ECCV 2026 Papers

Despite their remarkable performance in real-world imagesuper-resolution, diffusion models remain challenged by the perception–distortion trade-off between structural fidelity and perceptual realism.Existing methods typically rely on a globally predefined restoration tra-jectory, applying spatially uniform noise and a fixed temporal schedule.Such rigid trajectories often lead to a dilemma: reliable structures may bedistorted, while severely degraded textures may become over-smoothed.To address this, we propose AdaBridge-SR, built upon a novel spatio-temporal adaptive bridge matching (ST-ABM) formulation that castsrestoration as a controlled bridge between degraded and clean imagedistributions. ST-ABM decouples the bridge stochasticity into spatialand temporal controls, adaptively determining where to inject stochas-ticity and when to allocate generative capacity conditioned on the inputdegradation. We instantiate these controls with a lightweight AdaptiveBridge Controller that predicts a spatial stochasticity map and a timereparameterization. Consequently, AdaBridge-SR unifies deterministicrestoration and stochastic exploration within a single model. Extensiveexperiments demonstrate that it achieves a superior perception–distortionbalance, supporting efficient deterministic one-step inference and high-quality stochastic few-step bridge refinement. Our code and models areavailable at https://github.com/W-JG/AdaBridge-SR.

Oriented object detection for aerial images has attractedgrowing attention due to the increasing demand in extensive applica-tions. However, the oriented box annotation is relatively expensive, andthus point-supervised oriented object detection has emerged as a practi-cal paradigm. Owing to the lack of size and angle annotations, existingmethods leverage various complementary information, but usually sufferfrom extra training stages or multiplied computational overhead. In thispaper, we propose three simple yet effective loss terms, and set a simplebaseline with the placement prior. In particular, the placement prior in-dicates that objects in aerial images are placed in a context, rather thanexhibiting arbitrary distribution. Therefore, we propose the angle priorloss, considering that the angles of objects are usually aligned in a neigh-borhood. Furthermore, we propose the size prior loss, considering thatthe sizes of objects are usually proportional for each class in an image.Besides, we propose to derive oriented boxes from the Voronoi Water-shed regions without training, which can jointly regularize the estimatedangles and sizes. By employing only loss terms, our baseline is not onlysimple and easy to extend but also readily transferable to other methods.We conduct extensive experiments on representative UAV datasets andremote sensing datasets, which indicates the effectiveness of our method.

Predictive Coding (PC) offers a biologically motivated al-ternative to backpropagation via local weight updates, yet routing er-ror between layers still relies on an autograd Jacobian-transpose (J⊤ )product—the last non-local operation in PC. We show that this depen-dency is largely avoidable. For any layer f (x)=Act(Norm(L(x))) withfrozen normalization statistics, the exact J⊤ factors into three locallyavailable terms, J⊤ v = L⊤ (s ⊙ σ ′ (pre-act) ⊙ v), where σ ′ is the activationderivative and s=γ/σrun is the normalization gain. Prior weight-feedbackmethods omitted both corrections; restoring them closes the transportgap for this layer class. Note that locality here holds up to three as-sumptions, which we state upfront—weight symmetry (L⊤ mirrors theforward operator, as assumed by all PC), a soft spectral-norm controlthat is not synapse-local, and a nearest-neighbour approximation forMaxPool. Substituting the identity into PC yields WF-Act-PC, whichremoves the autograd backward pass from error transport. On CIFAR-10/100 (50 epochs, 5 seeds), WF-Act-PC is the only PC method whoseaccuracy improves with depth, surpassing iPC—the strongest classicalPC baseline—by 2.7–22.3 pp on CIFAR-10. With both methods tunedper architecture, it matches or exceeds a comparably-tuned backpropa-gation baseline on the deeper CIFAR-10 architectures (VGG-9: 93.57%vs. 92.43%; ResNet-18: 92.76% vs. 91.54%) and on the harder Tiny-ImageNet benchmark, while trailing tuned BP on the deeper CIFAR-100VGG cells. Our WF-Act-PC implementation is publicly available athttps://github.com/jlshen025/pcax.

Constructing world models for AI agents in visual environ-ments has long been a significant challenge, chiefly due to the di!cultyof achieving precise alignment of the dynamics between the world modeland the environment, given noisy pixel-level observations. Existing worldmodels solely rely on video or image generation, which are exceedinglycomplex and expensive to train, yet still su"er from hallucinations andviolations of basic rules due to misalignment. In this paper, we show thata few rules su!ce to convert a pretrained large language model (LLM)to an accurate world model for a specific visual environment. We pro-pose a training-free approach to e!ciently learn these complementaryrules from multimodal agent trajectories collected in the visual environ-ment. In particular, we convert the multimodal input into structuredtext and extract symbolic rules using an LLM by comparing agents’ truetrajectories with world model predictions if they have conflicts. We usethis neurosymbolic approach to induce, update, and prune rules dur-ing exploration, resulting in a precise world model. It allows us to builda novel model-predictive control (MPC) agent, “WALL-EVE”, with-out training any policy. Before taking each action, WALL-EVE utilizesan LLM as a look-ahead optimizer to determine the next k-step actionsthrough interactions with our world model. On challenging tasks in ALF-World and Minecraft, WALL-EVE achieves higher success rates thanexisting methods, while reducing inference times and the number of to-kens required for reasoning. In ALFWorld, WALL-EVE surpasses thestate-of-the-art method RAFA with only 17% of RAFA’s token usage.In Minecraft, WALL-EVE outperforms baselines by 8-30% in successrate but costs 8–20 fewer replanning rounds and only 60–85% tokens.Code is available here.

The increasing diversity and density of traffic participantshave made modern driving environments highly complex, posing greatchallenges to reliable perception in autonomous driving. Single-vehicleperception is fundamentally limited by sensor range and occlusions, whilevehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) cooperationonly partially mitigate these issues due to their fixed or ground-levelviewpoints. In contrast, aerial platforms offer global and flexible sens-ing perspectives that can dynamically adjust their positions to providebroader spatial coverage and reduce blind zones. These advantages mo-tivate the exploration of aerial-ground cooperative perception (AGCP).However, progress in this area is hindered by the lack of public datasetsand standardized benchmarks. To bridge this gap, we introduce VVSim,a large-scale dataset for AGCP that provides synchronized multimodaldata and state information from both vehicles and unmanned aerial ve-hicles (UAVs). VVSim consists of about 61k fully annotated frames span-ning diverse interaction scenarios (e.g., cut-in, lane change), 5 weatherconditions (sunny, foggy, rainy, cloudy, snowy), and 11 scenes (e.g., city,town, university, highway, mountain). In total, it comprises 255k Li-DAR sweeps and 3.5M images (RGB, semantic segmentation, and depth)collected from vehicles and UAVs, along with detailed annotations for2D/3D bounding boxes, object trajectories, and agent states. To supportunified evaluation, we further propose VVFormer, a novel architecturethat integrates multi-agent and multimodal features from UAVs and ve-hicles. Extensive experiments on VVSim demonstrate that VVFormerachieves superior performance, significantly outperforming strong V2Vand V2I cooperative baselines on 3D perception tasks. The code is avail-able at https://github.com/LOTEAT/vvdetection3d.

Camera pose estimation from image streams is a critical com-ponent of spatial world models that integrate perception into planningand decision-making. Nearly all Visual Odometry (VO) and SimultaneousLocalization and Mapping (SLAM) systems have focused on datasetscontaining raw, uncompressed videos. Many working systems insteaduse ubiquitous hardware units to efficiently compress and decode videostreams, saving orders of magnitude in storage and bandwidth. However,this lossy compression introduces visual artifacts that hinder the per-formance of traditional tracking systems. We present VOCA, a causalstereo visual-odometry method that exploits codec information to im-prove tracking performance. We achieve state-of-the-art performance oncausal VO for relative trajectory error, efficiency, and absolute trajec-tory error on compressed streams. This work highlights the potential ofleveraging widely available video codec information for vision tasks.

Camouflaged Object Detection (COD) aims to identify andsegment camouflaged objects in complex environments, which are oftenconcealed because their color and texture are similar to the background.Several existing COD methods introduce depth maps to boost detec-tion performance via learning complementary RGB-D features, ignor-ing modality-specific characteristics of concealed objects in the depthdomain. To address this issue, we propose a depth collaborative net-work, called VCP-DCN, to mine distinguishable multi-modality featuresbeyond visual concealed prototype in depth domain. Specifically, VCP-DCN progressively performs multi-modality alignment, interaction, andfusion for the COD task. In the alignment stage, we propose a SeparablePrototype Embedding (SPE) module to learn modality-consistency andmodality-specific RGB/depth prototype tokens through prototype con-trastive learning. Furthermore, we develop a Multi-modality Dual Atten-tion (MDA) module to enhance the cross-modal feature representationthrough local response maps between modality-consistency RGB/depthprototype tokens and visual tokens on the interaction stage. Finally,we design a Depth Adaptive Injection (DAI) module to adaptively mea-sure contribution of RGB/depth features with a decision-making mecha-nism, which calculates similarity distance between RGB/depth modality-specific prototype tokens and modality-consistency ones on the fusionstage. Extensive experiments demonstrate the effectiveness of our VCP-DCN on three authoritative datasets. The code is available at https://github.com/duan-song/VCPDCN.

Human-object interaction detection (HOID) has traditionally been formulated as a supervised detection problem over predefined interaction categories. While such paradigms achieve strong performance on closed-set benchmarks, they fundamentally entangle interaction understanding with dataset-specific supervision, limiting their ability to generalize to open-world and compositional scenarios. Recent HOI detectors attempt to leverage MLLMs through prompting strategies to transfer interaction-specific knowledge. However, such prompt-based approaches primarily focus on extracting discriminative representations from pretrained models, while underexploring their inherent multimodal reasoning capabilities. As a result, they struggle to provide informative contextual reasoning for ambiguous and open-world interaction scenarios. In this work, we present AgentHOI, a training-free, agentic framework that transfers the generalist multimodal reasoning capabilities of foundation models to HOI detection in the wild. Instead of learning interaction classifiers, AgentHOI modularly orchestrates complementary vision foundation modules to perform open-ended semantic reasoning and spatial grounding in a coordinated manner. To address the challenges of incomplete interaction discovery and ambiguous localization in complex scenes, we introduce two key mechanisms: (1) Context-aware Multi-round Reasoning, which progressively refines interaction hypotheses to ensure exhaustive and compositional HOI discovery, and (2) Multifaceted Interaction Localization, which enhances grounding precision by generating instance-specific descriptions that integrate semantic, spatial, and appearance cues. Extensive experiments demonstrate that AgentHOI achieves superior performance over state-of-the-art supervised and weakly supervised methods in real-world settings, despite requiring no HOID data for training. Code is available at https://github.com/ oceanflowlab/AgentHOI.

While point cloud foundation models have significantly ad-vanced point cloud video understanding, existing parameter-efficient fine-tuning (PEFT) methods still suffer from two critical limitations: pro-hibitive annotation costs for large-scale point cloud datasets and severememory bottlenecks. In this paper, we aim to mine richer supervisionsignals from existing data rather than blindly scaling datasets. A furtherkey principle is that the memory footprint of fine-tuning must be dras-tically reduced compared to full fine-tuning, which remains elusive forcurrent PEFT techniques. Driven by these challenges, we identify threecore desiderata: data-, parameter-, and memory efficiency, and presentPoinTriE, a unified framework that excels along all three dimensions.For pre-training, pseudo-motion trajectories are synthesized via rigidtransformations, paired with text corpora and 2D projections derivedfrom raw point clouds. We then propose a Geometric-Motion DualityNetwork optimized via multimodal contrastive learning, rigid rotationprediction, and motion distribution divergence to produce dense self-supervision. During fine-tuning, we freeze the pretrained backbone andonly update a lightweight Spatio-temporal Side Network built with LoRAunits. Equipped with a gradient flow masking strategy, PoinTriE simul-taneously reduces memory consumption and parameter overhead. Exten-sive experiments confirm that PoinTriE establishes new state-of-the-artresults on action recognition and semantic segmentation tasks.

Vision-language-action (VLA) models that generate contin-uous action chunks via flow matching lack an internal signal for judgingwhether a given prediction is reliable. Distribution shift and long-horizonrollouts can push backbone representations away from the region the ac-tion head decodes reliably, yet the policy has no mechanism to detector react to this drift. We observe that the cost of transporting observa-tion features to the action representation in a shared feature space risesprecisely when such drift occurs, providing a per-step reliability esti-mate without extra supervision. Building on this observation, we proposeDiG (Discrepancy Gate), a lightweight plug-in module for flow-matchingVLA policies. DiG computes a sliced Wasserstein transport cost betweenbackbone features and the action expert’s own input projection, maps itthrough an exponential gate, and uses the gate to modulate both a resid-ual feature refinement and the training loss. At inference time, the gateenables DiG-Refine, an iterative refinement process that corrects actionchunks before execution. Experiments on both simulation and real-worldscenarios show that DiG consistently improves success rates, with thelargest gains under distribution shift and on long-horizon tasks.

Multimodal large language models (LLMs) are increasinglyused to generate dermatology diagnostic narratives directly from images.However, reliable evaluation remains the primary bottleneck for respon-sible clinical deployment. We introduce a novel evaluation frameworkthat combines DermBench, a meticulously curated benchmark, withDermEval, a robust automatic evaluator, to enable clinically meaning-ful, reproducible, and scalable assessment. We build DermBench, whichpairs 4000 real-world dermatology images with expert-certified diagnos-tic narratives and uses an LLM-based judge to score candidate narrativesacross clinically grounded dimensions, enabling consistent and compre-hensive evaluation of multimodal models. For individual case assessment,we train DermEval, a reference-free multimodal evaluator. Given an im-age and a generated narrative, DermEval produces a structured critiquealong with an overall score and per-dimension ratings. This capabil-ity enables fine-grained, per-case analysis, which is critical for identi-fying model limitations and biases. Experiments on a diverse dataset of4500 cases demonstrate that DermBench and DermEval achieve closealignment with expert ratings, with mean deviations of 0.251 and 0.117(out of 5) respectively, providing reliable measurement of diagnostic abil-ity and trustworthiness across different multimodal LLMs. The bench-mark resources are publicly available at https://github.com/yuhos16/DermBench.

Semantic matching aims to establish pixel-level correspon-dences between instances of the same category, and previous approachesbased on 2D foundation models have achieved promising results on thistask. However, they fail to achieve geometry-grounded dense semanticmatching, which encompasses geometric awareness, manifold preserva-tion, and cross-image invisibility reasoning. This new matching task si-multaneously needs geometry-aware descriptors and holistic dense match-ing mechanisms, but existing approaches rarely provide both. To bridgethis gap, we leverage VGGT, a 3D geometric foundation model that in-herently provides strong geometry-grounded priors to support both capa-bilities in a unified framework. However, directly transferring faces twochallenges: task heterogeneity, where VGGT originally matches cross-view images of the same instance, not cross-instance variations in shapeor appearance; and scarce dense matching annotations for task adapta-tion. To address these challenges, we propose an approach that (i) retainsVGGT’s intrinsic capabilities by reusing early feature stages, fine-tuninglater ones, and adding a semantic head for bidirectional correspondences;and (ii) adapts VGGT for semantic matching under data scarcity throughcycle-consistent training strategy, synthetic data augmentation, and pro-gressive training recipe. Extensive experiments demonstrate that our ap-proach achieves superior geometric awareness, manifold preservation, andmatching reliability, outperforming previous baselines.

Concept personalization injects a user-specified concept intoa pretrained diffusion model from only a few reference images, enablingcustomized content creation. However, existing personalization methodsfrequently suffer from concept-dominant failures—while the personal-ized concept is well preserved, other prompt-specified contexts (e.g.,background, attributes, and interactions) are not properly generated,thereby limiting controllability. To address this problem, in this paper, wepropose Target-Oriented Perturbation-based Augmentation (TOPA), areference images-only augmentation framework. TOPA optimizes addi-tive perturbations on reference images to provide token–region atten-tion guidance and further reduces background entanglement via subject-isolation compositing. Unlike prior solutions that modify training ob-jectives or architectures, TOPA requires no changes to personalizationpipelines, and its augmented datasets can be directly applied to existingmethods. Extensive experiments across multiple personalization methodsdemonstrate consistent improvement on context adherence while preserv-ing concept fidelity.

Multimodal survival analysis aims to improve cancer prognosis using heterogeneous biomedical data, such as histopathology images and genomic profiles. A common strategy is to align representations across modalities so that shared signals can be captured. However, strong cross-modal alignment can also remove modality-specific evidence that is critical for survival prediction. In this paper, we revisit multimodal survival learning from a simple observation: effective models should first discover shared patterns across modalities, and then preserve modalityspecific signals. This motivates a representation learning principle that we refer to as Together Then Apart. Based on this idea, we propose TTA, a framework that balances cross-modal alignment and representation distinctiveness. TTA first performs prototype-based alignment to capture shared survival-related structures between modalities. It then encourages modality-specific distinctiveness through an anchor-guided contrastive objective. To further account for modality imbalance and noisy correspondences, we model cross-modal interactions using unbalanced optimal transport. We evaluate the proposed approach on multiple TCGA cancer cohorts with paired histopathology and genomic data. TTA consistently improves survival prediction over recent multimodal survival models. Moreover, the learned prototype structures reveal interpretable cross-modal patterns associated with clinical outcomes. Code is available at https://github.com/Y-Research-SBU/TTA.

Vision-language agents that orchestrate specialized tools forimage restoration (IR) have emerged as a promising method, yet most ex-isting frameworks operate in a training-free manner. They rely on heuris-tic task scheduling and exhaustive tool traversal, resulting in sub-optimalrestoration paths and prohibitive computational cost. We argue that thecore bottleneck lies in the absence of a learned policy to make decision,as a vision-language model cannot efficiently handle degradation-awaretask ordering and tool composition. To this end, we propose TIR-Agent, atrainable image restoration agent that performs a direct tool-calling pol-icy through a two-stage training pipeline of supervised fine-tuning (SFT)followed by reinforcement learning (RL). Two key designs underpin effec-tive RL training: (i) a random perturbation strategy applied to the SFTdata, which broadens the policy’s exploration over task schedules andtool compositions, and (ii) a multi-dimensional adaptive reward mech-anism that dynamically re-weights heterogeneous image quality metricsto mitigate reward hacking. To support high-throughput, asynchronousGPU-based tool invocation during training, we further develop a glob-ally shared model-call pool. Experiments on both in-domain and out-of-domain degradations show that TIR-Agent outperforms 12 baselines,including 6 all-in-one models, 3 training-free agents, and 3 proprietarymodels, and achieves over 2.5× inference speedup by eliminating redun-dant tool executions.

Artistic image synthesis aims to recreate the expressive vi-sual identity of a target artist, yet existing methods often fail to cap-ture an artist’s global style. Conventional style transfer methods trans-fer the style of one or a few reference artworks to a content image ina One-to-One manner, making them effective for artwork-level styliza-tion but limited in representing the broader stylistic distribution of anartist. Text-to-image diffusion models conditioned on artist names, suchas ‘∼ in Van Gogh style’, offer greater flexibility, but they often sufferfrom text-induced bias and reproduce patterns from only a few iconicworks. To address these limitations, we introduce Global Style Transfer(GST), an artistic image synthesis paradigm, in a Many-to-One man-ner,that aggregates multiple artworks from a target artist and transferstheir shared global style to a single content image. For GST, we proposeGlobal Style Guidance (GSG), which learns a residual global style offset∆⟨t in the intermediate feature space, or h-space, of a diffusion modelunder a fixed prompt. By learning artist-level style semantics purely fromvisual statistics, GSG mitigates text-dependent artistic bias. We furtherpropose Content Alignment Guidance (CAG), a training-free perceptualguidance mechanism that preserves the semantic structure of the contentimage while allowing artist-specific geometric deformation. Experimentson WikiArt demonstrate that GST achieves superior stylistic fidelity,content preservation, and output diversity compared to existing styletransfer and diffusion-based artistic synthesis methods.

Unified models (UMs) combine visual understanding (I2T)and generation (T2I) in a single framework, alongside broader unimodalcapabilities. We focus on the cross-modal pair T2I and I2T, where cross-consistency—what a model understands, it should be able to generate—is a core promise of unification and a practical necessity in applicationsthat compose both capabilities. Yet, existing benchmarks evaluate themin isolation: FID/GenEval for T2I; MME/MMBench for I2T. We showthat this gap is consequential: models that score competitively on thesebenchmarks can fail severely when their own understanding and gener-ation are composed, progressively losing core entities, attributes, spa-tial relations, and counts, resulting in semantic drift. To quantify drift,we introduce the Semantic Drift Protocol (SDP), inspired by the Tele-phone Game: starting from a caption or image, we iteratively alternateI2T and T2I over multiple generations and measure how faithfully se-mantics are preserved. We propose two complementary metrics: MeanCumulative Drift (MCD), an embedding-based measure of overall con-tent retention across three representation spaces, and Multi-GenerationGenEval (MGG), which extends GenEval’s object-level compliance scor-ing across generations. To stress-test models beyond COCO-style data,we create a benchmark of 400 image-text pairs sampled from NoCapsand DOCCI that emphasizes novel objects and fine-grained descriptions.Applying SDP to seven recent unified models reveals that drift behaviorvaries dramatically and is not predicted by single-pass scores: BAGELretains high semantic fidelity over multiple generations, while VILA-Uand Janus variants collapse within five generations, despite comparableisolated metrics. We identify six recurring failure modes and find thatdegradation is typically catastrophic rather than gradual: once a criticalerror occurs, subsequent generations compound it. Overall, SDP exposesfailure modes that single-pass benchmarks miss, enabling a more faithfulassessment of unified model reliability. Code and benchmark resourcesare available at https://github.com/mollahsabbir/telephone-game-semantic-drift.

Real-time co-speech gesture generation requires producing 3D motion clip by clip as speech streams in. Existing methods are fundamentally open-loop: each clip is synthesized conditioned only on past context, with no mechanism to verify or correct trajectory plausibility. Small per-clip errors therefore compound silently, causing the well-known drift problem--motion that gradually deviates from natural distributions over minute-scale horizons. We make a key observation: drift stems not from poor local motion quality--modern diusion models already produce convincing short clipsbut from the absence of forward constraints that tell each clip where it should arrive. Supplying even a single plausible key pose at a clip's tail as a destination anchor is sufficient to suppress drift dramatically. Building on this insight, we propose StreamTalk, a closed-loop streaming framework that introduces a periodic generate-retrieve-refine feedback cycle. At inference, Streaming Pose-Guided Generation (SPG) first produces a coarse clip, retrieves a plausible tail key pose from a speaker-specic motion database, and refines the clip with this anchor before passing it to the next window. To enable the model to exploit such sparse anchors effectively, we introduce Stochastic Anchor Masking (SAM) during training, which independently masks random pose and translation frames so the model learns to inpaint complete motion from partial boundary conditions. A part-aware DiT architecture further disentangles hand, body, and translation streams to prevent global displacement from interfering with local articulation. Extensive experiments on the BEAT2 benchmark demonstrate that StreamTalk achieves state-of-the-art motion quality (FGD), signicantly suppresses long-horizon drift compared to open-loop baselines, and runs in real time at 76 FPSenabling practical minute-scale streaming co-speech gesture generation. Code will be released upon acceptance.

Embodied agents perceive the world as continuous streams,yet prevailing benchmarks evaluate spatial reasoning offline, presuppos-ing global spatial omniscience and overlooking real-time spatial intelli-gence. We present StreamSpatialBench, a testbed for streaming 3Dvisual-spatial reasoning built on three pillars: fine-grained online tempo-ral perspectives (past / current / future anchored to the query times-tamp), dynamic multi-agent interactions, and comprehensive 3D spatialreasoning. It comprises 5,280 millisecond-grounded QA pairs over 2,003egocentric/scene videos across 10 sub-tasks. Benchmarking 18 represen-tative MLLMs against humans reveals a substantial gap, rooted in threedeficiencies: limited spatiotemporal memory, absent incremental updat-ing, and poor adaptation to dynamic, multi-agent scenes. To close thisgap, we propose Stream3D-Mem, which (i) maintains a bounded mem-ory buffer with selective consolidation via round-decayed compressionto prune redundancy while preserving recent context, and (ii) distillscognitive-map knowledge into compact implicit latent tokens, bypassinglatency-heavy explicit map generation. Together they offer a rigoroustestbed and a practical step toward continuous, human-like 3D spatialintelligence for robotics and AR/VR. Code: https://github.com/OPPO-Mente-Lab/StreamSpatial.