ECCV 2026 Papers

Video generation models are increasingly capable of producing realistic videos, but they still struggle to generate videos that follow basic physical laws. Compounding this is a lack of reliable granular evaluation methods for localizing and specifying physical law violations in videos. We address this by introducing Physics Question Scene Graph (PQSG), a hierarchical question-based evaluation pipeline. PQSG evaluates generated videos by checking their faithfulness to a prompt across objects, actions, and adherence to physical laws using a graph-based hierarchy of questions generated by a vision-language model (VLM), guided by high-quality in-context examples. By representing questions as a graph, PQSG introduces logical dependencies within questions, ensuring that each query is contextually valid. Moreover, PQSG provides granular assessments of which qualities of the video violate physical plausibility constraints. We validate PQSG by creating FinePhyEval, a dataset with physics-based prompts and corresponding generated videos from diverse state-of-the-art video generation models (Sora 2, Veo 3, and Wan 2.1), with each video annotated across multiple categories by humans. Using FinePhyEval, we measure the correlation between PQSG’s fine-grained scores and human judgments, showing higher overall correlations than prior work. We also find that PQSG ranks closedsource models higher than Wan 2.1 on physical realism. Lastly, we show that the annotations we provide in FinePhyEval can also be used for subtask evaluation: we benchmark two strong VLMs on generating and answering questions, finding that while models can create human-like questions, they still fall short of human performance in answering them.

Multimodal Large Language Models (MLLMs) have shown promising capabilities in generating Scalable Vector Graphics (SVG) via direct code synthesis. However, existing paradigms typically adopt an open-loop “blind drawing” approach, where models generate symbolic code sequences without perceiving intermediate visual outcomes. This methodology severely underutilizes the powerful visual priors embedded in MLLMs’ vision encoders, treating SVG generation as a disjointed textual sequence modeling task rather than an integrated visuo-spatial one. Consequently, models struggle to reason about partial canvas states and implicit occlusion relationships, which are visually explicit but textually ambiguous. To bridge this gap, we propose Render-in-the-Loop, a novel generation paradigm that reformulates SVG synthesis as a stepwise, visual-context-aware process. By rendering intermediate code states into a cumulative canvas, the model explicitly observes the evolving visual context at each step, leveraging on-the-fly feedback to guide subsequent generation. However, we demonstrate that applying this visual loop naively to off-the-shelf models is suboptimal due to their inability to leverage incremental visual-code mappings. To address this, we first utilize fine-grained path decomposition to construct dense multi-step visual trajectories, and then introduce a Visual Self-Feedback (VSF) training strategy to condition the next primitive generation on intermediate visual states. Furthermore, a Render-and-Verify (RaV) inference mechanism is proposed to effectively filter degenerate and redundant primitives. Our framework, instantiated on a multimodal foundation model, outperforms strong open-weight baselines on the standard MMSVGBench. This result highlights the remarkable data efficiency and generalization capability of our Render-in-the-Loop paradigm for both Text-to-SVG and Image-toSVG tasks.

Few-shot classification from a frozen backbone collapses when the target domain lies far from the pre-training distribution, because global embeddings map unrelated classes to nearly identical CLS tokens. Local patch descriptors are more resilient: an edge or texture gradient activates similar features whether it appears in a photograph or an X-ray. Yet part-level matching remains largely unexplored for few-shot recognition. The few methods that attempt it formulate matching as optimal transport or bipartite assignment, requiring iterative solvers whose cost grows rapidly with shots. We take a different approach. Instead of solving a matching optimization, we formulate patch-based classification as probabilistic evidence accumulation under a discriminative variational model. Each query patch carries a Bernoulli latent variable; minimizing the resulting assignment free energy yields a closed-form sigmoid gate and a softplus evidence score whose per-patch contribution saturates at −log(1−π), bounding clutter influence without iterative solvers or learned parameters. We evaluate extensively across twelve datasets spanning satellite, medical, fine-grained, and natural-image domains. Without any gradient updates, VPG outperforms the best fine-tuned baseline on five of ten dataset-settings and exceeds the previous training-free state of the art by +5.35 points on CDFSL. With lightweight episodic fine-tuning, VPG reaches 71.06% on CDFSL and 81.8% on four unseen Meta-Dataset domains.

Millimeter-wave (mmWave) radar provides privacy-preserving sensing and is valuable for human action recognition (HAR). Existing mmWave point cloud datasets are limited in scale and mostly collected under homogeneous single-source settings, preventing current methods from handling real-world distribution shifts caused by heterogeneous radar sources, such as different devices and frequency bands. To address this, we introduce UniMM-HAR, the largest and first mmWave point cloud HAR dataset for heterogeneous multi-source scenarios, standardizing three distinct radar configurations to realistically evaluate crosssource generalization. We further propose the Doppler-aware Point Cloud Network (DAP-Net) to tackle heterogeneity challenges. DAP-Net enhances intra-modal representations and performs cross-modal alignment to learn source-invariant action semantics. Leveraging action-consistent spatio-temporal Doppler patterns as anchors, the Dual-space Doppler Reparameterization (D²R) module performs sample-adaptive geometric densification and Doppler-guided feature recalibration, while the Text Alignment Module (TAM) provides stable semantic anchors via a pretrained textual space. Experiments show that DAP-Net significantly outperforms existing methods under heterogeneous radar settings, achieving state-of-the-art accuracy and strong cross-source robustness. Code and dataset are publicly available at https://github.com/jolin830/DAP-Net.

The generation of plausible layouts is a critical step in textdriven 3D scene synthesis. While Large Language Models (LLMs) have shown promise in high-level organization in this task, their native finegrained spatial placement remains a significant bottleneck, often requiring in-context guidance or complex agentic reflections to fix errors. This paper focuses directly on strengthening the core spatial ability by finetuning LLM to generate structured scene code (codematic indoor scene generation). We identify and address two fundamental issues hindering the LLM from doing this: data scarcity and poor instruction following. First, to mitigate data scarcity, we consolidate multiple heterogeneous datasets, spanning procedural, real-world, and professionally designed scenes, into a unified, large-scale training corpus. This dataset comprises 52K rooms and 280k data pairs, featuring diverse descriptions at both the scene and object levels. Second, even with abundant data, LLMs exhibit poor instruction following due to an information imbalance, where the tokenheavy layout history overwhelms the concise user prompt. To resolve this, we propose SceneSpinner, a framework that introduces a language-based planning stage to provide high-level reasoning and a novel Conditional Mutual Information (CMI) regularization to explicitly force the model to focus on user instructions. Experiments demonstrate that our approach significantly improves the ability of LLMs to generate plausible, diverse, and instruction-aligned 3D layouts.

Diffusion models achieve state-of-the-art generative performance but are fundamentally bottlenecked by their slow, iterative sampling process. While diffusion distillation techniques enable high-fidelity, few-step generation, traditional objectives often restrict the student’s performance by anchoring it solely to the teacher. Recent approaches have attempted to break this ceiling by integrating Reinforcement Learning (RL), typically through a simple summation of distillation and RL objectives. In this work, we propose a novel paradigm by re-conceptualizing distribution matching as a reward, denoted as R . This unified perspective bridges the algorithmic gap between Diffusion Matching Distillation (DMD) and RL, providing several primary benefits: (1) Enhanced Optimization Stability: We introduce Group Normalized Distribution Matching (GNDM), which adapts standard RL group normalization to stabilize R estimation. By leveraging group-mean statistics, GNDM establishes a more robust and effective optimization direction. (2) Seamless Reward Integration: Our reward-centric formulation inherently supports adaptive weighting mechanisms, allowing for the fluid combination of DMD with external reward models. (3) Improved Sampling Efficiency: By aligning with RL principles, the framework readily incorporates Importance Sampling (IS), leading to a significant boost in sampling efficiency. Extensive experiments demonstrate that GNDM outperforms vanilla DMD, reducing the FID by 1.87. Furthermore, our multi-reward variant, GNDMR, surpasses existing baselines by striking an optimal balance between aesthetic quality and fidelity, achieving a peak HPS of 30.37 and a low FID-SD of 12.21. Ultimately, R provides a flexible, stable, and efficient framework for real-time, high-fidelity synthesis. Code is available at: https://github.com/Flq2002/Rdm.

We investigate a challenging task of dynamic scene geometry estimation, which requires representing both spatial and temporal features. Typically, existing methods align two features into a unified latent space to model scene geometry. However, this unified paradigm suffers from potentially mismatched representations due to the heterogeneous nature between spatial and temporal features. In this work, we propose 4D-VGGT, a general foundation model with divide-and-conquer spatiotemporal representation for dynamic scene geometry. Our model is divided into three aspects: 1) Multi-setting input. We design an adaptive visual grid that supports input sequences with arbitrary numbers of views and time steps. 2) Multi-level representation. We propose a cross-view global fusion for spatial representation and a cross-time local fusion for temporal representation. 3) Multi-task prediction. We append multiple task-specific heads to spatiotemporal representations, enabling a comprehensive visual geometry estimation for dynamic scenes. Under this unified framework, these components enhance the feature discriminability and application universality of our model for dynamic scenes. In addition, we integrate multiple geometry datasets to train our model and conduct extensive experiments to verify the effectiveness of our method across various tasks on multiple dynamic scene geometry benchmarks. Our code is released at https://github.com/Haonan-Wang-aurora/4D-VGGT.

Despite impressive performance on standard benchmarks, multimodal large language models (MLLMs) face critical challenges in real-world clinical environments where medical images inevitably suffer various quality degradations. Existing benchmarks exhibit two key limitations: (1) absence of large-scale, multidimensional assessment across medical image quality gradients and (2) no systematic confidence calibration analysis. To address these gaps, we present MedQ-Deg, a comprehensive benchmark for evaluating medical MLLMs under image quality degradations. MedQ-Deg provides multi-dimensional evaluation spanning 18 distinct degradation types, 30 fine-grained capability dimensions, and 7 imaging modalities, with 24,894 question-answer pairs. Each degradation is implemented at 3 severity degrees, calibrated by expert radiologists. We further introduce Calibration Shift metric, which quantifies the gap between a model’s prediction-consistency-based perceived confidence and actual performance to assess metacognitive reliability under degradation. Our comprehensive evaluation of 40 mainstream MLLMs reveals several critical findings: (1) overall model performance degrades systematically as degradation severity increases, (2) models exhibit a Dunning-Kruger-like overconfidence pattern, maintaining inappropriately high confidence despite severe accuracy collapse, and (3) models display markedly differentiated behavioral patterns across capability dimensions, imaging modalities, and degradation types. We hope MedQ-Deg drives progress toward medical MLLMs that are robust and trustworthy in real clinical practice.

Multimodal learning of image and tabular data is a challenging problem, especially given the unstructured feature representations involved. We address this by framing feature sequencing as a Column Permutation Problem and showing how reordering columns can reduce dispersion and improve training stability for multimodal learning of image and tabular data. We propose iSyncTab, an approach that enforces these learned sequential constraints through a joint classification and order-consistency objective. We introduce Neural Synchrony-guided Paired Feature Sequencing (NS-PFS), a neuroscience-inspired NeuroAI approach that aligns image and tabular feature clusters through a Hungarian assignment of their synchrony matrix, derived from cluster energy and centroid similarity. The resulting cross-modal feature sequence provides a coherent feature alignment that preserves structural synchrony between modalities. We then introduce an Order-aware Memoryaugmented Transformer (OMT) with a custom loss function that enforces these learned feature sequence constraints. Experiments on multimodal benchmarks demonstrate that systematically derived feature sequences consistently boost predictive performance and model robustness. Our findings underscore feature sequencing as an important inductive bias for structured multimodal learning of image and tabular data. Neural synchrony-driven sequencing enhances cross-modal coherence and representation quality for image and tabular data.

Vision-Language Models (VLMs) incur substantial computational overhead and inference latency due to the large number of visual tokens introduced by high-resolution image and video inputs. Existing parameter-free token compression methods typically rely on token selection or merging, yet they risk discarding substantial visual information or distorting the original representation distribution, resulting in pronounced performance degradation at high compression ratios. In response, we aim to explore a more effective and efficient visual token compression strategy, with a promising direction in the frequency domain. Motivated by the success of frequency-domain transforms in image compression (e.g., JPEG), we systematically analyze the frequency redundancy in visual representations and uncover a nonuniform distribution of semantic information across frequency bands. Building upon this, we introduce Fourier Compressor, an effective, parameter-free, and highly generalizable module that removes redundancy from visual representations within the frequency domain. Implemented via FFT with O(n2 logn) complexity and no additional parameters, Fourier Compressor introduces negligible computational overhead while preserving semantic fidelity. Extensive experiments on imagebased benchmarks demonstrate that our method achieves a favorable performance-efficiency trade-off, retaining over 96% of the original accuracy while reducing inference FLOPs by up to 83.8% and boosting generation speed by 31.2%. It consistently outperforms existing parameterfree methods and even surpasses some parameterized approaches. Importantly, Fourier Compressor generalizes consistently across both LLaVA and Qwen-VL architectures, and further extends to video understanding tasks, highlighting its practical applicability for efficient VLMs.

Existing zero-shot 6D object pose estimation methods predominantly rely on multi-stage pipelines with iterative render-and-compare refinement or coarse-to-fine hypothesis selection, leading to substantial computational overhead and implementation complexity. Generative modeling on geometric manifolds offers an attractive alternative by directly representing pose posteriors. In this work, we present UF0-6D, a Unified Flow-based Zero-shot framework for 6D object pose estimation without refinement. UF0-6D reformulates pose estimation as conditional Riemannian flow matching on SE(3), learning an instance-conditioned posterior via geodesic-consistent bridge velocities and symmetry-aware Riemannian optimal transport supervision. A single ODE integration pass replaces coarse-to-refine cascades and iterative alignment, unifying modelbased and model-free settings via a geometry-conditioning design that maps both CAD renders and reference-view renders into a shared pointtoken representation. Our RGB-D variant achieves a mean AR of 81.2 on BOP benchmark datasets, surpassing the previous state-of-the-art by +6.0 AR, while our RGB variant reaches 70.6 AR at only 0.379s per image, establishing a new accuracy–speed Pareto frontier.

Magnetic Resonance Imaging (MRI) is often acquired with anisotropic resolution to reduce scan time, producing stair-step artifacts along the through-plane direction. In through-plane MRI superresolution, an efficiency–fidelity trade-off arises: feed-forward regressors are fast but oversmooth at large slice-thicknesses, while sampling-based methods improve fidelity at high inference cost. We propose DRIFT, a two-stage thickness-conditioned rectified flow framework for throughplane MRI super-resolution with continuous input slice-thickness. Stage 1 employs an Anatomical Projection Network (APN) to map low-resolution patches to a coarse high-resolution manifold, providing a deterministic anatomical initialization that shortens the residual transport of Stage 2 and stabilizes slice-wise refinement. Stage 2 refines details via rectified flow and introduces a Physics-Aware Difficulty (PAD) metric derived from slice-thickness induced through-plane bandwidth deficit to guide an Adaptive Integration Scheduler (AIS), allocating ODE steps by thickness. A Consistent Endpoint Trajectory Alignment (CETA) loss enforces thickness-consistent reconstructions. Experiments show that DRIFT outperforms super-resolution baselines while reducing inference cost. Code, models, and interactive demos are available at https://yoonseokchoiai.github.io/drift-eccv2026/.

The widespread adoption of consumer 360° cameras has created a growing demand for robust in-the-wild panoramic 3D reconstruction. Existing methods remain fragile on long-horizon trajectories with strong ego-motion and dynamic interference. Quantifying this gap at scale is challenging because available benchmarks are often short, constrained, or captured in controlled settings, limiting evaluation and progress. To address this, we introduce OmniLife360, a benchmark for 3D reconstruction from in-the-wild 360° captures. It spans diverse scenes, environments, and motion dynamics captured by individuals across real-world activities, enabling rigorous evaluation of reconstruction robustness in complex natural conditions. Extensive evaluations reveal that state-ofthe-art methods excel in controlled settings but suffer significant degradation in dynamic scenarios. Motivated by these findings, we propose Omni4DGS, a motion-aware Gaussian Splatting framework that combines motion decomposition with lightweight supervision to better handle motion variations, dynamic distractors, and long-horizon pose drift. Experiments show that Omni4DGS consistently improves reconstruction quality and stability in dynamic 360° scenarios, reducing motion-induced artifacts and producing cleaner renderings under strong scene motion. The benchmark is available at https://huggingface.co/datasets/ AutoLab-SJTU/OmniLife360.

Traditional Image Aesthetic Assessment (IAA) methods mainly rely on regressing absolute Mean Opinion Scores (MOS). However, such a paradigm overlooks the inherently dynamic nature of human aesthetic perception, which relies on subconscious comparison against implicit visual references. Consequently, the lack of causal reasoning regarding aesthetic differences prevents models from learning generalizable aesthetic principles, thus limiting their generalization across diverse scenarios. In this work, we rethink the IAA task and propose Relative Editinduced Difference Aesthetic learning (RED-Aes), a novel framework that leverages controllable image editing models to simulate the human aesthetic reasoning process. Instead of fitting absolute score distributions, RED-Aes explicitly learns the visual factors that drive aesthetic changes. To support this paradigm, we construct the RED-20k dataset, which comprises editing-based image pairs, quantitative aesthetic differences, and Chain-of-Thought (CoT) reasoning. Furthermore, we introduce a three-stage training strategy guided by a relative ranking consistency reward, optimizing the model solely via relative supervision. Extensive experiments demonstrate that RED-Aes achieves state-of-the-art performance on multiple public benchmarks, exhibiting superior generalization capabilities.

Spatial visual perception is a fundamental requirement in physical-world applications like autonomous driving and robotic manipulation, driven by the need to interact with 3D environments. Capturing pixel-aligned metric depth using RGB-D cameras would be the most viable way, yet it usually faces obstacles posed by hardware limitations and challenging imaging conditions, especially in the presence of specular or texture-less surfaces. In this work, we argue that the inaccuracies from depth sensors can be viewed as “masked” signals that inherently reflect underlying geometric ambiguities. Building on this motivation, we present MDM, a depth completion model which leverages visual context to refine depth maps through masked depth modeling and incorporates an automated data curation pipeline for scalable training. It is encouraging to see that our model outperforms top-tier RGB-D cameras in terms of both depth precision and pixel coverage. Experimental results on a range of downstream tasks further suggest that MDM offers an aligned latent representation across RGB and depth modalities. Code, checkpoints, and 3M RGB-depth pairs (including 2M real data and 1M simulated data): https://github.com/robbyant/lingbot-depth.

3D sequential affordance reasoning requires a model to locate a logical series of functional parts based on language instructions. Current Multimodal Large Language Models (MLLMs) tackle this by generating a special <SEG> token to guide the final mask prediction. However, compressing complex reasoning into a single token loses crucial low-level structural details. This leads to severe geometric ambiguity, causing the predicted masks to have inaccurate boundaries. To solve this, we propose the Geo-DPO framework, which addresses the ambiguity from two directions. First, we introduce the Hierarchical Geometry Adapter (HGA) to structurally inject multi-scale 3D physical features back into the <SEG> token. Second, we propose Contrastive Affordance Preference Optimization (CAPO), a reward-based training objective that forces the segmentation to align with real geometric boundaries. Extensive experiments on the standard SeqAfford dataset demonstrate that our method achieves state-of-the-art performance, generating highly precise masks and effectively resolving geometric ambiguity. Code is available here.

Universal person re-identification (ReID) aims to retrieve pedestrian identities across diverse real-world scenarios, including severe occlusions, clothing changes, and cross-modality shifts, within a unified model. However, existing 2D representations fundamentally struggle with spatial ambiguities due to a lack of depth and topological awareness, while naively introducing monocular 3D priors often causes severe negative transfer due to geometric estimation noise under extreme visual degradation. To safely harness the clothing-invariant and canonical structural properties of 3D geometry, we propose UniGeo, a Universal Monocular 3D-Enhanced ReID framework driven by a ConsistencyAware Reliability Gate and Dual-Stream Residual Fusion. Specifically, the processing of 3D information is strategically decoupled into geometric extraction and dynamic utilization. To provide pure structural compensation, we project monocular 3D parameters into kinematic joint representations, explicitly capturing instance-level geometric topology to resolve appearance-based ambiguities. To robustly incorporate these cues without perturbing the reliable 2D feature space, we isolate the 3D prior as a late-stage structural residual; modulated by the consistency-aware gate, this mechanism adaptively filters geometric noise and enables controlled fallback to the pure 2D baseline. Extensive experiments show that our method improves challenging, structure-sensitive scenarios while preserving competitive performance on clean domains. Code is available at https://github.com/BohanSu/UniGeo.

As web agents increasingly demonstrate capabilities in automated task execution, the development of robust evaluation frameworks for assessing their navigation and task completion performance has emerged as a critical research priority. However, existing benchmarks exhibit several fundamental limitations. First, they suffer from insufficient scale and limited domain diversity, thereby constraining comprehensive evaluation of cross-domain generalization. Second, prevailing LLM-asJudge evaluation methodologies inadequately capture fine-grained interaction semantics, particularly regarding precise query formulation and filtering operations. Third, current benchmarks predominantly emphasize navigation success metrics while neglecting critical requirements for realworld deployment scenarios. To address these limitations, we introduce WebRetriever, a large-scale benchmark encompassing 800 websites and 1,550 tasks across diverse domains, including consumer, professional, and enterprise sectors, with comprehensive coverage of user intent patterns. We propose NavEval (Navigation Evaluation), a novel LLM-as-Judge framework that leverages rich interaction context beyond visual screenshots, achieving state-of-the-art alignment with human judgment across multiple evaluation datasets. Furthermore, we establish three complementary evaluation protocols that collectively provide holistic assessment of web agent capabilities: navigation proficiency, knowledge-assisted interaction, and end-to-end task completion with information extraction. Extensive experimental analysis reveals substantial performance disparities across evaluation protocols, demonstrating that navigation success alone serves as an insufficient predictor of real-world application effectiveness. WebRetriever delivers fine-grained diagnostic insights into agent capabilities and establishes a rigorous foundation for advancing web agent research and development.

Autoregressive (AR) visual generation has emerged as a powerful paradigm for image and multimodal synthesis, owing to its scalability and generality. However, existing AR image generation suffers from severe memory bottlenecks due to the need to cache all previously generated visual tokens during decoding, leading to both high storage requirements and low throughput. In this paper, we introduce LineAR, a novel, training-free progressive key-value (KV) cache compression pipeline for AR image generation. By exploiting the intrinsic characteristics of visual attention, LineAR manages the cache from a 2D line-level perspective, preserving the visual dependency regions while progressively evicting less-informative tokens under inter-line attention guidance, leveraging the spatial continuity prior of images to ensure that each eviction step is harmless to subsequent line generation. Experiments across seven AR image generation models validate that LineAR achieves lossless or even improved generation quality with only a few cached lines. It improves ImageNet FID from 2.77 to 2.68 on LlamaGen-XL and COCO FID from 23.85 to 22.86 on Janus-Pro-1B at 1/6 budget ratio, and also improves DPG and HPSv2.1 scores on Lumina-mGPT-768 with just 1/8 KV cache. Additionally, LineAR achieves significant memory and throughput gains across devices, e.g., up to 67.06% memory reduction and 4.17× speedup on LlamaGen-XL, and 73.65% memory reduction and 4.34× speedup on Janus-Pro-1B, evaluated on NVIDIA RTX PRO 6000 GPUs.