ECCV 2026 论文列表

Test-time adaptation (TTA) methods are typically evaluated by task accuracy, leaving open the question of whether adaptation also changes which features the model relies on. We examine this question by tracking concept sensitivity throughout adaptation using TCAV directional derivatives. Across five TTA methods and five benchmarks spanning natural image and medical imaging domains, we observe recurring shifts in concept sensitivity: adaptation can reduce sensitivity to task-relevant or pathology-related concepts while increasing sensitivity to nuisance concepts such as backgrounds or institutional artifacts—a phenomenon we term Concept Sensitivity Drift, even when accuracy improves. On Waterbirds, where the causal/spurious split is defined by construction, this drift coincides with degraded worst-group accuracy; on cross-hospital medical adaptation, it coincides with increased reliance on hospital-specific artifacts rather than pathology. Motivated by this observation, we propose CS-TTA, a source-concept-supervised but target-label-free plug-in regularizer that mitigates excessive conceptsensitivity drift during adaptation. CS-TTA requires no architectural changes and can be added to any existing TTA objective. Experiments show that CS-TTA provides consistent worst-group accuracy gains on spurious-correlation benchmarks, improves AUC by 1.4–1.8 points on cross-hospital medical adaptation, and gives consistent overall accuracy gains on CIFAR-10-C and ImageNet-C.1234

The ability to render scenes at adjustable fidelity from a single model, known as level of detail (LoD), is crucial for practical deployment of 3D Gaussian Splatting (3DGS). Existing discrete LoD methods expose only a limited set of operating points, while concurrent continuous LoD approaches enable smoother scaling but often suffer noticeable quality degradation at full capacity, making LoD a costly design decision. We introduce Matryoshka Gaussian Splatting (MGS), a training framework that enables continuous LoD for standard 3DGS pipelines without sacrificing full-capacity rendering quality. MGS learns a single ordered set of Gaussians such that rendering any prefix, the first k splats, produces a coherent reconstruction whose fidelity improves smoothly with increasing budget. Our key idea is stochastic budget training: each iteration samples a random splat budget and optimises both the corresponding prefix and the full set. This strategy requires only two forward passes and introduces no architectural modifications. Experiments across four benchmarks and six baselines show that MGS matches the full-capacity performance of its backbone while enabling a continuous quality–speed trade-off from a single model. Extensive ablations on ordering strategies, training objectives, and model capacity further validate the designs.

We present BlenderFusion, a generative visual compositing framework that recomposes objects, camera, and background to synthesize new scenes. It follows a layering-editing-compositing pipeline that (i) segments and lifts up visual inputs into editable 3D entities (layering), (ii) edits them in Blender with 3D-grounded control (editing), and (iii) fuses them into a coherent scene using a generative compositor (compositing). The generative compositor extends a pre-trained diffusion model to a novel dual-stream architecture, processing both the original (source) and edited (target) scenes in parallel, and is fine-tuned on video frames with two important training strategies: (i) source masking, enabling flexible image modifications like background replacement or object replacement; (ii) simulated object jittering, facilitating robust disentangled control over objects and the camera. Extensive experiments on synthetic and real-world datasets show that BlenderFusion significantly outperforms prior methods in precise 3D-aware control and complex compositional scene editing. The framework also generalizes to unseen data and fine-grained editing operations beyond the training distribution. See the project page for demos and more results: blenderfusion.github.io.

Video personalization aims to generate personalized videos by preserving the motion and background of a reference video while transferring the subject appearance from a given subject image. Such video personalization has achieved tremendous progress alongside recent advancements in video diffusion transformers (DiT). Existing DiT-based approaches predominantly rely on training-based methods that require auxiliary control signals, while independent control over motion, subject appearance, and background remains challenging due to the entangled nature of the unified spatiotemporal attention mechanism. In this paper, we uncover a key structural insight that the spatio-temporal attention matrix of DiT exhibits an inherent functional decomposition—intra-frame blocks primarily encode spatial information, while inter-frame blocks encode motion dynamics. Building on this observation, we introduce SA-V2V, a novel training-free video personalization framework that disentangles and enables independent control over these visual attributes. We propose two guidance approaches: (1) For motion guidance, we present stochastic temporal injection, which leverages inter-frame attention maps to precisely control dynamic trajectories.(2) For appearance guidance, we propose a Target-aware Feature Guidance to enable independent control over subject and background appearance by modulating the synthesized features with masked attention features from the subject image and the reference video, respectively. Extensive experiments demonstrate that the proposed SA-V2V significantly outperforms existing approaches in motion fidelity, subject adherence, and background preservation. Our code is available at https://parksoobin1219.github.io/SA-V2V/

Album highlight recommendation (AHR) focuses on identifying images from personal collections that are most suitable for personal enjoyment and social sharing. However, due to the inherent subjectivity and ambiguity involved in selecting highlights, AHR research has long been hindered by limited data availability, resulting in relatively slow progress in the field. To mitigate this issue, we first engage multiple annotators to re-annotate the public PEC and CUFED album datasets, producing high-quality labels for 910 albums comprising 87K images. These binary annotations are then aggregated into finegrained highlight scores to reduce individual annotation bias. To assess how well public consensus aligns with personal preferences, we further collect 10K images from 100 volunteers and invite the album owners themselves to provide personalized highlight annotations. This newly constructed dataset is named C-AHR-100. In addition to contributing these three benchmark datasets, we propose Highlight4U, an agentic multi-modal large language model (MLLM) framework specifically optimized for AHR. Highlight4U first identifies highlight-worthy reference images through localized scoring of individual images, and then refines its recommendations by incorporating contextual and semantic information across the reference set. Extensive experiments demonstrate that Highlight4U effectively captures public highlight perception and can further generalize to accommodate the personalized preferences of individual album owners. The source code and dataset are provided in the https://github.com/naninasm/Finding-Highlight-Images-InYour-Albums-From-Benchmark-To-MLLM.

Interactive long video generation requires prompt switching to introduce new subjects or events, while maintaining perceptual fidelity and coherent motion over extended horizons. Recent distilled streaming video diffusion models reuse a rolling KV cache for long-range generation, enabling prompt-switch interaction through re-cache at each switch. However, existing streaming methods still exhibit progressive quality degradation and weakened motion dynamics. We identify two failure modes specific to interactive streaming generation: (i) at each prompt switch, current cache maintenance cannot simultaneously retain KV-based semantic context and recent latent cues, resulting in weak boundary conditioning and reduced perceptual quality; and (ii) during distillation, unbounded time indexing induces a positional distribution shift from the pretrained backbone’s bounded RoPE regime, weakening pretrained motion priors and long-horizon motion retention. To address these issues, we propose Anchor Forcing, a cache-centric framework with two designs. First, an anchor-guided re-cache mechanism stores KV states in anchor caches and warm-starts re-cache from these anchors at each prompt switch, reducing post-switch evidence loss and stabilizing perceptual quality. Second, a tri-region RoPE with region-specific reference origins, together with RoPE re-alignment distillation, reconciles unbounded streaming indices with the pretrained RoPE regime to better retain motion priors. Experiments on long videos show that our method improves perceptual quality and motion metrics over prior streaming baselines in interactive settings.

This paper presents MAGiSt3R, a multi-agent 3D reconstruction framework performing reconstruction and camera tracking for monocular RGB videos at almost 10 FPS. MAGiSt3R relies on a feedforward model from the 3R family to process RGB videos and regress local point maps, and on a merging model, MAGMA, that combines local maps at both intra-agent and inter-agent levels to obtain the final, global point map. Furthermore, MAGiSt3R performs pose graph optimization to mitigate cumulative camera drift occurring along the feedforward pipeline. We evaluate MAGiSt3R on both synthetic and realworld datasets, demonstrating its superior reconstruction and camera tracking accuracy compared to state-of-the-art feed-forward approaches.

Recent advances in self-supervised visual representation learning have demonstrated the effectiveness of predictive latent-space objectives for learning transferable features. In particular, Image-based JointEmbedding Predictive Architecture (I-JEPA) learns representations by predicting latent embeddings of masked target regions from visible context. However, it predicts target regions in parallel and all at once, lacking ability to order predictions meaningfully. Inspired by human visual perception, which attends selectively and progressively from primary to secondary cues, we propose DSeq-JEPA, a Discriminative Sequential Joint-Embedding Predictive Architecture that bridges latent predictive and autoregressive self-supervised learning. Specifically, DSeq-JEPA integrates a discriminatively ordered sequential process with JEPA-style learning objective. This is achieved by (i) identifying primary discriminative regions using an attention-derived saliency map that serves as a proxy for visual importance, and (ii) predicting subsequent regions in discriminative order, inducing a curriculum-like semantic progression from primary to secondary cues in pre-training. Extensive experiments across tasks – image classification (ImageNet), fine-grained visual categorization (iNaturalist21, CUB, Stanford Cars), detection/segmentation (MS-COCO, ADE20K), and low-level reasoning (CLEVR) – show that DSeq-JEPA consistently learns more discriminative and generalizable representations compared to I-JEPA variants.

Few-shot semantic segmentation (FSS) often degrades when support and query images exhibit large pose or scale differences since conventional prototype matching operates in geometry-entangled feature spaces. To mitigate this brittleness, we propose a lightweight encoderdecoder framework that disentangles object representations into a geometryinvariant semantic code and a low-dimensional modulation code capturing instance-specific geometric variation. The reconstruction decoder is trained with equivariance constraints so that geometric information gets encoded by the modulation pathway while preserving semantic consistency in a separate branch. As a result, robustness to pose-scale discrepancies is improved before segmentation without explicit cross-image alignment. To further strengthen supervision from limited supports, we introduce a refinement module that synthesizes lightweight augmented views of each support and performs cyclic ensemble refinement to generate more stable predictions. Predictions from multiple supports are then fused using spatially adaptive reliability weighting, producing cleaner and better-aligned query predictions. Across standard FSS benchmarks, our method consistently improves performance - particularly under large viewpoint changes. Ablations confirm that the disentanglement combined with cyclic ensemble and spatial refinement are critical to the gains. Project Website.

Visibility in underwater environments degrades rapidly under turbid conditions, yet the effects on computer-vision models remain unclear. This issue is compounded by reliance on synthetic turbidity datasets, which may misrepresent real-world information loss. To address this gap, we introduce the Turbid Underwater Baseline (TUB) dataset, comprising 1,320 images captured under extreme turbidity and over 16,000 high-confidence ground-truth segmentation masks. We additionally propose PCD, a metric derived from phase congruency maps that is invariant to contrast and aims to capture the loss of structural information in real turbidity. We show that PCD correlates strongly with the performance of instance segmentation models on both real and synthetic turbid images, whereas common metrics in the field show weak to no correlation at all. The dataset and relevant code can be found on the project page: https://vap.aau.dk/pcd

The Mueller matrix encodes rich structural and physical information, but its rapid and accurate estimation remains challenging. While physics-based polarimetric Bidirectional Reflectance Distribution Function (pBRDF) models provide useful formulation constraints, their application is limited by inaccessible parameters and model mismatch. To address this challenge, we propose a physics-guided two-stage framework for single-shot linear Mueller matrix recovery. Given one set of four polarization images and an object mask under a fixed known incident polarization state, the first stage estimates pBRDF-related parameters and uses an analytical pBRDF model to construct a structured Mueller initialization. A second network then refines this initialization through residual correction. Evaluated on a newly constructed hybrid real-synthetic dataset, our method improves matrix reconstruction, observed-state rendering, held-out forward Stokes prediction, and physical-plausibility diagnostics over adapted pBRDF fitting baselines. Downstream validation through shape from polarization (SfP) and material classification further provides secondary evidence that the recovered matrices preserve useful polarimetric cues for geometric and material analysis.

Depth foundation models (DFMs) offer strong learned priors for 3D perception from single RGB images but lack physical depth cues, leading to ambiguities in metric scale. We introduce metalenses, an emerging class of ultrathin planar optical elements, as a solution to physically encode missing metric depth cues via nanophotonics. In this paper, we bridge the gap between metalens and DFMs to achieve accurate metric monocular depth sensing. In a single monocular shot, our metalens embeds depth-dependent positional shifts into two polarized optical wavefronts. With an input adaptation strategty, we enable direct fine-tuning that aligns a pretrained DFM with the optical signals. To scale the training data, we further develop a comprehensive simulation pipeline that synthesizes metalens responses from RGB-D datasets, incorporating physical factors to minimize the sim-to-real gap. Experiments demonstrate that this approach outperforms both monocular metric depth estimation and depth-from-defocus baselines, showing an effective pathway for accurate monocular metric depth sensing.

Image morphing aims to produce a smooth and semantically consistent transition between two input images. Existing diffusionbased morphing methods either require expensive per-pair optimization or rely on implicit spatial alignment, which easily fails under large layout discrepancies. To address these limitations, we propose AlignMorph, a novel tuning-free diffusion framework guided by the principle of transport-then-denoise. We explicitly decouple geometric alignment from generative denoising to avoid structural entanglement. Our framework consists of two core components. (1) Global Semantic Transport, which achieves diffusion-compatible semantic alignment via entropic optimal transport and reliability-aware latent warping; and (2) Coordinate-Aligned Generation, which uses a symmetric bi-phase attention handoff to maintain consistent spatial coordinates throughout denoising. Without any tuning, AlignMorph effectively eliminates ghosting and achieves superior structural coherence and temporal smoothness on morphing benchmarks. Code is available at https://github.com/ 51xOne/Alignmorph.

Multi-task model merging aims to consolidate several task-specific experts into a unified model, yet static merging consistently suf-fers from parameter interference. While dynamic merging models aimto bridge this gap, many works rely on the costly storage and loadingof redundant expert components at inference. In this work, from theperspective of task expert, we view parameter interference as parame-ter perturbation introduced to each expert during merging process. Weshow that such parameter perturbations can be modeled as affine trans-formation, which can be approximated as additive offsets. Motivatedby these, we propose Recover Task eXpert (ReTeX), a frameworkthat predicts those offsets, in order to undo parameter interference andrecover task-expert performance from a single merged checkpoint. Torecover the appropriate expert when task identity is unknown, we in-troduce a router-free task identifier based on SVD subspace signaturescomputed offline before inference. At inference, the identifier selects thetask whose subspace yields the smallest projection residual for a giveninput. As a result, ReTeX recovers over 95% of individual-expert per-formance in both vision and NLP domains, while significantly improvinggeneralization to unseen tasks. Crucially, we also show that the param-eter offset prediction leads to emergent adaptive interpolation of expertknowledge for out-of-distribution (OOD) tasks. ReTeX adaptively in-terpolates seen expert knowledge to handle unseen tasks. Our code isavailable at https://github.com/BAIKLAB/ReTeX

Text-guided image editing aims to modify visual content ac-cording to a target prompt while preserving the background. Recentinversion-free image editing frameworks such as FlowEdit have demon-strated strong editing capability without requiring inversion. Empirically,FlowEdit can achieve substantial semantic changes under appropriatehyperparameter settings. However, we observe that under certain globalattribute shifts, the editing trajectory may not effectively move awayfrom the source distribution in the early timesteps. Our analysis sug-gests that in the high-noise regime, the dominant manifold-seeking flowtoward the data manifold can reduce the influence of the text-conditioneddirection, leading to limited global modification while background struc-tures remain only moderately preserved. Inspired by this observation,we propose an inversion-free, frequency-aware semantic compensationstrategy that strengthens the effective signal in the early stage of gener-ation, while maintaining structural consistency in the background. Theproposed method improves global editing capacity without sacrificingbackground fidelity.

Camera-based perception systems for autonomous drivingare typically developed and evaluated using fixed sensor rigs, while real-world vehicle fleets exhibit substantial variation in camera placement,orientation, field of view, and camera count. This mismatch introducesa cross-rig domain gap in which only the geometric observation pro-cess changes. To study this effect under controlled conditions, we intro-duce Plentiful CARLA Camera Rigs, a benchmark that renders identicaldriving scenes under 14 systematically designed camera rigs. This setupenables direct analysis of cross-rig generalization without confoundingchanges in scene content or appearance. Using the benchmark, we ana-lyze cross-rig transfer behavior of representative multi-view perceptionarchitectures and observe substantial performance shifts induced by ge-ometric rig variation. To facilitate structured analysis, we further intro-duce two calibration-based descriptors derived from rig metadata: RigVariance, capturing internal rig diversity, and Rig Contrastive Distance,measuring geometric discrepancy between rigs. Our experiments showthat geometric rig differences strongly correlate with relative cross-rigperformance shifts and that Rig Contrastive Distance provides a reliableproxy for ranking transfer difficulty between sensor rigs.

Vision-language models offer a promising approach for zero-shot anomaly detection (ZSAD). However, due to object-centric bias,normal and anomalous text prototypes exhibit a high semantic over-lap. While enforcing strict orthogonality between them improves dis-criminability, mapping highly contiguous visual inputs onto drasticallyorthogonal prototypes introduces a geometric dilemma, disrupting thepre-trained structural continuity. To address this problem, we proposeProximity-CLIP, a framework that visually calibrates the semantic mar-gin to guide visual adaptation. First, we introduce a visually-calibratedsemantic proximity learning mechanism that uses a bounded dynamicregularization to learn an appropriate semantic margin, ensuring dis-criminative separation while preserving structural alignment. Second, wedesign an Anomaly Query Module (AQM) driven by these text priors.Using the calibrated anomalous prototype as a semantic query, the AQMactively retrieves localized defect cues from contextual visual patches,mitigating the dilution of subtle anomalies during global pooling. Exten-sive experiments demonstrate that Proximity-CLIP outperforms currentstate-of-the-art methods across multiple ZSAD benchmarks with mini-mal architectural modifications.

Plane segmentation from a single RGB image remains chal-lenging due to imprecise region grouping and geometrically inconsistentsupervision, often leading to over-segmentation and false planar detections.We propose instead a pixel-wise planarity prediction framework for ro-bust monocular plane segmentation. Building on a pretrained monoculargeometric backbone predicting depth and surface normals, we introducea dedicated planarity head that estimates per-pixel planarity confidence.During inference, predicted depth, normals, and planarity are combinedin a lightweight region-growing procedure that enforces geometric consis-tency when forming plane segments. We further analyze existing planeground-truth annotations and demonstrate substantial geometric incon-sistencies under strict distance thresholds. Across multiple datasets, ourmethod achieves improved geometric precision and segmentation qualitycompared to prior state-of-the-art approaches, while improving computa-tional efficiency.

We study learning from limited supervision (LLS) for hyperspectral imagery (HSI): learning transferable representations from abundant unlabeled HSIs in the target sensing domain and then adapting with only a handful of labels per class. Joint-Embedding Predictive Architectures (JEPAs) are appealing for LLS because they learn via latent prediction rather than pixel reconstruction; however, they remain unexplored for HSI. Generic masking or cropping is ill-suited to HSI: band masking corrupts high-dimensional spectra, spatial masking removes entire pixel-level semantics, and dense spatial–spectral extraction can yield highly correlated, degenerate prediction targets. We propose MorphJEPA, the first JEPA tailored to HSI, which uses morphology as the pretext signal by predicting embeddings of pixel-aligned targets generated by per-band morphological opening/closing. This induces a directional structural bottleneck that preserves core topology while suppressing high-frequency nuisances. MorphJEPA couples this objective with a compact disentangled spatial–spectral encoder fused by crossattention and a context-anchored Sketched Isotropic Gaussian Regularizer (SIGReg) to prevent collapse and maintain a well-conditioned latent manifold. Extensive 1/5/10-shot evaluations show consistent gains in intra-scene settings (Houston 2013, Trento) and under cross-scene temporal/spatial shifts (Houston 2013 → 2018, HyRank) for both zero-shot transfer and low-shot adaptation. Project page: https://github.com/ amartya-ray/MorphJEPA.