ECCV 2026 论文列表

Compositional text-to-3D generation aims to create multiobject scenes from text descriptions, yet existing approaches often suffer from severe inter-object penetrations during joint optimization, where entangled object representations lead to physically implausible scenes. To address this, we propose CompSAG to enable penetration-free and complete compositional 3D generation. Our key insight is a Gaussian surface offset (GSO) mechanism that mimics real-world repulsive forces to effectively resolve inter-object penetrations. The proposed GSO calculates offset values by aggregating repulsive forces from an object proxy surface, pushing apart the Gaussians in collision regions. Furthermore, during appearance optimization, we design a Gaussian surface rectification module to suppress the spiky geometric artifacts introduced by SDS. These are built upon a layout-guided scene initialization driven by a fine-tuned LLM, which provides precise spatial grounding from the text. Extensive experiments across diverse multi-object scenarios demonstrate that CompSAG produces complete, spatially coherent, and penetrationfree 3D scenes, outperforming prior compositional generation baselines.

Feed-forward 3D reconstruction (F3R) transformers have recently achieved remarkable success. However, scaling them to long image sequences remains challenging, as the quadratic complexity of cross-view global attention quickly becomes the dominant computational bottleneck. While recent efforts attempt to improve efficiency through compressed or sparse attention, they fail to fully exploit the inherent sparsity and dynamic behavior of global attention. In this work, we present a comprehensive analysis of global attention across multiple F3R transformers and reveal that attention patterns are highly heterogeneous, dynamic, and extremely sparse across layers and attention heads. Motivated by these findings, we propose SAF3R, a training-free dynamic sparse attention framework tailored to F3R transformers. SAF3R integrates tailored sparse attention mechanisms with offline head profiling and an efficient online adaptation strategy to match input-dependent attention behaviors. Extensive experiments demonstrate that SAF3R achieves high sparsity ratios while preserving camera pose estimation and 3D reconstruction quality, translating into substantial end-to-end speedup on F3R transformers compared to existing methods. Code is available at https://github.com/jndeng/SAF3R.

While many tasks require reasoning across multiple images, open-source Large Multimodal Models (LMMs) remain unreliable in these settings. We analyze multi-image LMMs and identify a phenomenon we term attention fragmentation: in each image, tokens at similar background locations act as attention sinks, absorbing disproportionate attention. Causal masking further skews this effect, as earlier images accumulate more sink attention than later images. Using an entropy score over per-image attention, we find that visual focus remains highly dispersed across images rather than isolating key evidence. By applying Pinsker’s inequality, we establish a theoretical bound showing that this high-entropy dispersion, combined with stronger early-image sinks, strictly reduces the usable non-sink attention available to earlier images, providing a mechanistic link to image order sensitivity. Motivated by this diagnosis, we propose Attention Remasking (AR), a post-training edit that blocks sink keys and opens a sparse set of cross-image links, routing the recovered attention to task-relevant tokens. AR improves accuracy and reduces order sensitivity across multi-image benchmarks, narrowing the gap between open-source LMMs and leading commercial models.

We present TimeWalker, a novel framework that models realistic, full-scale 3D head avatars of a person on lifelong scale. Unlike current human head avatar pipelines that capture a person’s identity only at the momentary level (i.e., instant photography, or short videos), TimeWalker constructs a person’s comprehensive identity from unstructured data collection over his/her various life stages, offering a paradigm to achieve full reconstruction and animation of that person at different moments of life. At the heart of TimeWalker’s success is a novel neural parametric model that learns personalized representation with the disentanglement of shape, expression, and appearance across ages. Central to our methodology are the concepts of two aspects: 1) We track back to the principle of modeling a person’s identity in an additive combination of his/her average head representation in the canonical space, and moment-specific head attribute representations driven from a set of neural head basis. To learn the set of head basis that could represent the comprehensive head variations of the target person in a compact manner, we propose a Dynamic Neural Basis-Blending Module (Dynamo). It dynamically adjusts the number and blend weights of neural head bases, according to both shared and specific traits of the target person over ages. 2) We introduce Dynamic 2D Gaussian Splatting (DNA-2DGS), an extension of Gaussian splatting representation, to model head motion deformations like facial expressions without losing the realism of rendering and reconstruction of the full head. DNA-2DGS includes a set of controllable 2D oriented planar Gaussian disks that utilize the priors from a parametric morphable face model, and move/rotate with the change of expression. Through extensive experimental evaluations, we show TimeWalker’s ability to reconstruct and animate avatars across decoupled dimensions with realistic rendering effects, demonstrating a way to achieve personalized “time traveling” in a breeze. Project page: https://TimeWalker2026.github.io/.

Spatial tracing, as a fundamental embodied interaction ability for robots, is inherently challenging as it requires multi-step metricgrounded reasoning compounded with complex spatial referring and realworld metric measurement. However, existing methods struggle with this compositional task. To this end, we propose RoboTracer, a 3D-aware VLM that first achieves both 3D spatial referring and measuring via a universal spatial encoder and a regression-supervised decoder to enhance scale awareness during supervised fine-tuning (SFT). Moreover, RoboTracer advances multi-step metric-grounded reasoning via reinforcement fine-tuning (RFT) with metric-sensitive process rewards, supervising key intermediate perceptual cues to accurately generate spatial traces. To support SFT and RFT training, we introduce TraceSpatial, a largescale dataset of 30M QA pairs, spanning outdoor/indoor/tabletop scenes and supporting complex reasoning processes (up to 9 steps). We further present TraceSpatial-Bench, a challenging benchmark filling the gap to evaluate spatial tracing. Experimental results show that RoboTracer surpasses baselines in spatial understanding, measuring, and referring, with an average success rate of 79.1%, and also achieves SOTA performance on TraceSpatial-Bench by a large margin, exceeding Gemini- 2.5-Pro by 36% accuracy. Notably, RoboTracer can be integrated with various control policies to execute long-horizon, dynamic tasks across diverse robots (UR5, G1 humanoid) in cluttered real-world scenes. Please see the project page at https://zhoues.github.io/RoboTracer.

Group Relative Policy Optimization (GRPO) enables stable and preference-oriented updates via group-wise comparisons for posttraining video generation. However, GRPO directly optimizes rewardinduced advantages. Under sustained optimization, the reward score can lose fidelity as a proxy for true video quality, consistent with the phenomenon described by Goodhart’s Law. This leads to two recurring issues: (i) shortcut-driven optimization under composite objectives and (ii) reward saturation within prompt groups. To address these issues, we introduce TaRoS, a Target-Robust Reward Signaling framework for Video generation GRPO. TaRoS leverages component level performance assessment together with intra-group sparsity to organize multi-aspect rewards towards optimization objectives. In addition, it adaptively downweights components that exhibit saturation, thereby preserving effective optimization directions and mitigating redundancy. This maintains meaningful optimization directions and preserves within-group ranking separation, thereby preventing reward hacking and leading to more reliable policy updates. Extensive experiments show consistent improvements in visual fidelity, motion coherence, and text-video alignment over strong baselines.

Recent text-to-image models such as DALL·E-3 excel at following diverse prompts yet remain blind to individual aesthetic preferences. We study personalized image generation, where models must align outputs with a user’s implicit visual preferences based on a few historically preferred images and a short prompt. To this end, we introduce PIPBench, the first profile-inclusive benchmark for evaluating personalized image generation. We further propose a novel data construction pipeline that leverages psychological and demographic profiling dimensions for both real-user data collection and scalable agentbased data generation. Using PIPBench, we conduct a thorough evaluation of representative line of methods. Our experiments reveal key limitations in existing methods, suggesting new challenges and opportunities for personalized text-to-image synthesis. Project page: https: //wuyuhang05.github.io/PIPBench/

Event cameras provide asynchronous visual sensing, making them a compelling modality for agile edge AI systems. For practical deployment, event-based vision algorithms must sustain sufficiently high throughput to keep pace with the incoming event stream. Yet existing approaches often fail to do so, as they are rarely designed in close consideration of the underlying sensor and compute hardware. We propose a new event-based feature detection and tracking algorithm that operates at high throughput on an embedded CPU, without GPU acceleration. Our pipeline is motion-defined: it activates only when sufficient motion is detected and dynamically adapts its computational load to the observed motion. At its core is a novel slice-parallel formulation of event-based corner detection and tracking that operates directly on live event streams. The pipeline is enabled by our Motion-Defined Concurrent Slices architecture, which efficiently processes event data and maximizes utilization of available compute resources. On event streams generated by handheld motion in typical scenes, our method achieves throughput above the average event rate, supporting responsive event-based vision on the edge. Extensive benchmarking on standard datasets and live demonstrations confirms the effectiveness of the proposed method. 1

Non-prehensile manipulation is often used as a preparatory step for robotic grasping, yet existing approaches typically require a predefined target object pose. In practice, however, objects admit multiple graspable configurations and the desired pose is not known in advance. We reformulate non-prehensile manipulation for grasping as optimizing an object centric graspability objective rather than reaching a specific pose. We construct a graspable set from synthesized grasps and define a graspability field that measures how suitable an object configuration is for successful grasp execution. The scalar measure provides a dense learning signal for reinforcement learning and determines when to terminate manipulation. This yields a closed-loop manipulation-to-grasp pipeline driven by a single policy. Experiments in simulation and on a real robot show that the policy reliably reconfigures objects into graspable states and transitions to grasping without external planners or manually specified stopping conditions. The predicted graspability distance correlates with real world grasp success, which indicates that the learned representation captures grasp feasibility of object configurations.

Fitting an underlying body model to 3D clothed human assets has been extensively studied, yet most approaches focus on either single-modal inputs such as point clouds or multi-view images alone, often requiring known metric scale—a constraint that is frequently unavailable, particularly for AIGC-generated assets where scale distortion is prevalent. We propose OmniFit, where “Omni” signifies our method’s ability to seamlessly handle diverse multi-modal inputs (e.g., full scans, partial depth, image captures) while remaining scale-agnostic across both real and synthetic assets. Our key innovation is a simple yet effective conditional transformer decoder that directly transforms surface points into dense body landmarks, which are subsequently used for SMPL-X parameter fitting. Additionally, an optional plug-and-play image adapter enriches geometric details with visual cues to address potential incompleteness. We further introduce a dedicated scale predictor to resize subjects into canonical proportions. Remarkably, OmniFit substantially outperforms state-of-the-art methods by 57.1%–80.9% across daily and loose clothing scenarios, making it the first body fitting method to surpass multi-view optimization baselines and the first to achieve millimeter-level accuracy on CAPE and 4D-DRESS benchmarks.

State-of-the-art 3D reconstruction models, whether from visual, range, or both, tend to underperform on thin objects. This is partially due to the small amount of space such objects occupy in RGB images and in 3D point clouds. To test the extent of their errors, we collected the first thin object dataset comprising synchronized RGB images, depth maps, and tactile response maps, where each frame is associated with inertial measurements, camera pose and calibration, and groundtruth depth and segmentation maps obtained from laser scanning of thin objects. We hypothesize that tactile data can aid in the reconstruction of thin objects as their response maps provide local shape and deformation information. Our dataset, termed VisTa3D, comprises of 387 scenes covering 70 thin objects over 17 environments. We benchmarked current 3D reconstruction models on VisTa3D and found that, indeed, they exhibit low fidelity on thin objects. To test if tactile data can help, we introduce the first visual-range-tactile 3D reconstruction model as a baseline. Code and data: https://huggingface.co/datasets/shaniaguo/VisTa3D.

Probe-based Confocal Laser Endomicroscopy (pCLE) enables real-time in vivo optical biopsies. However, acquiring large-scale, high-quality pCLE images is prohibitively expensive and technically complex, causing severe training data scarcity. Furthermore, when applied to pCLE image generation, existing cross-modal translation models consistently introduce artifacts and cellular structure distortions. To address this, we propose Y-diff, a novel generative framework translating widely accessible H&E-stained pathology images into the scarce pCLE modality with high fidelity, providing robust data support for computational pathology. Y-diff innovatively introduces a knowledge distillation mechanism into this task paradigm, efficiently decoupling the learning of pCLE-domain textures from that of H&E pathology staining domain spatial structures. Specifically, a Teacher model captures distinct pCLE optical textures via color encoding, while a Student model extracts and distills this knowledge under strict H&E spatial constraints. Extensive experiments demonstrate that Y-diff consistently outperforms baselines, while significantly alleviating conventional translation artifacts and structural distortions. What’s more, the synthesized data further expands the scale of this rare modality, reducing mean absolute percentage error to 13.84% in downstream cell counting tasks. Source code is available at: https://github.com/Hdw2agon/Y_diff.

3D point cloud semantic segmentation remains a fundamental challenge for autonomous driving and robotic perception. While Transformerbased architectures have achieved significant progress by capturing longrange dependencies, existing methods rely on deterministic point-wise predictions that fail to maintain semantic consistency across geometric voids such as occlusions and sparse boundaries. This can be attributed to the isotropic nature of conventional discrete feature aggregation, which lacks the directional awareness necessary to propagate reliable information across these discrete gaps. To address this, we propose Uncertainty-driven Gaussian Sphere Propagation (UGSP), a framework that transitions from discrete point processing to continuous geometric field reconstruction. By leveraging approximate Bayesian inference to identify reliable semantic anchors and high-uncertainty voids, a Spherical Harmonics (SH)-based aggregation mechanism is introduced that models the local scene as a collection of anisotropic Gaussian spheres. This approach enables the interpolation of semantic information along continuous spatial paths, allowing for direction-aware semantic propagation that effectively recovers structural integrity in high-uncertainty regions. Extensive experiments on indoor and outdoor benchmarks demonstrate the superiority of UGSP in semantically ambiguous and geometrically complex scenarios. The source code is publicly available at https: //github.com/LENGYI1221/UGSP.

What does it take to build a visual reasoner that works across charts, science, spatial understanding, and open-ended tasks? The strongest vision-language models (VLMs) suggest that broad visual reasoning is within reach, yet their closed data and reinforcement learning (RL) pipelines make their gains difficult to study, reproduce, or extend. We introduce Vero, a family of fully open VLMs that match or exceed existing open-weight models across diverse visual reasoning tasks. We scale RL data and rewards across six broad task categories, constructing Vero-600K, a 600K-sample dataset from 59 datasets, and designing task-routed rewards that handle heterogeneous answers. Across VeroEval, our 30-benchmark suite, Vero-600K outperforms existing RL datasets under controlled comparisons. Applied to five starting models, Vero variants gain 2.9–5.4 points on average over their initial models. Notably, Vero-Qwen3I-8B, trained on the Instruct model, surpasses Qwen3-VL-8B-Thinking by 3.8 points on average without additional distillation. Systematic ablations reveal that different task categories elicit distinct reasoning patterns and that broad gains depend on learning them jointly rather than in isolation. All data, code, and models are released.

Zero-shot Human-Object Interaction (HOI) detection has achieved significant progress by leveraging the open-vocabulary capabilities of Vision-Language Models. However, the generalization of existing methods is severely restricted by a two-fold representational bottleneck. First, at the feature level, essential fine-grained interaction cues are frequently overshadowed by high-level global semantics. Second, at the decision level, models exhibit a strong background shortcut bias, often relying on scene co-occurrences rather than the actual humanobject interactions. To address these challenges, we propose a novel zeroshot HOI detection framework, DRIC, which is explicitly formulated for Disentangling and Reusing Interaction Cues. Specifically, we introduce the Context-Orthogonal Residual Disentangling (CORD) module, which projects features onto a global prototype subspace to decouple fine-grained residual details from global semantic biases, subsequently reintegrating them for enhanced local discriminability. Furthermore, the proposed Stratified Local Fusion (SLF) mechanism mitigates background shortcuts by aggregating multi-depth features from the image encoder of CLIP. This recovers essential contact-level details, guiding the model toward actual interactions. Extensive experiments demonstrate that DRIC consistently outperforms state-of-the-art methods across zero-shot evaluation settings, validating the robustness of the proposed framework in handling interaction ambiguity and cross-scene transfer.

Unpaired image dehazing models cast the restoration process as unsupervised domain translation, effectively avoiding the rigid physical assumptions of traditional methods. However, these methods frequently suffer from severe semantic hallucinations and content distortion. This stems from the significant semantic gap between unpaired clean and hazy domains, forcing the discriminator to overfit to high-level semantics rather than low-level clarity. In this paper, we propose Degradation-Agnostic Clarity Learning (DCL), a novel paradigm that steers the network to learn genuine low-level clarity. We introduce two core insights: (1) Explicitly injecting low-level degradations provides strong surface statistical shortcuts that effectively suppress high-level semantic interference; (2) Clarity is a relative concept, and various degraded variants of a clean image share a common intrinsic clarity direction. To realize this, we innovatively formulate discriminator training as a Multi-Task Learning (MTL) problem. First, we employ Adversarial AutoAugment to dynamically mine unlearned, diverse degraded clean samples by maximizing the discriminator’s loss. Second, we utilize Pareto Optimization to extract the shared common descent direction across degraded variants, thereby preventing the model from overfitting to any specific degradation type. Furthermore, we extend our core hypothesis to an Enhance-to-Resist self-supervised texture enhancement paradigm that benefits both dehazing and lowlight enhancement tasks. Extensive experiments on real-world datasets demonstrate that DCL significantly suppresses semantic hallucinations and achieves state-of-the-art performance in unpaired dehazing.

As video corpora continue to expand in both scale and task complexity, there is increasing demand for approaches that retrieve relevant videos from large-scale corpora (inter-video reasoning) and subsequently perform fine-grained, query-conditioned tasks (intra-video reasoning) within the retrieved content, such as temporal grounding. However, existing approaches typically treat retrieval as a preprocessing step, and consequently, when the initial retrieval fails, there is no mechanism to refine the search, leading to the failure of subsequent fine-grained intra-video reasoning. Moreover, while recent agentic frameworks have advanced video understanding, they typically assume that the queryrelevant video is already given, focusing exclusively on intra-video reasoning tasks. To address these limitations, we propose VideoSearch-R1, an agentic framework for iterative video retrieval and reasoning through multi-turn interaction with a video search engine. Specifically, we introduce Soft Query Refinement (SQR) to refine search query tokens in a continuous latent space rather than rewriting queries in the discrete text space, enabling more efficient and fine-grained adjustments. SQR and its reasoning process are trained using Group Relative Policy Optimization (GRPO), guided by task-level reward signals derived from retrieval and downstream tasks. Building upon this, VideoSearch-R1 achieves stateof-the-art performance across three datasets on Video Corpus Moment Retrieval (VCMR), iteratively retrieving videos from large-scale corpora, refining search queries, and performing precise query-conditioned temporal grounding within the retrieved content. Our analyses show that SQR effectively refines the original query, requiring significantly fewer generated tokens than explicit text-level query refinement. Code and model checkpoints are publicly available at mlvlab.github.io/VideoSearch-R1.

Geometric reconstruction of indoor scenes poses unique challenges due to their inherent characteristics, such as large textureless surfaces, complex occlusions, and intricate geometric structures. While recent 2D/3D Gaussian Splatting-based methods have achieved impressive performance in general scene reconstruction, their direct application to indoor settings often leads to noisy, incomplete, or geometrically inconsistent results. In this paper, we propose a novel framework IndoorSplat to bridge this gap and enable high-quality indoor scene geometry reconstruction based on 2D Gaussian Splatting (2DGS). To this end, we first revisit the importance of Gaussian initialization in the indoor scene reconstruction process and introduce a carefully designed dense initialization strategy. This strategy leverages local geometric priors to densify point clouds estimated by an advanced feed-forward model and utilizes them to initialize the structure of Gaussian primitives. In addition, to better capture the rich structural details in indoor scenes, we introduce an iterative adaptive densification strategy that accelerates convergence in complex regions by applying more aggressive splitting and cloning policies to Gaussians with consistent activation. Extensive experiments demonstrate that our method achieves state-of-the-art performance by producing more geometrically accurate and complete surface reconstructions compared to existing approaches.

Vision-language models increasingly support everyday communication and decision-making from images. Their outputs shape how people talk about disability and can affect how people with disabilities are treated. Disability is common, but it is almost entirely absent from vision-side, controlled evaluations of model behavior. We ask a simple question: when a prompt is not about physical ability, should a model change its judgment just because visible limb-deficiency cues are present? A reliable model should base judgments and descriptions on task-relevant visual evidence, not on the mere presence of a disability cue. We introduce InclusiveCFImageBias, a benchmark of 924 images and more than 22k image-level queries. Each example contains two image versions that show the same person in the same scene, with and without visible limb-deficiency cues. We keep the background and activity fixed, and only edit cues such as a prosthesis or a residual limb. We query both versions with the same prompt and compare the outputs directly. Our evaluation spans both open-weight families (Qwen3-VL, DeepSeek-VL2, Gemma-3, and Ministral-3) and proprietary models (GPT-5 and Gemini-2.5). Many models shift both structured judgments and open-ended wording when limb-deficiency cues are visible. In our tests, these cues often push decisions toward more affirmative answers and higher ratings, and they make free responses less neutral and more evaluative. Higher scores are not automatically fairer. The key issue is that outputs change under a minimal, controlled visual substitution. InclusiveCFImageBias makes this cue-driven instability measurable. More importantly, it aims to draw community attention to disability-related VLM bias and provide a clear target for building more consistent disability-facing multimodal systems.