception and reasoning abili-ties. Yet, it remains unclear whether they can detect, localize, and explainfine-grained errors in modern high-fidelity video generations. We intro-duce Spotlight, a novel benchmark to rigorously assess whether cur-rent VLMs can precisely localize and explain nuanced video-generationerrors. Spotlight comprises 600 videos generated by state-of-the-artT2V models (Veo3, Seedance, LTX-2), annotated with over 1,600 fine-grained error localizations and explanations spanning physics, semanticsand anatomy. Our experiments reveal that current VLMs lag behindhumans significantly, with humans outperforming our best baselines bynearly 2× on the task. Our analysis highlights key areas of improvementfor utilizing VLMs as automated video evaluators, including the need formore robust perception and hallucination mitigation. Data and code isavailable at https://spotlight-video.github.io.

Agentic multimodal large language models (MLLMs) (e.g.,OpenAI o3 [41] and Gemini Agentic Vision [10]) achieve remarkablereasoning capabilities through the iterative invocation of visual tools.However, the cascaded perception, reasoning, and tool-calling loops intro-duce significant sequential overhead. This overhead, termed agentic depth,incurs prohibitive latency and seriously limits system-level concurrency.To this end, we propose SpecEyes, an agentic-level speculative accelera-tion framework that breaks this sequential bottleneck. Our key insight isthat a lightweight MLLM can plan a tool-free execution path to directlyanswer many queries, bypassing the expensive tool-use loop. To regulatethis speculative planning, we introduce a cognitive gating mechanismbased on answer separability, which quantifies the model’s confidence inself-verification without requiring oracle labels. Furthermore, we designa heterogeneous parallel funnel that exploits the small model’s statelessconcurrency to mask the large model’s stateful serial execution, therebymaximizing system throughput. Extensive experiments on V* Bench,HR-Bench, and POPE demonstrate that SpecEyes achieves 1.1 − 3.35×speedup over the baseline while preserving or even improving accuracy,thereby boosting serving throughput under concurrent workloads.

Vision–language models (VLMs) have recently shown promise for assisting clinical reasoning in dermatological diagnosis. However, their trustworthiness and clinical utility remain limited by three key challenges: heterogeneous datasets with inconsistent diagnostic labels and concept annotations, the lack of grounded diagnostic rationales for reliable reasoning supervision, and limited scalability when transferring knowledge from small, densely annotated datasets to large collections with sparse labels. To address these challenges, we propose Skin-R1, a dermatology-oriented VLM that integrates textbook-grounded clinical reasoning supervision with reinforcement learning (RL) to improve the accuracy and robustness of diagnostic prediction. First, we construct a textbook-based reasoning generator that synthesizes hierarchy-aware and differential-diagnosis (DDx) diagnostic trajectories derived from authoritative dermatology knowledge. Second, these trajectories are used for supervised fine-tuning (SFT), establishing a clinically grounded reasoning foundation for the model. Finally, we introduce an RL training framework that incorporates the hierarchical structure of dermatological diseases into the reward design, enabling the model to generalize grounded diagnostic reasoning to large-scale datasets with sparse annotations. Extensive experiments across multiple dermatology benchmarks demonstrate that Skin-R1 consistently improves diagnostic accuracy and robustness compared to state-of-the-art Med-VLM baselines. Ablation studies further highlight the critical role of grounded reasoning supervision introduced during the SFT stage. Code and model weight are available at https://github.com/l593191569/Skin-R1.

Inferring physical properties such as mass, stiffness, and elas-ticity from a single image is essential for simulation and embodied AI, yetmost existing approaches rely on multi-view reconstruction or physics-based supervision. We introduce SiPhy, a unified framework for single-image physical property reasoning that aligns 3D-aware visual cues, depthwith language-based material knowledge. From one RGB image, SiPhysamples pseudo-voxel points, extracts CLIP features, and grounds themto material candidates proposed by an VLM. A part-based contrastiveaggregator enforces region consistency, while a heaviness-aware refine-ment improves thickness and volume estimation for dense objects.Across ABO-500, MVImgNet-100, and PhysXNet-100, SiPhy achievesstate-of-the-art single-image performance, surpassing multi-view recon-struction methods by improving mass MnRE by up to 93% (vs. PUGS),reducing density MAE by 35.5% (vs. NeRF2Physics), and loweringYoung’s modulus error by 23.5%. We further validate SiPhy on realhand–object interaction datasets, demonstrating its potential as a dataannotation engine for physical understanding from single-view imagery.Codes are available at https://github.com/DominoAI-Lab/SiPhy-ECCV-2026.

Embodied Question Answering (EQA) requires an agent toexplore unseen environments, gather visual evidence, and answer natural-language questions. Efficient exploration requires both global planningand task-aware prioritization of informative regions. However, existingfrontier-based methods typically rely on greedy, step-wise decisions, whichoften cause redundant motion and frequent Vision-Language Model (VLM)queries. We propose SAFE-EQA, a semantic-aware framework that im-proves exploration efficiency through globally coherent path planning andreduced VLM interaction. SAFE-EQA incrementally builds a landmarkgraph to capture scene topology and a query-conditioned semantic mapto estimate task relevance of different regions. To support comprehen-sive, question-oriented exploration, we formulate planning as a Semantic-Aware Traveling Salesman Problem (TSP), which optimizes the visita-tion order over discovered landmarks. We further introduce an adaptivereplanning mechanism that selectively interrupts local navigation andre-optimizes trajectories when higher-priority frontiers emerge. Exten-sive experiments on OpenEQA, EXPRESS-Bench and HM-EQA showthat SAFE-EQA matches or outperforms state-of-the-art EQA methodswhile reducing VLM token consumption by 27.8–45.6%.