ECCV 2026 Papers

Unified unsupervised anomaly detection aims to train a sin-gle model to detect and localize diverse anomalies across multi-class sam-ples, representing one of the most challenging tasks in anomaly detec-tion. Existing methods typically rely on highly compressing inputs intoa low-dimensional latent space to subsequently reconstruct anomalousfeatures. However, low-dimensional latent spaces critically limit infor-mation capacity, leading to unavoidable feature loss during compressionand significantly degrading the quality of reconstruction and represen-tational power of the model. Therefore, we indicate that dimensionalitycompression in latent space is not a necessary requirement for anomalydetection, and the editability introduced by compression is not inher-ently essential for models reconstructing anomalies. Thus, we propose anovel anomaly detection framework without latent space compression,called HLRAD (High-dimensional Latent Representation for UnifiedAnomaly Detection). Unlike prior methods, HLRAD innovatively en-ables dimension expansion rather than compression in the latent space,and the method effectively avoids feature loss from compression to ensurehigh-quality reconstruction while constructing a semantically enrichedhigh-dimensional latent representation space. Furthermore, by embed-ding the expanded high-dimensional semantic features into the latentrepresentation, HLRAD enables the model to more fully capture the se-mantic feature distribution of normal samples, significantly enhancingcomplex anomaly detection performance. We conducted extensive ex-periments on major anomaly detection benchmark datasets, includingMVTec-AD, VisA, and Real-IAD. In unified settings, HLRAD outper-forms state-of-the-art methods.

To overcome the black-box nature of predictive AI and thehallucination risks of generative models, retrieval-based models o!er aninterpretable, evidence-based paradigm for pathology clinical workflow.However, real-world clinical queries are inherently interleaved (e.g., pathol-ogy images and text). Current dual-encoders su!er from an Archi-tectural Mismatch, lacking the mechanism to fuse such composedqueries. To address this, we formalize the task of Pathology ComposedRetrieval (PCR). While Multimodal Large Language Models (MLLMs)o!er deep-fusion capabilities, directly applying them exposes a TaskMismatch and a Domain Mismatch. To resolve these challenges,we propose HOMIE, a model-agnostic adaptation framework that trans-forms any generative MLLM into a specialized pathology retrieval expert.Evaluated on our newly introduced PCR Benchmark, a lightweight 2B-parameter HOMIE variant substantially outperforms existing paradigms,surpassing specialized 7B pathology MLLMs and dual-encoders by largemargins on composed retrieval, while maintaining strong performanceon traditional simple retrieval. The project page is available at https://qfchou.github.io/HOMIE_page/.

Adversarial attacks on point clouds require effective constraints to ensure imperceptibility. However, existing methods often overlook the intrinsic hierarchical organization of 3D shapes, thereby limiting their ability to preserve structural coherence. In this paper, we propose HHA, a novel framework that generates hierarchy-aware adversarial perturbations by leveraging hyperbolic geometry. HHA first decomposes the input point cloud into semantic and geometric substructures to capture its multi-scale organization. Then, each substructure is embedded into hyperbolic space, where localized constraints limit distortion and maintain hierarchical consistency. This hyperbolic regularization keeps perturbations aligned with the underlying structure and thereby enhances imperceptibility. Extensive experiments validate that HHA produces adversarial point clouds with improved structural coherence and imperceptibility, outperforming state-of-the-art methods.

Building interactive digital twins requires recovering both 3Dgeometry and the kinematic structures that govern how objects articu-late. Yet existing methods for articulated object reconstruction requireexplicitly observable motion from multiple articulation states. We intro-duce a rest-state formulation that reconstructs articulated objects froma single closed configuration, an inherently ill-posed setting where geom-etry, semantics, and motion priors compensate for the absence of motioncues. Our framework adopts an explicit mesh as an intermediate repre-sentation for cross-model verification and fusion, reconciling noisy out-puts from vision-language and segmentation models into spatially con-sistent part structures. To estimate joint parameters without observedmotion, we use a video diffusion model to synthesize articulation hy-potheses and validate them through geometric consistency. Our approachachieves accurate part decomposition and physically plausible articu-lation, performing competitively with motion-observing reconstruction-based, generation-based, and modular pretrained-model baselines.

In real-world clinical practice, medical images face open-world shifts: (i) long-tailed rare diseases, (ii) subtle lesions dominated bynormal anatomy, and (iii) hierarchical taxonomies. Yet most open-worldparadigms assume flat, balanced label spaces, leaving these biomedi-cal demands unresolved. We introduce Generalized Biomedicine Dis-covery (GBD) and a unified benchmark spanning long-tail, anomaly,and taxonomy-aware discovery. Our key insight is that dominant knownpatterns form a visual manifold that masks subtle novelty. Inspired byexpert diagnosis, we propose SCAN (Surprise-evoked ComplementaryAccommodatioN), which follows a cognition-inspired perceptual progres-sion: it applies predictive suppression to filter expected norms, trig-gers surprise-evoked salience to highlight unexpected deviations, andperforms complementary accommodation to integrate these shifts intoglobal representations. Extensive experiments show that SCAN improvesnovel concept discovery while generally preserving established clinicalknowledge, and it plugs into existing architectures to better navigatethe known–unknown trade-off in medical imaging. Code is available atgithub.com/lytang63/generalized-biomedicine-discovery.

Novel view synthesis has recently been revolutionized by 3DGaussian Splatting (3DGS), which enables real-time rendering throughexplicit primitive rasterization. However, existing methods tie visual fi-delity strictly to the number of primitives: quality downscaling is achievedonly through pruning primitives. We propose the first inherently scal-able primitive for radiance field rendering. Fourier Splatting employsscalable primitives with arbitrary closed shapes obtained by parameter-izing planar surfels with Fourier encoded descriptors. This formulationallows a single trained model to be rendered at varying levels of de-tail simply by truncating Fourier coefficients at runtime. To facilitatestable optimization, we employ a straight-through estimator for gradi-ent extension beyond the primitive boundary, and introduce HYDRA,a densification strategy that decomposes complex primitives into sim-pler constituents within the MCMC framework. Our method achievesstate-of-the-art rendering quality among planar-primitive frameworksand comparable perceptual metrics compared to leading volumetric rep-resentations on standard benchmarks, providing a versatile solution forbandwidth-constrained high-fidelity rendering.

Understanding camera movement in natural language is crit-ical for training and evaluating video generation models, among other ap-plications. However, we demonstrate that existing vision-language mod-els (VLMs) fail this task in surprising ways, frequently confusing trans-lation with rotation, left with right, and object movement with cameramovement. To address these limitations, we establish natural languagecamera movement understanding as a standalone research task. We in-troduce a two-level cinematographic taxonomy and an extensive, atomicbenchmark featuring both real and synthetic videos. Furthermore, we cu-rate a large-scale, multi-source training set enhanced by targeted cameramovement augmentation. Our fine-tuned VLM-8B outperforms Gemini3.1 Pro by 10% and 11% on our benchmark’s real and synthetic videos,respectively. Despite these gains, a significant gap remains relative to hu-man performance, underscoring the need to promote and facilitate futureresearch on natural language camera movement understanding.

Audio-video generation has recently gained unprecedentedresearch attention, aiming to synthesize high-quality sounding video con-tent with fine-grained synchronization and semantic alignment betweenthe auditory and visual components. The preceding methods predomi-nantly adopt a dual-branch design with separate tokenization and gen-eration modules per modality, neglecting the representation gap whilenecessitating intensive computational resources for proper training. In-spired by recent advancements in one-dimensional visual tokenization,we present AVTok, a novel unified tokenizer designated for holisticaudio-video generation. AVTok features a dual-stream transformer-basedarchitecture with shared encoder-decoder and modal-specific learnablequeries to efficiently and effectively encode an audio-video pair into acompact one-dimensional latent representation with a unified codebook.To cope with the heterogeneous information imbalance that hinders AV-Tok from exploiting aligned audio-visual information, we devise a hier-archical training strategy to progressively realize reconstruction capabil-ities for each modality. Extensive experiments demonstrate that AVTokexcels both in audio-video reconstruction and when integrated into down-stream pipelines for audio-to-video, video-to-audio, and class-conditionaljoint audio-video generation. AVTok paves the way for the challenge ofjoint audio-video tokenization and provides a potential direction to buildunified large multimodal models for audio-video generation.

Reconstructing animatable 3D human avatars from monocularvideo is a fundamental problem in computer vision with broad applicationsin AR/VR and digital content creation. Existing approaches typically coupleparametric body models with neural rendering or 3D Gaussian splattingand optimize all body regions jointly from short videos, which often de-grades fidelity in the visible areas. To overcome this limitation, we introduceFlexiAvatar, a unified framework that explicitly optimizes only the visiblebody regions, effectively eliminating artifacts arising from unobserved limbs.Our method integrates occlusion-robust SMPL-X tracking with part-specificresidual refinement to capture high-frequency geometric and appearancedetails. To complete entirely unseen regions (e.g., back views), we leverage adiffusion-based approach to generate texture consistent with the observedappearance. Experiments on full-body (NeuMan, ZJU-MoCap, WildAvatar),upper/half-body (talk-show clips), and head-only (INSTA) inputs show thatFlexiAvatar delivers consistently higher reconstruction quality, outperformingstate-of-the-art methods by an average PSNR improvement of approximately3% across datasets. Finally, by restricting optimization to observed regions,our method reduces the effective number of Gaussians that must be opti-mized and rendered, leading to reduced runtime and memory overhead inpartial-visibility scenarios. The project page can be found here.

Visually-guided acoustic highlighting seeks to rebalance au-dio in alignment with the accompanying video, creating a coherent au-dio–visual experience. While visual saliency and enhancement have beenwidely studied, acoustic highlighting remains underexplored, often leadingto misalignment between visual and auditory focus. Existing approachesuse discriminative models, which struggle with the inherent ambiguityin audio remixing, where no natural one-to-one mapping exists betweenpoorly-balanced and well-balanced audio mixes. To address this limitation,we reframe this task as a generative problem and introduce a ConditionalFlow Matching (CFM) framework. A key challenge in iterative flow-basedgeneration is that early prediction errors — in selecting the correct sourceto enhance — compound over steps and push trajectories off-manifold. Toaddress this, we introduce a rollout loss that penalizes drift at the finalstep, encouraging self-correcting trajectories and stabilizing long-rangeflow integration. We further propose a conditioning module that fusesaudio and visual cues before vector field regression, enabling explicit cross-modal source selection. Extensive quantitative and qualitative evaluationsshow that our method consistently surpasses the previous state-of-the-artdiscriminative approach, establishing that visually-guided audio remixingis best addressed through generative modeling. Qualitative samples areavailable at the project page.

Recent real-world image super-resolution (SR) methods often adapt text-to-image (T2I) backbones with ControlNet-style branches or spatial conditioning tokens, which increases memory and computes with resolution and often constrains training to a fixed scale. We propose Fill2SR, which repurposes a masked-inpainting Diffusion Transformer for SR without extra spatial branches. Our Inpainting-Interface Evidence Adapter (IIEA) writes the low-quality (LQ) observation into the native masked-image slot under a full-image mask, turning inpainting into a reverse-degradation conditional rectified flow trained with LoRA-only tuning. We further introduce RCDT, an offline pipeline that distills degradation descriptors from unpaired real images and transfers them onto clean targets using frozen open-source models. Fill2SR supports mixed-resolution training up to QHD and yields stable performance across 512/1024/2048 outputs. On synthetic benchmarks, our base model with IIEA achieves the best LPIPS on DIV2K and LSDIR; adding RCDT trades a small LPIPS drop for consistently stronger no-reference quality on RealLQ250 and RealPhoto60. Fill2SR remains memory-predictable, running 15362 inference on a single 32GB GPU and extending to multimegapixel outputs via tiled restoration.

We propose Event-LiDAR* , a 3D eventification frameworkthat converts conventional LiDAR scans into temporally sparse repre-sentations for efficient point cloud processing. Unlike dense multi-scanprocessing, naive frame differencing, or correspondence-based residuals,3D event extraction is formulated as a temporal estimation problemunder sparse, viewpoint-dependent observations. Short-term geometricevolution across consecutive scans is approximated with a first-ordergeometric model, and deviations unexplained by this model are treatedas events. This yields compact, information-preserving representationsthat suppress redundancy while retaining changes unpredictable by thefirst-order model. Event-LiDAR is applied to LiDAR-based 3D objectdetection with existing backbones, and an event-aware network designis further introduced to reallocate modeling capacity toward the inputstage for sparse inputs. On nuScenes, under the 1F+9T setting witha 72% point reduction, Event-LiDAR maintains accuracy comparableto full-scan baselines while achieving 23% faster end-to-end inferenceon PTv3, driven in part by a 32% speedup in its feature extractionbackbone. The gains generalize across backbones, with a 9% end-to-endspeedup on CenterPoint under the same setting. Training is likewiseaccelerated by 17% under the same setting. Event-LiDAR thus serves asa low-latency, architecture-agnostic geometric preprocessor inspired byevent-based sensing, providing a practical front-end for efficient 3D pointcloud perception.

Despite recent progress in 3D-LLMs, they remain limitedin accurately grounding language to visual and spatial elements in 3Denvironments. This limitation stems in part from training data that fo-cuses on language reasoning rather than spatial understanding due toscarce 3D resources, leaving inherent grounding biases unresolved. Toaddress this, we propose 3D scene editing as a key mechanism to gener-ate visual counterfactuals that mitigate these biases through fine-grainedspatial manipulation, without requiring costly scene reconstruction orlarge-scale 3D data collection. Furthermore, to make these edits targetedand directly address the specific weaknesses of the model, we introduceDEER-3D, an error-driven framework that diagnoses grounding failuresand generates targeted counterfactual training supervision via a struc-tured “Decompose, Diagnose, Edit, and Retrain” loop. Specifically, givena grounding failure, DEER-3D first identifies the predicate-level error(e.g., attribute or spatial relation). It then performs minimal predicate-aligned scene edits, such as recoloring or repositioning, and constructsaligned question–answer pairs that explicitly target the failed predicate,forming targeted counterfactual training examples. We evaluate our edit-ing pipeline across multiple benchmarks for 3D grounding and scene un-derstanding tasks, consistently demonstrating improvements across allgrounding datasets through iterative refinement (4-6% gains). DEER-3D underscores the effectiveness of targeted, error-driven scene editingin bridging linguistic reasoning with spatial grounding in 3D LLMs.

We present EmbodiedHead, a speech-driven talking headframework that equips LLMs with real-time visual avatars for conver-sation. A practical embodied avatar must achieve real-time generation,unified listening-speaking behavior, and high rendered visual quality si-multaneously. Our framework couples the first Rectified-Flow DiffusionTransformer (DiT) for this task with a differentiable renderer, enablingdiverse, high-fidelity generation in as few as four sampling steps. Priorlistening-speaking methods rely on dual-stream audio, introducing an in-terlocutor look-ahead dependency incompatible with causal user–LLMinteraction. We instead adopt a single-stream interface with explicitper-frame listening-speaking state conditioning and a Streaming AudioScheduler, suppressing spurious mouth motion during listening while en-abling seamless turn-taking. A two-stage training scheme of coefficient-space pretraining and joint image-domain refinement further closes thegap between motion-level supervision and rendered quality. Extensive ex-periments demonstrate state-of-the-art visual quality and motion fidelityin both speaking and listening scenarios.

Diffusion models for multi-agent trajectory prediction arelimited by iterative denoising, which causes inference latency that hinderstheir use in time-critical settings like autonomous driving. Fast-samplingvariants using DDIM and informed initial noise distribution partiallyalleviate this issue, but they either fail to achieve true single-step gen-eration or are constrained by the chosen noise distribution. ConsistencyModels (CMs) offer high-quality one-step generation by mapping noisedirectly to data, but are difficult to train from scratch. We propose EC-Traj , an enhanced CM pipeline with improved training and conditionalgeneration for trajectory prediction. Our framework extends the student-teacher consistency training scheme: the student produces standard out-puts, while the teacher explicitly fuses its predictions with parts of theground truth to give stronger supervision. We also exploit CMs’ directdenoising for top-K multi-shot generation during training. Combiningconditional generation with this enhanced consistency objective yieldsfaster inference and improved prediction accuracy, establishing compet-itive new benchmarks on the large-scale Argoverse 2 dataset.3

Feed-forward 3D reconstruction for autonomous driving hasdemonstrated strong performance and promising generalization capabil-ities. However, pixel-level methods often suffer from multi-view inconsis-tencies and content duplication. While voxel-level approaches alleviatethese issues by predicting 3D Gaussian primitives in voxel space, theyremain limited by resolution and computational overhead. We introducean efficient feed-forward framework for scene reconstruction from uncal-ibrated images. The key insight is the explicit decoupling of instanta-neous generation necessity and long-term geometric persistence duringGaussian generation. Specifically, we introduce Generation Confidencedriven by reprojection errors to enable demand-aware Gaussian expan-sion, and Reuse Confidence constructed from depth uncertainty and tem-poral decay to model the long-term stability. To balance fidelity andsparsity, we employ differentiable voxelization to aggregate neighboringfeatures and suppress spatial redundancy, followed by confidence-awaresoft-gating pruning. Gaussian distillation is further introduced to enablea compact set of backbone Gaussians to approximate the full scene repre-sentation. Extensive experiments on multiple datasets demonstrate thatour method achieves strong reconstruction quality and robustness whilemaintaining a highly compact scene representation.

A pivotal step in autonomous driving simulation involves in-serting foreground vehicles with predefined trajectories into simulatedscenes. This process enhances scene diversity and facilitates the creationof various corner cases for testing and improving autonomous drivingmodels. However, existing methods often rely on pre-reconstructed 3Dassets, which frequently lead to lighting inconsistencies between the in-serted foreground and the background. Moreover, the reliance on limited,manually-curated 3D assets hinders large-scale deployment. To addressthese challenges, we propose DriveWeaver, a novel framework for con-trollable vehicle insertion in autonomous driving simulation. Specifically,for a masked target insertion area, DriveWeaver performs video inpaint-ing conditioned on vehicle point clouds to generate high-quality, tempo-rally consistent vehicles. This video-inpainting-based approach ensuresseamless blending between the foreground and background, while thereadily available point cloud conditions enable superior generalization.To support long-term generation, we further design a global-to-local hier-archical inpainting strategy, ensuring the consistent identity and appear-ance of the inserted vehicles. Meanwhile, we extract explicit 3D Gaussianrepresentations of the inserted vehicles through an urban reconstructionpipeline to enable real-time rendering for autonomous driving simula-tion. Extensive experiments across diverse datasets demonstrate thatour method outperforms existing baselines in visual realism and geomet-ric consistency, providing a robust tool for scalable autonomous drivingscene augmentation.

We introduce the Noise-Tilted Reverse Kernel (NTRK), areward-guided diffusion sampler that injects reward gradients through thenoise term, leaving the pretrained reverse kernel unchanged and requiringonly a single sample per step. Reward-guided sampling at inference timehas greatly expanded the versatility of pretrained diffusion models. Yet ex-isting methods face a trade-off. Gradient-based guidance shifts the reversemean, steering generation but pushing intermediate states outside theregion that the model was trained on and degrading quality. Search-basedmethods preserve quality but gain no gradient signal. No prior methodachieves both. NTRK resolves this by keeping the reverse mean fixed andbiasing the noise term toward high reward. This is enabled by a whiteningoperator, the central mechanism behind NTRK, which converts rewardgradients into noise-compatible perturbations without losing their guidingsignal. Across various reward alignment tasks, NTRK outperforms recentstate-of-the-art baselines without losing sample quality. Remarkably, onaesthetic generation, NTRK surpasses the reward of the best baseline at500 NFEs using only 25 NFEs, a 20× reduction in compute.

Although recent robot perception research emphasizes train-ing on data from diverse environments to improve generalization, mostexisting methods still rely on centralized learning, which is inefficient anddifficult to scale across heterogeneous robot platforms. Federated learn-ing (FL) offers an alternative by enabling distributed training withoutraw data transfer, but it suffers from severe performance degradation un-der domain shifts caused by heterogeneity across clients. In real roboticdeployments, data distributions often overlap across platforms, environ-ments, and sensing conditions, making it difficult to partition clientsinto clearly separated domains. However, this characteristic breaks theassumption of clearly separable client domains commonly used in clus-tered FL. To address this gap in robot perception, particularly in depthestimation, we introduce two realistic and unexplored non-IID scenariosthat reflect heterogeneity in terms of platform, environment, and depthdistribution. We then propose FeDepth, a descriptor-based clusteredFL framework that models client relationships through soft clustering.Unlike hard clustering methods that assume clearly separated clusters,FeDepth allows clients to participate in multiple clusters, capturing con-tinuous and ambiguous domain transitions commonly observed in roboticenvironments. Extensive experiments demonstrate that FeDepth con-sistently improves robustness over standard FL and clustered FL base-lines across multiple depth estimation architectures, providing a practicaland effective solution for federated robot perception. Our project pageis available at https://vision3d-lab.github.io/fedepth/.