ECCV 2026 Papers

World models have driven remarkable progress in physicalperception and dynamic forecasting, yet current mainstream implemen-tations predominantly rely on visible-light (RGB) observations, whichlimits their all-weather reliability. Developing a thermal world model isessential, but it is hindered by the unique distribution of thermal imageryand the extreme scarcity of continuous video datasets. In this paper, weidentify a fundamental challenge in extending standard masked modelingto the thermal domain: intra-frame spatial homogenization. Due to ther-mal equilibrium, distinct physical entities often exhibit near-zero temper-ature gradients. Consequently, unconstrained self-attention blurs seman-tic boundaries and collapses object topologies, severely degrading down-stream dynamics forecasting. To overcome this, we propose Thermo-JEPA, a geometry-grounded pre-training framework based on Learn-ing Using Privileged Information (LUPI). We extract intra-frame spatialtopologies from aligned RGB data and inject them as a confidence-gatedstructural prior into the thermal teacher’s attention mechanism. Thisexplicitly enforces semantic segregation without corrupting fine-grainedintra-object thermal gradients. Concurrently, we release RGBT-World,the largest unified, spatiotemporally aligned RGB-Thermal video corpustailored for generative pre-training. Extensive experiments demonstratethat Thermo-JEPA effectively prevents representation collapse, achiev-ing state-of-the-art zero-shot dynamics alignment on pure thermal inputsand significantly outperforming massive generic video foundation mod-els.

We present a framework for document-centric backgroundgeneration with multi-page editing and thematic continuity. To ensuretext regions remain readable, we employ a latent masking formulationthat softly attenuates updates in the diffusion space, inspired by smoothbarrier functions in physics and numerical optimization. In addition, weintroduce Automated Readability Optimization (ARO), which automat-ically places semi-transparent, rounded backing shapes behind text re-gions. ARO determines the minimal opacity needed to satisfy percep-tual contrast standards (WCAG 2.2) relative to the underlying back-ground, ensuring readability while maintaining aesthetic harmony with-out human intervention. Multi-page consistency is maintained througha summarization-and-instruction process, where each page is distilledinto a compact representation that recursively guides subsequent gen-erations. This design reflects how humans build continuity by retain-ing prior context, ensuring that visual motifs evolve coherently acrossan entire document. Our method further treats a document as a struc-tured composition in which text, figures, and backgrounds are preservedor regenerated as separate layers, allowing targeted background editingwithout compromising readability. Finally, user-provided prompts allowstylistic adjustments in color and texture, balancing automated consis-tency with flexible customization. Our training-free framework producesvisually coherent, text-preserving, and thematically aligned documents,bridging generative modeling with natural design workflows.

Dataset distillation (DD) has proven to reduce training costwhile preserving accuracy. While promising, the factors that make onedistilled dataset more effective than another remain poorly understood.In this work, we investigate this question through the lens of discretevisual tokenizers. Whereas many prior DD efforts emphasize matchingglobal data distributions, we suggest that the effectiveness depends onwhich semantic concepts are captured and how they are composed. Dis-crete visual tokenizers provide a finite vocabulary that enables directstatistical analysis of such compositional structure. Through quantita-tive analysis of token-level statistics, we introduce the structural scoreto measure the adequacy of token compositions. We observe that dis-tilled datasets with balanced token composition yield higher validationperformance. On the other hand, divergence from the original data doesnot necessarily harm performance. We further show that samples withhigh structural scores in the discrete token space can effectively guidediffusion-based DD. Our findings highlight the importance of token com-position in dataset effectiveness, offering a principled complement to dis-tributional similarity considerations in DD.

Remote photoplethysmography (rPPG) enables non-contactphysiological monitoring by capturing subtle skin color variations inducedby cardiac cycles. Despite its promise, rPPG remains highly sensitive toenvironmental illumination. Existing illumination-aware methods sufferfrom two key limitations: (1) limited out-of-distribution (OOD) gen-eralization due to domain discrepancies between training and testingdata, and (2) reliance on the Lambertian assumption, which neglectsnon-Lambertian specular highlights that frequently corrupt facial skinsignals. In this paper, we propose SHINE-PPG (Specular-Highlight In-trinsic Network for rPPG Estimation), a novel framework that leveragesnon-Lambertian intrinsic decomposition to decouple facial videos intoillumination, reflectance, and specular components in a self-supervisedmanner. By isolating physiological information within the intrinsic re-flectance, our method effectively suppresses both ambient lighting vari-ations and surface highlights to recover high-fidelity rPPG signals. Tofurther enhance robustness, we introduce an adversarial illuminationenhancement strategy that dynamically synthesizes challenging unseenlighting conditions during training, significantly improving OOD general-ization. Extensive experiments on five benchmark datasets demonstratethat SHINE-PPG consistently outperforms previous methods, particularlyunder complex and dynamic illumination scenarios. The code is availableat https://github.com/Edmond-Yang/SHINE-PPG.

We present SD3.5-Fla.sh, an efficient few-step distillation framework that brings high-quality image generation to accessible consumer devices. Our approach distills computationally prohibitive rectified flow models through a reformulated distribution matching objective tailored specifically for few-step generation. We introduce two key innovations: “timestep sharing” to reduce gradient noise and “split-timestep fine-tuning” to improve prompt alignment. Combined with comprehensive pipeline optimizations like text encoder restructuring and specialized quantization, our system enables both rapid generation and memory-efficient deployment across different hardware configurations. This democratizes access across the full spectrum of devices, from mobile phones to desktop computers. Through extensive evaluation including large-scale user studies, we demonstrate that SD3.5-Flash consistently outperforms existing few-step methods, making advanced generative AI truly accessible for practical deployment.

Knowledge Distillation (KD) transfers knowledge from a largeteacher model to a smaller student by aligning their predictive distri-butions. However, conventional KD formulations - typically based onKullback–Leibler divergence - assume that the teacher provides reliabletargets. In practice, teacher predictions are often noisy or overconfident,and existing correction-based approaches rely on ad-hoc heuristics andextensive hyper-parameter tuning, which hinders generalization. We in-troduce REDistill (Robust Estimator Distillation), a simple yet prin-cipled framework grounded in robust statistics. REDistill replaces thestandard KD objective with a power divergence loss, a generalizationof KL divergence that adaptively downweights unreliable teacher out-put. This formulation provides a unified and interpretable treatmentof teacher noise, requires only logits, integrates seamlessly into existingKD pipelines, and incurs negligible computational overhead. Extensiveexperiments on CIFAR-100 and ImageNet-1k demonstrate that REDis-till consistently improves student accuracy in diverse teacher-student ar-chitectures. Remarkably, it achieves these gains without model-specifichyper-parameter tuning, underscoring its robustness and strong general-ization to unseen teacher–student pairs.

We study the problem of physically plausible shadow castingwhen animating 3D Gaussian Splatting (3DGS) avatars, either individu-ally or in multi-avatar and object-interaction scenarios, within existing3DGS scenes. In contrast to prior methods that rely on binary hit tests andmesh-based shadow casters, our method performs shadow computationentirely in Gaussian space, without requiring any mesh reconstruction. Weintroduce RAGA, a Ray-Traced Gaussian Shadow Casting formulationbased on exact ray–Gaussian line integrals. For each occluding Gaussian,we integrate the opacity profile along the shadow ray and normalize bythe theoretical maximum integral, producing a weight that captures howthe ray traverses the occluder rather than merely whether an intersectionoccurred. To reduce temporal variance from clothing deformations in ani-mated avatars, we further introduce an avatar proxy representation thatstabilizes shadow casting while preserving visual fidelity. We implement*Work partly done as intern at Snap Inc.RAGA using custom CUDA kernels integrated with the NVIDIA OptiXframework; as such, our shadow tracer runs at rates of about 50 FPS.We evaluate on single-avatar, multi-avatar, and avatar–object interactionscenarios across multiple datasets, demonstrating substantially improvedshadow realism, temporal stability, and scene coherence. Our project pageis available at https://miraymen.github.io/raga/.

Heterogeneity in data distributions and model architectures poses significant challenges in federated learning (FL). While various heterogeneous FL (HtFL) approaches have been proposed to address these challenges, prototype-based FL (PBFL) has emerged as a promising framework that exchanges prototypes—per-class mean activations from the penultimate layer—rather than model parameters. However, existing PBFL methods suffer from suboptimal prototype separation in the feature space, which limits their discriminative capacity. We propose Prototype Normalization (ProtoNorm), a novel PBFL framework that systematically addresses this limitation. Inspired by the Thomson problem in classical physics, ProtoNorm optimizes global prototype configurations on a unit hypersphere, maximizing separation between class prototypes. Extensive experimental evaluation demonstrates that ProtoNorm consistently outperforms existing HtFL methods across diverse heterogeneous settings. Importantly, ProtoNorm maintains communication efficiency while operating entirely on the server side with less computation than existing PBFL methods, making it well-suited for resource-constrained environments.

Polarimetric imaging enables advanced vision applicationssuch as normal estimation and de-reflection by capturing unique surface-material interactions. However, existing applications (alternatively calleddownstream tasks) rely on datasets constructed by naïvely regroupingraw measurements from division-of-focal-plane sensors—where pixels ofthe same polarization angle are extracted and aligned into sparse im-ages without proper demosaicking. This reconstruction strategy resultsin suboptimal, incomplete targets that limit downstream performance.Moreover, current demosaicking methods are task-agnostic, optimizingonly for photometric fidelity rather than utility in downstream tasks.Towards this end, we propose PolarAPP, the first framework to jointlyoptimize demosaicking and its downstream tasks. PolarAPP introducesa feature alignment mechanism that semantically aligns the representa-tions of demosaicking and downstream networks via meta-learning, guid-ing the reconstruction to be task-aware. It further employs an equivalentimaging constraint for demosaicking training, enabling direct regressionto physically meaningful outputs without relying on rearranged data.Finally, a task-refinement stage fine-tunes the task network using thestable demosaicking front-end to further enhance accuracy. Extensiveexperimental results demonstrate that PolarAPP outperforms existingmethods in both demosaicking quality and downstream performance.Code is available here.

Adverse weather image restoration aims to recover clear vis-ibility from degraded images in complex weather conditions. Existingworks attempt to address this problem by modeling relationships betweenpixels, however, this paradigm defies the spatially non-uniformity factof degradations and learns non-discriminative features from semantic-conflict regions. In this paper, we propose SSR, a Semantic-center guildedState space model for image Restoration. The key idea of SSR is to shiftthe conventional scanning strategy of pixel-serial to semantic-guildedone. Specifically, we introduce a Superpixel-guided Selective Scan Mech-anism (S3 M), which first partitions the image into perceptually coher-ent regions via superpixel clustering and then performs relations mod-eling within the semantic-related regions. Moreover, a Region-level Gat-ing Mechanism (RGM) is developed to perform intra-region calibra-tion by modulating degradation outliers within each semantic super-pixel unit along the channel dimension. Extensive experiments on 6well-established benchmarks demonstrate that SSR performs favorablyagainst state-of-the-art models with competitive computational cost. Thesource code is publicly available at https://github.com/LIDAYU-DayuLi/SSR.

Recent progress in CLIP-like vision-language models (VLMs)has greatly advanced medical image analysis. However, most existingmedical VLMs still rely on coarse image-text contrastive objectives andfail to capture the systematic visual knowledge encoded in well-definedmedical phenotype ontologies. To address this limitation, we constructPhenoKG, the first large-scale, phenotype-centric multimodal knowl-edge graph that encompasses around 524K high-quality image-text pairslinked to more than 3,000 phenotypes. Building upon PhenoKG, we pro-pose PhenoLIP, a pretraining framework that explicitly incorporatesstructured phenotype knowledge into medical VLMs through a two-stage process. We first learn a knowledge-enhanced phenotype embed-ding space that captures the hierarchical structure of phenotype ontolo-gies and then distill this textual knowledge into multimodal pretrainingvia a teacher-guided knowledge distillation objective. To support eval-uation, we further introduce PhenoBench, an expert-verified bench-mark designed for phenotype recognition, comprising over 7,800 image-caption pairs covering more than 1,000 phenotypes. Extensive experi-ments demonstrate that PhenoLIP outperforms previous state-of-the-artbaselines, improving upon BIOMEDICA in phenotype classification ac-curacy by 8.06% and in cross-modal retrieval by 15.03%, underscoringthe value of integrating phenotype-centric priors into medical VLMs forstructured and interpretable medical image understanding.

Industrial anomaly inspection faces a major challenge dueto the lack of real-world anomaly samples. While generative models areused to create anomaly data, existing methods still struggle when han-dling small-scale anomalies. This failure occurs because extreme down-sampling in diffusion models causes the information of small anomaliesto be lost in the latent space. To address this, we introduce UniScale,a unified training and inference framework for high-fidelity industrialanomaly generation across arbitrary scales. During training, we intro-duce an Error-Suppressed Multi-Scale Training (EMT) strategy, whichenables the model to learn the rich location-aware textures of anomalies,while suppressing upsampling-induced interpolation errors in texture ac-quisition, ensuring the model is capable of learning small-scale anomalies,while remaining effective for regular scale anomalies. For inference, wepropose Generation-then-Fusion Denoising. It decouples anomaly genera-tion from background integration, preventing small anomalies from beingoverwhelmed. Extensive experiments demonstrate that our method out-performs state-of-the-art competitors in both anomaly generation qualityand downstream detection performance. It achieves a relative IS(a) im-provement of 45.86% (from 1.81 to 2.64) on VisA and 37.70% (from 1.22to 1.68) on MVTec AD 2, while also improving the downstream pixel-level IoU by 4.22% on VisA and AUROC by 6.55% on MVTec AD 2.Code is available at https://github.com/HUST-SLOW/UniScale.

Self-supervised learning (SSL) has become a leading paradigmfor skeleton-based action recognition, yet the choice of prediction targetremains a central limitation. Existing masked modeling methods typi-cally reconstruct raw joint coordinates, emphasizing low-level detail andremaining sensitive to noise, while recent non-reconstruction variantspredict learned latent features that improve semantics but capture lit-tle explicit motion structure. We propose Path-JEPA, a self-supervisedpredictive learning framework that adopts path signatures as the predic-tion target for skeleton sequences. Rather than predicting coordinatesor abstract embeddings, Path-JEPA predicts latent representations ofcontinuous-time, multi-scale geometric descriptors computed over joint,edge, and chain paths, capturing motion across multiple spatial and tem-poral scales. These targets encode displacement, signed area, and higher-order interactions, yielding a geometry-grounded representation of hu-man motion that is naturally robust to temporal resampling and frame-rate variation. To make such targets effective within a JEPA framework,we introduce signature-augmented masking, which propagates joint-spacemasks to all dependent signature tokens, forcing the model to infer themissing motion geometry from broader anatomical and temporal context.Because signatures summarize motion over intervals rather than indi-vidual frames, they also yield greater computational efficiency on longsequences. Extensive experiments on NTU RGB+D 60, NTU RGB+D120, and PKU-MMD show that Path-JEPA learns stronger representa-tions than prior masked prediction methods, achieving state-of-the-artperformance across diverse downstream tasks while exhibiting improvedrobustness to irregular sampling. Project website:Path-JEPA.

Reconstructing 3D geometry from streaming video requirescontinuous inference under bounded resources. Recent geometric foun-dation models achieve impressive reconstruction quality through all-to-all attention, yet their quadratic cost confines them to short, offline se-quences. Causal-attention variants such as StreamVGGT enable single-pass streaming but accumulate an ever-growing KV cache, exhaustingGPU memory within hundreds of frames and precluding the long-horizondeployment that motivates streaming inference in the first place. Wepresent OVGGT, a training-free framework that bounds both memoryand compute to a fixed budget regardless of sequence length. Our ap-proach combines Self-Selective Caching, which leverages FFN residualmagnitudes to compress the KV cache while remaining fully compatiblewith FlashAttention, with Dynamic Anchor Protection, which shieldscoordinate-critical tokens from eviction to suppress geometric drift overextended trajectories. Extensive experiments on indoor, outdoor, andultra-long datasets show that OVGGT processes arbitrarily long videoswithin a constant VRAM envelope while achieving state-of-the-art 3D ge-ometric accuracy. The code is available at github.com/VAISR/OVGGT.

Facial landmark detectors trained on real human faces oftenfail to generalize effectively to stylized domains such as caricatures andartistic portraits, necessitating unsupervised domain adaptation (UDA).Although self-training is a widely used UDA strategy to bridge do-main gaps, it frequently breaks down under large domain shifts as it isprone to amplifying confident yet erroneous pseudo-label predictions. Tothis end, we propose PriSM (Parsing and Style-Mixed Consistency), anovel method for robust landmark pseudo-label validation. PriSM lever-ages two complementary signals: a Parsing Network enforces high-levelstructural alignment through face parsing consistency, while a StyleMixNetwork enforces fine-grained geometric constraints by reducing the re-construction error between the input face and another face synthesizedfrom the pseudo-label’s structure and the input face’s appearance. Ex-tensive experiments on the challenging CariFace and ArtiFace bench-marks under the UDA setting demonstrate that PriSM significantly out-performs existing state-of-the-art methods and exhibits strong gener-alizability to unseen domains. The source code is available at: https://github.com/jerrychild999922/PriSM-UDA.

Differentiable rendering of planar rational splines remains largely underexplored, despite their widespread use in vector graphics and design. Existing differentiable vector renderers primarily focus on Bézier curves and rely on analytic rasterization, which can suffer from gradient instability and limited flexibility. We propose NURBS Splatting, a unified framework that represents planar rational curves as continuous Gaussian fields. By sampling Gaussians along the curve parameter domain and inside closed regions, rendering is reformulated as a smooth accumulation process with stable gradients. Our method naturally supports long splines, rational weights, non-uniform knots, and closed-region filling. We demonstrate its effectiveness in calligraphy reconstruction, vectorization frameworks, and long-spline image abstraction, showing improved stability and reconstruction quality over existing approaches.

Multi-modality image fusion (MMIF) enhances scene representation by exploiting complementary cues from different modalities. Adverse weather, however, causes significant image degradation, disrupting feature representation and requiring simultaneous feature restoration and cross-modal complementarity. Existing methods often struggle with effective representation learning under such conditions, limiting their practical performance. To address these challenges, we propose a mask-guided MMIF method that integrates feature restoration and interaction. We first introduce "Pseudo Ground Truth" to simplify training, promoting faster and more effective feature learning. Then, we design a mask generation mechanism based on the mapping relationship between the fused result and the source images, quantifying the relative contribution of each modality during the fusion process. By incorporating the proposed mask-guided cross-modal cross-attention mechanism, the network is encouraged to selectively attend to informative features during modality interaction, mitigating the risk of overfitting to the static distribution of the "Pseudo Ground Truth". Additionally, we propose a mask-guided learning strategy and a task-coupled degradation-aware learning strategy to balance feature restoration and interaction. Extensive experiments on synthetic and real-world datasets demonstrate that our method surpasses state-of-the-art approaches in visual quality, quantitative metrics, and downstream tasks. The source code is available at https://github.com/ixilai/AMG-Fuse.

We present Split-then-Merge (StM), a controllable genera-tive video composition framework that minimizes reliance on annotateddatasets and handcrafted rules. Instead of requiring manual supervision,StM decomposes unlabeled videos into dynamic foreground and back-ground layers. By self-composing these elements, the model learns tosynthesize complex interactions between moving subjects and diversescenes, capturing the intricate dynamics necessary for high-_x001C_delity videogeneration. Speci_x001C_cally, StM introduces a transformation-aware train-ing pipeline utilizing multi-layer fusion and augmentation to address af-fordance challenges in video composition. An identity-preservation lossfurther maintains foreground _x001C_delity during blending. Extensive exper-iments show that StM outperforms state-of-the-art methods on bothquantitative benchmarks and qualitative evaluations, including humanstudies and Vision-Language Model assessments. Finally, we release StM-50K, the _x001C_rst multi-layer video dataset, to facilitate future research ingenerative video composition. Data, code, and more details are availableat our project page.Fig. 2: Video Composition. Given input foreground and background videos, image-based methods (a)_x0015_(b) use only the _x001C_rst frame, while (c)_x0015_(e) take full video inputs.(a) Object insertion [59] followed by Image-to-Video (I2V) and (b) end-to-end I2Vcomposition SkyReels [13] fails to retain motion due to lack of video access. (c) Manualcopy-paste preserves motion but violates a_x001B_ordance (swan placed on ground). (d) Naivegenerative composition yields appearance and motion drift (e.g ., black swan turnswhite). (e) Our method preserves identity and motion, and achieves a_x001B_ordance-awareplacement with realistic blending (swan placed in water with wave and shadows).

Diffusion models have recently achieved remarkable successin high-fidelity image synthesis, yet their application to visual text gener-ation and editing remains relatively underexplored. Unlike general imagegeneration, visual text tasks demand precise structural regularity andlegibility, which may pose additional challenges for small-scale text andnon-Latin scripts such as Chinese. Existing UNet-based models oftenstruggle to produce clear and coherent text, while DiT-based models,though more expressive, are typically limited to a single task, whichmay lead to redundant training pipelines, inconsistent visual styles, andreduced cross-task generalization. To address these challenges, we pro-pose InnoText, a unified DiT-based framework capable of performingboth text generation and editing within a single model. We introduce aFont Size-Aware Modulation (FSAM) module to enhance representationsacross font scales, a Small-Character Aware Augmentation strategy toimprove fine-grained fidelity, and a Task-Specific Region Weighted Lossfor adaptive optimization. To support training and evaluation, we alsoconstruct a high-quality bilingual (English-Chinese) visual text datasetcovering diverse fonts, sizes, and backgrounds. Experimental results demon-strate that our method achieves superior generation accuracy and editingquality, producing visually appealing and realistic text images.