ECCV 2026 Papers

Composed Image Retrieval (CIR), a task towards personalizing visual intelligence, aims to retrieve a target image from a database based on users’ multimodal query, which contains a reference image and a modification text. The text specifies how to alter the reference image to form a “mental image”, based on which CIR should find the target image in the database. The fundamental challenge of CIR is that this “mental image” is not physically available and is only implicitly defined by the query. The contemporary literature pursues zero-shot methods and uses a Large Multimodal Model (LMM) to generate a textual description for a given multimodal query, and then employs a Vision-Language Model (VLM) for textual-visual matching to search for the target image. In contrast, we address CIR from first principles by directly generating the “mental image” for more accurate matching. Particularly, we prompt an LMM to generate a “mental image” for a given multimodal query and propose to use this “mental image” to search for the target image. As the “mental image” has a synthetic-to-real domain gap with real images, we also generate a synthetic counterpart for each real image in the database to facilitate matching. In this sense, our method uses LMM to construct a “paracosm”, where it matches the multimodal query and database images. Hence, we call this method Paracosm. Notably, Paracosm is a training-free zero-shot CIR method. It significantly outperforms existing zero-shot methods on challenging benchmarks, achieving state-of-the-art performance for zero-shot CIR.

Vision-Language Models (VLMs) have advanced Visual Document Retrieval (VDR) by enabling OCR-free, multi-vector representations. However, these models generate excessive tokens to preserve dense visual details, which severely limits retrieval scalability. Existing compression methods mitigate this overhead but rely on isolated heuristics that fail to balance token importance, semantic redundancy, and coverage capacity. We propose a visual token compression framework formulated as an Unbalanced Optimal Transport (UOT) problem. Our approach unifies these three objectives into a single alternating optimization process. Furthermore, we theoretically show that minimizing the transport cost upper-bounds the MaxSim retrieval error, ensuring ranking stability. Extensive experiments on the ViDoRe V1 and V3 benchmarks demonstrate that our method preserves fine-grained document semantics and improves the efficiency-effectiveness trade-off over state-of-the-art baselines. Code is available at https://github.com/shhhhhyy/UnbalancedOptimal-Transport-for-EVDR.

Long video understanding is fundamentally constrained by massive data volume and high information density. While agent-based methods mitigate context length bottlenecks, they predominantly rely on static, "one-size-fits-all" multi-agent systems (MAS). These rigid architectures are computationally redundant and lack the flexibility to handle diverse, instance-level reasoning demands. To address this, we propose DyMAC (Dynamic MAS Construction), a novel framework that dynamically generates instance-specific MAS for each video-question pair. By logically abstracting the MAS into a directed computational graph via automated code generation, DyMAC exhibits exceptional adaptability, seamlessly tailoring its collaborative architecture to the specific complexity of each input. Furthermore, we employ a streamlined two-stage training strategy—Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR)—to enhance the model’s dynamic construction and decision-making capabilities. Extensive experiments across four mainstream benchmarks demonstrate that DyMAC achieves a new state-of-the-art, outperforming the second-best method by 3.3% on average. Notably, it significantly improves inference efficiency compared to existing static approaches, further highlighting its superior adaptability and robustness in complex video scenarios.

Multimodal large models are increasingly used to generate scalable vector graphics (SVG), but reliable evaluation remains underexplored. Existing protocols are often code-centric or borrow raster-image metrics after rendering SVGs, which fail to reflect human perception and overlook SVG-specific qualities such as geometry and spatial composition. We introduce SVGEval, a vision-grounded multimodal benchmark for human-aligned SVG quality assessment. SVGEval explicitly incorporates visual renderings to evaluate whether models can judge the rendered outcome rather than only inspect SVG code, and provides highquality annotations obtained via multi-round human labeling with expert refinement. Systematic evaluations across representative multimodal models reveal a clear gap: models perform relatively well on semantic alignment and aesthetics, yet struggle on geometryand layout-related judgments. Building on SVGEval, we train an explainable SVG quality scorer that outputs multi-aspect scores with textual rationales. Ablations show that explicit visual grounding and reasoning supervision are crucial, especially for spatial and geometric assessment. SVGEval offers a reliable testbed and practical scorer for evaluating and improving SVG generation in the era of multimodal models.

While the shortage of explicit action data limits VisionLanguage-Action (VLA) models, human action videos offer a scalable yet unlabeled data source. A critical challenge in utilizing large-scale human video datasets lies in transforming visual signals into ontologyindependent representations, known as latent actions. However, the capacity of latent action representation to derive robust control from visual observations has yet to be rigorously evaluated. We introduce the Latent Action Representation Yielding (LARY) Benchmark, a unified framework for evaluating latent action representations on both highlevel semantic actions (what to do) and low-level robotic control (how to do). The comprehensively curated dataset encompasses over one million videos (1,000 hours) spanning 151 action categories, alongside 620K image pairs and 595K motion trajectories across diverse embodiments and environments. Our experiments reveal two crucial insights: (i) General visual foundation models, trained without any action supervision, consistently outperform specialized embodied latent action models. (ii) Latent-based visual space is fundamentally better aligned to physical action space than pixel-based space. These results suggest that general visual representations inherently encode action-relevant knowledge for physical control, and that semantic-level abstraction serves as a fundamentally more effective pathway from vision to action than pixel-level reconstruction.

Deploying 3D point cloud analysis in privacy-sensitive and resource-constrained settings faces two coupled barriers: data cannot be centralized for training, and the trained model must run on limited edge hardware. We present a multi-seed benchmark that jointly evaluates federated learning (FL) and knowledge distillation (KD) for 3D point cloud classification. It spans thirteen FL algorithms and ten KD objectives, supporting their full 130-pair teacher–objective cross-product per dataset; every standardized configuration is repeated over three random seeds for 504 training runs in total, with the complete combined grid evaluated at multi-seed scale on the clinical dataset. We characterize federated degradation and the combined-pipeline pitfall on ModelNet40, then validate them on a real-world clinical craniosynostosis dataset of patient head shapes, where the privacy and edge-deployment stakes are concrete. We report three findings. First, under extreme non-independent and identically distributed (non-IID) label skew, standalone FL degrades sharply: on ModelNet40 the strongest method reaches only 76.32% against a 92.26% centralized reference, on the clinical data the best reaches 75.83% against 100%, and the four server-side optimizers collapse to near the chance level; the best algorithm differs by dataset, so none is universally robust. Second, distillation compresses the teacher into a student 74.51% smaller and roughly twice as fast at inference, with five of the seven objectives evaluated on ModelNet40 matching or surpassing the 92.44% teacher. Third, the combined pipeline exposes an evaluation pitfall: when distillation keeps a hard-label cross-entropy term on a labeled proxy split, a collapsed federated teacher at 8.50% paired with Logit-MSE still yields a 92.94% student. This 84.4-point gap reflects the proxy labels rather than the federated model, and the hard-label term reuses the very labels whose privacy motivated federation. Objectives without a hard-label term instead track teacher quality (r ≈ 0.99 on the clinical grid) and collapse when the teacher does. We therefore recommend evaluating FL-KD pipelines with label-free distillation, so that the reported accuracy reflects the federated teacher rather than the proxy.

Audio-visual deception detection aims to predict whether a person is lying by integrating visual and acoustic modalities, which has two main challenges: 1) the modality conflict problem and 2) heterogeneous cue representation difficulty. However, existing approaches 1) often overlook the differences across modalities for different individuals; and 2) typically rely on a single encoder to handle diverse and individual-specific cues, which limits models’ representation capacity for heterogeneous cues. To address these challenges, we propose MoMCE, a novel model with mixture of modality and cue experts for deception detection. It consists of two key components: 1) Prompt-aware Mixture of Modality Experts, which employs a learnable prompt routing mechanism to generate adaptive instance-aware modality weight distributions for dynamic modality adjustment. In addition, we propose a consistency-aware expert weighting loss. For samples with high crossmodal consistency, it encourages balanced contributions across modalities. In contrast, for samples with strong conflicts, it reduces the entropy of the modality weight distribution to focus on more reliable modalities. 2) Prompt-aware Mixture of Cue Experts, which captures heterogeneous and diverse deceptive cues within each modality. This module introduces multiple experts with distinct semantic biases on top of a shared backbone to model different deceptive patterns. Additionally, we introduce a cue expert diversity loss to balance learning across multiple cue experts, promoting effective representation of diverse deceptive cues. Extensive experiments demonstrate that MoMCE adapts to variations in both cross-modal contributions and cue heterogeneity, achieving substantial improvements in deception detection performance. The code is available at this link.

Aerial–Ground Person Re-Identification (AGPReID) matches identities across unmanned aerial vehicles (UAVs) and ground cameras but suffers from extreme viewpoint gaps. Most existing methods rely on coarse binary aerial/ground labels and rigid orthogonality to disentangle view and identity features. We identify an important limitation of prior methods: the oversimplified binary-label design cannot fully capture continuous aerial viewpoint shifts. Moreover, rigid orthogonality constraints may further suppress identity cues. To address these issues, we propose Query-guided View-aware Adaptive Modulation (QVAM) for identity-preserving cross-view matching. Specifically, a View-aware Decoder (VAD) distills fine-grained viewpoint cues from patch tokens using learnable view queries. Guided by these cues, Adaptive Feature Modulation (AFM) predicts query-conditioned masks to suppress view-biased responses while preserving identity discrimination. A Cross-View Prototype Alignment (CVPA) loss further aligns modulated features at batch and memory levels with dual-view memory banks. Extensive experiments on AG-ReID, AG-ReIDv2, and CARGO show that QVAM achieves stateof-the-art performance, improving the previous best results by +10.26% Rank-1/+10.83% mAP on CARGO-ALL and +2.62% Rank-1/+2.37% mAP on AG-ReID A→G. The code is available at https://github.com/ Sakuraandroxy/QVAM.

Recently, vision foundation models (VFMs) have demonstrated strong power for versatile downstream task analysis. This paper starts from the problem of how to efficiently deploy VFMs at the cloud to support various user requests with arbitrary multi-task combinations that end in any scalable fashion. To achieve feature data transmission between frontend-cloud, traditional feature coding employs the direct compression-reconstruction paradigm. However, this can result in potential redundancy: (1) task-irrelevant information, and (2) repeated coding of cross-task shared knowledge, leading to an inflexible and redundant scheme. To this end, we propose a novel feature coding paradigm, termed reconfigurable multi-task feature compression, which aims to efficiently and adaptively compress the intermediate features to support the requested targeted tasks. Correspondingly, we propose a unified Reconfigurable Feature Compression framework, RFC, by feature factorization and recomposition. Specifically, the original feature is first factorized into multiple task-specific features with light-weight adapters. Then to efficiently compress these separate features, we consolidate them in a task-conditional auto-regressive manner, leveraging previously encoded task features as hyperprior conditions to reduce the redundancy of shared information in the current task feature. At the cloud side, a task-attentive recomposition module is further developed to fulfill the multi-task inference power within a single forward pass. Finally, we construct a comprehensive benchmark of reconfigurable feature compression to verify the effectiveness of RFC. Our project page can be found at https://jhang2020.github.io/Projects/RFC/RFC.html.

Multimodal Large Language Models have demonstrated impressive video understanding, yet their ability to reason over long-form narratives is often masked by visual-centric evaluations and inefficient context processing. Existing benchmarks over-rely on visual heuristics while marginalizing auditory cues, effectively reducing models to "silent observers" that bypass genuine cross-modal reasoning. Moreover, standard dense sampling creates an evidence–context trade-off: increasing frames to capture evidence inevitably leads to attention distraction and token explosion. To bridge these gaps, we present VideoHolmesV2, a novel benchmark designed for Deep Audio-Visual Coupling. Unlike previous works, it enforces an Evidence-Based Evaluation, requiring models to justify answers with precise spatio-temporal audio-visual evidence, thereby reducing confounding effects of guessing and hallucinated evidence. To support this, we introduce: (1) a MultiModel Cross-Verification pipeline to ensure task rigor; (2) a Spatiotemporal Evidence-Aware Metric for fine-grained calibration. Furthermore, we propose an Audio-Text Guided Token Compression framework. By fusing task intent with auditory anchors, our method distills highvalue reasoning cues to mitigate long-context noise. In our evaluation, even strong proprietary models achieve below 60% accuracy, while our approach outperforms over comparable open-source omni-models.

Open-World Object Detection (OWOD) requires detectors to identify previously unseen objects as unknown and incrementally incorporate them into the set of known categories, while preserving previously acquired knowledge. Existing frameworks rely heavily on exemplar replay to mitigate catastrophic forgetting, but in some real applications, storing raw data conflicts with data access restrictions and leads to data exposure risks, while incurring significant memory overhead. In this paper, we propose REAL-OW, a novel rehearsal-free framework that decouples incremental knowledge through a collaborative adapter architecture based on Low-Rank Adaptation (LoRA). Specifically, we deploy General Adapters (GAs) in the backbone to enable the significance-aware refinement of cross-task universal representations, while Specific Adapters (SAs) in the decoder provide orthogonal storage for task-specific expertise. To resolve representation drift in objectness modeling under rehearsal-free constraints, we introduce Dual-Stage Objectness Modeling (DSOM), which alternates between feature aggregation and boundary consolidation to stabilize objectness distributions while maintaining the separation between known and unknown categories. Furthermore, DSOM is supported by a Calibrated Gaussian Negative Log-Likelihood (CG-NLL) distance tailored for the dispersed feature distributions inherent in rehearsal-free settings. Extensive evaluations demonstrate that REAL-OW achieves state-of-the-art performance, surpassing existing exemplar replay methods in both detection precision and unknown discovery. Our approach establishes a new baseline for rehearsal-free OWOD.

Egocentric AI assistants seek to infer user actions and intentions from long first-person videos to facilitate real-world tasks ranging from daily activities to industrial operations. However, collecting large-scale annotated egocentric videos that capture variations in objects, tools, environments and procedures is costly and impractical, motivating the need for data-efficient learning. Existing data augmentation approaches either offer limited visual diversity through simple transformations or rely on computationally expensive generative models that may introduce artifacts. We propose LogFA (Local-global Featurespace Augmentation), a novel framework that performs data augmentation directly in the feature space for Temporal Action Segmentation (TAS). LogFA efficiently generates semantically consistent variations without pixel-level synthesis by modifying pre-extracted video features at both local and global levels. Locally, LogFA leverages visionlanguage models with a Prompt-based Feature Enhancement strategy to create diverse action-level variations through text-guided feature modifications. Globally, LogFA constructs a Generalized Directed Acyclic Graph to model procedural dependencies and sample alternative action sequences. Experiments on egocentric video benchmarks show LogFA significantly improves model generalization to unseen environments while maintaining low computational and data collection costs. Code will be released at https://github.com/ZijiaLewisLu/ECCV2026- LogFA.

3D Gaussian Splatting (3DGS) has emerged as a powerful technique for real-time LiDAR and camera synthesis in autonomous driving simulation. However, simulating LiDAR with 3DGS remains challenging for extrapolated views beyond the training trajectory, as existing methods are typically trained on single-traversal sensor scans, suffer from severe overfitting and poor generalization to novel ego-vehicle paths. To enable reliable simulation of LiDAR along unseen driving trajectories without external multi-pass data, we present LiDAR-EVS, a lightweight framework for robust extrapolated-view LiDAR simulation in autonomous driving. Designed to be plug-and-play, LiDAR-EVS readily extends to diverse LiDAR sensors and neural rendering baselines with minimal modification. Our framework comprises two key components: (1) pseudo extrapolated-view point cloud supervision with multi-frame LiDAR fusion, view transformation, occlusion curling, and intensity adjustment; (2) spatially-constrained dropout regularization that promotes robustness to diverse trajectory variations in real-world driving. Extensive experiments demonstrate that LiDAR-EVS achieves SOTA performance on extrapolated-view LiDAR synthesis across three datasets, making it a promising tool for data-driven simulation, closed-loop evaluation, and synthetic data generation in autonomous driving. Code and demos are available at: https://lastbasket.github.io/LiDAR-EVS_page/.

Perceptual video compression leverages generative priors to reconstruct realistic textures and motions at low bitrates. However, existing perceptual codecs often lack native support for variable bitrate and progressive delivery, and their generative modules are weakly coupled with entropy coding, limiting bitrate reduction. Inspired by the next-scale prediction in the Visual Auto-Regressive (VAR) models, we propose ProGVC, a Progressive-based Generative Video Compression framework that unifies progressive transmission, efficient entropy coding, and detail synthesis within a single codec. ProGVC encodes videos into hierarchical multi-scale residual token maps, enabling flexible rate adaptation by transmitting a coarse-to-fine subset of scales in a progressive manner. A Transformer-based multi-scale autoregressive context model estimates token probabilities, utilized both for efficient entropy coding of the transmitted tokens and for predicting truncated fine-scale tokens at the decoder to restore perceptual details. Extensive experiments demonstrate that as a new coding paradigm, ProGVC delivers promising perceptual compression performance at low bitrates while offering practical scalability at the same time.

Through pre-training on extensive text and image datasets, current multi-modal large language models (MLLMs) achieve strong performance on general tasks. However, circuit schematics present a unique challenge for MLLMs due to their dense component layouts and distinct topological logic, demanding fine-grained structural parsing to extract the electrical semantics. To address this, we propose Circuit-MLLM, a multimodal reasoning framework that reformulates circuit topology analysis as a process of device localization, path tracing, and sequential reasoning within the latent space. We introduce a circuit knowledge mining mechanism that deeply aligns the model’s latent representations with structurally rich features derived from multi-granularity circuit vision experts, enabling the model to effectively internalize topological semantics. Building upon these internalized semantics, we devise a topology-guided sequencing strategy that decouples reasoning from the rigid raster-scan order, enforcing stepwise inference along the circuit’s topological logic in latent space. Across diverse circuit analysis tasks, Circuit-MLLM consistently outperforms strong baselines, notably achieving a 25% higher average score than GPT-5.1, which demonstrates the effectiveness of our framework in circuit schematic topology analysis. Code is publicly available at https://github.com/IC-Yuan/Circuit-MLLM.

3D Gaussian Splatting (3DGS) enables real-time, photorealistic novel view synthesis, making it a highly attractive representation for model-based video tracking. However, leveraging the differentiability of the 3DGS renderer “in the wild” remains notoriously fragile. A fundamental bottleneck lies in the compact, local support of the Gaussian primitives. Standard photometric objectives implicitly rely on spatial overlap; if severe camera misalignment places the rendered object outside the target’s local footprint, gradients strictly vanish, leaving the optimizer stranded. We introduce SpectralSplats, a robust tracking framework that resolves this "vanishing gradient" problem by shifting the optimization objective from the spatial to the frequency domain. By supervising the rendered image via a set of global complex sinusoidal features (Spectral Moments), we construct a global basin of attraction, ensuring that a valid, directional gradient toward the target exists across the entire image domain, even when pixel overlap is completely nonexistent. To harness this global basin without introducing periodic local minima associated with high frequencies, we derive a principled Frequency Annealing schedule from first principles, gracefully transitioning the optimizer from global convexity to precise spatial alignment. We demonstrate that SpectralSplats acts as a seamless, drop-in replacement for spatial losses across diverse deformation parameterizations (from MLPs to sparse control points), successfully recovering complex deformations even from severely misaligned initializations where standard appearancebased tracking catastrophically fails.

Editing multiple elements of an image in a single forward pass is a practical alternative to multi-turn image manipulation, offering improved efficiency and potentially better harmonization. However, when several instructions target different regions, semantic interference often leads to attribute leakage and poor edit disentanglement, especially as the number of edits increases. In this work, we propose MICE (Multi-Instance Concurrent Editing), a training-free strategy for scalable multi-instance image editing with Multimodal Diffusion Transformers. MICE modifies the additive bias of joint attention to regulate interactions between instance-specific edit instructions, latent, and context tokens identified via user-provided segmentation masks. Specifically, MICE allows intra-instance attention, penalizes interactions between neighboring region tokens, and suppresses unrelated cross-instance attention. As a result, our method enforces attribute binding while preserving global visual consistency. We evaluate MICE on LoMOE-Bench and introduce MICE-Bench, a more challenging benchmark with an average of 8.5 concurrent edits per image. The experiments demonstrate that our approach outperforms strong baselines and recent competitors in terms of visual quality preservation and faithfulness to the editing instructions.

Affordance grounding aims to localize image regions that support a specific action, serving as a core capability for physical intelligence and embodied perception. Previous studies have primarily relied on weakly supervised learning with action labels from exocentric images. However, these methods often struggle with visually ambiguous exocentric images containing co-occurring actions; moreover, they fail to distinguish semantically similar actions because existing methods typically rely on brief action phrases that lack rich semantic details for action-specific localization. Although large vision-language models (LVLMs) encode rich action semantics and their action-conditioned textual outputs implicitly contain spatial cues, they do not directly provide action-specific spatial localization. To address these problems, we propose TokAG, a zero-shot affordance grounding framework that exploits the token-level semanticspatial signals in LVLMs to localize action-relevant regions without external supervision. We observe that attention maps associated with different LVLM output tokens vary significantly, with many attending to irrelevant regions such as the background. Thus, we introduce a spatialaware token-selection mechanism to systematically evaluate each output token and select the one whose attention maps exhibit dominant activation over the target object, instead of relying on arbitrary attention maps. By extracting these object-focused attention maps, we transform the LVLM’s implicit semantic signals into zero-shot affordance heatmaps. Our zero-shot framework consistently outperforms prior weakly supervised approaches across multiple benchmarks, improving NSS by 10.7% on the unseen split of AGD20K and by 29.7% on HICO-IIF. The code and models will be made publicly available.

Event cameras emit asynchronous events in response to environmental appearance changes. The scarcity of real-world event datasets makes simulation essential. However, most simulators infer event timestamps from frame sequences, forcing many threshold crossings to share a small set of discrete times; a failure mode we term timestamp batching that worsens under fast motion and occlusion. We present TIDES, a continuous-time event simulator built on dynamic Gaussian splatting. Because TIDES operates on an explicit 3D scene representation with learnt geometry and motion, it can derive per-pixel intensity dynamics directly from the scene, rather than by differencing rendered frames. This enables accurate threshold-crossing prediction, including multiple crossings per rendering step, without temporal upsampling or frame interpolation. The same 3D scene model reveals where objects partially occlude one another; TIDES uses this to guide adaptive time stepping, concentrating computation only in regions where occlusion dynamics make simple models of brightness change unreliable. Finally, we model finite sensor bandwidth using a tile-level arbiter whose throughput, jitter, and event drops reproduce realistic sensor artifacts. Across paired RGB-event benchmarks, TIDES attains state-of-the-art event-stream fidelity. We also show that events simulated by TIDES transfer more effectively to real downstream tasks than competitors’.