ECCV 2026 Papers

Humans can effortlessly reason about scenes across different viewpoints, yet it remains unclear whether Vision–Language Models (VLMs) possess similar cross-view spatial abilities. Satellite-street scene pairs, with their complex contexts and extreme viewpoint variations, provide an ideal testbed. Motivated by this, we introduce CVSBench, a large-scale benchmark for evaluating cross-view spatial reasoning through satellite-street pairs. This benchmark supports multiple tasks, including cross-view VQA, cross-view grounding, and viewpoint identification. CVSBench comprises 3,297 cross-view image groups with 9,468 object-level annotations and 40,679 question–answer (QA) pairs, enabling systematic and controlled evaluation of cross-view spatial reasoning. Extensive evaluations reveal that advanced VLMs struggle to maintain object-level and layout consistency under drastic viewpoint changes. To bridge this gap towards human-like spatial cognition, we investigate two categories of approaches: spatially grounded reasoning and the incorporation of cognitive map inputs. Our findings demonstrate that language-only reasoning yields marginal improvements, while incorporating visual spatial imagination via a 3D scene imagination pipeline substantially improves cross-view reasoning. These results highlight the necessity of explicit visual-spatial representations for robust spatial cognition in VLMs. Our data and code are released at https://huggingface.co/datasets/zlyzlyzly/CVSBench.

Music-driven dance video generation aims to synthesize expressive human motion that is temporally aligned with music while maintaining high visual fidelity. Despite recent progress, existing methods still struggle to generate dance videos that simultaneously exhibit expressive motion and high visual quality. This limitation primarily arises from two factors: (1) Dataset. The lack of large-scale and high-quality datasets and effective data collection pipelines specifically tailored for dance video generation; and (2) Method. The absence of principled framework-level solutions for effectively integrating music as a complementary conditioning signal into the Video Generation Foundation Models. To address the dataset limitation, we introduce CIPE-Dance, a large-scale Internetsourced dance video dataset, equipped with Choreograph Informed text annotations and constructed via a Progressive Expert pipeline. To the best of our knowledge, CIPE-Dance is the largest dataset for dance video generation to date, comprising 300k high-quality clips (over 400 hours) and covering diverse dancers, environments, and dance genres. To overcome the method limitation, we propose OmniDance, a frameworklevel recipe for integrating music into a TI2V foundation model without sacrificing its original controllability or visual fidelity. Motivated by the complementary roles of text (low-frequency semantics) and music (highfrequency temporal dynamics), OmniDance co-designs a depth-aware specialization model architecture, an anchored easy-to-hard curriculum learning strategy, and modality-specialized time-dependent CFG strategy, achieving unified TI2V/MI2V/MTI2V generation. Extensive experiments on the CIPE-Dance dataset demonstrate that OmniDance achieves state-of-the-art performance across TI2V, MI2V, and MTI2V tasks, while exhibiting robust multimodal integration capability. Project is available at https://github.com/AMAP-ML/OmniDance.

Injection molding is the cornerstone of mass-producing plastic components. While current algorithms can automate mold design for basic geometries using standard two-piece molds, complex parts featuring undercuts, side holes, or re-entrant features present a significant challenge. These geometries often necessitate auxiliary components beyond the primary upper and lower molds. In practice, designing these intricate assemblies is a laborious process that relies heavily on expert knowledge. Furthermore, the scarcity of public datasets has hindered the development of effective learning-based solutions. To bridge these gaps, we introduce MoldCAD, a curated dataset that pairs complex single-body CAD parts with industry-standard mold assemblies. Each entry includes the upper and lower molds, parting surfaces, demolding orientations, and necessary auxiliary components. The dataset comprises 4,934 CAD models and over 3,850 mold assemblies, totaling more than 23k individual models. Building upon this dataset, we propose a comprehensive pipeline that predicts demolding orientations, identifies auxiliary components, and constructs parting surfaces to derive a complete, manufacturing-ready mold assembly for downstream CAD/CAM workflows. Our results demonstrate a promising path toward fully automated industrial mold design and contribute to the broader advancement of manufacturing-aware CAD generation. Code is available at https://github.com/tb2-sy/AIMold.

The segmentation of thin linear structures is inherently topology-critical, where minor local errors can sever long-range connectivity. While recent State-Space Models (SSMs) offer efficient long-range modeling, their isotropic serialization (e.g., raster scanning) creates a geometry mismatch for anisotropic targets, causing state propagation across rather than along the structure trajectories. To address this, we propose FGOS-Net, a framework based on frequency-geometric disentanglement. We first decompose features into a stable topology carrier and directional high-frequency bands, leveraging the latter to explicitly correct spatial misalignments induced by downsampling. Building on this calibrated topology, we introduce frequency-aligned scanning that elevates serialization to a geometry-conditioned decision, preserving directionconsistent traces. Coupled with an active probing strategy to selectively inject high-frequency details and suppress texture ambiguity, FGOS-Net consistently outperforms strong baselines across four challenging benchmarks. Notably, it achieves 91.3% mIoU and 97.1% clDice on DeepCrack while running at 80 FPS with only 7.87 GFLOPs.

Face recognition systems (FRSs) are increasingly deployed in critical real-world services for authentication, such as banking applications and airport identity checks, necessitating stringent security configurations. Consequently, the security vulnerabilities of FRSs have garnered significant attention. While existing studies have extensively explored FRS security, prior analyses have primarily focused on mediumsecurity threshold settings, which are not directly applicable to FRSs operating under high-security constraints. In this paper, we propose the first successful impersonation attack against FRSs under high-security threshold settings. Among various threat models, we focus on a practical and challenging scenario: score-based impersonation attacks under strict rate limits. To precisely evaluate the feasibility of such attacks, we provide a principled mathematical analysis characterizing the gaps in each stage of the attack pipeline. Our method significantly enhances impersonation capabilities in score-based attacks, even under elevated decision thresholds. On the LFW benchmark, with a budget of only 100 confidence score queries per identity, our attack achieves an impersonation success rate exceeding 92% against Amazon Rekognition at a confidence score threshold of 99—recommended setting for law enforcement scenarios. We further observe consistently robust performance across multiple open-source FRSs evaluated at similarly stringent decision thresholds.

Learning from inaccurate annotations has gained significantattention due to the high cost of precise labeling. However, despite thepresence of erroneous labels, models trained on noisy data often retainthe ability to make accurate predictions. This intriguing phenomenonraises a fundamental yet largely unexplored question: why models canstill extract correct label information from inaccurate annotations re-mains unexplored. In this paper, we conduct a comprehensive investiga-tion into this issue. By analyzing weight matrices from both empiricaland theoretical perspectives, we find that label inaccuracy primarily ac-cumulates noise in lower singular components and subtly perturbs theprincipal subspace. Within a certain range, the principal subspaces ofweights trained on inaccurate labels remain largely aligned with thoselearned from clean labels, preserving essential task-relevant information.We formally prove that the angles of principal subspaces exhibit mini-mal deviation under moderate label inaccuracy, explaining why modelscan still generalize effectively. Building on these insights, we proposeLIP, a lightweight plug-in designed to help classifiers retain principalsubspace information while mitigating noise induced by label inaccu-racy. Extensive experiments on tasks with various inaccuracy conditionsdemonstrate that LIP consistently enhances the performance of existingalgorithms. We hope our findings can offer valuable insights to under-stand of model robustness under inaccurate supervision.

Our work addresses 3D hand trajectory prediction in egocentric interaction, where future hand motion is inferred from visual observations, past motion, spatial context, and intent. Real-world actions follow stage-aware interaction structures (e.g., approach, manipulate) describing how the hand interacts with objects over time. However, prior works typically treat trajectory prediction as continuous signal regression, decoupling motion from semantic supervision and ignoring interaction structure. Without stage-aware cues to infer intent, models struggle to separate purposeful motion from egocentric noise and generalize across diverse interactions. We therefore present EgoMAN, a unified framework for interaction-structured 3D hand trajectory prediction that models hand motion as stage-aware interactions between the hand and surrounding objects. EgoMAN introduces a novel Trajectory-Token Interface where a small set of tokens encodes interaction stages, temporal progression, and 6DoF pose, enabling interaction stage-aware reasoning to guide efficient long-horizon 3D trajectory generation while preserving physical interpretability. To support this formulation, we construct the EgoMAN dataset with 219K 6DoF trajectories, stage-aware annotations, and 3M semantic, spatial, and motion QA pairs. Experiments show that EgoMAN improves trajectory accuracy, smoothness, and generalization, enabling interaction-structured reasoning for egocentric hand motion prediction for applications in robotics and assistive systems.

With the increase in model parameters and training data, the instruction following and generalization capabilities of Large VisionLanguage Models (LVLMs) have been significantly improved. Based on the Mixture-of-Experts (MoE) architecture, LVLMs expand their parameter capacity while maintaining the inference cost. However, traditional MoE methods employ a Top-k static routing strategy, which fails to account for variations in the input and to adaptively select the number of experts, resulting in suboptimal resource utilization. In this paper, we propose viewing token routing as an information encoding task, framing dynamic routing as a Minimum Description Length (MDL) problem in encoding. By validating the connection between MDL and gating entropy in the MoE scenario, we introduce Gating Entropy-based Uncertaintyaware Adaptive Routing (GeMoE) for MoE. Unlike traditional static or heuristic-based dynamic routing methods, GeMoE explicitly models the trade-off between model complexity and performance. By using gating entropy to assess the complexity of tokens, GeMoE adaptively determines the number of experts each token should engage. On a wide range of backbones and benchmarks, our method achieves 99.5% average performance retention compared to the original static routing, while improving average expert activation sparsity by 36.5%. The code will be publicly available at https://github.com/caichaoxiang/GeMoE.

Metric feed-forward 3D reconstruction for panoramic data remains under-explored due to the lack of large-scale panoramic RGBD training data. We present Realsee3D, a hybrid dataset of 10K indoor scenes (1K real, 9K synthetic) with 299K panoramic viewpoints and precise metric annotations, and Argus, a feed-forward network trained on it for metric panoramic 3D reconstruction. In the sparse unordered capture setting of Realsee3D, a poorly chosen coordinate anchor can cause global pose drift. Argus addresses this with a learned covisibility module that selects the geometrically optimal reference view to anchor the metric world frame. To further improve multi-task learning, we decompose the bidirectional pixel-to-world mapping into interpretable sub-steps with per-step supervision and cross-coordinate joint constraints, reinforcing geometric consistency across prediction branches. On the Realsee3D benchmark, Argus achieves state-of-the-art metric performance in camera pose estimation, depth estimation, and point cloud reconstruction. Project page: https://argus-paper.realsee.ai.

Sign Language Production (SLP) faces a fundamental tradeoff: direct text-to-pose models suffer from regression-to-the-mean effects, while dictionary-retrieval methods produce disjointed transitions. To resolve this, we propose a novel training paradigm that leverages sparse keyframes to capture the underlying kinematic distribution of human signing. By generating dense motion from discrete anchors, our approach mitigates regression-to-the-mean while ensuring fluid articulation. To achieve this at scale, we introduce FAST, an ultra-efficient sign segmentation model that automatically mines precise temporal boundaries. We then present SignSparK, a Conditional Flow Matching (CFM) framework that utilizes these temporal anchors to synthesize 3D signing sequences. This keyframe-driven formulation also unlocks Keyframe-toPose (KF2P) generation, making precise spatiotemporal editing of signing sequences possible. Furthermore, SignSparK scales across four distinct sign languages, constituting the largest multilingual SLP framework to date, and integrates 3D Gaussian Splatting for photorealistic rendering. Extensive evaluations demonstrate that SignSparK achieves state-of-the-art across diverse SLP tasks and multilingual benchmarks. Our code is available at https://github.com/JianHe0628/SignSparK.

Monocular depth estimation (MDE) is a fundamental yet inherently ill-posed task. Recent vision foundation models (VFMs), particularly DINO-based transformers, have significantly improved accuracy and generalization for dense prediction. Prior works generally follow a unified paradigm: sampling a fixed set of intermediate transformer layers at uniform intervals to build multi-scale features. This common practice implicitly assumes that geometric information is uniformly distributed across layers, which may underutilize the structural 3D cues encoded in VFMs. In this study, we present a systematic layer-wise analysis of DINOv3, revealing that 3D information is distributed nonuniformly: deeper layers exhibit stronger depth predictability and better capture inter-sample geometric variation. Motivated by this, we introduce a Last-Layer-Centric Feature Recombination (LFR) module to enhance geometric expressiveness. LFR treats the final layer as a geometric anchor and adaptively selects complementary intermediate layers according to a minimal-similarity criterion. Selected features are fused with the last-layer representation via compact linear adapters. Extensive experiments show that LFR module consistently improves MDE accuracy and achieves state-of-the-art performance. Our analysis sheds light on how geometric knowledge is organized within VFMs and offers an efficient strategy for unlocking their potential in dense 3D tasks. Code is released at https://github.com/book-book24/LFR.

We present V-JEPA 2.1, a family of self-supervised models that learns dense, high-quality, and temporally consistent representations for visual scenes in both images and videos. V-JEPA 2.1 combines four key ingredients: (i) a Dense Predictive Loss, a masking-based objective in which all tokens—visible context and masked tokens alike— contribute to the training loss, encouraging explicit spatial and temporal grounding; (ii) Deep Self-Supervision, which applies the selfsupervised objective hierarchically at multiple intermediate encoder layers to improve representation quality; (iii) Multi-Modal Tokenizers that support unified training over images and videos; and (iv) effective model and data scaling. Empirically, V-JEPA 2.1 achieves state-ofthe-art results on a range of benchmarks: 7.71 mAP on Ego4D for shortterm object-interaction anticipation, 40.8 Recall@5 on EPIC-Kitchens for high-level action anticipation, and a 20% improvement in real-robot grasping success rate over VJEPA-2 AC. The model also achieves stateof-art performance in robotic navigation (5.687 ATE on Tartan Drive), depth estimation (0.307 RMSE on NYUv2 with a linear probe), and global recognition (77.7% on Something-Something-V2).

Personalized Text-to-Image (PT2I) generation aims to pro-duce customized images based on reference images. A prominent interestpertains to the integration of an image prompt adapter to facilitate zero-shot PT2I without test-time fine-tuning. However, current methods grap-ple with three fundamental challenges: 1. the elusive equilibrium betweenConcept Preservation (CP) and Prompt Following (PF), 2. the difficultyin retaining fine-grained concept details in reference images, and 3. therestricted scalability to extend to multi-subject personalization. To tacklethese challenges, we present Dynamic Image Prompt Adapter (DynaIP),a cutting-edge plugin to enhance the fine-grained concept fidelity, CP·PFbalance, and subject scalability of state-of-the-art T2I multimodal diffu-sion transformers (MM-DiT) for PT2I generation. Our key finding is thatMM-DiT inherently exhibit decoupling learning behavior when injectingZ. Wang and T. Chu—Equal contribution. ∗ Corresponding author.reference image features into its dual branches via cross attentions. Basedon this, we design an innovative Dynamic Decoupling Strategy that re-moves the interference of concept-agnostic information during inference,significantly enhancing the CP·PF balance and further bolstering thescalability of multi-subject compositions. Moreover, we identify the vi-sual encoder as a key factor affecting fine-grained CP and reveal thatthe hierarchical features of commonly used CLIP can capture visual in-formation at diverse granularity levels. Therefore, we introduce a novelHierarchical Mixture-of-Experts Feature Fusion Module to fully leveragethe hierarchical features of CLIP, remarkably elevating the fine-grainedconcept fidelity while also providing flexible control of visual granularity.Extensive experiments across single- and multi-subject PT2I tasks verifythat our DynaIP outperforms existing approaches, while requiring onlysingle-subject training datasets.

Point cloud registration aligns multiple unposed point clouds into a common reference frame and is a core step for 3D reconstruction and robot localization when no initial pose guess is available. In this work, we cast point cloud registration as conditional generation: a learned, continuous point-wise velocity field transports noisy points to a registered scene, from which the pose of each view is recovered. Unlike prior methods that perform correspondence matching to estimate pairwise transformations and then optimize a pose graph for multi-view registration, our model directly generates the registered point cloud, yielding both efficiency and point-level global consistency. By scaling the training data and conducting test-time rigidity enforcement, our approach achieves state-of-the-art average performance on existing pairwise registration benchmarks and on our proposed cross-domain multi-view registration benchmark. The superior zero-shot performance on this benchmark demonstrates that our method generalizes across view counts, scene scales, and sensor modalities even with low overlap.

Implicit Neural Representations (INRs) encode signals as the weights of a coordinate-based neural network and have recently been proposed as an alternative domain for downstream learning. While promising, classification directly in weight space remains challenging due to the high dimensionality and complex structure of INR parameters. Furthermore, the way discriminative information is distributed across INR weights remains poorly understood. We propose a hierarchical Mixture-of-Experts (HMoE) Transformer that processes INR weights using conditional computation aligned with the structure of the underlying implicit network. Coupled with a meta-learning framework that shapes INR parameters for downstream tasks, our model achieves state-of-the-art accuracy across standard benchmarks, ranging from low-resolution datasets to high-resolution ImageNet-1K. To gain insight into how INRs encode discriminative information, we develop weight-space attribution and pruning methods that identify parameters most relevant for classification. These analyses reveal how class-specific structure emerges within INR layers and support the suitability of MoE architectures for weight-space learning. Our approach advances both the performance and interpretability of weight-space classifiers.

Image-conditioned image editing (ICIE) guides edits with a reference image to convey visual attributes—such as style, tone, or object identity—that are difficult to specify through text alone. Despite growing practical demand, no large-scale unified ICIE dataset currently exists in the open-source ecosystem; existing efforts cover only a narrow subset of edit types at small scale, typically constructed via costly forward synthesis. We reformulate ICIE data construction as a reference-completion problem: high-quality text-conditioned image editing (TCIE) datasets already supply the input image, instruction, and edited target, and can be augmented with aligned reference images through type-specific synthesis and lightweight instruction adaptation. Building on this insight, we propose a scalable pipeline equipped with weak-instruction augmentation and five-dimensional VLM-based post-filtering, and use it to construct RCEdit-500K—the first large-scale unified open ICIE dataset comprising 477K quadruplets across six edit categories (add, remove, replace, background, style, alter) with both concrete and abstract reference types. Training on RCEdit-500K consistently improves reference-guided editing: LoRA adaptation on diffusion models yields up to +1.22 average gain, and an autoregressive model without native editing ability acquires competitive ICIE performance, demonstrating that data availability is the primary bottleneck for open-source ICIE.

Driven by next-token prediction, NLP shifted from taskspecific models into powerful generalist foundation models. What, then, is the equivalent catalyst needed to achieve a general-purpose model in computer vision? In this paper, we contend that large-scale text-tovideo generation serves as a strong pre-training paradigm for computer vision, providing the necessary spatiotemporal priors, vision-language alignment, and scalability required for general visual intelligence. We introduce GenCeption, which leverages a pre-trained video generative diffusion backbone to define a feed-forward perception model, capable of performing various vision tasks steered by text instructions. Empirical results demonstrate that GenCeption achieves state-of-the-art performance across a diverse suite of tasks, including depth, surface normal, and camera pose estimation, expression-referring segmentation, and 3D keypoint prediction, often matching or surpassing specialized models (e.g., DepthAnything V3, SegmentAnything V3). Furthermore, the video generative pretrained backbone outperforms alternative pretraining paradigms (e.g., V-JEPA, and Video MAE) under comparable settings. Importantly, GenCeption exhibits preliminary data and model scaling properties along with exceptional data efficiency where it achieves comparable performance with leading models like D4RT and VGGT-Ω with 7× to 500× less training data. Finally, GenCeption also exhibits intriguing emergent behaviors: a model trained exclusively on synthetic human videos generalizes to real-world footage and out-of-distribution object categories (e.g., animals and robots). These findings suggest that video generation is not merely a synthesis tool, but a foundational path toward generalist vision intelligence for the physical world.

Image-guided style transfer aims to apply the artistic characteristics of a style image to a content image while preserving its semantic structure and layout. Despite advances in diffusion-based methods, existing approaches often face challenges in disentangling content and style, particularly when independently optimized adapters are naively combined, causing conflicts between adapters and limiting controllability over the content-style balance in inference. We further demonstrate that training-free structural guidance directly derived from the content image through the internal attention of pre-trained model outperforms a dedicated content LoRA adapter in terms of structural fidelity and computational efficiency. Building on these observations, we propose AnyStyle, a streamlined framework for image-guided style transfer. The framework adopts a unified single-adapter paradigm for coherent style capture from the style image and incorporates training-free structural guidance from the content image, thus avoiding complex entanglement between multiple adapters and improving controllability and stability. Extensive experiments show that our method delivers competitive quantitative performance and significantly improved perceptual quality. Code is available at https://github.com/Yvan1001/AnyStyle.

Achieving simultaneous preference alignment and distillation acceleration in video diffusion models remains an open challenge. Existing methods optimize the two objectives over mismatched representation spaces, where improving one objective often compromises the other. To overcome this, we propose Reward Lightning, a unified framework that aligns and accelerates a video diffusion model within a single shared representation. Its central principle is homology: both objectives are evaluated on identical latent features, which mitigates the gradient conflicts that arise when they are optimized over disjoint representations. As a foundational component, we first introduce a latent reward model (LRM) that scores videos directly in the latent space, without decoding back to the pixel space. Building on the LRM, homologous preference distillation (HPD) reuses this shared backbone to perform adversarial distillation and preference alignment jointly, yielding few-step generators that remain faithful and well aligned. Extensive experiments demonstrate that the LRM surpasses pixel-level and latent-level reward baselines by 11.0% and 14.7% in preference accuracy, and that Reward Lightning generates high-fidelity videos in merely 1 to 4 steps, improving the average VBench score by 2.1% while leading in text alignment, motion quality, and visual quality. Project page: https://reward-lightning.github.io.