BEGIN:VCALENDAR VERSION:2.0 PRODID:Linklings LLC BEGIN:VTIMEZONE TZID:America/Los_Angeles X-LIC-LOCATION:America/Los_Angeles BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:19700308T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:19701101T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260805T205609Z LOCATION:Hall K DTSTART;TZID=America/Los_Angeles:20260719T180000 DTEND;TZID=America/Los_Angeles:20260719T204500 UID:siggraph_SIGGRAPH 2026_sess237@linklings.com SUMMARY:Papers Fast Forward DESCRIPTION:Sponsored by Adobe Research\n\nFabricTryOn: Taming Image Editi ng Models for Garment Re-Texturing\n\nWe study fabric try-on as a garment re-texturing task, replacing fabric materials while preserving garment geo metry and illumination. Leveraging modern image editing models and a real- image data curation pipeline, our framework performs fabric removal and ap plication directly in image space, enablin...\n\n\nJun Ma (Zhejiang Sci-Te ch University, Style3D Research); Qian He (State Key Lab of CAD and CG, Zh ejiang University; Style3D Research); Gaofeng He and Huang Chen (Style3D R esearch); Chen Liu (State Key Lab of CAD and CG, Zhejiang University; Styl e3D Research); Xiaogang Jin (State Key Lab of CAD and CG, Zhejiang Univers ity); Yin Yang (University of Utah, Style3D Research); and Huamin Wang (St yle3D Research)\n---------------------\nDreamActor-M2: Universal Character Image Animation via Spatiotemporal In-Context Learning\n\nWe present Drea mActor-M2, a universal character animation framework that regards motion c onditioning as spatiotemporal in-context learning. It leverages video foun dation model priors and enables end-to-end pose-free motion transfer from raw videos. Without explicit pose estimation, it achieves stron...\n\n\nMi ngshuang Luo (ByteDance; Institute of Computing Technology, Chinese Academ y of Sciences); Shuang Liang, Yuxuan Luo, Zhengkun Rong, and Tianshu Hu (B yteDance); Ruibing Hou and Hong Chang (Institute of Computing Technology, Chinese Academy of Sciences); Yong Li (School of Computer Science and Engi neer, Southeast University); and Yuan Zhang and Mingyuan Gao (ByteDance)\n ---------------------\nMACE-Dance: Motion-Appearance Cascaded Experts for Music-Driven Dance Video Generation\n\nMACE-Dance generates dance videos f rom music using cascaded motion and appearance experts. It first produces expressive 3D dance motion from music, then animates a reference image int o a temporally coherent video. The paper also introduces MA-Data, a large- scale dataset and evaluation protocol for t...\n\n\nKaixing Yang (Renmin U niversity of China); Jiashu Zhu (Alibaba Group, AMAP); Xulong Tang (Malou Tech Inc); Ziqiao Peng (Renmin University of China); Xiangyue Zhang (Wuhan University); Puwei Wang (Renmin University of China); Jiahong Wu and Xian gxiang Chu (Alibaba Group, AMAP); Hongyan Liu (Tsinghua University); and J un He (Renmin University of China)\n---------------------\nComposing Peopl e Together: Iterative Pose-Image Generation for Multi-Person Interaction S cenes\n\nWe present a method for generating realistic multi-person interac tion scenes from text. By modeling human pose and appearance together and composing scenes step by step, our approach produces images that better ca pture complex interactions, with improved accuracy and greater diversity c ompared to ex...\n\n\nWenxuan Peng, Bharath Hariharan, and Hadar Elor (Cor nell University)\n---------------------\nHumanFlow: Controllable Human Ima ge Generation via Flow Matching\n\nHumanFlow is a unified flow-matching fr amework for controllable full-body human image generation. It introduces a Control Encoder, Token-ControlNet, and topology-aware loss for structural consistency, along with the large-scale MiCoGen dataset, achieving improv ed controllability and fidelity over ex...\n\n\nWenzhuo Fan, Hongsheng Zhe ng, and Jianchi Sun (Wuhan University); Fei Fang (Wuhan Textile University ); Hong Ding (Guangxi University of Finance and Economic); and Chunxia Xia o (Wuhan University)\n---------------------\nBoundary-aware Neural Model R eduction for PDEs\n\nNeural eigenanalysis enables reduced-order modeling a cross shape families but is limited to Neumann boundaries. We extend it to Dirichlet, Robin, and mixed conditions by treating boundary configuration s as inputs, forming a unified shape–boundary space for consistent spectra l analysis and simu...\n\n\nLi Liao, Pengfei Shen, and Yifan Peng (The Uni versity of Hong Kong)\n---------------------\nLearning Laplacian Eigenspac e with Mass-Aware Neural Operators on Point Clouds\n\nNEO is a neural fram ework for fast Laplace-Beltrami spectral analysis on 3D point clouds. Rath er than computing eigenvectors with expensive iterative solvers, it predic ts low-frequency eigenspaces directly from geometry, enabling near-linear scaling, robustness to irregular sampling, and accurate ze...\n\n\nZherui Yang (University of Science and Technology of China); Tao Du (Tsinghua Uni versity, Shanghai Qi Zhi Institute); and Ligang Liu (University of Science and Technology of China)\n---------------------\nEfficient Multiscale Lan czos Eigenpair Extraction\n\nWe extend the implicitly restarted Lanczos me thod to a multiscale context using arbitrary multigrids (algebraic, geomet ric) and we demonstrate the gain in performance and robustness on a variet y of application scenarios.\n\n\nTheo Braune (Centre National de la Recher che Scientifique - Laboratoire d'informatique de l'École Polytechnique (LI X), Adobe Research) and Jérémie Dumas and Jean-Marc Thiery (Adobe)\n------ ---------------\nComponent Modes Synthesis Method with Multiple Partitions for Large-scale Eigenvalue Problem\n\nTo solve ultra-large-scale modal an alysis problems, we extend Component Mode Synthesis (CMS) methods to use m ultiple partitions to generate more efficient local eigenbasis functions a nd reduce the computational cost of the low-frequency interface modes. We obtained the lowest 1500 eigenmodes of a mo...\n\n\nChongyao Zhao (State K ey Lab of CAD and CG, Zhejiang University); Junzhou Yin (Independent Resea rcher); and Hujun Bao and Jin Huang (State Key Lab of CAD and CG, Zhejiang University)\n---------------------\nIris3D: 3D Generation via Synchronize d Diffusion Distillation\n\nIris3D is a novel 3D generation system that co mbines feedforward reconstruction models with a new Synchronized Diffusion Distillation (SDD) method to produce detailed geometry and vivid textures . By addressing view inconsistency and over-smoothing issues in traditiona l Score Distillation Sampling (S...\n\n\nYixun Liang, Weiyu Li, and Rui Ch en (The Hong Kong University of Science and Technology); Fei-Peng Tian (Li ght Illusions); Jiarui Liu, Ying-Cong Chen, and Ping Tan (The Hong Kong Un iversity of Science and Technology); Xiao-Xiao Long (The Hong Kong Univers ity of Science and Technology, Nanjing University); and Yixun Liang\n----- ----------------\nCubePart: An Open-Vocabulary Part-Controllable 3D Genera tor\n\nCubePart is an open-vocabulary 3D generative framework offering exp licit semantic part-based control over mesh structure. Guided by a text pr ompt and custom part schema, it synthesizes coherent multi-part meshes, wh ich are ready for direct integration into game engines and animation pipel ines withou...\n\n\nYiheng Zhu, Kangle Deng, Jean-Philippe Fauconnier, Ina ki Navarro Oiza, Daiqing Li, Ava Pun, Yinan Zhang, Peiye Zhuang, and Xiaox ia Sun (Roblox); Maneesh Agrawala (Roblox, Stanford University); and Kiran Bhat and Tinghui Zhou (Roblox)\n---------------------\nPixal3D: Pixel-Ali gned 3D Generation from Images\n\nPixal3D is a pixel-aligned paradigm that solves the fidelity bottleneck in image-to-3D synthesis. By utilizing a b ack-projection conditioning scheme to explicitly lift 2D features into vie w-aligned 3D volumes, it eliminates ambiguity, enabling scalable generativ e reconstruction of highly accurate 3D...\n\n\nDong-Yang Li (Tsinghua Univ ersity); Wang Zhao, Yuxin Chen, and Wenbo Hu (Tencent); Meng-Hao Guo (Tsin ghua University); Fang-Lue Zhang (Victoria University of Wellington); Ying Shan (Tencent); and Shi-Min Hu (Tsinghua University)\n------------------- --\nInvSculpt: Inverse Sculpting Modeling via Controlled 3D Generation and a Vector Displacement Field\n\nWe propose InvSculpt, which decomposes a m esh into a base shape and reusable details represented by a vector displac ement field. It uses text-guided 2D priors and a 3D flow model for mask-fr ee detail removal while preserving identity. The method enables accurate d ecomposition and high-fidelity detai...\n\n\nHengyu Meng and Lanjiong Li ( The Hong Kong University of Science and Technology (Guangzhou), Tencent Li ghtSpeed Studio); Zhijing Shao (The Hong Kong University of Science and Te chnology (Guangzhou)); Yingda Yin, Lingting Zhu, Zeyu Hu, and Xin Wang (Te ncent LightSpeed Studio); Ligang Liu (Laoshan Laboratory); and Zeyu Wang ( The Hong Kong University of Science and Technology (Guangzhou), The Hong K ong University of Science and Technology)\n---------------------\nVecSet-E dit: Unleashing Pre-trained LRM for Mesh Editing from Single Image\n\nWe p ropose VecSet-Edit, a novel pipeline leveraging a pre-trained Large Recons truction Model for high-fidelity 3D mesh editing. Using only 2D image cond itions and strategic token manipulation based on our exploration of VecSet features, it enables precise localized edits while preserving original de ...\n\n\nTeng-Fang Hsiao, Bo-Kai Ruan, Yu-Lun Liu, and Hong-Han Shuai (Nat ional Yang Ming Chiao Tung University)\n---------------------\nProx-E: Fin e-Grained 3D Shape Editing via Primitive-Based Abstractions\n\nProx-E is a training-free framework for fine-grained 3D shape editing that uses a pri mitive-based geometric proxy to explicitly control structural changes. By editing this proxy with a VLM and guiding a 3D generative model accordingl y, it enables localized, identity-preserving edits that outperform ...\n\n \nEtai Sella (Tel Aviv University, Snap); Hao Phung (Cornell Tech); Nitay Amiel (Technion - Israel Institute of Technology); Or Litany (Technion - I srael Institute of Technology, NVIDIA); Or Patashnik (Tel Aviv University, Snap); and Hadar Averbuch-Elor (Cornell Tech)\n---------------------\nSha peUP: Scalable Image-Conditioned 3D Editing\n\nShapeUp introduces a scalab le framework for high-fidelity 3D asset editing. By casting editing as an image-conditioned, latent-to-latent translation, it enables precise local and global transformations while maintaining native 3D consistency. Our ap proach ensures seamless identity preservation witho...\n\n\nInbar Gat (Aig ency.ai, Tel Aviv University); Dana Cohen Bar and Guy Levy (Tel Aviv Unive rsity); Elad Richardson (Runway); and Daniel Cohen-Or (Tel Aviv University )\n---------------------\nInvisible Holographic Window: Full-color 3D Imag e Reconstruction from Transparent Surface-relief Computer-generated Hologr ams\n\nWe demonstrate an invisible holographic window: a transparent surfa ce‑relief computer-generated hologram formed on glass. Using a scattering‑ suppressed encoding scheme and spatial‑frequency band‑division multiplexin g, we achieve high‑transmittance, crosstalk‑free ...\n\n\nRyo Higashida, M asato Miura, Teruyoshi Nobukawa, Yuta Yamaguchi, Ken-ichi Aoshima, and Nob uhiko Funabashi (Japan Broadcasting Corporation (NHK)) and Masahiro Yamagu chi (Institute of Science Tokyo)\n---------------------\nPAColorHolo: A Pe rceptually-Aware Color Management Framework for Holographic Displays\n\nWe present PAColorHolo, a perceptually-aware color management framework that enables accurate color reproduction in holographic displays. Our approach jointly addresses system-level color distortions through color space tran sformation, adaptive illumination control, and a color-restoration neural ne...\n\n\nChun Chen, Minseok Chae, Seung-Woo Nam, Myeong-Ho Choi, Minseon g Kim, Eunbi Lee, Yoonchan Jeong, and Jae-Hyeung Park (Seoul National Univ ersity) and Chun Chen and Myeong-Ho Choi\n---------------------\nSingle-Vi ew Holographic Volumetric 3D Printing with Coupled Differentiable Wave-Opt ical and Photochemical Optimization\n\nWe introduce Single-View Holographi c Volumetric Additive Manufacturing (SHVAM), a layer-free 3D printing tech nique that creates microscale structures in seconds using holographic ligh t patterns. Our coupled optical-chemical optimization framework enables ~1 0 μm lateral features from a mechanica...\n\n\nFelix Wechsler, Riccardo Ri zzo, and Christophe Moser (École Polytechnique Féderale de Lausanne (EPFL) )\n---------------------\nHoloPathTracer: Fast and Accurate Wave Path Trac ing for Holography\n\nTo bring physically accurate 3D cues to holography, HoloPathTracer moves beyond prior computer-generated holography pipelines that separate radiance rendering from wave propagation. By tracing complex wave paths directly through the scene, it enables phase-only holograms wi th natural defocus, reflec...\n\n\nWenbin Zhou and Xiangyu Meng (The Unive rsity of Hong Kong); Jiankai Xing (The University of Hong Kong, Tsinghua U niversity); Xin Liu (The University of Hong Kong); Suyeon Choi (Stanford U niversity, Seoul National University); and Yifan Peng (The University of H ong Kong)\n---------------------\nComplex-Valued Holographic Radiance Fiel ds\n\nComplex-valued holographic radiance fields optimize holographic 3D s cenes without relying on intensity-based intermediaries. Leveraging multi- view images, our method uses complex-valued Gaussian primitives encoding a mplitude and phase aligned with scene geometry. This eliminates expensive single-vie...\n\n\nYicheng Zhan (University College London), Dong-Ha Shin and Seung-Hwan Baek (POSTECH), Kaan Akşit (University College London), and Yicheng Zhan\n---------------------\nMorphSkein: A Shape-Changing Afterim age Display Preserving Pixel Density During Surface-Area Changes Across Tr oposkein-Based Shapes\n\nShape-changing displays typically lose pixel dens ity as surface area expands, limiting usability. We introduce MorphSkein, a shape-changing afterimage display that preserves a minimum density (1.44 px/cm in our prototype) across naturally occurring axisymmetric shapes of varying surface area generat...\n\n\nMaxime Daniel (ESTIA Institute of Te chnology, University of British Columbia) and Shariff AM Faleel and Pouran g Irani (University of British Columbia)\n---------------------\nCoherentR aster: Efficient 3D Gaussian Splatting for Light Field Displays\n\nWe prop ose CoherentRaster, an efficient 3D Gaussian Splatting-based light field r endering framework for glasses-free 3D displays. CoherentRaster introduces Cross-view Coherent Attribute Reuse to eliminate redundant computations a cross neighboring viewpoints and View-coherent Remapping to restore war... \n\n\nGyujin Sim and Seungjoo Shin (Pohang University of Science and Techn ology); Hosung Jeon, Gwangsoon Lee, and Hyon-Gon Choo (Electronics and Tel ecommunications Research Institute (ETRI)); and Sunghyun Cho (Pohang Unive rsity of Science and Technology)\n---------------------\nDualBrep: A Dual- Field Continuous Representation for B-rep Modelling\n\nDualBrep presents a novel continuous representation for CAD models, resolving modeling challe nges caused by heterogeneous data structures. It uses dual scalar fields: a Signed Distance Function for geometry and an Unsigned Distance Field for topology. Compressing these into a single latent space ena...\n\n\nYilin Liu, Pradeep Jayaraman, Chinthala Reddy, Xiang Xu, and Hooman Shayani (Aut odesk Research)\n---------------------\nAutoregressive B-Rep Shape Generat ion with Parametric Surfaces\n\nParaCAD is a point-cloud-conditioned autor egressive framework for B-Rep generation that directly predicts native par ametric surface types and continuous parameters via a surface-centric toke nization, preserving CAD semantics and downstream editability beyond grid/ point-based representations.\n\n\nDafei Qin and Rui Xu (University of Hong Kong); Zeyu Shen (Institute of Software, Chinese Academy of Sciences; Uni versity of Chinese Academy of Sciences); Kaichun Qiao, Hongyang Lin, and Q ixuan Zhang (ShanghaiTech University); Huaijin Pi (University of Hong Kong ); Lan Xu and Jingyi Yu (ShanghaiTech University); Wenping Wang (Texas A&M University); and Taku Komura (University of Hong Kong)\n----------------- ----\nB-repLer: Language-guided Editing of CAD Models\n\nCurrent language- guided CAD editing struggles with dataset shortages and relies on construc tion history. We introduce B-repLer, a novel framework enabling text-based CAD edits in a learned latent space without requiring construction histor y. We also present BrepEDIT-240K, the first large-scale editi...\n\n\nYili n Liu and Niladri Shekhar Dutt (University College London (UCL)); Changjia n Li (University of Edinburgh); and Niloy J. Mitra (University College Lon don (UCL), Adobe Research)\n---------------------\nFlatten the Complex: Jo int B-Rep Generation via Compositional k-Cell Particles\n\nGenerating CAD B-Reps is challenging due to complex topology-geometry entanglement. We so lve this by reformulating B-Reps into compositional k-cell particles. Our flow matching framework unifies topological entities, enabling the joint g eneration, precise 3D reconstruction, and local in-painting of ...\n\n\nJu nran Lu, Yuanqi Li, Hengji Li, Jie Guo, and Yanwen Guo (Nanjing University , State Key Laboratory for Novel Software Technology)\n------------------- --\nBrepForge: Factorized B-rep Synthesis via Wireframe Composition and Bo undary-Conditioned Surface Instantiation\n\nBrepForge is a generative fram ework for CAD B-rep modeling that factorizes synthesis into wireframe cons truction and boundary-conditioned surface generation. This two-stage desig n improves topological validity and geometric precision, enabling the crea tion of complex and realistic CAD models.\n\n\nJing Li, Yihang Fu, and Fal ai Chen (University of Science and Technology of China)\n----------------- ----\nImg2CADSeq: Image-to-CAD Generation via Sequence-Based Diffusion\n\n Img2CADSeq is a multi-stage pipeline converting single-view images into pa rametrically editable 3D CAD models. Using a coarse-to-fine hierarchical c odebook and VQ-Diffusion , it bridges 2D visual features with 3D semantics , generating standard STEP files ready for direct commercial software use .\n\n\nShiyu Tan, Zixuan Zhao, Hao Gao, Zhiheng Chen, and Xiaolong Yin (Sc hool of Software and BNRist, Tsinghua University) and Enya Shen (School of Software and BNRist, Tsinghua University; Haihe Lab of ITAI)\n----------- ----------\nFast and Exact Winding Numbers for Triangle Meshes\n\nGenerali zed winding numbers are a useful measure for inside-outside classification . This paper introduces a novel method for computing winding numbers for t riangle meshes, which leverages a ray-mesh intersection and an elementary sum over boundaries. Our algorithm gives the exact answer and is more ...\ n\n\nPeiyuan Xie, Christian Hafner, and Chris Wojtan (Institute of Science and Technology Austria (ISTA))\n---------------------\nSpatially Accelera ted Winding Numbers for Curved Geometry\n\nExisting winding number methods for curved geometry, such as SVG and CAD datasets, scale linearly with th e number of NURBS primitives. We propose a sublinear approach that leverag es a BVH to agglomerate winding number contributions from distant boundary components while maintaining accurate point co...\n\n\nJacob Spainhour, B rad Whitlock, and Kenneth Weiss (Lawrence Livermore National Laboratory)\n ---------------------\nRobust Containment Queries over Collections of Trim med NURBS Surfaces via Generalized Winding Numbers\n\nWe introduce an accu rate and precise generalized winding number method for trimmed NURBS patch es based on integrating along boundary curves. For points near a patch, we add a correction term determined via line-surface intersections. The meth od enables robust inside-outside containment queries that ...\n\n\nJacob S painhour and Kenneth Weiss (Lawrence Livermore National Laboratory) and Ja cob Spainhour\n---------------------\nThe Antipodal Method: Fast, Accurate , and Robust 3D Generalized Winding Numbers\n\nWe introduce a new formulat ion and algorithm for fast, precise computation of generalized winding num bers applicable to both meshes and parametric surfaces. Our approach expre sses the winding number as the sum of ray-surface intersections and a boun dary integral. It achieves significant speedups ove...\n\n\nCedric Martens (Université de Montréal), Philip Trettner (Shaped Code GmbH), and Mikhail Bessmeltsev (Université de Montréal)\n---------------------\nADS: Random Sampling of Occupancy Functions using Adaptive Delaunay Scaffolding\n\nWe present ADS, a method that delivers both pseudo-random surface samples and isosurface connectivity from occupancy functions, unlike prior approaches that typically provide only one. ADS achieves this with orders of magnitu de fewer function evaluations at comparable accuracy than prior approaches ,...\n\n\nSuzuran Takikawa, Leo Foord-Kelcey, and Oliver Oxford (The Unive rsity of British Columbia); Nicholas Vining (NVIDIA, The University of Bri tish Columbia); and Alla Sheffer (The University of British Columbia)\n--- ------------------\nDifferentiable Voxelization of Surface Representations \n\nWe provide an efficient technique to differentiate voxel values with r espect to the positions of the surface that generated the voxel grid. This allows to solve problems, such as intersection detection, that can be eas ily solved on regular voxel grids for surface representations such as tria ngle mes...\n\n\nTobias Djuren, Ugo Finnendahl, Markus Worchel, Hendrik Me yer, and Marc Alexa (TU Berlin)\n---------------------\nUncertainty-aware geometry processing on Gaussian Process Implicit Surfaces\n\nWe present a geometry processing framework enabling computations directly on probabili stic representations of shapes. In contrast to classical geometry processi ng pipelines our approach considers uncertainty in the input data and acco unt for the distribution of plausible geometries, enabling a prin...\n\n\n Baptiste Genest and David Coeurjolly (CNRS, LIRIS)\n---------------------\ nHybrid Gaussian Wang Tiles for Class-aware Authoring and Rendering\n\nWe present a multi-class Gaussian Splatting Wang Tiles framework for synthesi zing large, detailed scenes from multiple exemplars. Our method enables sm ooth transitions via hybrid tile mixing, supports real-time rendering with hybrid CPU–GPU sorting, and provides an interactive authoring tool f...\n \n\nYunfan Zeng (The Hong Kong University of Science and Technology), Li M a (Eyeline), and Pedro V. Sander (The Hong Kong University of Science and Technology)\n---------------------\nFaster 3D Gaussian Splatting Convergen ce via Structure-Aware Densification\n\nWe present a structure-aware densi fication method for 3D Gaussian Splatting that adapts primitives based on local texture frequency. By combining multi-scale analysis and anisotropic splitting, it preserves fine details while avoiding unnecessary density, enabling faster training and higher-quality ...\n\n\nLinjie Lyu (Max Planc k Institute for Informatics); Ayush Tewari (Cambridge University); and Jia nchun Chen, Thomas Leimkuehler, and Christian Theobalt (Max Planck Institu te for Informatics)\n---------------------\nGaussian Point Splatting\n\nWe propose Gaussian point splatting, a stochastic method for rendering massi ve 3DGS scenes. By sampling pixel-sized opaque points, splatting them atom ically, and applying stochastic transparency, we eliminate the need for so rting. Our approach distributes workload evenly across GPU threads, enabli ng...\n\n\nJoris Rijsdijk, Christoph Peters, Michael Weinmann, and Ricardo Marroquim (Delft University of Technology)\n---------------------\nSHARP- GS: Scalable High-fidelity Accelerated Rendering Pipeline for Ultra-high R esolution 3D Gaussian Splatting\n\nThis work introduces a high-performance 3D Gaussian Splatting framework that unlocks real-time 8K rendering for i mmersive VR. By overcoming architectural bottlenecks, it achieves a massiv e 2.55× speedup and sustains over 250 FPS, delivering high-fidelity, 120 H z stereoscopic experiences without comp...\n\n\nJunRan Ding, WeiHang Liu, YuKe Li, YiZhou Wang, AnTong Li, QiHan Ding, Xin Lou, and JingYi Yu (Shang haiTech University)\n---------------------\nA LoD of Gaussians: Out-of-Cor e Training and Rendering for Seamless Ultra-Large Scene Reconstruction\n\n Gaussian Splatting struggles with large scenes due to memory limits and ch unking artifacts. We introduce "A LoD of Gaussians," a new framework enabl ing ultra-large scene training and rendering on a single consumer GPU. Thr ough out-of-core streaming and Level-of-Detail, it achieves seamless multi -sca...\n\n\nFelix Windisch, Thomas Köhler, Lukas Radl, Mattia D'urso, and Michael Steiner (Graz University of Technology); Dieter Schmalstieg (Univ ersity of Stuttgart, Graz University of Technology); and Markus Steinberge r (Graz University of Technology)\n---------------------\nMobile3DGS³: Acc elerate Mobile 3DGS Rendering via Gradient-Aware Super-Sampling and Frame Interpolation\n\nWe propose Mobile3DGS³, the first mobile 3DGS acceleratio n paradigm solving high-res real-time rendering pain points via gradient-a ssisted post-rendering. It features TSGC, GASS, GAFI for alternating super -sampling/interpolation, delivering high-fidelity efficiency.\n\n\nFan Gao , YIbo Zhao, Changhao Song, Jiarui Wen, Yuxuan Song, Youcheng Cai, and Lig ang Liu (University of Science and Technology of China)\n----------------- ----\nTopologically Consistent Multi-view 3D Head Reconstruction via Coars e-Guided Layered Surface Sampling\n\nFrom calibrated multi-view images, SH ELLS reconstructs 18k-vertex 3D heads in 0.08 seconds. It aggregates DinoV 2 features via projective surface-aware feature sampling, allowing a trans former to predict dense semantic meshes 3.5x faster with 88% less GPU memo ry than state-of-the-art methods. The res...\n\n\nTimo Bolkart, Daoye Wang , and Prashanth Chandran (Google)\n---------------------\nEmotion Manipula tion for Talking-Head Videos via Facial Landmarks\n\nPreserving lip synchr onization during emotion manipulation in video is a major challenge. This paper proposes a method using a pretrained StyleGAN and a latent-based lan dmark editing module. It modifies the latent edit direction with facial la ndmarks, enabling fast, high-quality emotion changes whil...\n\n\nKwanggyo on Seo (Flawless AI); Rene Jotham Culaway (Burt Intelligence); ByeongUk Le e (Amazon Robotics); Junyong Noh (KAIST, Visual Media Lab); and Kwanggyoon Seo\n---------------------\nLearning a Delighting Prior for Facial Appear ance Capture in the Wild\n\nA high-quality facial appearance capture metho d from casual smartphone videos with a powerful delighting prior.\n\n\nYux uan Han, Xin Ming, Tianxiao Li, and Zhuofan Shen (Tsinghua University); Qi xuan Zhang and Lan Xu (ShanghaiTech University); and Feng Xu (Tsinghua Uni versity)\n---------------------\nBringing Diversity from Diffusion Models to Semantic-Guided Face Asset Generation\n\nThis paper introduces a framew ork for generating high-quality, semantically controllable facial assets. By leveraging a diffusion model's diversity, we develop a face asset synth esis pipeline to produce paired albedo and geometry data with semantic lab els. These pairs are then used to train a disent...\n\n\nYunxuan Cai, Sita o Xiang, Zongjian Li, Haiwei Chen, and Yajie Zhao (University of Southern California Institute for Creative Technologies) and Yunxuan Cai\n-------- -------------\nEchoAvatar: Real-time Generative Avatar Animation from Audi o Streams\n\nWe introduce a novel framework designed to generate continuou s, coherent full-body motion from arbitrary audio streams with low latency . Our approach is a unified streaming architecture capable of synthesizing continuous motion from incremental audio inputs. And it serves as a plug- and-play solution...\n\n\nBohong Chen, Yumeng Li, Yinglin Xu, Youyi Zheng, Yanlin Weng, and Kun Zhou (State Key Laboratory of CAD & CG, Zhejiang Uni versity)\n---------------------\nSee-through: Single-image Layer Decomposi tion for Anime Characters\n\nWe introduce a framework that automates the t ransformation of static anime illustrations into manipulatable 2.5D models . Our approach decomposes a single image into fully inpainted, semanticall y distinct layers with inferred drawing orders — up to 19 layers including hair, face, eyes, clothing...\n\n\nJian Lin and Chengze Li (Saint Francis University); Haoyun Qin (University of Pennsylvania, Spellbrush / Shitaga ki Lab); Kwun Wang Chan (Saint Francis University); Yanghua Jin (Spellbrus h); and Hanyuan Liu; Chun Wang, Stephen Choy; and Xueting Liu (Saint Franc is University)\n---------------------\ngCDT: A Highly Parallel GPU Algori thm for Large-Scale Constrained Delaunay Triangulation\n\nWe present a hig hly parallel GPU algorithm for 2D constrained Delaunay triangulation that robustly handles arbitrary valid constraints. It computes triangulations o f up to 10 million vertices in 0.1 seconds on an RTX 4090, substantially o utperforming prior GPU methods and widely used CPU implementat...\n\n\nPen g Fan (Zhejiang University); Min Tang (Zhejiang University, Zhejiang Sci-T ech University); Ruofeng Tong (Zhejiang University); Lili He (Zhejiang Sci -Tech University); Peng Du (Zhejiang University); and Hailong Li (Shenzhen Poisson Software Co., Ltd.)\n---------------------\nScalable GPU Construc tion of 3D Voronoi and Power Diagrams\n\nWe present a GPU algorithm for co nstructing large-scale 3D Voronoi and power diagrams. By combining directi onal geometric culling with hierarchical best-first traversal, our method efficiently handles tens of millions of points across diverse spatial dist ributions. Applied to mesh-based neural rende...\n\n\nBernardo Taveira and Carl Lindström (Chalmers University of Technology, Zenseact); Maryam Fate mi (Zenseact); and Lars Hammarstrand and Fredrik Kahl (Chalmers University of Technology)\n---------------------\nGPU-accelerated Certified Hausdorf f Distance Between Triangle Meshes\n\nWe present a GPU-accelerated certifi ed algorithm for the directed Hausdorff distance between triangle meshes. By replacing recursive branch-and-bound with a parallel wavefront pipeline , it delivers millisecond-scale, tolerance-controlled results on models wi th millions of triangles, enabling interac...\n\n\nHaopeng Fan (Zhejiang U niversity); Min Tang (Zhejiang University, Zhejiang Sci-Tech University); Leonardo Sacht (Universidade Federal de Santa Catarina); Qiang Zou (State Key Lab of CAD and CG, Zhejiang University); and Ruofeng Tong and Peng Du (Zhejiang University)\n---------------------\nPQ-Free HD: Priority-Queue-F ree Hausdorff Distance for Triangle Meshes on GPU\n\nWe present a method f or computing error-bounded Hausdorff distance between triangle meshes on t he GPU. By eliminating the serial bottleneck in traditional branch-and-bou nd algorithms, our approach achieves high-throughput performance and enabl es geometry processing applications that require guarante...\n\n\nZhihao H u and Renjie Chen (University of Science and Technology of China)\n------- --------------\nManifold k-NN: Accelerated k-NN Queries for Manifold Point Clouds\n\nWe present Manifold k-NN, a dynamic programming-based framework for accelerated k-nearest neighbor search on manifold point clouds. By re cursively partitioning insertion history via Voronoi successor lists, our method achieves 1–10× speedup over kd-trees in volume-to-surface queries, while sup...\n\n\nPengfei Wang, Qinghao Guo, Haisen Zhao, and Shiqing Xin (Shandong University); Shuangmin Chen (Qingdao University of Science and T echnology); Changhe Tu (Shandong University); and Wenping Wang (Texas A&M University)\n---------------------\nLocality-Aware Automatic Differentiati on on the GPU for Mesh-Based Computations\n\nWe present a GPU system for a utomatic differentiation for functions defined on triangle meshes that exp loits locality and sparsity to deliver efficient derivative computation. F rom simple user-defined energy terms, the system produces gradients, Jacob ians, and sparse Hessians, accelerating simulatio...\n\n\nAhmed Mahmoud, R ahul Goel, Jonathan Ragan-Kelley, and Justin Solomon (Massachusetts Instit ute of Technology (MIT))\n---------------------\nMV-S2V: Multi-View Subjec t-Consistent Video Generation\n\nMV-S2V is the first work to explore video generation with multi-view subject consistency. It presents a simple yet effective framework, including a tailored data curation pipeline and a Tem porally-Shifted RoPE (TS-RoPE) specially designed for multi-view reference conditioning, achieving superior 3D ...\n\n\nZiyang Song (The Hong Kong P olytechnic University), Xinyu Gong (The University of Texas at Austin), Ba ngya Liu (University of Wisconsin-Madison), and Zelin Zhao (Georgia Instit ute of Technology)\n---------------------\nOmniRoam: World Wandering via L ong-Horizon Panoramic Video Generation\n\nThis work introduces a controlla ble panoramic video generation framework designed for long-horizon scene w andering. It utilizes a global-to-local preview-and-refine pipeline to syn thesize high-fidelity, spatially and temporally coherent panoramic videos from a single input panoramic image or video.\n\n\nYuheng Liu (University of California Irvine); Xin Lin (University of California San Diego); Xinke Li (City University of Hong Kong); Baihan Yang (University of California San Diego); Chen Wang (University of Pennsylvania); Kalyan Sunkavalli, Yan nick Hold-Geoffroy, Hao Tan, and Kai Zhang (Adobe Research); Xiaohui Xie ( University of California Irvine); and Zifan Shi and Yiwei Hu (Adobe Resear ch)\n---------------------\nActCam: Zero-Shot Joint Camera and 3D Motion C ontrol for Video Generation\n\nActCam is a zero-shot method providing join t control over camera trajectories and character acting. It integrates int o preexisting artistic workflows by leveraging traditional cinematography skills for fine-grained control. Requiring no additional training, the fra mework avoids costly finetuning and...\n\n\nOmar El Khalifi, Thomas Rossi, Oscar Fossey, and Thibault Fouque (Kinetix); Ulysse Mizrahi (Kinetix, Tel Aviv University); Philip Torr (University of Oxford); Ivan Laptev and Fab io Pizzati (MBZUAI); and Baptiste Bellot-Gurlet (Kinetix)\n--------------- ------\nHL-OutPaint: Coarse-to-Fine Video Outpainting for High-Resolution Long-Range Videos\n\nWe introduce HL-OutPaint, a coarse-to-fine framework for high-resolution long-range video outpainting. HL-OutPaint incorporates a novel global-local frame swapping mechanism that enables both long-term and short-term temporal consistency. Experimental results show that our a pproach produces spatiall...\n\n\nJeongeun Park, Janghyeok Han, and Geonun g Kim (POSTECH); Hyun-Seung Lee, Kyuha Choi, and Youngseok Han (Visual Dis play Business, Samsung Electronics); and Sunghyun Cho (POSTECH)\n--------- ------------\nUniVidX: A Unified Multimodal Framework for Versatile Video Generation via Diffusion Priors\n\nUniVidX is a unified multimodal video g eneration framework that repurposes video diffusion priors to flexibly gen erate and translate across visual modalities. Instantiated as UniVid-Intri nsic and UniVid-Alpha, it supports tasks such as inverse rendering, religh ting, matting, and inpainting, while ac...\n\n\nHouyuan Chen (HKUST), Hong Li (Beihang University), Xianghao Kong (HKUST), Tianrui Zhu (Nanjing Univ ersity), Shaocong Xu (Beijing Academy of Artificial Intelligence), Weiqing Xiao (Nanjing University), Yuwei Guo and Chongjie Ye (The Chinese Univers ity of Hong Kong), Lvmin Zhang (Stanford University), Hao Zhao (Tsinghua U niversity), and Anyi Rao (HKUST)\n---------------------\nPoints as Tori: F ast Pointwise Signed Distance for Point Clouds\n\nSigned distance computat ion has been either fast or robust, but not both. We explain why, using a theory that unifies signed distance with reconstruction. Our method direct ly estimates signed distance from point clouds, providing fast pointwise e valuation at arbitrary spatial resolution, without req...\n\n\nNicole Feng and Ioannis Gkioulekas (Carnegie Mellon University) and Keenan Crane (Car negie Mellon University, Roblox Research)\n---------------------\nSuperSDF :Sparse SDF Super-Resolution for Surface Extraction\n\nSuperSDF is a learn ing-based method for signed distance field super-resolution that reconstru cts high-fidelity meshes from coarse inputs, without mesh supervision or a uxiliary surface representations. Using a sparse voxel network near the su rface, our approach learns how to directly refine the input...\n\n\nSagar Panwar and Nissim Maruani (INRIA), Céline Loscos (L Research), and Mathieu Desbrun and Pierre Alliez (INRIA)\n---------------------\nSAND: Spatially Adaptive Network Depth for Fast Sampling of Neural Implicit Surfaces\n\nS AND is an efficient neural implicit surface representation framework that accelerates sampling by adapting network depth to spatial importance and local geometric complexity. By combining a volumetric depth map with a tai led MLP for adaptive termination, it significantly improves inference-time qu...\n\n\nChuanxiang Yang (Shandong University), Junhui Hou (City Univer sity of Hong Kong), Yuan Liu (Hong Kong University of Science and Technolo gy), Siyu Ren (City University of Hong Kong), Guangshun Wei (Shandong Univ ersity), Taku Komura (University of Hong Kong), Yuanfeng Zhou (Shandong Un iversity), and Wenping Wang (Texas A&M University)\n---------------------\ nThe PhaseTree: Multiphase Signed Distance Fields\n\nWe introduce the Phas eTree, a novel hierarchical construction tree representation for compactly modeling objects composed of multiple phases or materials. An object is d efined as a single tree that combines phase-aware primitives and operators , yielding a unified multiphase signed distance represent...\n\n\nEric Gal in and Pierre Hubert-briere (Université Claude Bernard Lyon 1, LIRIS); Mar ie-Paule Cani (Centre National de la Recherche Scientifique - Laboratoire d'informatique de l'École Polytechnique (LIX)); Adrien Peytavie (Universit é Claude Bernard Lyon 1, LIRIS); and Eric Guérin and Hugo Schott (INSA, Ly on; LIRIS)\n---------------------\nDual Contouring of Signed Distance Data \n\nWe propose an algorithm to reconstruct explicit polygonal meshes from discretely sampled Signed Distance Function (SDF) data, which is especiall y effective at recovering sharp features. Building on the Dual Contouring of Hermite Data method, we solve a quadratic optimization problem to place the mes...\n\n\nXiana Carrera and Ningna Wang (Columbia University); Chri stopher Batty (University of Waterloo); Oded Stein (Technion, University o f Southern California); and Silvia Sellán (Columbia University)\n--------- ------------\nDual Contouring over Expanded Cubes (DCx) for Zero-Level Set Extraction from Neural Unsigned Distance Functions\n\nDCx extends the Dua l Contouring method which only extracts manifold surfaces from signed dist ance fields, to enable the surface extraction from unsigned distance field s. This is achieved by looking up a carefully designed table over expanded 2x2x2 cubes, allowing complex topologies to be extracted, ...\n\n\nQingch ao Bao and Xuhui Chen (Institute of Software, Chinese Academy of Sciences; University of Chinese Academy of Sciences); Jingpeng Yin (Dalian Universi ty of Technology); Fei Hou and Wencheng Wang (Institute of Software, Chine se Academy of Sciences; University of Chinese Academy of Sciences); Hong Q in (Stony Brook University); and Ying He (Nanyang Technological University (NTU))\n---------------------\nSubgrid Marching Tetrahedra\n\nSubgrid Mar ching Tetrahedra is an iso-surfacing method that reconstructs high quality surfaces while capturing fine topological features below grid resolution. Using grid edge intersections locations with the implicit function, it pr oduces manifold meshes with the usual benefits from classical march...\n\n \nHossein Baktash (Carnegie Mellon University); Mark Gillespie (INRIA, Sac lay; University of Utah); and Keenan Crane (Carnegie Mellon University, Ro blox)\n---------------------\nAtomSlicer: Constant-Thickness Field-Aligned Non-Planar Slicing and Continuous Toolpaths for FFF\n\nAtomSlicer is a 3D printing method for fused filament fabrication that generates non-planar layers and continuous toolpaths aligned with user-defined fields. It enabl es better control of layer orientation, constant thickness, and near-conti nuous deposition, helping improve print quality, reduce inte...\n\n\nGiova nni Cocco, Vincent Belle, Eric Garner, Sylvain Lefebvre, and Xavier Cherma in (Université de Lorraine, CNRS, Inria, LORIA)\n---------------------\nFr eeShell: A Context-Free 4D Printing Technique for Fabricating Complex 3D T riangle Mesh Shells\n\nFreeShell introduces a robust thermal-shrinkage-act uated 4D printing technique for fabricating freeform thin-shell surfaces. By printing triangular tiles connected by shrinkable connectors using a si ngle material, heating triggers the transformation from flat structures in to 3D shells. An optimized ...\n\n\nChao Yuan, Shengqi Dang, Xuejiao Ma, a nd Nan Cao (Tongji University) and Chao Yuan\n---------------------\nMecha nical Cloaking of Halftoned Imagery\n\nWe explore a new direction in mecha nical cloaking: halftoning an image using a porous structure that behaves like a linear, isotropic material and visually matches an image. For an ex ternal observer, this creates the surprising effect where the object appea rs mechanically homogeneous while its porou...\n\n\nJonàs Martínez (Inria) ; Brisard Sébastien (Aix Marseille Université); Kostas Danas (LMS, CNRS, E cole Polytechnique); Eric Garner (Inria); Sid Kumar (TU Delft); and Sylvai n Lefebvre (Inria)\n---------------------\nShellular Metamaterial Design v ia Compact Electric Potential Parametrization\n\nThis work introduces a co mpact, expressive design space for shellular metamaterials, along with a f ast GPU-based homogenization pipeline that evaluates elastic properties in near real time. This enables interactive exploration and inverse design, achieving diverse geometries and a broad range of mec...\n\n\nTianyi Huang , Chang Liu, and Bohan Wang (National University of Singapore)\n---------- -----------\nDeepMill++: Neural Guidance Meets Rasterization for Efficient Accessibility Analysis\n\nWe introduce DeepMill++, a conservative and hig hly efficient framework for cutter accessibility analysis on arbitrary tri angular meshes. DeepMill++ reformulates accessibility and occlusion detect ion as a rasterization-based visibility and depth pooling problem. DeepMil l++ achieves up to 9.5× speedup...\n\n\nFanchao Zhong, Yao Zhang, and Guan ze Xin (Shandong University); Peng-Shuai Wang (Peking University); and Lin Lu, Changhe Tu, and Haisen Zhao (Shandong University)\n------------------ ---\nUniformly Deployable Kirigami on Arbitrary Planar Graphs\n\nWe prese nt a method for exploring the space of uniformly deployable hinged kirigam i structures, as a constrained embedding space of an arbitrary planar grap h. The design space can be used for desired deployment properties, such as conformal behavior and fully-closed deployed configuration, and for i...\ n\n\nAviv Segall, Jing Ren, and Olga Sorkine-Hornung (ETH Zurich)\n------- --------------\nIsoGami: Rigid-Deployable Kirigami Materials From Isohedra l Tilings\n\nWe present a computational method to design rigid-deployable Kirigami sheets based on isohedral tilings. By exploring combined topology and geometry spaces, our approach identifies self-deployable structures w ith controlled expansion. Our approach generates diverse designs, demonstr ated through an i...\n\n\nGuo Han, Juan Montes Maestre, Numerow Logan, Ron an Hinchet, Stelian Coros, and Bernhard Thomaszewski (ETH Zürich)\n------- --------------\nTwinPose: Person-Specific Subspaces for Multi-View 3D Pose Estimation\n\nTwinPose is an observation-driven framework for real-time m ulti-person 3D motion capture from sparse multi-view inputs in complex sce nes. By constructing instance-aware "twin poses" that unify 2D pose semant ics and multi-view geometric consistency, it enables robust, efficient, an d detector-agnostic...\n\n\nWenwu Yang, Tianyi He, Jiwei Ding, Xun Wang, a nd Rong Zhang (Zhejiang Gongshang University) and Kun Zhou (Zhejiang Unive rsity)\n---------------------\nAMOR: Airborne Motion Reconstruction via Ho motopy-Aware Trajectory Optimization\n\nExisting Human-Mesh-Recovery metho ds struggle with dynamic airborne motions, often producing unstable and un realistic results. We present a refinement approach that enforces physical consistency by selecting reliable motion segments, applying homotopy-awar e trajectory optimization. Experiments on in...\n\n\nChanha Kim and Jungda m Won (Seoul National University)\n---------------------\nFLASHand: Feed-f orward reLightable and Animatable Single-view Hand Reconstruction\n\nWe pr opose FLASHand, the first feed-forward model to reconstruct high-fidelity, relightable, and animatable 3D hand avatars from a single RGB image. By l everaging NIMBLE priors and mesh-based 2D Gaussian Splatting, FLASHand ach ieves instant, personalized reconstruction with disentangled appearance, . ..\n\n\nLing-Xiao Zhang and Lin Gao (Institute of Computing Technology, Ch inese Academy of Sciences; University of Chinese Academy of Sciences); Wei -Hong He (South China University of Technology); Yu-Xuan Yang (East China Normal University); Yunbing Xing (Institute of Computing Technology, Chine se Academy of Sciences); Yu-Kun Lai (Cardiff University); and Yiqiang Chen (Institute of Computing Technology, Chinese Academy of Sciences)\n------- --------------\nAGILE: Hand-object Interaction Reconstruction from Video v ia Agentic Generation\n\nAGILE reconstructs 3D hand-object interactions fr om monocular videos via agentic generation. It produces complete, simulati on-ready object meshes and robustly estimates object pose without SfM, ena bling accurate and stable results on challenging real-world data.\n\n\nJin -Chuan Shi, Binhong Ye, Tao Liu, Xiaoyang Liu, Yangjinhui Xu, Junzhe He, Z eju Li, and Hao Chen (Zhejiang University) and Chunhua Shen (Zhejiang Univ ersity, Zhejiang University of Technology)\n---------------------\nEgoForc e: Forearm-Guided Camera-Space 3D Hand Pose from a Monocular Egocentric Ca mera\n\nEgoForce enables real-time 3D hand pose, shape, and absolute posit ion recovery from a single head-mounted RGB camera. Designed for lightweig ht smart glasses, it uses forearm cues and camera-aware ray-space lifting to provide precise tracking across diverse camera models and device setups .\n\n\nChristen Millerdurai (DFKI, Max Planck Institute for Informatics); Shaoxiang Wang and Yaxu Xie (DFKI); Vladislav Golyanik (Max Planck Institu te for Informatics); and Didier Stricker and Alain Pagani (DFKI)\n-------- -------------\nOmniHands: Robust Motion Capture of Interactive Hands via A Versatile Transformer\n\nOmniHands is a universal architecture for hand r econstruction across diverse scenarios, supporting single-hand reconstruct ion, two-hand interaction reconstruction, temporal inputs, and multi-view inputs. It achieves this within a unified framework by first tokenizing th e features of each hand and th...\n\n\nDixuan Lin (Beijing Normal Universi ty, Stanford University); Yuxiang Zhang and Mengcheng Li (Tsinghua Univers ity); Wei Jing, Qi Yan, and Qianying Wang (Lenovo Group Ltd); Yebin Liu (T singhua University); Hongwen Zhang (Beijing Normal University); and Dixuan Lin\n---------------------\nStreaming of rendered content with adaptive f rame rate and resolution\n\nTo improve the perceived quality of streamed c ontent while reducing rendering costs, we exploit the spatiotemporal limit s of the human visual system and adaptively adjust both frame rate and res olution based on scene content and motion. Those are controlled with a neu ral network, trained on a datase...\n\n\nYaru Liu and Joseph March (Univer sity of Cambridge) and Rafał K. Mantiuk (University of Cambridge, Meta)\n- --------------------\nGabor Fields: Orientation-Selective Level-of-Detail for Volume Rendering\n\nWe propose using Gabor kernels as volumetric densi ty primitives to enable compact, cheap and continuous level-of-detail to p rimitive-based physically-based volume rendering with single and multiple scattering. We show state-of-the-art results in regression quality, runtim e performance and introduce ...\n\n\nJorge Condor, Nicolai Hermann, Mehmet Ata Yurtsever, and Piotr Didyk (Università della Svizzera Italiana)\n---- -----------------\nLightOpt: Lights Optimization for Real-time Rendering\n \nLightOpt is a differentiable optimization framework that automatically r educes the number of lights in real-time game scenes while preserving visu al appearance. By optimizing and restructuring real-time light sources, it lowers lighting cost, reduces overlap, and improves cross-platform perfor mance ...\n\n\nTuo Chen (Tsinghua University), Luyan Cao and Kui Wu (LIGHT SPEED), and Shimin Hu (Tsinghua University)\n---------------------\nDPF: D ifferentiable Polyphase Filtering for Large-Kernel Approximation\n\nWe int roduce a novel Polyphase Filtering framework offering the fastest approxim ation for real-time, large-kernel image and video processing. Overcoming f ixed-resolution processing constraints, our method efficiently handles com plex, spatially-variant kernels. It significantly outperforms existing t.. .\n\n\nZhe Cao, Zhizhen Wu, and Yuchi Huo (State Key Lab of CAD and CG, Zh ejiang University); Zhonggui Chen (School of Informatics, Xiamen Universit y); and Rui Wang (State Key Lab of CAD and CG, Zhejiang University)\n----- ----------------\nGeneralized Spherical Harmonics Products using Spherical Grids\n\nWe introduce the Generalized Spherical Harmonics Product, a nove l method supporting inputs with different SH orders and flexible output tr uncation. Using Spherical Grids as an intermediate representation, we redu ce SH products to simple point-wise multiplications, achieving 3.5–6.5× sp eedup ov...\n\n\nDi An, Jiaqi Wu, and Bowen Xu (Tsinghua University); Ling qi Yan (Mohamed bin Zayed University of Artificial Intelligence); and Kun Xu (Tsinghua University)\n---------------------\nGauSmoke: Hybrid Physics- Optical Gaussian Splatting for Sparse Smoke Reconstruction\n\nWe present a physics-aware method for reconstructing dynamic fluids from sparse-view v ideos. By integrating volumetric rendering with physically guided Gaussian optimization, it enforces consistency in density and motion, reducing art ifacts and improving realism. The approach achieves accurate, stab...\n\n\ nWenran Zhang and Yuxiang Cai (Nankai University); Letian Huang (State Key Lab for Novel Software Technology, Nanjing University); Dongwei Ye and Ji e Guo (Nanjing University); and Ren Bo (Nankai University)\n-------------- -------\nLagrangianSplats: Divergence-Free Transport of Gaussian Primitive s for Fluid Reconstruction\n\nLagrangianSplats reconstructs physically pla usible 3D fluid velocity fields from sparse video by combining divergence- free kernels with Lagrangian Gaussian transport. A sliding-window optimiza tion enables efficient long-range supervision, producing state-of-the-art transport consistency and physical...\n\n\nNingxiao Tao (School of Intelli gence Science and Technology, Peking University); Mengyu Chu (Peking Unive rsity, State Key Laboratory of General Artificial Intelligence); and Baoqu an Chen (Peking University)\n---------------------\nFast VEM Fluid Simulat ion\n\nFastVEM is an efficient boundary-conforming fluid simulation framew ork. It combines Virtual Element discretization, a simulation-friendly bod y-fitted grid construction strategy, and a tailored geometric multigrid me thod to achieve robust, high-fidelity, and efficient fluid–boundary intera ction...\n\n\nRunze Zhang and Bo Ren (Nankai University)\n---------------- -----\nDiffSurFlow: Efficient and Robust Differentiable Fluid Optimization via Surrogate Strategy on Flow Map\n\nWe present DiffSurFlow, an efficien t and robust framework for long-horizon and vortex-rich fluid optimization . Powered by a physics-informed surrogate gradient strategy that exploits flow map "gradient highways" while pruning computationally intensive high- order gradient passes, it achieves signific...\n\n\nYuhao Quan, Hui Wang, Weile Lian, Zhi Wang, and Xubo Yang (Shanghai Jiao Tong University)\n----- ----------------\nGeneric Variational Spacetime Optimization of Vortex Cor e Manifolds\n\nWe introduce a highly efficient variational framework for c omputing optimal vortex cores in 3D unsteady flows. By unifying diverse de tection criteria through time-preintegrated Lagrangians, our method accura tely reconstructs spacetime vortex manifolds by solving Euler-Lagrange equ ations in just one ...\n\n\nXingdi Zhang, Peter Rautek, and Markus Hadwige r (King Abdullah University of Science and Technology (KAUST))\n---------- -----------\nSegviGen: Repurposing 3D Generative Model for Part Segmentati on\n\nSegviGen is a framework for 3D part segmentation which unifies three settings in one architecture: interactive part segmentation, full segment ation, and 2D segmentation map–guided full segmentation. It improves over sota by 40% on interactive segmentation and by 15% on full segmentation, w hile...\n\n\nLin Li (Renmin University of China); Haoran Feng (Tsinghua Un iversity); Zehuan Huang and Haohua Chen (Beihang University); Wenbo Nie (B eijing Jiaotong University); Shaohua Hou, Keqing Fan, and Pan Hu (Beihang University); Sheng Wang and Buyu Li (Bambu Lab); and Lu Sheng (Beihang Uni versity)\n---------------------\nSimArt: Decomposing Monolithic Meshes int o Sim-ready Articulated Assets via MLLM\n\nSIMART is a unified multimodal large language model framework that converts static 3D meshes into simulat ion-ready articulated assets. By jointly modeling part decomposition and k inematics with a sparse 3D VQ-VAE, it reduces token complexity and enables scalable, high-fidelity object generation for ...\n\n\nChuanrui Zhang (Na nyang Technological University, Singapore); Minghan Qin, Yuang Wang, Baife ng Xie, and Hang Li (ByteDance Seed); and Ziwei Wang (Nanyang Technologica l University, Singapore)\n---------------------\nNeural Cellular Automata: From Cells to Pixels\n\nWe pair Neural Cellular Automata with a lightweig ht Neural Field that decodes coarse cell states and local coordinates into high-resolution appearance, decoupling lattice size from output resolutio n. Trained end-to-end with task-specific losses, our model synthesizes tex tures and grows morphologies ...\n\n\nEhsan Pajouheshgar, Yitao Xu, and Al i Abbasi (EPFL); Alexander Mordvintsev (Google Research); and Wenzel Jakob and Sabine Süsstrunk (EPFL)\n---------------------\nNeural Particle Autom ata: Learning Self-Organizing Particle Dynamics\n\nNeural Particle Automat a extend Neural Cellular Automata from fixed grids to dynamic particles. E ach particle carries a continuous position and internal state, updated by a shared neural rule with SPH-based local perception. Backed by optimized CUDA kernels, NPA learn scalable self-organizing dynami...\n\n\nEhsan Pajo uheshgar (EPFL), Hyunsoo Kim (Korea Advanced Institute of Science and Tech nology (KAIST)), Sabine Süsstrunk and Wenzel Jakob (EPFL), and Jinah Park (Korea Advanced Institute of Science and Technology (KAIST))\n------------ ---------\nPixTex: Consistent 3D Texturing via Pixel-Space Multi-View Diff usion\n\nWe propose PixTex, the first pixel-space multi-view diffusion fra mework for texture generation. By bypassing latent compression and utilizi ng a coarse-to-fine strategy, it ensures lossless geometric guidance and s uperior pixel-level multi-view consistency, achieving state-of-the-art, hi gh-fidelity, ...\n\n\nYuqing Zhang (State Key Lab of CAD&CG, Zhejiang); Ya n-Pei Cao (VAST); Hao Xu, Yiqian Wu, Sirui Lin, and Yuqing Wang (State Key Lab of CAD&CG, Zhejiang University); Ding Liang and Yuan-Chen Guo (VAST); and Xiaogang Jin (Zhejiang University; State Key Laboratory of CAD&CG, Zh ejiang University)\n---------------------\nGenerative 3D Gaussians with Le arned Density Control\n\nDeG is an image-to-3D generative framework for Ga ussian splatting that learns to place detail adaptively. It evaluates each Gaussian’s effect on rendering loss, then reinforces density in important regions and suppresses less useful ones, enabling variable-resolution, hi gh-quality 3D asset gen...\n\n\nRunjie Yan (Institute for Interdisciplinar y Information Sciences, Tsinghua University; VAST) and Yan-Pei Cao, Peng W ang, Ding Liang, and Yuan-Chen Guo (VAST)\n---------------------\nLLM-enha nced Scene Graph Learning for Household Rearrangement\n\nWe present an LLM -enhanced scene graph framework for autonomous household rearrangement. By transforming scene graphs into affordance-enhanced graphs that encode sce ne-grounded object functionality and user preferences, our method detects misplaced objects and plans correct placements. We demonstrat...\n\n\nWenh ao Li, Shilong Zou, Zhinan Yu, Zheng Zhou, Wenxuan Li, and Chenyang Zhu (N ational University of Defense Technology); Ruizhen Hu (Shenzhen University ); Kai Xu (Institute of AI for Industries Chinese Academy of Science); and Wenhao Li\n---------------------\nCasLayout: Cascaded 3D Layout Diffusion for Indoor Scene Synthesis with Implicit Relation Modeling\n\nCasLayout i s a cascaded diffusion framework that decomposes 3D indoor scene generatio n into four interpretable stages. This design reduces data requirements wh ile enabling flexible, controllable synthesis and seamless integration wit h LLMs and VLMs for zero-shot tasks.\n\n\nYingrui Wu (Institute of Automat ion, Chinese Academy of Sciences; School of Artificial Intelligence, Unive rsity of Chinese Academy of Sciences); You-Kang Kong (Tsinghua University, Microsoft Research Asia); Mingyang Zhao (State Key Laboratory of Mathemat ical Sciences, Academy of Mathematics and Systems Science, Chinese Academy of Sciences); Weize Quan and Dongmin Yan (Institute of Automation, Chines e Academy of Sciences); and Yang Liu (Microsoft Research Asia)\n---------- -----------\nRaster2Seq: Polygon Sequence Generation for Floorplan Reconst ruction\n\nRaster2Seq is an anchor-based autoregressive formulation that f rames floorplan reconstruction as scalable, sequential polygon prediction, transforming rasterized floorplan images into vectorized format. Through using our proposed labeled polygon sequence representation, the method joi ntly encodes sp...\n\n\nHao Phung and Hadar Averbuch-Elor (Cornell Univers ity)\n---------------------\nToward Richer Material Generation via Procedu ral Data Enhancement\n\nGenerative material models are limited by simple P BR data. We augment single-lobe GGX materials into layered, multi-lobe mod els capturing richer effects (e.g., dust, clearcoat). These are encoded as neural materials in a shared 6D latent space. The resulting dataset enabl es generative models to prod...\n\n\nYunchen Yu (Cornell University, NVIDI A); Jacob Munkberg, Jon Hasselgren, and Chris Cummings (NVIDIA); Steve Mar schner (Cornell University, NVIDIA); and Andrea Weidlich (NVIDIA)\n------- --------------\nInfiniteDiffusion: Bridging Learned Fidelity and Procedura l Utility for Open-World Terrain Generation\n\nWe introduce InfiniteDiffus ion, a training-free algorithm enabling diffusion models to generate spati ally infinite outputs with the core properties of procedural noise: seed-c onsistency (determinism) and constant-time random access. We incorporate a novel architecture for terrain generation to produ...\n\n\nAlexander Gosl in (Independent)\n---------------------\nRCGP: Resource Contracts for Grap hics Programming\n\nRCGP introduces a type system for graphics programming that statically enforces resource contracts, the agreements on type, layo ut, binding, and synchronization between shaders, pipelines, and host code . Through contracts, modules, combinators, and witnesses, RCGP catches des criptor mismatches, layo...\n\n\nVenkataram Sivaram (Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Tec hnology (MIT)); Sai Praveen Bangaru (NVIDIA); Ravi Ramamoorthi and Tzu-Mao Li (University of California San Diego); and Jonathan Ragan-Kelley and Fr edo Durand (Computer Science and Artificial Intelligence Laboratory (CSAIL ), Massachusetts Institute of Technology (MIT))\n---------------------\nVi deo Analysis and Generation via a Semantic Progress Function\n\nThis paper introduces a Semantic Progress Function - a one-dimensional representatio n capturing the semantic pace of a given video. Building on this, a semant ic linearization procedure is introduced to retime sequences and control t he rate at which semantic change unfolds, yielding smoother, more co...\n\ n\nGal Metzer and Sagi Polaczek (Tel Aviv University), Arash Mahdavi-Amiri (Simon Fraser University), and Raja Giryes and Daniel Cohen-Or (Tel Aviv University)\n---------------------\nGo-with-the-Track: Video Compositing a nd Motion Control with Point Tracking\n\nWe present Go-with-the-Track, a v ideo generation framework conditioned on multiple reference images and poi nt-tracks that jointly anchor the generated frames and the references. Our approach enables precise control over motion and multi-reference composit ing, unlocking a wide range of applications, ...\n\n\nKoichi Namekata (Eye line Labs, University of Oxford); Yash Kant (Eyeline Labs, Netflix); Zhizh eng Liu (Eyeline Labs; University of California, Los Angeles); Ryan Burger t (Eyeline Labs, Stony Brook University); Yuancheng Xu (Eyeline Labs, Netf lix); Kuan Heng Lin (Eyeline Labs, Columbia University); Emmett Steven (Ne tflix); Julien Philip and Li Ma (Eyeline Labs); Andrea Vedaldi (University of Oxford); and Paul Debevec and Ning Yu (Eyeline Labs, Netflix)\n------- --------------\nReRoPE: Repurposing RoPE for Relative Camera Control\n\nRe RoPE is a plug-and-play framework for controllable video generation that a chieves precise camera control by injecting relative pose information into the underutilized low-frequency bands of standard Rotary Positional Embed dings (RoPE).\n\n\nChunyang Li, Yuanbo Yang, Jiahao Shao, and Hongyu Zhou (Zhejiang University); Katja Schwarz (Independent); and Yiyi Liao (Zhejian g University)\n---------------------\nFreeOrbit4D: Training-Free Arbitrary Camera Redirection for Monocular Videos via Foreground-Complete 4D Recons truction\n\nFreeOrbit4D is a training-free framework for camera redirectio n from monocular videos. By decoupling foreground and background reconstru ction and completing foreground geometry via multi-view diffusion, it buil ds a foreground-complete 4D proxy that guides video generation, enabling f aithful and temp...\n\n\nWei Cao and Hao Zhang (University of Illinois Urb ana-Champaign); Fengrui Tian (University of Pennsylvania); Yulun Wu, Yingy ing Li, and Shenlong Wang (University of Illinois Urbana-Champaign); Ning Yu (Eyeline Labs, Netflix); and Yaoyao Liu (University of Illinois Urbana- Champaign)\n---------------------\nUCM: Unified Modeling of Camera Control and Memory with Time-aware Positional Encoding Warping for World Models\n \nWe propose UCM, a novel framework for the unified modeling of long-term scene consistency and camera controllability in video generation-based wor ld models. Trained on over 500K monocular videos, UCM achieves high-fideli ty world exploration capability with excellent generalizability to open-wo rld en...\n\n\nTian-Xing Xu and Zi-Xuan Wang (Tsinghua University), Guangy uan Wang and Li Hu (Alibaba Group), Zhongyi Zhang (University of Science a nd Technology of China), Peng Zhang and Bang Zhang (Alibaba Group), and So ng-Hai Zhang (Tsinghua University)\n---------------------\nCameraSquad: Ac hieving Content Consistency in Parallel Multi-Trajectory Camera-Controlled Video Generation\n\nCameraSquad is a multi-trajectory camera control fram ework that supports both single-trajectory and parallel multi-trajectory v ideo generation. By decoupling content and camera control mechanisms and u sing a dual-mode cross-view attention mechanism, it ensures viewpoint cons istency and camera contro...\n\n\nZhufeng Xu and Xuan Gao (Institute of Co mputing Technology, Chinese Academy of Sciences; University of Chinese Aca demy of Sciences); Bailin Deng (School of Computer Science and Informatics , Cardiff University); Yikang Ding, Xiaoqiang Liu, Haoxian Zhang, and Peng fei Wan (Kling Team, Kuaishou Technology); Hongbo Fu (The Hong Kong Univer sity of Science and Technology); and Lin Gao (Institute of Computing Techn ology, Chinese Academy of Sciences; University of Chinese Academy of Scien ces)\n---------------------\nInfant Vision: Reconfiguring Perception in Mi xed Reality\n\nThis work interweaves vision science with perceptual experi ence in art. Leveraging MR technology, it renders the infant's mysterious vision into a first-person experiential artwork. For the SIGGRAPH communit y, the project serves as a case study of how scientific theories and domai n knowledge can be ...\n\n\nWei Chen Yen (national yang ming chiao tung un iversity), Seth Riskin (Massachusetts Institute of Technology (MIT)), and Chun-Cheng Hsu (national yang ming chiao tung university)\n--------------- ------\nResonance: Meditative Neural Rhythms as Collective Spatial Experie nce\n\nResonance transforms a traditionally solitary meditative practice i nto a shared spatial experience by externalizing a meditator's neural rhyt hms as architectural-scale light and motion. Rather than treating neurofee dback as private data, the work situates cognition within an immersive env ironment th...\n\n\nRuipeng Wang (Massachusetts Institute of Technology (M IT); Critical Matter Group, Media Lab); Yuxiang Cheng and Zhiyan Xing (Har vard University; Critical Matter Group, Media Lab); and Behnaz Farahi (Med ia Lab, Massachusetts Institute of Technology (MIT); Critical Matter Group , Media Lab)\n---------------------\nThe Adaptation Threshold in Mixed Rea lity Theater: Cognitive Integration and Narrative Reality\n\nMixed-reality technologies increasingly enter theatrical and artistic practice, and que stions of perception, presence, and staging become as important as technic al feasibility. This paper focuses on audience experience, stylization, an d artistic intention. The project demonstrates that mixed reality...\n\n\n Nils Gallist (University of Applied Sciences Upper Austria); David Gochfel d (University of York); and Lino Brunmayr (Media Interaction Lab, Universi ty of Applied Sciences Upper Austria, Hagenberg)\n---------------------\nE dges of Immersion: Data Translation into Ecoaesthetic Experience\n\nClimat e knowledge is increasingly shaped by complex data that remain experientia lly distant. This paper examines how immersive ecoaesthetic practices tran slate environmental data into situated sensory experience. Through a case study, it introduces edges of immersion, showing how artistic strategies.. .\n\n\nRasa Smite (ZHdK Zurich University of the Arts, RIXC Center for New Media Culture) and Raitis Smits (Art Academy of Latvia, RIXC Center for N ew Media Culture)\n---------------------\nNexus: Native Mesh Generation wi th Diffusion\n\nNexus is a revolutionary two-stage diffusion framework for generating high-fidelity 3D polygon meshes. By replacing slow sequential modeling with geometry structure and topology diffusion, Nexus effortlessl y handles complex, non-manifold geometries and achieves more controllable and faster mesh gene...\n\n\nHanxiao Wang (Institute of Automation, Chines e Academy of Sciences; University of Chinese Academy of Sciences); Ying-Ti an Liu and Yuan-Chen Guo (VAST); Qi-Yuan Feng (CS Dept, Tsinghua Universit y); Zi-Xin Zou and Ding Liang (VAST); Biao Zhang (Xi'an Jiaotong Universit y); and Yan-Pei Cao (VAST)\n---------------------\nMatérn Noise for Triang ulation-Agnostic Flow Matching on Meshes\n\nThis paper tackles the task of learning to generate signals over triangle meshes in a triangulation-agno stic manner, meaning the trained model can be applied to different mesh tr iangulations effectively. It proposes a specific noise distribution which is triangulation agnostic on meshes, to be used ...\n\n\nTianshu Kuai (Uni versity of Montreal, Mila); Arman Maesumi and Daniel Ritchie (Brown Univer sity); and Noam Aigerman (University of Montreal, Mila)\n----------------- ----\nTempo3D: Efficient Temporal-Aware Fine-Tuning and Multi-View Latent Aggregation for 3D Generation\n\nTempo3D is a resource-efficient framework for high-fidelity 3D generation from a single image. By innovatively enha ncing detail learning and resolving single-view ambiguities without expens ive retraining, it delivers state-of-the-art geometric fidelity. Additiona lly, Tempo3D empowers users with natu...\n\n\nHuizhi Zhu and Jiongming Qin (Wuhan University, School of Computer Science); Yusen Wang (DFRD, Dongfen g Motor Corporation Research&Development Institute); and Chunxia Xiao (Wuh an University, School of Computer Science)\n---------------------\nPEGAsus : 3D Personalization of Geometry and Appearance\n\nPEGAsus enables persona lized 3D shape generation by learning reusable geometry and appearance con cepts from a single reference shape, supporting both global and region-wis e personalization. The learned concepts can be combined with text to synth esize novel personalized 3D assets outperforming existi...\n\n\nJingyu Hu (The Chinese University of Hong Kong), Bin Hu (The University of Hong Kong ), Ka-Hei Hui (Autodesk Research), Haipeng Li (The Hong Kong University of Science and Technology), Zhengzhe Liu (Lingnan University), Daniel Cohen- Or (Tel Aviv University), and Chi-Wing Fu (The Chinese University of Hong Kong)\n---------------------\nAniGen: Unified $S^3$ Fields for Animatable 3D Asset Generation\n\nAniGen generates animatable 3D assets from a single image by jointly producing geometry, skeletons, and skinning weights. Its unified S³ field representation enables consistent, high-quality rigged a sset creation across diverse categories, including animals, humans, cartoo n characters, and articulat...\n\n\nYi-Hua Huang (The University of Hong K ong (HKU)), Zi-Xin Zou (VAST), Yuting He (The Chinese University of Hong K ong), Chirui Chang (The University of Hong Kong (HKU)), Cheng-Feng Pu (Tsi nghua University), Ziyi Yang (The University of Hong Kong (HKU)), Yuan-Che n Guo and Yan-Pei Cao (VAST), and Xiaojuan Qi (The University of Hong Kong (HKU))\n---------------------\nMeshFlow: Mesh Generation with Equivariant Flow Matching\n\nMeshFlow is a powerful diffusion-based framework utilizi ng equivariant modeling to transform random triangle soups into high-quali ty meshes in less than one second.\n\n\nQi Sun (City University of Hong Ko ng), Kiyohiro Nakayama (Stanford University), Jing Yan (Cornell Tech), Qix ing Huang (The University of Texas at Austin), Alexander Rush (Cornell Tec h), Leonidas Guibas and Gordon Wetzstein (Stanford University), Jing Liao (City University of Hong Kong), and Guandao Yang (The University of Texas at Austin)\n---------------------\nStrips as Tokens: Artist Mesh Generatio n with Native UV Segmentation\n\nWe propose Strips as Tokens (SATO), a nov el framework with a token ordering strategy inspired by triangle strips. B y constructing the sequence as a connected chain of faces that explicitly encodes UV boundaries, our method naturally preserves the organized edge f low and semantic layout characteristi...\n\n\nRui Xu and Dafei Qin (Univer sity of Hong Kong); Kaichun Qiao (ShanghaiTech University); Qiujie Dong an d Huaijin Pi (University of Hong Kong); Qixuan Zhang, Longwen Zhang, Lan X u, and Jingyi Yu (ShanghaiTech University); Wenping Wang (Texas A&M Univer sity); and Taku Komura (University of Hong Kong)\n---------------------\nM TPano: Multi-Task Panoramic Scene Understanding via Label-Free Integration of Dense Prediction Priors\n\nMTPano is a robust multi-task foundation mo del for panoramic scene understanding. Trained via a label-free pipeline u sing perspective dense priors, it adopts a dual-stream architecture that r esolves feature conflicts between depth, segmentation, and surface normal estimations, achieving state-of-the...\n\n\nJingdong Zhang (Texas A&M Univ ersity); Xiaohang Zhan and Lingzhi Zhang (Adobe); Yizhou Wang (Northeaster n University, Adobe); and Zhengming Yu, Jionghao Wang, Wenping Wang, and X in Li (Texas A&M University)\n---------------------\nSmoothMotionVectors: Optimizing Your Content for Video Codecs in Free View Video Compression\n\ nConventional video codecs exploit spatial and temporal self-similarity to compress structured signals. We bring this principle to dynamic scene rec onstruction with 3D Gaussian Splatting by organizing deformation outputs t hrough a tailored pipeline. The resulting motion vectors are smoother acro ss sp...\n\n\nMingyang Song (Disney Research Studios, ETH Zürich); Yang Zh ang (Disney Research Studios); Siyu Tang (ETH Zürich); and Tunc Ozan Aydin (Disney Research Studios)\n---------------------\nCAGS: Color-Adaptive Vo lumetric Video Streaming with Dynamic 3D Gaussian Splatting\n\nCAGS is a n ovel color-adaptive system for 3D Gaussian Splatting that enables high-qua lity, immersive volumetric video streaming. By utilizing vector quantizati on and reference images for color restoration, CAGS overcomes bandwidth li mitations, delivering significantly faster speeds and superior visu...\n\n \nDaheng Yin (Simon Fraser University); Yili Jin (McGill University, Simon Fraser University); Jianxin Shi (Nankai University, Simon Fraser Universi ty); Isaac Ding and Miao Zhang (Simon Fraser University); Fangxin Wang (Th e Chinese University of Hong Kong); Zhaowu Huang (Fuzhou University, South east University); Cong Zhang and Jiangchuan Liu (Simon Fraser University); and Fang Dong (Southeast University)\n---------------------\nTSMC: Time-v arying 4D Scene Mesh Compression\n\nTSMC is a compression framework for ti me-varying scene meshes that separates static and dynamic regions, leverag es mesh-enclosed volumes to create reference meshes for motion prediction, and efficiently compresses dynamic displacements, enabling bandwidth-effi cient 4D scene compression and real-time...\n\n\nGuodong Chen (Northeaster n University), Libor Váša (University of West Bohemia), Amrita Mazumdar (N VIDIA), and Mallesham Dasari (Northeastern University)\n------------------ ---\nSparse-to-Complete: From Sparse Image Captures to Complete 3D Scenes\ n\nWe introduce S2C-3D, a novel sparse-view 3D reconstruction framework f or high-fidelity, complete scene reconstruction. It features three compone nts: a specialized diffusion model for scene-specific image restoration, a training-free view-consistency conditioned sampling process in the diffus ion mod...\n\n\nYiyang Shen (State Key Lab of CAD and CG, Zhejiang Univers ity); Yin Yang (The University of Utah); and Kun Zhou and Tianjia Shao (St ate Key Lab of CAD and CG, Zhejiang University)\n---------------------\nAr ticulate That Object Part (ATOP): 3D Part Articulation from Text and via M otion Personalization\n\nWe present a few-shot method for articulating sta tic 3D objects using text-guided motion personalization. Leveraging diffus ion models, we generate plausible part motions and adapt them to input obj ects via image prompting. Differentiable rendering with score distillation transfers multi-view motion ...\n\n\nAditya Vora, Sauradip Nag, Kai Wang, and Hao (Richard) Zhang (Simon Fraser University) and Aditya Vora\n------ ---------------\nControllable Texture Tiling via Diffusion Transformers wi th Transformed Rotary Embeddings\n\nWe introduce a novel Diffusion Transfo rmer framework for high-fidelity, controllable texture tiling. By leveragi ng Coordinate-Transformed Rotary Embeddings, our method enables precise ma nipulation of texture frequency, scale, and orientation. It seamlessly int egrates reference patterns into images w...\n\n\nJunrong Huang and Zhiyuan Zhang (City University of Hong Kong), Rui Tang (Manycore Tech Inc.), Hong bo Fu (Hong Kong University of Science and Technology), and Jing Liao (Cit y University of Hong Kong)\n---------------------\nVideoNeuMat: Neural Mat erial Extraction from Generative Video Models\n\nVideoNeuMat extracts reus able neural materials from video diffusion models. It fine-tunes a large v ideo model as a virtual gonioreflectometer and uses a Large Reconstruction Model (LRM) to reconstruct compact neural materials in a single pass. The resulting materials are realistic, diverse, and gene...\n\n\nBowen Xue (U niversity of Manchester); Saeed Hadadan (NVIDIA); Zheng Zeng (University o f California Santa Barbara, NVIDIA); Fabrice Rousselle (NVIDIA); Zahra Mon tazeri (University of Manchester); and Milos Hasan (NVIDIA)\n------------- --------\nPureSample: Neural Materials Learned by Sampling Microgeometry\n \nPureSample is a neural BRDF representation that learns material appearan ce directly from forward random walks on microgeometry. It enables efficie nt BRDF evaluation, importance sampling, and pdf evaluation using a flow-b ased sampler and a lightweight view-dependent albedo network, supporting c omplex...\n\n\nZixuan Li, Zixiong Wang, and Jian Yang (Nankai University); Miloš Hašan (NVIDIA); and Beibei Wang (Nanjing University)\n------------- --------\nFast and Accurate Gaussian Process Modelling of Real-World Mater ials\n\nWe propose a BRDF modeling method that provides accurate and compa ct representations for isotropic and anisotropic BRDFs. We propose new the oretical developments enabling tractable Gaussian Process BRDF regression, leading to analytical BRDF representations. State-of-the-art methods can be outperfor...\n\n\nArnau Colom (Pompeu Fabra University Interactive Tec hnologies Group (GTI)); Christian Bouville (Institut de Recherche en Infor matique et Systèmes Aléatoires (IRISA)); Julien Pettre (Institut National de Recherche en Informatique et en Automatique (INRIA) Rennes University, CNRS, IRISA); Kadi Bouatouch (Institut de Recherche en Informatique et Sy stèmes Aléatoires (IRISA)); Ricardo Marques (Pompeu Fabra University Inte ractive Technologies Group (GTI)); and Arnau Colom\n---------------------\ nSuccessive Height Preintegration for a Height-Interdependent Path Formula tion of Multiple Bounces in Smith Microfacet BRDFs\n\nClassic microfacet m odels for rendering account only for single-bounce light, leading to energ y loss. We present a novel multiple-bounce Smith microfacet BRDF with a he ight-interdependent formulation, successive height preintegration, and an improved BRDF evaluation method. Our model can conserve en...\n\n\nSiyuan Zhang (Chiba University), Takuya Funatomi (Kyoto University), Yuki Fujimur a and Yasuhiro Mukaigawa (Nara Institute of Science and Technology), and H iroyuki Kubo (Chiba University)\n---------------------\nRobust In-Engine T exture Optimization from Inconsistent Generative Targets\n\nWe present a p ractical method for optimizing textures directly inside production renderi ng engines from inconsistent AI-generated images. By handling geometric mi smatch and unreliable details, our approach avoids blur and ghosting, enab ling robust texture refinement without requiring differentiable ...\n\n\nT aejoon Kim, Seung-Uk Yoon, Seong-Jae Lim, Bon-Woo Hwang, Kinam Kim, and Se ung Wook Lee (Electronics and Telecommunications Research Institute (ETRI) )\n---------------------\nSpatiotemporal FLIP for Fast Free-Surface and Tw o-Phase Simulation With Very Large Time Steps\n\nWe present ST-FLIP, a spa tiotemporal extension of the Fluid-Implicit Particle (FLIP) method for fre e-surface and two-phase simulation. ST-FLIP augments particles with 4D spa tiotemporal coordinates and supports time steps up to an order of magnitud e larger than in existing solvers, delivering several...\n\n\nBernhard Bra un and Rene Winchenbach (Technical University Munich), Jan Bender (RWTH Aa chen University), and Nils Thuerey (Technical University Munich)\n-------- -------------\nBuoyancy-driven Phase Separation in the Material Point Meth od\n\nThe Material Point Method struggles with separating materials due to a shared background grid that couples velocities. We introduce a hybrid a pproach using separate velocity grids and a unified pressure model, enabli ng buoyancy-driven phase separation for immiscible fluids.\n\n\nMehrnaz Ay azi, Craig Schroeder, and Tamar Shinar (University of California Riverside )\n---------------------\nVolume-Preserving LBM-MPM Coupling for Air-Water -Sand Mixtures\n\nWe present a physically-based framework for simulating s and–water–air mixtures by coupling LBM fluids with MPM granular sand under a unified formulation. A water retention model with built-in volume conse rvation enables stable, realistic simulation of mixtures across diverse, m ultiscale ...\n\n\nXiaoyu Xiao (Shanghai Jiao Tong University); Haoxiang W ang (Department of Automation, Tsinghua University); Xiaokang Yang (Shangh ai Jiao Tong University); Mathieu Desbrun (INRIA, Ecole Polytechnique); an d Wei Li (Shanghai Jiao Tong University)\n---------------------\nA Nonloca l Monolithic Variational Framework for Free Surface Flows\n\nWe propose a unified nonlinear optimization framework that achieves a monolithic coupli ng of incompressibility, viscosity, and surface tension within a single so lver. This first particle-based solver resolves their interdependence via position-based nonlocal viscosity formulations, enhancing stabil...\n\n\n Shusen Liu, Yuzhong Guo, Lixin Ren, Ying Qiao, and Xiaowei He (Institute o f Software, Chinese Academy of Sciences)\n---------------------\nStochasti c geomorphological transport for terrain erosion simulation\n\nGeomorpholo gical transport is the long-distance transport of quantities which are key elements of terrain erosion. We propose a new stochastic algorithm for th e efficient simulation of geomorphological transport with momentum conserv ation, which enables us, for the first time, to model dynamic emerg...\n\n \nNicholas McDonald (erosiv Studio GmbH) and Guillaume Cordonnier (Inria, Université Côte d'Azur)\n---------------------\nMixwell: Sharp 2D Fluid Br ushes for Progressive Physics-Based Mixing\n\nMixwell introduces sharp 2D fluid brushes and GPU-accelerated analytical methods for progressive, reso lution-independent physics-based mixing. Derived from potential flow aroun d cylindrical tines, Mixwell evaluates drift per sample without grids or i ntermediate resampling, enabling real-time, arbitr...\n\n\nDoug James (Sta nford University) and Ethan James (N/A)\n---------------------\nGesture Fl ux: Unveiling Budaixi Hand Gestures through Mixed Reality Performance\n\nT his work presents Gesture Flux, an immersive performance system integratin g mixed-reality gesture recognition, Budaixi (traditional Taiwanese glove puppetry), and contemporary dance. Through three interaction modes, hidden puppeteering gestures are transformed from direct articulation to percept ual...\n\n\nYi Jen Lin, Wei-Chen Yen, and Chun-Cheng Hsu (National Chiao T ung University)\n---------------------\nExpanded Voices: Voice Cloning, La tent Spaces, and Embodied Translation\n\nThis work treats AI models not on ly as technical systems, but as perceptual, embodied, and cultural artifac ts. By combining voice cloning and latent space visualization in an intera ctive installation, the project demonstrates how AI representations can be come accessible and critically examined. It o...\n\n\nTomas Andrade Weber (Barcelona Supercomputing Center); Maria Arnal Dimas (Independent); and Ra quel Barrachina Llobet, Sol Bucalo Mana, Jerónimo Calderón, Fernando Cucch ietti, Thaleia Ntiniakou, Adrià Espinoza Gómez, Paula Fernández Vergara, D avid Garcia-Povedano, Alex Gil, Roger Gonzalez March, Othmane Hayoun Mya, Marc Heras Villacampa, Míriam Herrero-Valea, Guillermo Marín, Paula Méndez Gordillo, and Sara Tolosa Alarcón (Barcelona Supercomputing Center)\n---- -----------------\nThe Imagery from Calligraphy: Encoding and Decoding Pic tographs Via Interactive Artistic Exploration\n\nThe Imagery from Calligra phy is a two-part new media artwork that explores how Chinese seal script can encode and decode cultural meaning. Through interactive writing and im mersive performance film, it links calligraphy, embodied action, and Xuan- paper craft into a participatory experiential archive...\n\n\nJiayang Huan g, Kang Zhang, and David Yip (The Hong Kong University of Science and Tech nology (Guangzhou))\n---------------------\nBruSHŪ: Cross-Modal Translatio n of Implicit Micro-Actions in Chinese Calligraphy\n\nBruSHŪ is a sensor-e mbedded calligraphy brush that translates fleeting micro-actions, contact boundaries, turns, and hesitations into sparse sound events and an evolvin g ink-like visual field. Rather than scoring or reconstructing strokes, it makes tacit embodied decisions perceptible and discu...\n\n\nTiancheng LI U (The Hong Kong University of Science and Technology (Guangzhou), Univers ity of Amsterdam); Shumeng Zhang (The Hong Kong University of Science and Technology (Guangzhou), University of Trento); and Nicolò Merendino (The H ong Kong University of Science and Technology (Guangzhou))\n-------------- -------\nCreativity ≠ Generativity: A Case Study of Attentive Machine Lear ning in Dance Performance\n\nDance expresses through translation. It trans forms intention and story into movement through tacit and relational knowl edge. When computational systems enter, they interpret movement as machine -legible forms, shaping how dance is perceived. This paper presents a case study of human–machine co-...\n\n\nZiyu Xu and Zhuodi Cai (University of Washington - DXARTS)\n---------------------\nGuidestar-Free Adaptive Optic s with Asymmetric Apertures\n\nWe introduce and demonstrate the first clos ed-loop adaptive optics system capable of optically correcting aberrations in real-time without a guidestar or a wavefront sensor. We enable these c apabilities by combining asymmetric apertures, machine learning–based PSF estimation and phase retrieva...\n\n\nWeiyun Jiang and Haiyun Guo (Rice Un iversity); Christopher A. Metzler (University of Maryland, College Park); Ashok Veeraraghavan (Rice University); and Weiyun Jiang\n----------------- ----\nMAROON: A Dataset for the Joint Characterization of Near-Field High- Resolution Radio-Frequency and Optical Depth Imaging Techniques\n\nRobust perception benefits from complementary sensors, yet close-range setups rem ain scarce. We characterize optical and near-field radar imagers via multi modal calibration, collecting data from RGB-D sensors and one mmWave MIMO radar. We evaluate depth sensing across materials, geometries, and dist... \n\n\nVanessa Wirth, Johanna Bräunig, Nikolai Hofmann, and Martin Vossiek (Friedrich-Alexander-Universität Erlangen-Nürnberg); Tim Weyrich (Friedric h-Alexander-Universität Erlangen-Nürnberg, University College London); Mar c Stamminger (Friedrich-Alexander-Universität Erlangen-Nürnberg); and Vane ssa Wirth\n---------------------\nNon-line-of-sight imaging with arbitrary relay surface geometries via 3D Gaussian Transient Rendering\n\nWe presen t a LOS-guided NLOS imaging pipeline for arbitrary relay surface geometrie s, representing hidden scenes with 3D Gaussian primitives and developing a n efficient differentiable transient renderer. Real experiments on complex relay surfaces captured with our custom system show improved robustn...\n \n\nYi Wang, Ziyu Zhan, and Yuran Wang (Tsinghua University); Hao Wang (Ci ty University of Hong Kong); and Qiang Liu, Zuoqiang Shi, Lingyun Qiu, and Xing Fu (Tsinghua University)\n---------------------\nBroadband Hyperspec tral 3D Imaging using Dispersed Structured Light\n\nWe present a broadband hyperspectral 3D (BH3D) imaging system spanning visible to SWIR (450–1500 nm). Using a single-spectrograph stereo setup, our method jointly reconst ructs accurate 3D geometry and material properties, enabling robust analys is of real-world scenes beyond the visible spectru...\n\n\nSuhyun Shin and Yunseong Moon (POSTECH), Ryota Maeda (University of Hyogo), David Lindell and Kyros Kutulakos (University of Toronto), and Seung-Hwan Baek (POSTECH )\n---------------------\nGenPIE: A Time-Resolved Plenoptic Imager\n\nWe p resent a physically grounded generative imaging system to resolve high-dim ensional, time-resolved plenoptic light transport from sparse measurements . Using decoupled laser-detector capture, 3D foundation-model priors, and differentiable transient rendering, we reveals indirect light transport fo ...\n\n\nZiheng Wang (The Chinese University of Hong Kong, Shenzhen); Siyu an Shen and Huanyu Xu (ShanghaiTech University); Kaichun Qiao, Longwen Zha ng, and Qixuan Zhang (ShanghaiTech University, Deemos Technology); Qilin S un (The Chinese University of Hong Kong, Shenzhen; Point Spread Technology ); and Shiying Li and Jingyi Yu (ShanghaiTech University)\n--------------- ------\nPolicy-based Foveated Imaging and Perception\n\nUltra-high-resolut ion sensors capture fine details critical for visual perception, but acqui ring all pixels at full resolution is often infeasible under realistic con straints. We introduce a real-time, task-aware foveated imaging framework that learns a sensor attention policy to dynamically allocat...\n\n\nHowar d Xiao, Jan Ackermann, Boyang Deng, and Gordon Wetzstein (Stanford Univers ity)\n---------------------\n8DNA: 8D Neural Asset Light Transport by Dist ribution Learning\n\nWe pre-bake complex light scattering in 3D assets int o neural 8D light transport functions by learning distributions from forwa rd path-traced samples. Unlike previous far-field methods, our representat ion remains accurate under near-field illumination, while delivering lower variance and faster infe...\n\n\nLiwen Wu and Haolin Lu (University of Ca lifornia San Diego); Bing Xu (University of California San Diego, NVIDIA); Miloš Hašan (NVIDIA); and Ravi Ramamoorthi (University of California San Diego)\n---------------------\nA Generalizable Light Transport 3D Embeddin g for Global Illumination\n\nWe introduce a scalable 3D embedding that lea rns global illumination directly from 3D scene configurations. This design scales to environments with millions of triangles, enabling the first gen eralizable GI learning on complex, high-fidelity indoor scenes, far beyond prior limits. It bypasses per-sc...\n\n\nBing Xu (University of Californi a San Diego, NVIDIA); Mukund Varma T, Cheng Wang, and Tzu-mao Li (Universi ty of California San Diego); Lifan Wu and Bartlomiej Wronski (NVIDIA); Rav i Ramamoorthi (University of California San Diego); and Marco Salvi (NVIDI A)\n---------------------\nRao-Blackwellized Markov chain Monte Carlo Ligh t Transport\n\nWe introduce a novel Rao-Blackwellization technique for var iance reduction in Markov chain Monte Carlo light transport. We improve bo th Metropolis Light Transport and the current state-of-the-art, Jump Resto re Light Transport, consistently outperforming traditional variance reduct ion, accelerating c...\n\n\nSascha Holl (Max Planck Institute for Informat ics, Saarland University); Gurprit Singh (Advanced Micro Devices (AMD)); a nd Hans-Peter Seidel (Max Planck Institute for Informatics)\n------------- --------\nRagged Neighborhood Attention for Spatiotemporal Neural Denoisin g of Deep Monte Carlo Renderings\n\nRaNAD is the first spatiotemporal neur al denoiser for deep Monte Carlo renderings. It handles semi-structured de ep images with variable bin counts per pixel through ragged neighborhood a ttention, and exploits temporal context across up to seven frames, greatly advancing in both quality and temporal ...\n\n\nXianyao Zhang, Gerhard Rö thlin, Tunc Ozan Aydin, Farnood Salehi, and Marios Papas (Disney Research Studios)\n---------------------\nPhotons × Force: Differentiable Radiation Pressure Modeling\n\nWe present a system for optimizing spacecraft design s under radiation pressure, the force that photons exert on objects. It co mbines a fast Monte Carlo simulation, a neural proxy of it, and gradient-b ased optimization to find spacecraft geometry, or operational parameters t hat minimize travel time o...\n\n\nCharles Constant and Santosh Bhattarai (University College London (UCL)), Elizabeth Bates (Alan Turing Institute) , and Marek Ziebart and Tobias Ritschel (University College London (UCL))\ n---------------------\nDifferentiable Neutron Transport\n\nWe present a d ifferentiable Monte Carlo framework for multi-energy group neutron transpo rt, extending ideas from differentiable rendering to nuclear engineering. The method enables gradient-based sensitivity analysis and inverse design. \n\n\nXi Deng, Maosen Tang, Michael Czekanski, David Bindel, and Steve Mar schner (Cornell University)\n---------------------\nGPC: Large-Scale Gener ative Pretraining for Transferable Motor Control\n\nWe present Generative Pretrained Controllers, a framework for training reusable generative contr ollers for physically simulated characters. GPC leverages FSQ discretizati on with end-to-end RL to learn reusable motor skills from large-scale moti on data. Once trained, the generative controller can be ...\n\n\nYi Shi (N VIDIA, Simon Fraser University); Yifeng Jiang and Chen Tessler (NVIDIA); a nd Xue Bin Peng (NVIDIA, Simon Fraser University)\n---------------------\n SMP: Reusable Score-Matching Motion Priors for Physics-Based Character Con trol\n\nSMP builds reusable and modular reward models for training motor c ontrollers. Once constructed from a motion dataset, the priors can be reus ed across diverse tasks while preserving the behaviors in the data, withou t requiring access to the original dataset or retraining.\n\n\nYuxuan Mu, Ziyu Zhang, Yi Shi, and Dun Yang (Simon Fraser University); Minami Matsumo to and Kotaro Imamura (Sony Interactive Entertainment); Guy Tevet (Stanfor d University); Chuan Guo (Snap); Michael Taylor (Sony Interactive Entertai nment); Chang Shu and Pengcheng Xi (National Research Council Canada); and Xue Bin Peng (Simon Fraser University, NVIDIA)\n---------------------\nMo tionBricks: Scalable Real-Time Motions with Modular Latent Generative Mode l and Smart Primitives\n\nMotionBricks is a real-time generative framework that transforms interactive motion control for animation and robotics. By combining a large-scale latent backbone with intuitive "smart primitives, " it delivers high-quality, zero-shot motion synthesis at 15,000 FPS, allo wing users to effortlessly bui...\n\n\nTingwu Wang, Olivier Dionne, Michae l De Ruyter, David Minor, and Davis Rempe (NVIDIA); Kaifeng Zhao (NVIDIA, ETH Zürich); Mathis Petrovich, Ye Yuan, Chenran Li, Zhengyi Luo, Brian Rob ison, Xavier Blackwell, and Bernardo Antoniazzi (NVIDIA); Xue Bing Peng (N VIDIA, Simon Fraser University); Yuke Zhu (NVIDIA, The University of Texas at Austin); and Simon Yuen (NVIDIA)\n---------------------\nDeep Motion W arping via Phase-Conditioned Diffusion Autoencoder\n\nWe present a deep mo tion warping framework utilizing a phase-conditioned diffusion autoencoder . By explicitly disentangling motion into root velocity, phase, and style, it enables intuitive, high-level character animation editing including ro ot warping, exaggeration, time warping, and style transfer...\n\n\nBowen Z heng (Zhejiang University, Tencent VISVISE); Linjun Wu (Zhejiang Universit y); Xinwei Jiang, Yujin Chai, and Zijiao Zeng (Tencent VISVISE); He Wang ( University College London (UCL)); and Xiaogang Jin (Zhejiang University)\n ---------------------\nMultiAct: Text-to-Motion Generation from Composite Text via Tailored Attention Guidance\n\nText-to-motion models struggle wit h composite prompts, often collapsing to a single action. We present Multi Act, an inference-time, unpaired framework that enhances pretrained genera tors by amplifying cross-attention for underrepresented tokens. A lightwei ght decision scheme selects optimal paramete...\n\n\nNathan Sala (Tel Aviv University); Ofir Abramovich and Ariel Shamir (Reichman University); Dani el Cohen-Or (Tel Aviv University); Andreas Aristidou (University of Cyprus , CYENS Centre of Excellence); and Sigal Raab (Tel Aviv University)\n----- ----------------\nAutoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation\n\nARDY is a real-time autoregressive diffusion model for interactive 3D human motion generation. Utilizing a no vel hybrid motion representation, ARDY empowers users to interactively syn thesize high-fidelity animations directed by dynamic text prompts and flex ible spatial constraints, such as full-body...\n\n\nKaifeng Zhao (ETH Züri ch, NVIDIA); Mathis Petrovich, Haotian Zhang, and Tingwu Wang (NVIDIA); Si yu Tang (ETH Zürich); and Davis Rempe (NVIDIA)\n---------------------\nA R obust and Efficient Intersection Algorithm for NURBS Surfaces: Handling Sm all Loops and Tangent Intersections\n\nThis paper presents a robust and ef ficient method for computing NURBS surface intersection curves, with empha sis on small loops and degenerate cases. By combining winding number theor y with subdivision on the parametric domain, the method reliably detects s tarting points of all branches, improving t...\n\n\nJieyin Yang and Xiaoho ng Jia (State Key Laboratory of Mathematical Sciences Academy of Mathemati cs and Systems Science Chinese Academy of Sciences University of Chinese A cademy of Sciences) and Jieyin Yang\n---------------------\nStress-Aware P anelization of Freeform Surfaces\n\nWe present a variational surface ratio nalization algorithm for creating structurally robust polyhedral approxima tions of freeform surfaces. Coupling a continuous panelization flow with d ifferentiable shell stress analysis, our method lets structural performanc e guide the panel decomposition, achievi...\n\n\nXinzhuo Hu (University of California Davis), Roi Poranne (University of Haifa), and Julian Panetta (University of California Davis)\n---------------------\nFree-form Surface Approximation Using Rotational Patches\n\nWe propose a method to approxim ate free-form surfaces using rotational patches for architectural rational ization. By segmenting the input surface and optimizing B-spline-based pat ch layouts, followed by post-processing, we generate smooth, seamless quad meshes with locally repeated building elements...\n\n\nYuanpeng Liu (RMIT University), Yi Min Xie (Hohai University), Ting-Uei Lee (RMIT University ), Ziqi Wang (The Hong Kong University of Science and Technology), Nico Pi etroni (University of Technology Sydney), and Yuanpeng Liu\n-------------- -------\nContinuity-Enhancing Degree Elevation and Splits\n\nWe present tw o novel methods for automatically enhancing the continuity of piecewise-po lynomial curves with negligible computational overhead. They provide the f irst practical interface for modeling curves with curvature (or higher) co ntinuity using the familiar Bezier handles of piecewise cubic cur...\n\n\n Cem Yuksel (University of Utah)\n---------------------\nNeuPPS: Neural Pie cewise Parametric Surfaces\n\nWe present Neural Piecewise Parametric Surfa ces (NeuPPS) that models complex surface geometries with a coarse patch la yout containing arbitrary n-sided surface patches, offering enhanced flexi bility, high fitting precision, and ensured continuity between adjacent pa tches. Applications to surface fit...\n\n\nLei Yang (The University of Hon g Kong); Yongqing Liang (Center for Robotics and Embodied Intelligence (CR EO)); Xin Li (Texas A&M University); Congyi Zhang (University of Texas Dal las); Guying Lin (Carnegie Mellon University); Cheng Lin (Macau University of Science and Technology); Alla Sheffer (University of British Columbia) ; Scott Schaefer, John Keyser, and Wenping Wang (Texas A&M University); an d Lei Yang\n---------------------\nNeuBase: Spline Surfaces with Neural Ba sis Functions\n\nWe introduce NeuBase, a neural parametric surface represe ntation that both accurately fits target surfaces with fine geometric deta il and supports intuitive real time surface deformation.\n\n\nAnshul Mendi ratta (Texas A&M University); Lei Yang (The University of Hong Kong (HKU)) ; and Xin Li, John Keyser, Scott Schaefer, and Wenping Wang (Texas A&M Uni versity)\n---------------------\nLight Architecture: Translating the AI Bl ack Box into Immersive Experience\n\nAs AI becomes an opaque "Black Box," Light Architecture physically "translates" invisible algorithmic structure s instead of merely exploiting generative outputs. Synthesizing kinetic li ght and spatial audio into a Gesamtkunstwerk, this "Experiential AI" offer s a sensory alternative to traditional Ex...\n\n\nYiyun Kang (Yiyun Kang S tudio, Korea Advanced Institute of Science and Technology (KAIST))\n------ ---------------\nMachine Civilization: Exploring Machine Alterity through a Generative Semiotic Installation\n\nMachine Civilization offers SIGGRAPH community a critical lens for examining the gap between explainability an d intelligibility in the age of intelligence. This work explored non-human meaning-making and comprehension limits in human-machine interaction. It raises an ethical question: can we engage ...\n\n\nJia-Qi Shi, Yixiong Wan g, Jiawen Zheng, Hourun Wang, Yufeng Zeng, and Chen Liang (Hong Kong Unive rsity of Science and Technology, Guangzhou)\n---------------------\nGorgon Loop: An Interactive Art Installation Revealing Algorithmic Judgment thro ugh Machine Vision and Generative Language\n\nGorgon Loop makes algorithmi c judgement in public space visible as a collective, embodied process. Thr ough machine vision, language models, and group commentary, it exposes how limited data, bias, and automation shape social evaluation, offering a cr itical lens on the cultural consequences of AI per...\n\n\nShuang Li (伦敦大学 金史密斯学院) and Patrick Hartono (RMIT University Vietnam)\n------------------- --\nFrom Computer Vision to the Vision of Law: Human Overs[a]ight\n\nThe a rtwork HUMAN OVERS[A]IGHT: THE OPS ROOM is relevant to the SIGGRAPH commun ity as it introduces the critical issues of lawmaking in the field of Arti ficial Intelligence. By bridging these disciplines, this translation aims to open a broader discourse on the topic from an innovative standpoint in. ..\n\n\nKristina Tica and Joaquín Santuber (Johannes Kepler University Lin z)\n---------------------\nSpatio-Temporal Control Variates with ReSTIR fo r Real-Time Rendering\n\nWe present ReSTCV, a real-time path tracing metho d that integrates spatio-temporal control variates into ReSTIR. By improvi ng sample reuse beyond single-sample estimation, our approach significantl y reduces noise, especially chromatic noise under complex lighting, while maintaining efficiency and re...\n\n\nZhong Shi and Cunhao Wu (CS Dept, Ts inghua University); Lifan Wu (NVIDIA); and Kun Xu (CS Dept, Tsinghua Unive rsity)\n---------------------\nToF ReSTIR: Time-of-Flight Rendering with S patio-temporal Reservoir Resampling\n\nWe present a novel method for simul ating time-of-flight imaging in complex, dynamic scenes. By reusing and co rrecting light paths, our approach enables interactive performance and sup ports applications such as shape reconstruction and navigation under pract ical real-time constraints.\n\n\nJuhyeon Kim, Wojciech Jarosz, and Adithya Pediredla (Dartmouth College)\n---------------------\nReSTIR BDPT: Bidire ctional ReSTIR Path Tracing with Caustics\n\nReSTIR accelerates path traci ng by reusing samples between pixels. However, existing methods are limite d by the sampling quality of path tracing, making them inefficient for sce nes with caustics and hard-to-reach lights. We develop a ReSTIR variant in corporating bidirectional path tracing that signi...\n\n\nTrevor Hedstrom (University of California San Diego); Markus Kettunen, Daqi Lin, and Chris Wyman (NVIDIA); Tzu-Mao Li (University of California San Diego); and Trev or Hedstrom\n---------------------\nReal-Time Level-of-Detail rendering wi th ReSTIR\n\nWe propose a method to overcome a limitation in ReSTIR path t racing by introducing a surface point mapping to reuse samples across fram es containing meshes with different topologies and enabling real-time LoD rendering with ReSTIR.\n\n\nYu-Chen Wang (University of California Irvine) ; Markus Kettunen, Daqi Lin, Chris Wyman, and Lifan Wu (NVIDIA); and Shuan g Zhao (University of Illinois Urbana-Champaign)\n---------------------\nM ulti-Layer Reservoir Splatting for Temporal Reuse under Disocclusion\n\nTh is work addresses disocclusion noise in ReSTIR by extending the framework to multiple screen-space layers. This enables occluded samples in previous frames to contribute to the final image, improving temporal reuse in diso ccluded regions. Depth ranges and active domains are introduced to reduce co...\n\n\nPengpei Hong and Song Zhang (University of Utah); Daqi Lin, Mar kus Kettunen, and Chris Wyman (NVIDIA); and Cem Yuksel (University of Utah )\n---------------------\nForget Superresolution, Sample Adaptively (when Path Tracing)\n\nWe introduce an end-to-end adaptive sampling and denoisin g pipeline for sparse real-time path tracing. Our method trains stably des pite discrete sampling decisions and uses perceptual, tonemapping-aware op timization to place samples where they matter most. It samples fine struct ures, specular highli...\n\n\nMartin Balint, Corentin Salaün, Hans-Peter S eidel, and Karol Myszkowski (Max Planck Institute for Informatics)\n------ ---------------\nHairGPT: A Unified Autoregressive Framework for 3D Realis tic Hairstyle Synthesis\n\nHairGPT introduces a paradigm shift toward stra nd-as-language modeling, reformulating 3D realistic hair generation as a d ual-decoupled autoregressive process where strands are treated as the fund amental generative units.\n\n\nHaimin Luo (ShanghaiTech University); min O uyang (ShanghaiTech University; Deemos Technology Co., Ltd.); and Lan Xu a nd Jingyi Yu (ShanghaiTech University)\n---------------------\nHairLRM: St rand-based Hair Modeling via Large Reconstruction Models\n\nHairLRM integr ates Large Reconstruction Model (LRM) geometric priors into strand generat ion. Leveraging an LRM mesh anchor and a Dual Orientation AutoEncoder, it transforms coarse geometry into high-fidelity strands, disentangling compl ex topologies to set a new benchmark for robust, accurate hair r...\n\n\nY uefan Shen (LIGHTSPEED); Yican Dong (State Key Laboratory of CAD & CG, Zhe jiang University); Xiufeng Huang (Hong Kong Baptist University); Zhongtian Zheng (LIGHTSPEED); Youyi Zheng (State Key Laboratory of CAD & CG, Zhejia ng University); and Kui Wu (LIGHTSPEED)\n---------------------\nHairPort: In-context 3D-aware Hair Import and Transfer for Images\n\nHairPort is a 3 D-aware hairstyle transfer method that faithfully moves a reference hairst yle onto a source face image, even under large pose and scale differences. By separating hair removal from synthesis and enforcing geometric consist ency through 3D reconstruction, HairPort achieves accurate, ide...\n\n\nAl ireza Heidari and Amirhossein Alimohammadi (Simon Fraser University), Wall ace Michel Pinto Lira and Adi Bar-Lev (Huawei Canada), and Ali Mahdavi-Ami ri (Simon Fraser University)\n---------------------\nCurvature Space Editi ng of Highly-Coiled Hair\n\nWe use the material curvatures of the super-he lix model to better analyze and edit tightly coiled hair. Within this curv ature space, we propose novel operations for hair ruffling, heuristics for tracking flips in handedness, and control armatures that enable intuitive ensemble strand editing.\n\n\nAlvin Shi (Yale University); Florence Berta ils-Descoubes (INRIA, France); A.M. Darke (University of California Santa Cruz); and Theodore Kim (Yale University)\n---------------------\nEfficien t Fur and Hair Multiple Scattering Using Volumetric Approximation\n\nWe in troduce a hybrid framework for rendering dense hair and fur. By preserving fine geometric details for direct illumination while approximating comple x multiple scattering within an equivalent anisotropic volume, our method achieves path-traced visual fidelity and soft glow of multiple scattering. ..\n\n\nRuike Hu (Nanjing University), Junqiu Zhu (Shandong University), M inghao Lin (Shandong Universityong), Ruian Zhang (Nanjing University), Lu Wang (Shandong University), Jie Guo and Yanwen Guo (Nanjing University), a nd Lingqi Yan (MBZUAI)\n---------------------\nOn-the-fly Repulsion in the Contextual Space for Rich Diversity in Diffusion Transformers\n\nModern T 2I DiTs are powerful yet often lack diversity. We identify a novel interve ntion point within DiTs: the Contextual Space at the transformer block lev el, where textual tokens are enriched by image content. We propose a metho d that repels a batch of samples in this space to produce diverse gene...\ n\n\nOmer Dahary (Tel Aviv University, Snap Research); Benaya Koren (Tel A viv University); and Daniel Garibi and Daniel Cohen-Or (Tel Aviv Universit y, Snap Research)\n---------------------\nBFS: Back-to-Front Layered Image Synthesis via Knowledge Transfer\n\nBFS is a novel generation-based frame work for layered image synthesis that produces foreground layers with real istic visual effects while harmonizing with a given background. Using a du al-branch diffusion model and two-stage training with unlayered data, it o vercomes data scarcity and achieves super...\n\n\nKyoungkook Kang (Samsung ) and Gyujin Sim and Sunghyun Cho (POSTECH)\n---------------------\nColorf ul-Noise: Training-Free Low-Frequency Noise Manipulation for Color-Based C onditional Image Generation\n\nColorful-Noise is a training-free image gen eration conditioning method that leverages FFT noise frequency manipulatio n. We show that low frequencies dominate output appearance and can encode structure from scribbles, color maps, or reference images — enabling anyon e to condition image generati...\n\n\nNadav Z. Cohen, Ofir Abramovich, and Ariel Shamir (Reichman University)\n---------------------\nInspiration Se eds: Learning Non-Literal Visual Combinations for Generative Exploration\n \nInspiration Seeds is a generative framework that combines two input imag es into diverse, visually coherent compositions—without text prompts. Trai ned on synthetic triplets using CLIP Sparse Autoencoders, it uncovers late nt visual relationships between inputs, supporting early-stage creative ex ...\n\n\nKfir Goldberg (Bria AI), Elad Richardson (Runway), and Yael Vinke r (Massachusetts Institute of Technology (MIT))\n---------------------\nCa nvas-to-Image: Compositional Image Generation with Multimodal Controls\n\n Canvas-to-Image is a unified framework enhancing diffusion models' composi tional and multimodal control through a single canvas interface. By encodi ng subject references, bounding boxes, and poses into one composite canvas , it enables integrated visual-spatial reasoning. This approach streamline s he...\n\n\nYusuf Dalva (Virginia Tech University); Gordon Guocheng Qian and Maya Goldenberg (Snap); Tsai-Shien Chen (Snap, University of Californi a Merced); Kfir Aberman and Sergey Tulyakov (Snap); Pinar Yanardag (Virgin ia Tech University); and Kuan-Chieh Jackson Wang (Snap)\n----------------- ----\nComboStoc: Combinatorial Stochasticity for Diffusion Generative Mode ls\n\nWe study an important factor in diffusion generative models: combina torial complexity. Existing training schemes insufficiently sample the joi nt space of dimensions and attributes, leading to weaker test time perform ance. We address this with ComboStoc, a simple stochastic process that bet ter exploi...\n\n\nRui Xu and Jiepeng Wang (University of Hong Kong), Hao Pan (Tsinghua University), Yang Liu and Xin Tong (Microsoft Research Asia) , Shiqing Xin and Changhe Tu (Shandong University), Taku Komura (The Unive rsity of Hong Kong), and Wenping Wang (Texas A&M University)\n------------ ---------\nMotion4Motion: Motion Transfer Across Subjects at Inference\n\n This work proposes a novel framework, Motion4Motion, for transferring a mo tion from one subject to another in a video. Without a skeleton or neural network training, Motion4Motion runs efficiently at inference.\n\n\nLing-H ao Chen (Tsinghua University, Stepfun); Zixin Yin (Hong Kong University of Science and Technology, Stepfun); and Duomin Wang, Xianfang Zeng, and Gan g Yu (Stepfun)\n---------------------\nSTyMo: Fast and Controllable Few-Sh ot Motion Style Transfer\n\nSTyMo is a few-shot approach that learns motio n style from only seconds of paired data and trains in one to two minutes. Our key insight is to decompose style into two components: a static compo nent capturing time-invariant posture, and a temporal component capturing frame-wise dynamics.\n\n\nJose Luis Ponton (Reality Labs, Meta; Universita t Politècnica de Catalunya (UPC)) and Alexander Winkler, Ladislav Kavan, Y uting Ye, and Petr Kadlecek (Reality Labs, Meta)\n---------------------\nS kinned Motion Retargeting with Spatially Adaptive Interaction Guidance\n\n We present a geometry-aware motion retargeting framework that preserves in teraction semantics by performing proximity matching over spatially adapti ve anchors. Unlike prior methods with static anchor definitions, our metho d dynamically repositions anchors to reachable regions on the target chara cter...\n\n\nSoojin Choi, Seokhyeon Hong, Chaelin Kim, Junghyun Nam, Junhy uk Jeon, and Junyong Noh (Korea Advanced Institute of Science and Technolo gy (KAIST))\n---------------------\nReActor: Reinforcement Learning for Ph ysics-Aware Motion Retargeting\n\nWe present a reinforcement learning-base d retargeting method that transfers human motion to diverse morphologies, including humanoids and a quadruped, without artifacts such as foot-slidin g, hovering, or self-penetration. A bilevel optimization framework refines upper-loop retargeting parameters def...\n\n\nDavid Müller, Agon Serifi, Sammy Christen, Ruben Grandia, Espen Knoop, and Moritz Bächer (Disney Rese arch)\n---------------------\nAdaptive Interpolation-Synthesis for Motion In-Betweening on Keyframe-Based Animation\n\nWe propose a motion in-betwee ning method tailored to stylized keyframe-based animation production. Our Adaptive Interpolation–Synthesis layer dynamically blends learned interpol ation with direct pose synthesis, mirroring the animator's workflow. Combi ned with a domain-based input keypose train...\n\n\nAnton Raël, Julien Bou cher, and Antoine Lhermitte (Animaj)\n---------------------\nLayerInbetwee n: Occlusion-Aware Stroke Correspondence and Inbetweening with Automatic L ayering\n\nLayerInbetween is an occlusion-aware framework for vector strok e correspondence and automatic inbetweening. It performs automatic layerin g to guide stroke tracing and correspondence finding for occluded strokes, and to resolve occlusion with layers in the inbetween frames. The vector- based method en...\n\n\nHaoran Mo (The Hong Kong University of Science and Technology (Guangzhou)); Zhongyue Guan (The Hong Kong University of Scien ce and Technology (Guangzhou), Tencent); Yixin Hu (Tencent America); and Z eyu Wang (The Hong Kong University of Science and Technology (Guangzhou), The Hong Kong University of Science and Technology)\n--------------------- \nM-ABD: Scalable, Efficient, and Robust Multi-Affine-Body Dynamics\n\nWe present a multibody simulation framework leveraging affine body dynamics w ith a co-rotational formulation, enabling pre-factorized system matrices u nder implicit integration. Mapping primal coordinates onto a dual space vi a KKT, our method exactly enforces joint constraints, achieving interactiv e...\n\n\nZhiyong He and Dewen Guo (University of Utah); Minghao Guo (Comp uter Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT)); Yili Zhao (USC); Wojciech Matusik (Comput er Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts I nstitute of Technology (MIT)); Hao Su (University of California San Diego) ; Chenfanfu Jiang (University of California Los Angeles); Peter Yichen Che n (The University of British Columbia, New York University); and Yin Yang (University of Utah)\n---------------------\nHeterogeneous Subspace Correc tions for GPU Deformable Multibody Dynamics\n\nHeterogeneous Subspace Corr ections (HSC) accelerates GPU simulations of complex multibody systems. By decoupling deformable and rigid/affine components into separate Neumann-p reconditioned Krylov iterations and utilizing an Adaptive Cross Approximat ion (ACA) for low-rank coupling, it overcomes ill-c...\n\n\nDewen Guo (Pek ing University, University of Utah); Zhendong Wang (Style3D Research); Min chen Li (Carnegie Mellon University, Genesis AI); Sheng Li and Guoping Wan g (Peking University); Huamin Wang (Style3D Research); Chenfanfu Jiang (Un iversity of California Los Angeles); and Yin Yang (University of Utah)\n-- -------------------\nDistributed Affine Body Dynamics with Adaptive Consen sus\n\nAffine Body Dynamics within Incremental Potential Contact enables a ccurate simulation of near-rigid, extremely stiff solids with strict non-p enetration. We present a distributed ABD solver based on consensus ADMM, w here nodes solve local subproblems in parallel and synchronize shared bodi es through ...\n\n\nJiafeng Liu, Wenhui Zhou, and Xinming Pei (State Key L ab of CAD and CG, Zhejiang University); Yifan Peng (The University of Hong Kong); Huamin Wang (Style3D Research); Yin Yang (University of Utah); and Lei Lan and Weiwei Xu (State Key Lab of CAD and CG, Zhejiang University)\ n---------------------\nBetter Bending: Analysis, Construction and Verific ation of Discrete Bending Models for Kirchhoff-Love Shells\n\nWhile thin s hells have been studied for decades, there is little consensus on how to d iscretize them. We systematically study bending models for Kirchhoff-Love shells. We develop new models and methods to address discovered simulation gaps, and make practical recommendations of when and how these mo...\n\n\ nZhen Chen and Danny Kaufman (Adobe Research) and Etienne Vouga (The Unive rsity of Texas at Austin)\n---------------------\nEfficient B-Spline Finit e Elements for Cloth Simulation\n\nWe present a quadratic B-spline FEM app roach for cloth simulation that captures smoother, more accurate deformati ons with rich wrinkle detail and few artifacts. Our optimized reduced inte gration scheme enables better performance compared to traditional linear F EM, while supporting robust, high-quali...\n\n\nYuqi Meng (Carnegie Mellon University, University of Utah); Yihao Shi (Carnegie Mellon University, Z hejiang University); Kemeng Huang (Carnegie Mellon University, University of Hong Kong); Zixuan Lu (University of Utah); Ning Guo (Zhejiang Universi ty); Taku Komura (University of Hong Kong); Yin Yang (University of Utah); and Minchen Li (Carnegie Mellon University, Genesis AI)\n---------------- -----\nInteractive Yarn-level Knitwear with Nested Douglas-Rachford Splitt ing\n\nWe propose a novel framework utilizing a generalized Douglas-Rachfo rd Splitting (DRS) scheme and hierarchical decomposition for high-resoluti on knitwear simulation. By convexifying complex yarn dynamics, it ensures robust, optimal convergence. With matrix-free GPU parallelization, our met hod achieve...\n\n\nChun Yuan, Zixuan Lu, Haoyang Shi, and Dewen Guo (Univ ersity of Utah); Huamin Wang (Style3D Research); Chenfanfu Jiang (UCLA); Z herong Pan and Kui Wu (LIGHTSPEED); and Yin Yang (University of Utah)\n--- ------------------\nSymX: Energy-based Simulation from Symbolic Expression s\n\nSymX is an open-source framework that turns succinct symbolic energy expressions into optimized code, delivering automatic first- and second-or der derivatives and parallel assembly for Newton-style optimization time i ntegrators. SymX facilitates building complex simulations, including high- order FEM...\n\n\nJosé Fernández-Fernández, Fabian Löschner, Lukas Westhof en, Andreas Longva, and Jan Bender (RWTH Aachen University) and José Ferná ndez-Fernández\n---------------------\nSurface Reconstruction from Unorien ted Points via Green’s 3rd Identity\n\nG3R reconstructs surfaces from unor iented point clouds by using Green's 3rd identity to recover globally cons istent normals by solving a well-conditioned linear system. It removes the tangential nullspace that weakens prior winding-number-based methods and delivers accurate orientations and higher-q...\n\n\nZhonghao Wu (Universit y of Science and Technology of China, Shanghai Innovation Institute) and D ong Xiao and Renjie Chen (University of Science and Technology of China)\n ---------------------\nA Bayesian Approach for Task-Specific Next-Best-Vie w Selection with Uncertain Geometry\n\nWe introduce a Bayesian framework f or task-specific next-best-view selection in 3D reconstruction. Unlike gen eral uncertainty-reduction methods, we selectively reduce uncertainty only in regions critical to the downstream task via task-specific acquisition functions on posterior distributions over i...\n\n\nJingsen Zhu (Cornell U niversity), Silvia Sellán (Columbia University), and Alexander Terenin (Co rnell University)\n---------------------\nCaRaFe: Camera-Radar Radiance Fi elds for Scene Reconstruction\n\nCaRaFe is the first method to jointly rec onstruct 3D scenes from both camera and radar data using a multi-modal sha red neural geometry representation. It can reconstruct large-scale urban e nvironments from a single drive-through and synthesize measurements from b oth sensors at novel viewpoints, leve...\n\n\nDavid Borts (ETH Zürich); Ju lian Ost (Princeton University, Torc Robotics); Shamik Basu (INRIA); Tim B roedermann (ETH Zürich); Andrea Ramazzina (Mercedes-Benz, Technical Univer sity of Munich); Christos Sakaridis (ETH Zürich); and Mario Bijelic and Fe lix Heide (Princeton University, Torc Robotics)\n---------------------\nPi xels2Peaks: Converting Terrain Images to Heightmaps\n\nPixels2Peaks automa tes the creation of 3D bare-earth terrain heightmaps from a single 2D phot ograph. Our method first estimates camera parameters and visible geometry; then uses a guided diffusion model to plausibly synthesize occluded regio ns. The result maintains geomorphological, hydrological, an...\n\n\nAryama an Jain (INRIA, Université Côte d'Azur); James Gain (University of Cape To wn); and Guillaume Cordonnier (INRIA, Université Côte d'Azur)\n----------- ----------\nMegaNorm: Local Patches Embedding for Efficient and Robust Poi nt Normal Orientation at Super-Large Scale\n\nMegaNorm is a fast, accurate , and robust method for orienting normals in non-watertight, scene-scale p oint clouds, scaling to ~50M points in under 10 minutes. It uses two feed- forward networks: one orients local patches in parallel, while a lightweig ht MLP evaluates pairwise consistency. Finally, a ...\n\n\nZhuodong Li, Ze ngke Liu, Fei Hou, Xuhui Chen, and Wencheng Wang (Institute of Software, C hinese Academy of Sciences; University of Chinese Academy of Sciences) and Ying He (Nanyang Technological University)\n---------------------\nResolu tion Where It Counts: Hash-based GPU-Accelerated 3D Reconstruction via Var iance-Adaptive Voxel Grids\n\nWe propose a variance-adaptive voxel grid th at adjusts size based on local SDF variance. Using a flat spatial hash tab le, it achieves constant-time access, memory efficiency, and full GPU para llelism. Additionally, a parallel quad-tree structure supports GPU-acceler ated Gaussian Splatting rendering,...\n\n\nLorenzo De Rebotti, Emanuele Gi acomini, Giorgio Grisetti, and Luca Di Giammarino (University of Rome La S apienza) and Lorenzo De Rebotti\n---------------------\nArtiFixer: Enhanci ng and Extending 3D Reconstruction with Auto-Regressive Diffusion Models\n \nArtiFixer enhances and extends 3D reconstructions efficiently and scalab ly. Given an initial reconstruction and reference views, it autoregressive generates novel content consistent with existing observations. ArtiFixer can produce hundreds of novel views in a single inference pass or serve as pseud...\n\n\nRiccardo De Lutio (NVIDIA); Tobias Fischer (NVIDIA, ETH Zür ich); Yen-Yu Chang (NVIDIA, Cornell University); Yuxuan Zhang and Zhangjie Wu (NVIDIA); Xuanchi Ren and Tianchang Shen (NVIDIA, University of Toront o); and Katarína Tóthová, Zan Gojcic, and Haithem Turki (NVIDIA)\n-------- -------------\nMix3R: Mixing Feed-forward Reconstruction and Generative 3D Priors for Joint Multi-view Aligned 3D Reconstruction and Pose Estimation \n\nWe present Mix3R, a method that mixes feed-forward reconstruction and generative 3D priors using a Mixture-of-Transformer architecture for multi -view aligned generative 3D reconstruction from sparse views.\n\n\nSiyou L in and Zhou Xue (Tsinghua University), Hongwen Zhang (Beijing Normal Unive rsity), Liang An (Tsinghua University), Dongping Li and Shaohui Jiao (Byte Dance Inc.), and Yebin Liu (Tsinghua University)\n---------------------\nD istance Field Rasterization for End-to-End Mesh Reconstruction\n\nSDFRaste r combines the efficiency of rasterization with signed distance fields for end-to-end mesh reconstruction from multi-view images. By optimizing a te trahedral SDF and extracting meshes during optimization, it produces accur ate, complete, and compact surfaces without heuristic post-processing.\n\n \nJinkai Cui, Kaiwen Song, Chumeng Niu, and Juyong Zhang (University of Sc ience and Technology of China)\n---------------------\nTopology Type Estim ation of Simulated 4D Image Data by Combining Downscaling and Convolutiona l Neural Networks\n\nThis project integrates a 4D camera model, synthetic 4D datasets, and CNNs trained on downscaled data to analyse its 4D topolog y. The results suggest that advanced 4D vision pipelines, particularly whe n paired with significant computational resources, could enable scientific breakthroughs unattainab...\n\n\nKhalil Hannouch and Stephan Chalup (The University of Newcastle) and Stephan Chalup\n---------------------\nNILE: Nested Interleaving of Low-Dimensional Elements\n\nWe introduce a novel mo dular modular meta-sampler architecture that bridges local subspace and gl obal sampling, allowing integrator designers to employ specialized low-dim ensional samplers while still achieving high-dimensional uniformity.\n\n\n Abdalla Ahmed (Shenzhen University); Matt Pharr (NVIDIA); Victor Ostromouk hov (Université Claude Bernard Lyon 1, CNRS - LIRIS); and Hui Huang (Shenz hen University)\n---------------------\nGradient-based Design of Non-Unifo rm Low-Discrepancy Samples\n\nThis paper introduces a first numerical opti mization method for approximately generating non-uniform low-discrepancy p oint sets with an arbitrary target distribution. It is based on a new diff erentiable estimator of non-uniform L2-discrepancy for a given point set.\ n\n\nXiangyu Li (Brown University), Ege Ciklabakkal (University of Waterlo o), Daniel Ritchie (Brown University), and Toshiya Hachisuka (University o f Waterloo)\n---------------------\nBeyond Positional Encoding: A 5D Spati o-Directional Hash Encoding\n\nThis paper introduces two new compact neura l encodings, the hash-sphere and hash-grid-sphere, which represent all-fre quency signals using hierarchical geodesic grids. While we demonstrated th eir effectiveness in neural path guiding, radiance field reconstruction, a nd incident radiance caching, these...\n\n\nPhilippe Weier (Meta, Saarland University); Lukas Bode (Meta); Philipp Slusallek (Saarland University); and Adrián Jarabo and Sébastien Speierer (Meta)\n---------------------\nHi -SPAD: Video-Rate Hyperspectral Imaging and Inference with Single-Photon C ameras\n\nThis work develops a hyperspectral imager using a SPAD image sen sor that achieves a spatial resolution of 1024 × 768 pixels across 123 spe ctral bands spanning visible and near-infrared wavelengths, at a temporal rate of 26 frames per second.\n\n\nHaejoon Lee (Carnegie Mellon University ), Mohit Gupta (University of Wisconsin-Madison), Vijayakumar Bhagavatula and Aswin Sankaranarayanan (Carnegie Mellon University), and Haejoon Lee\n ---------------------\nRobust Planar Maps for 3D Vectorization\n\nVectoriz ing 3D scenes into 2D vector images requires planar maps to define solid n on-overlapping regions. Existing planar map methods are slow and susceptib le to numerical instability. We introduce a robust and efficient method fo r constructing planar maps, using a spatial hierarchy as the fundament...\ n\n\nRobert Fuchs (Carnegie Mellon University) and Keenan Crane (Carnegie Mellon University, Roblox)\n---------------------\n2D Gaussian Splatting f or Bézier Spline Line Art Vectorization\n\nWe introduce a state-of-the-art vectorization method using depth and semantic features to extract strokes . By modeling Bézier curves via differentiable 2D Gaussian splatting, we e fficiently optimize geometry and texture. The framework extends to video f or temporal consistency, offering high-quality,...\n\n\nTianhao Chen (ETH Zurich), Clara Fernandez (Disney Research Studios), Marteinn Oskarsson and Chuck Tappan (Walt Disney Animation Studios), Olga Sorkine-Hornung and Ma rkus Gross (ETH Zurich), and Christopher Schroers and Abdelaziz Djelouah ( Disney Research Studios)\n---------------------\nMonte Carlo Rendering of Biharmonic Diffusion Curves\n\nBiharmonic diffusion curves are lightweight vector graphics primitives that offer artistic control over color gradien ts. We present their first Monte Carlo renderer, which builds upon recent advances in Monte Carlo geometry processing. We phrase the biharmonic prob lem as two coupled second-order pro...\n\n\nPaul Himmler and Tobias Günthe r (Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU))\n------------- --------\nReal-Time GPU Vector Graphics SDF Generation Based on Quadratic Stroke Rendering\n\nWe introduce a GPU-based algorithm for real-time signe d distance field generation from quadratic vector graphics. By treating SD F computation as parallel stroke rendering, the method achieves substantia l speed improvements over CPU approaches, enabling dynamic text, interacti ve user interfaces, and ...\n\n\nXuhai Chen and Guangze Zhang (Fujian Key Laboratory of Urban Intelligent Sensing and Computing, Xiamen University); Juan Cao (School of Mathematical Sciences, Xiamen University); and Zhongg ui Chen (Fujian Key Laboratory of Urban Intelligent Sensing and Computing, Xiamen University)\n---------------------\nStroke of Surprise: Progressiv e Semantic Illusions in Vector Sketching\n\nThis work introduces Progressi ve Semantic Illusions, a novel vector sketching task where a single sketch undergoes a dramatic semantic transformation through the sequential addit ion of strokes, and presents Stroke of Surprise, a generative framework th at optimizes vector strokes to satisfy distinct ...\n\n\nHuai-Hsun Cheng, Siang-Ling Zhang, and Yu-Lun Liu (National Yang Ming Chiao Tung University )\n---------------------\nNeuralSketch2Surf: Fast Neural Surfacing of Unor iented 3D Sketches\n\nWe introduce NeuralSketch2Surf, the first fast and r obust neural surfacing solution, capable of processing arbitrary unoriente d sketches at interactive rates.\n\n\nHongsheng Ye, Anandhu Sureshkumar, a nd Zhonghan Wang (LTCI - Telecom Paris, Institut Polytechnique de Paris); Stefanie Hahmann (University Grenoble Alpes, CNRS, INRIA, Grenoble INP, LJ K); Marie-Paule Cani (LIX-Ecole Polytechnique/CNRS, Institut Polytechnique de Paris); Georges-Pierre Bonneau (University Grenoble Alpes, CNRS, INRIA , Grenoble INP, LJK); and Amal Dev Parakkat (LTCI - Telecom Paris, Institu t Polytechnique de Paris)\n---------------------\nSketch2Arti: Sketch-base d Articulation Modeling of CAD Objects\n\nSketch2Arti is the first sketch- based articulation modeling system for CAD objects. Given a CAD model and simple 2D sketches, it identifies movable parts and predicts motion parame ters. It can also complete missing internal structures revealed by motion, enabling controllable, iterative, and general...\n\n\nYi Yang (University of Edinburgh, Tsinghua University); Hao Pan and Yijing Cui (Tsinghua Univ ersity); Alla Sheffer (University of British Columbia); and Changjian Li ( University of Edinburgh)\n---------------------\nImplicit Surface Compress ion -- with Good Old Discrete Cosine Transform and Motion Compensation\n\n We present a real-time 4D compression framework for volumetric capture usi ng the Truncated Signed Distance Field (TSDF) as coding domain. Combining DCT-based spatial coding with motion-compensated temporal prediction, our training-free GPU-accelerated system enables real-time encoding and decodi ng w...\n\n\nTao Jin, Shengxi Wu, and Tianshu Huang (Carnegie Mellon Unive rsity); Mallesham Dasari (Northeastern University); Srinivasan Seshan (Car negie Mellon University); and Anthony Rowe (Carnegie Mellon University, Bo sch Research)\n---------------------\n4D Human-Scene Reconstruction from L ow-Overlap Captures\n\nWe reconstruct dynamic 4D scenes with multiple peop le from just 4 sparse cameras with minimal overlap. Our method separately handles backgrounds using video generation and humans using body model pri ors, then harmonizes the result with a temporally consistent enhancement s tep, achieving high-quality...\n\n\nMinhyuk Hwang, Sangmin Kim, Seunguk Do , Daneul Kim, and Jaesik Park (Seoul National University)\n--------------- ------\nATGS: Anchored Temporal Gaussian Splatting for Long Volumetric Vid eo Representation\n\nATGS is a volumetric video reconstruction method that handles long sequences with complex motions and supports rendering from a rbitrary viewpoints, enabling immersive experiences.\n\n\nJiahao Wu and Ji e Liang (School of Computer Science, Peking University; Peking University, China); Die Hu (Peking University); Jiayu Yang (Pengcheng Laboratory, Pen gcheng lab); Kaiqiang Xiong (Peking University; Peking University, China); Xiang Li (Pengcheng lab, Peking University); Xiaoyun Zheng (Pengcheng Lab , Pengcheng lab); Chao Wang (Peng Cheng Lab; Peking University, China); an d Ronggang Wang (Peking University, Pengcheng lab)\n---------------------\ nVFAvatar: Feed-Forward 3D Avatar Reconstruction from Casual Image Collect ions\n\nWe introduce Visual-Fusion-Avatar (VFAvatar), a novel end-to-end f eed-forward framework that reconstructs 3D avatars by fusing visual cues f rom casual image collections in just a few seconds. \nVFAvatar integrates a pose-free reconstruction foundation model with a pretrained human genera tion prior in...\n\n\nShuo Huang, Zixuan Wang, Xiaoyu Qin, Shikun Sun, Jia yi Li, and Jia Jia (CS Dept, Tsinghua University)\n---------------------\n High-Fidelity 4D Cloth Capture Pipeline with a Two-Level Pattern\n\nWe pre sent a 4D (spatio-temporal) cloth capture system and a two-level pattern t hat achieves 1mm spatial resolution using only 16 RGB cameras. Our method produces temporally coherent sequences with no self penetration that faith fully capture fine wrinkles and folds even during complex motions with s.. .\n\n\nZiheng Liu (University of Utah); Anka Chen (NVIDIA); and Shu Chen, Yin Yang, Cem Yuksel, and Jenny Lin (University of Utah)\n---------------- -----\nIskra: A System for Inverse Geometry Processing\n\nIskra is a syste m for differentiating through solutions to geometry processing problems.\n It differentiates a broad class of algorithms, exploiting existing fast pr oblem-specific schemes common to geometry processing, including local-glob al and ADMM solvers. It is compatible with machine learning fra...\n\n\nAn a Dodik, Ahmed Mahmoud, and Justin Solomon (Massachusetts Institute of Tec hnology (MIT))\n---------------------\nSharpNet: Enhancing MLPs to Represe nt Functions with Controlled Non-differentiability\n\nWe propose SharpNet, an enhanced MLP architecture that augments the input with a feature funct ion to produce controllable derivative jumps. We show that SharpNet can pr eserve the sharpness in the intended regions while maintaining smoothness elsewhere, and that it outperforms baseline methods in sha...\n\n\nHanting Niu (Institute of Software, Chinese Academy of Sciences; University of Ch inese Academy of Sciences); Junkai Deng (Nanyang Technological University) ; Fei Hou and Wencheng Wang (Institute of Software, Chinese Academy of Sci ences; University of Chinese Academy of Sciences); and Ying He (Nanyang Te chnological University)\n---------------------\nSchrödinger Bridges on Dis cretized Geometric Domains\n\nWe introduce a spatially discrete formulatio n of the Schrödinger bridge problem on meshes and grids that enables struc ture-preserving interpolation between probability distributions. We demons trate our approach across mesh- and grid-based applications, including dis placement interpolation, shape int...\n\n\nLeticia Mattos Da Silva, Mohamm ad Sina Nabizadeh, and Justin Solomon (Massachusetts Institute of Technolo gy (MIT))\n---------------------\nExplicit flows for implicit surfaces\n\n Our paper presents a method for shape-morphing and deformation using an ex act mathematical flow. We introduce a neural architecture guaranteed to en code a flow, leveraging implicit supervision of explicit flows, ensuring t opological consistency and efficient forward/inverse computation without O DE s...\n\n\nCamille Buonomo (CNRS - LIRIS); Julie Digne (CNRS, LIRIS); an d Raphaëlle Chaine (CNRS - LIRIS)\n---------------------\nA Few-Step Gener ative Model on Cumulative Flow Maps\n\nWe propose a unified few-step gener ative framework based on cumulative flow maps for long-range transport in probability space. By connecting instantaneous updates with finite-time tr ansport, it applies broadly to diffusion and flow models, supports few-ste p and one-step generation, requires minimal...\n\n\nZhiqi Li, Duowen Chen, Yuchen Sun, and Bo Zhu (Georgia Institute of Technology)\n--------------- ------\nGenerative Modeling with Orbit-Space Particle Flow Matching\n\nOrb it-Space Geometric Probability Paths (OGPP) is a flow-matching framework f or generative modeling of particle systems that exploits shared geometric structure and uses probability path geometry to encode geometric informati on. It enables high-quality, efficient generation of particle configuratio n...\n\n\nSinan Wang, Jinjin He, Shenyifan Lu, Ruicheng Wang, Greg Turk, a nd Bo Zhu (Georgia Institute of Technology)\n---------------------\nBodyRe Lux: Temporally Consistent Full-Body Video Relighting\n\nBodyReLux is a vi deo diffusion framework for relighting full-body human performances with p hotorealistic, temporally consistent results. Trained on video relighting pairs captured on large LED Sphere using digital bi-pack technology, it en ables precise, dynamic lighting control and generalizes robus...\n\n\nLi M A, Mingming He, Xueming Yu, David George, and Ahmet Tasel (Eyeline Labs); Paul Debevec (Eyeline Labs, Netflix); and Julien Philip (Eyeline Labs)\n-- -------------------\nPixel Cube: Diffusion-based Portrait Video Relighting Through Realistic Lighting Reproduction\n\nPixel Cube is a novel lighting system for realistic lighting emulation and high-speed video relighting d ata acquisition. With a high-quality lighting-aware dataset, we train a di ffusion model for dynamic portrait relighting. Our model produces photorea listic, temporally consistent portrait videos un...\n\n\nYufan Zhang (Geor ge Mason University), Yu Ji (LightThought LLC), Ayo Ajiboye (George Mason University), Rundi Wu (Columbia University), Yu Guo (George Mason Universi ty), Changxi Zheng (Columbia University), and Jinwei Ye (George Mason Univ ersity)\n---------------------\nEasyVFX: Frequency-Driven Decoupling for R esource-Efficient VFX Generation\n\nEasyVFX is a resource-efficient framew ork for high-fidelity VFX synthesis. It leverages frequency-domain decompo sition to disentangle spatial detail and motion dynamics, reducing learnin g complexity. A frequency-aware MoE enables efficient training, while a te st-time adaptation strategy refines unse...\n\n\nYue Ma and Xu Ye (Hong Ko ng University of Science and Technology); Qinghe Wang (Dalian University o f Technology); Yucheng Wang, Hongyu Liu, and Yinhan Zhang (Hong Kong Unive rsity of Science and Technology); Xinyu Wang (Tsinghua University); Yuanpe ng Chen and Shanhui Mo (Independent); Paul Liang and Fangneng Zhan (Massac husetts Institute of Technology (MIT)); and Qifeng Chen (Hong Kong Univers ity of Science and Technology)\n---------------------\nLongE2V: Long-Horiz on Event-based Video Reconstruction, Prediction, and Frame Interpolation w ith Video Diffusion Models\n\nThis work introduces long-horizon event-base d video generation, a challenging task prone to severe temporal inconsiste ncy, and presents LongE2V, a novel approach leveraging video diffusion pri ors to jointly handle video reconstruction, prediction, and frame interpol ation, recovering high-quality vid...\n\n\nCheng-De Fan, Chun-Wei Tuan Mu, Chen-Wei Chang, Chin-Yang Lin, Kun-Ru Wu, Yu-Chee Tseng, and Yu-Lun Liu ( National Yang Ming Chiao Tung University)\n---------------------\nAutoregr essive Modeling of Film with Applications in Video Montage\n\nThis work in troduces an autoregressive transformer designed for video montage, transfo rming raw footage into cinematic sequences. We train it on large corpora o f movies to learn film grammar. Instead of generating pixels, we introduce a footage-constrained decoding algorithm to select shots from the...\n\n\ nMarcelo Sandoval-Castañeda (TTIC, Adobe); Fabian Caba Heilbron (Adobe); S hiry Ginosar (TTIC); Bryan Russell (Adobe); Josef Sivic (Adobe; CTU, Pragu e); Alexei Efros (University of California Berkeley); and Gregory Shakhnar ovich (TTIC)\n---------------------\nRelit-LiVE: Relight Video by Jointly Learning Environment Video\n\nRelit-LiVE produces physically consistent an d temporally stable relighting results without requiring prior knowledge o f the camera pose. It utilizes raw reference images to ensure realistic li ghting effects and introduces a joint prediction framework generating reli ghting videos with frame-level war...\n\n\nWeiqing Xiao (Nanjing Universit y, BAAI); Hong Li (Beihang University, BAAI); Xiuyu Yang (Tsinghua Univers ity); Houyuan Chen (HKUST, BAAI); Wenyi Li (University of Chinese Academy of Sciences China); Tianqi Liu (Huazhong University of Science and Technol ogy); Shaocong Xu (BAAI); Chongjie Ye (The Chinese University of Hongkong) ; Hao Zhao (Tsinghua University, BAAI); and Beibei Wang (Nanjing Universit y)\n---------------------\nGR3EN: Generative Relighting for 3D Environment s\n\nGR3EN enables relighting 3D environments to user-specified lighting c onditions. It provides full control over every light source in a 3D scene — enabling adding new light sources, changing the color of existing ones o r turning them off entirely. Shadows, reflections, and interreflections al l u...\n\n\nXiaoyan Xing (University of Amsterdam); Philipp Henzler (Googl e Research); and Junhwa Hur, Runze Li, Jonathan T. Barron, Pratul P. Srini vasan, and Dor Verbin (DeepMind)\n---------------------\nTaming optimizati on variance in compact neural shading networks\n\nWe present a training st rategy for small networks that improves stability, reducing run-to-run var iance by 88% and average loss by 38%. Multiple instances train in parallel on different batch slices, while weaker ones are periodically removed unt il only the best-performing instance remains, improvin...\n\n\nBenedikt Bi tterli, Petrik Clarberg, Chris Cummings, and Aaron Lefohn (NVIDIA); Steve Marschner (Cornell University, NVIDIA); and Jan Novák, Fabrice Rousselle, Andrea Weidlich, and Tizian Zeltner (NVIDIA)\n---------------------\nMulti -feature Radiance Baking Neural Networks for Instant Volumetric Rendering\ n\nWe propose MRBNN that delivers real-time photorealistic volume renderin g. We leverage reformulation of in-scattering integral and design an effic ient neural representation and sampling strategy. Our method achieves rend ering in a few milliseconds with the ability to handle complex cases such as spat...\n\n\nJiaming Liang (Peking University); Hongliang Yuan (XiaoMi company); and Meng Gai, Guoping Wang, and Sheng Li (Peking University)\n-- -------------------\nBounding Stratified Bernoulli Impulses for Ray Marchi ng Gaussian Process Implicit Surfaces\n\nIn this paper, we present an opti mized algorithm for rendering Gaussian Process Implicit Surfaces (GPISes). By using stratified Bernoulli impulses, we precompute point-level bounds stored in a bound map. We also represent region-level bounds with a spatia l data structure. Combining these bounds, we ...\n\n\nJunjie Chen and Zhim in Fan (Nanjing University); Ling-Qi Yan (Mohamed bin Zayed University of Artificial Intelligence); Junqiu Zhu (Shandong University); Yanwen Guo (Na njing University); Kun Zhou (State Key Lab of CAD&CG, Zhejiang University; Institute of Hangzhou Holographic Intelligent Technology); and Jie Guo (N anjing University)\n---------------------\nAdaptive Ray Marching for Rende ring Gaussian Process Implicit Surfaces\n\nThis paper accelerates Gaussian Process Implicit Surface (GPIS) rendering by introducing an online algori thm for incrementally sampling conditional Gaussian distributions along ra ys and an adaptive ray marching scheme that minimizes steps. Together, the se techniques reduce rendering error by up to 4...\n\n\nZhiqian Zhou (Univ ersity of California Irvine), Dario Seyb (Valve), and Shuang Zhao (Univers ity of Illinois Urbana-Champaign)\n---------------------\nToward Embodied Collective Memory by Remediating Vernacular Photographs into Spatial Exper ience\n\nDecember 2022 Beijing remediates vernacular photographs of China' s COVID-19 reopening into immersive VR experiences that evoke presence, em pathy, and reflection. Through an embodied artistic and technical practice , the work opens new perspectives on representing collective memory, disti nct from trad...\n\n\nMinzhi Shen, Mengtai Zhang, Yuan Zhang, Gen Mai, Zhi gang Wang, and Wenbo Wang (Tsinghua University)\n---------------------\nA Walled City: An Interactive AI Artwork for Decentralized Urban Memories an d Collective Presence\n\nIt proposes a new artistic model of collective vi rtual presence grounded in historical urbanism and decentralized networks. Drawing from the spatial logic and aesthetic influence of the Kowloon Wal led City, the project provides a speculative approach to participatory spa tial memory systems through i...\n\n\nWeidi Zhang (Arizona State Universit y), lijiaozi Cheng (The University of Sheffield), Yi Wang (Midjourney), an d Jieliang Luo (Minus AI)\n---------------------\nUnfinished Decay: Animat ing Historical Images from Gigapixel Digitizations via Fragment-Based AI W orkflows\n\nUnfinished Decay explores AI-assisted animation of historical images using gigapixel digitizations. Through a fragment-based workflow, m icroscopic surface details are animated to foreground materiality and deca y. The project reframes AI as an interpretive collaborator, offering a mat erial-sensitive ...\n\n\nJuergen Hagler (University of Applied Sciences Up per Austria; Ars Electronica, Linz) and Celine Pham (University of Applied Sciences Upper Austria; ANIMA PLUS, Linz)\n---------------------\nLarge-S cale Photogrammetric Documentation of St. John's Co-Cathedral: A Workflow for Cultural Heritage Preservation\n\nCultural heritage sites face threats from climate change, conflict, and degradation. Digital documentation has become a preservation necessity, demonstrated by Notre-Dame's 2019 fire. However, academic photogrammetry research remains disconnected from practi cal implementation by heritage institution...\n\n\nMatthew Kenely (SeyTrav el Ltd); Mark Bugeja (University of Malta); Andre Grima, Peter Pullicino, and Matthew Pullicino (Stargate Studios); and Dylan Seychell (University o f Malta)\n---------------------\nMeshFEM: A Block-accelerated Solver for N onlinear Finite Elements\n\nWe present a high-performance Newton-type solv er designed for the nonlinear finite element problems common in simulation and geometry processing. By leveraging the inherent block structure of sp atially discretized systems across all solver phases, we achieve significa nt speedups over existing direct...\n\n\nHaleh Mohammadian and Xinzhuo Hu (University of California Davis), Roi Poranne (University of Haifa), and J ulian Panetta (University of California Davis)\n---------------------\nFas t Sparse Matrix Permutation for Mesh-Based Direct Solvers\n\nWe present a fast sparse matrix permutation algorithm tailored to mesh-based direct sol vers. By exploiting mesh structure through patch-based nested dissection, our method reduces permutation overhead while preserving high-quality orde rings, leading to substantial end-to-end speedups in sparse Chole...\n\n\n Behrooz Zarebavani (University of Toronto); Ahmed Mahmoud and Ana Dodik (M assachusetts Institute of Technology (MIT)); Changcheng Yuan (Texas A&M Un iversity); Serban Porumbescu and John Owens (University of California Davi s); Maryam Mehri Dehnavi (University of Toronto, NVIDIA); and Justin Solom on (Massachusetts Institute of Technology (MIT))\n---------------------\nS tatistical Gradient Filtering for Geometry Optimization Under Limited Obse rvations\n\nWe present a novel technique that leverages statistical shape space for accurate and robust 3D geometry optimization under limited obser vations, such as non-rigid registration on partial scans and single-view i nverse rendering.\n\n\nWonjong Jang, Gwangjin Ju, and Seungyong Lee (POSTE CH)\n---------------------\nGMT: A Geometric Multigrid Transformer Solver for Microstructure Homogenization\n\nGMT is a Geometric Multigrid Transfor mer that significantly accelerates microstructure homogenization. Embeddin g Transformers into the multigrid hierarchy to capture multi-scale physica l interactions delivers a 160x speedup over SOTA GPU solvers while maintai ning engineering-grade accuracy at high r...\n\n\nYu Xing (Shandong Univer sity), Yang Liu (Microsoft Research Asia), Tianyang Xue (The University of Hong Kong), and Lin Lu (Shandong University)\n---------------------\nRobu st Biharmonic Skinning Using Geometric Fields\n\nA mesh-free approach for computing bounded biharmonic skinning weights. Our method works well even in presence of severe mesh quality degradation and can incorporate artisti c control via weight-painting. It relies on a differentiable field and inc orporates boundary handling via hardware accelerated ...\n\n\nAna Dodik, V incent Sitzmann, and Justin Solomon (MIT CSAIL); Oded Stein (University of Southern California); and Ana Dodik\n---------------------\nJGS2-GQ: Trai ning-free 2nd Jacobi with Gaussian Quadrature\n\nJGS2-GQ replaces data-dri ven Cubature with training-free Gaussian Quadrature, enabling stable, high -resolution GPU simulation. It handles frictional contact, maintains Newto n-like convergence, and outperforms JGS2 by 50% on novel deformations with out training.\n\n\nDewen Guo, Zixuan Lu, Zhiyong He, and Yuqi Meng (Univer sity of Utah); Bohan Wang (National University of Singapore); Lei Lan and Weiwei Xu (State Key Lab of CAD and CG, Zhejiang University); Chenfanfu Ji ang (UCLA); and Yin Yang (University of Utah)\n---------------------\nDich optic Foveation\n\nThis paper presents a psychophysical study of dichoptic foveation, applying blur to one eye and sharpening to the other, in virtu al reality. A perceptual model of interocular frequency differences across the visual field is developed and applied to enhance foveated rendering q uality in VR headsets.\n\n\nHenry Kam, Colin Groth, Jenna Kang, Pratham Sa raf, Qi Sun, and Kenneth Chen (New York University)\n--------------------- \nReality Check: How Avatar and Face Representation Affect the Perceptual Evaluation of Synthesized Gestures\n\nThis paper investigates how avatar a nd face representation influence the perceptual evaluation of synthesized co-speech gestures. Across three controlled experiments, it shows that vis ual presentation systematically biases judgments of motion quality and pro vides practical recommendations for fairer...\n\n\nHaoyang Du (Technologic al University Dublin), Yinghan Xu and John Dingliana (Trinity College Dubl in), Brian Keegan (Technological University Dublin), Rachel McDonnell (Tri nity College Dublin), and Cathy Ennis (Maynooth University)\n------------- --------\nDynamic Visual Dominance in Stereoscopic Foveation\n\nThis work investigates visual dominance as a dynamic property across the FOV. Using a blur detection task, we confirm dynamic dominance in the fovea and calib rate monocular dominance in the periphery. The resulting gaze-independent peripheral dominance map can be integrated into different rendering ...\n\ n\nDaniel Jiménez Navarro (Max Planck Institute for Informatics); Colin Gr oth, Kenneth Chen, and Qi Sun (New York University); Karol Myszkowski and Hans-Peter Seidel (Max Planck Institute for Informatics); and Ana Serrano (Universidad de Zaragoza)\n---------------------\nA Two-Millisecond Passth rough Headset for Perceptual Studies\n\nWe present a nearly perspective-co rrect video passthrough headset with just 2 milliseconds of end-to-end lat ency. We use this headset to study perceptual requirements for low-latency passthrough in head-mounted displays.\n\n\nEric Penner (Reality Labs Rese arch, Meta); Josephine D'Angelo (Reality Labs Research, Meta; University o f California, Berkeley); Clinton Smith and Nathan Matsuda (Reality Labs Re search, Meta); and Neethan Siva and Phillip Guan (Reality Labs, Meta)\n--- ------------------\nStyleID: A Perception-Aware Dataset and Metric for Sty lization-Agnostic Facial Identity Recognition\n\nStyleID is a perception-a ware dataset and evaluation framework for measuring facial identity consis tency under artistic stylization. By aligning identity similarity with hum an judgments across styles and strengths, it improves robustness beyond na tural-photo face encoders and better supports stylize...\n\n\nKwan Yun, Ch angmin Lee, AYeong Jeong, Youngseo Kim, Seungmi Lee, and Junyong Noh (Visu al Media Lab, KAIST)\n---------------------\nQuantifying reality of ultra- realistic 3-D displays - the effect of resolution and contrast\n\nWe built an ultrarealistic display system capable of reproducing 3D scenes with su ch fidelity that they are indistinguishable from their real counterparts. Then, we used the display to measure how reducing resolution and contrast affects perceived realism.\n\n\nJoseph Gerard March and Dounia Hammou (Uni versity of Cambridge), Simon J. Watt (Bangor University), and Rafał K. Man tiuk (University of Cambridge)\n---------------------\nIt’s Not Just a Pha se: Creating Phase-Aligned Peripheral Metamers\n\nWe present a method for enhancing the perceived quality of foveated images by extrapolating local image statistics, such as intensity, orientation and especially phase, fro m coarse to fine scales. In particular, we demonstrate that correct phase alignment across space and scales is crucial for achie...\n\n\nSophie Kerg aßner and Piotr Didyk (Università della Svizzera Italiana)\n-------------- -------\nRole-Aware Virtual Agents for Navigational Interaction guided by a Multimodal Large Language Model\n\nWe present a role-aware virtual agent navigation system that generates consistent, role-aligned behaviors. Usin g multimodal large language models (MLLMs), our approach interprets scene context, user state, and role-context instructions to produce navigation d ecisions and stylized paths, enabling age...\n\n\nMinyoung Kim (George Mas on University), Changyang Li (Goertek Alpha Labs), Cuong Nguyen (Adobe Res earch), and Lap-Fai Yu (George Mason University)\n---------------------\nE goRelight: Egocentric Human Capture and Illumination Recovery for Relighta ble and Photoreal Avatar Rendering\n\nTo enable seamless mixed-reality tel epresence in the wild, we present EgoRelight, a robust human performance c apture and HDR illumination recovery approach using a single portable head -mounted device. This dual capability drives the photorealistic and religh table full-body 3D avatars, blending users...\n\n\nJianchun Chen (Max Plan ck Institute for Informatics; Saarbrücken Research Center for Visual Compu ting, Interaction and AI); Yinda Zhang, Rohit Pandey, and Thabo Beeler (Go ogle); and Marc Habermann and Christian Theobalt (Max Planck Institute for Informatics; Saarbrücken Research Center for Visual Computing, Interactio n and AI)\n---------------------\nVoronoi Rooms: Dynamic Visibility Modula tion of Overlapping Spaces for Telepresence\n\nWe propose a multi-user Mix ed Reality telepresence system that aligns rooms to maximize shared space while visualizing non-shared spaces. Using Voronoi-based dynamic visibilit y modulation based on user proximity, our approach conveys spatial context of each user's environment, enabling seamless inte...\n\n\nTaehei Kim, Ji hun Shin, Hyeshim Kim, Hyuckjin Jang, Jiho Kang, and Sung-Hee Lee (KAIST) and Taehei Kim\n---------------------\nAtmospheric Haptics: Rendering Airf low and Temperature as Tactile Fields for Interactive Experiences in Virtu al Reality\n\nAtmospheric Haptics treats airflow and temperature as progra mmable, non-contact tactile fields for VR. Four perceptual primitives—Inte nsity, Temperature, Spatiality, and Dynamics—drive a three-layer rendering pipeline implemented in a handheld prototype. A user study confirms enhan ced re...\n\n\nShengyi Zhan (School of software technology, Zhejiang Unive rsity; Zhejiang Key Laboratory of Intelligent Systems and Equipment for D igital Creativity,Hangzhou,Zhejiang); Xueting Wu (College of computer scie nce and technology, Zhejiang University,Hangzhou, Zhejiang; Zhejiang Key L aboratory of Intelligent Systems and Equipment for Digital Creativity,Han gzhou,Zhejiang); and Ning Zou (College of computer science and technology, Zhejiang University,Hangzhou, Zhejiang; Future Design Laboratory, Innovat ion Centre of Yangtze River Delta, Zhejiang University，Jiaxing,Zhejiang)\n ---------------------\nRetrofitting Existing 3D Objects with Surface-Confo rming Capacitive Sensing\n\nWe present a computational fabrication pipelin e for retrofitting 3D objects with real-time multi-touch sensing without m odifying their interiors. From a 3D scan, our method routes and optimizes surface-conforming electrode layouts under geometric, fabrication, and har dware constraints. We guide asse...\n\n\nAndela Ilic, Junpeng Gao, Zhipeng Li, Yijing Jiang, Rachel Schuchert, and Manuel Meier (ETH Zürich); Philip p Herholz (Independent Researcher); and Christian Holz (ETH Zürich)\n----- ----------------\nGeneralized Aberrations for Processing-Aware Optical Des ign\n\nOptimal imaging performance requires designing optics with the down stream processing in the loop—especially as AI pipelines grow in importanc e. Yet the scalar nature of loss functions in the processing-aware setting breaks industry-standard lens design solvers. We generalize classical ray aber...\n\n\nGeoffroi Côté (Keysight Technologies, Princeton University); Ethan Tseng and Felix Heide (Princeton University); and Geoffroi Côté\n-- -------------------\nProbe-based Walk on Spheres for Efficient Path Reusin g\n\nA Walk on Spheres-based Monte Carlo PDE solver that reuses full rando m walk paths via fixed probes, achieving lower variance and faster converg ence on mixed-boundary Laplace, Poisson, and screened Poisson problems.\n\ n\nWanchao Huang, Yutian Zhu, Qing Fang, and Ligang Liu (University of Sci ence and Technology of China)\n---------------------\nGradient Domain Reco nstruction for Monte Carlo PDE Solvers\n\nWe present a gradient-domain Mon te Carlo framework for solving Poisson equations on complex domains. The m ethod directly estimates solution differences between query points and rec onstructs the final solution efficiently, reducing variance and improving convergence without introducing additional bia...\n\n\nJiaqi Wu and Xuejun Hu (CS Dept, Tsinghua University); Shuang Zhao (University of Illinois Ur bana-Champaign); and Kun Xu (CS Dept, Tsinghua University)\n-------------- -------\nMonte Carlo PDE Solvers for Nonlinear Radiative Boundary Conditio ns\n\nThis paper extends Monte Carlo PDE solvers to handle nonlinear radia tive boundary conditions using a Picard-style fixed-point iteration framew ork. It introduces a heteroscedastic regression-based denoising method for boundary estimates and demonstrates the approach on heat radiation simula tions with...\n\n\nAnchang Bao (School of Software and BNRist, Tsinghua Un iversity); Enya Shen (School of Software and BNRist, Tsinghua University; Haihe Lab of ITAI); and Jianmin Wang (School of Software and BNRist, Tsing hua University)\n---------------------\nWalk on Decomposed Subdomains: A H ybrid Monte Carlo–Deterministic Solver for Elliptic PDEs\n\nWe introduce a hybrid Monte Carlo-deterministic solver for elliptic PDEs on complex geom etries. Our method decomposes the domain into simple subdomains, estimates local solution operators with Monte Carlo and couples them via a sparse l inear system, delivering low-variance solutions orders of magnit...\n\n\nC lément Jambon, Mohammad Sina Nabizadeh, and Mina Konaković Luković (Massac husetts Institute of Technology (MIT))\n---------------------\nWalking on Spheres and Talking to Neighbors: Variance Reduction for Laplace's Equatio n\n\nWe introduce a novel variance reduction scheme for Walk on Spheres wi th Laplace's equation, demonstrate performance, and provide analytic guara ntees on performance.\n\n\nMichael Czekanski (Cornell University), Benjami n Faber (University of Wisconsin-Madison), Margaret Fairborn (Columbia Uni versity), Adelle Wright (University of Wisconsin-Madison), and David Binde l (Cornell University)\n---------------------\nNeural Quadrature Rule and Autoregressive Adaptive Sampling\n\nLearnable adaptive sampler and integra tor using Transformer.\n\n\nHaolin Lu, Liwen Wu, Zimo Wang, Tzu-Mao Li, an d Ravi Ramamoorthi (University of California San Diego)\n----------------- ----\nEchoes of the Prior: A Computational Phenomenology of Forgetting\n\n Echoes of the Prior visualizes entropy within neural networks, turning mac hine forgetting into an interactive 4D experience. Unlike traditional CG s imulating physical reality, this work reveals the aesthetics of algorithmi c degradation, transforming technical failure into a new artistic medium b ridg...\n\n\nGege Gao (University of Tuebingen, ETH Zürich); Bernhard Schö lkopf (Max Planck Institute for Intelligent Systems, ETH Zürich); and Andr eas Geiger (University of Tübingen, Tübingen AI Center)\n----------------- ----\nElectrospun Fields: 3D Nano-Fiber Material Computation as Design Met hod\n\nElectrospun Fields advances SIGGRAPH’s interest in material computa tion by framing electric fields as programmable boundary conditions and fi ber deposition as a material rendering of invisible forces. It contributes a reproducible UR20 robotic workflow for non-planar electrospinning, enab ling ...\n\n\nJustin Wan, Ayah Mahmoud, and Sergio Mutis (Massachusetts In stitute of Technology (MIT); Critical Matter Group, MIT Media Lab); Avanti ka Velho and Zhiyan Xing (Harvard University; Critical Matter Group, MIT M edia Lab); and Behnaz Farahi (Media Lab, Massachusetts Institute of Techno logy (MIT); Critical Matter Group, MIT Media Lab)\n---------------------\n Computational Visual Semiotics: The “L” Programming Language by Frieder Na ke\n\nThis study investigates Frider Nake's 1970 programming language "L" to explore the intersection of code and art. By examining how "L" translat es programming languages into imaginative and intellectual explorations, w e gain insights into the cultural dimensions of code and its potential as a medium f...\n\n\nAndres Burbano (Universitat Oberta de Catalunya, UOC)\n ---------------------\nTo Perform/ To Live: Decolonizing of Digital Music Instruments and Feminism with Human-AI Co-Created Embodied Experience of D aily Objects\n\nTo Perform/To Live brings decolonial, China-situated femin ist perspectives into sonic art. It repurposes domestic objects and routin es as interface logic, making gendered labor audible without romanticizing it and challenging interaction defaults. Performances are sensed and shap ed live, then captur...\n\n\nShumeng Zhang (The Hong Kong University of Sc ience and Technology (Guangzhou), University of Trento); Shiqi Lin (The Ho ng Kong University of Science and Technology (Guangzhou)); Tiancheng Liu ( The Hong Kong University of Science and Technology (Guangzhou), University of Amsterdam); Mingming Fan (The Hong Kong University of Science and Tech nology (Guangzhou), The Hong Kong University of Science and Technology); a nd Raul Masu (Conservatory of Trento, The Hong Kong University of Science and Technology (Guangzhou))\n---------------------\nData Materialization: Principles and Practices in Artistic Research\n\nDefining data materializa tion is relevant because it is an emerging practice that like SIGGRAPH lie s at the intersection of art and technology. Furthermore, as data collecti on explodes in every facet of society, data materialization offers a new f orm of epistemic translation--the transformation of a...\n\n\nCourtney Sta rrett (Texas A&M University) and Susan Reiser (University of North Carolin a Asheville)\n---------------------\nPrinting the Underdetermined: Materia lizing Multi-solutionness in Figurative Paintings\n\nAs generative models enter graphics pipelines, we need ways to see what gets assumed, invented, or discarded when an image becomes a 3D scene. Figurative paintings are a clear test case. Our workflow keeps this non-uniqueness explicit by sampl ing plausible completions, reconstructing each via 3D Gau...\n\n\nYutao Mi ng, Teng Xu, Youjia Wang, Yunyang Liu, and Fengmin Yang (ShanghaiTech Univ ersity, Crysta AI); Fuqiang Zhao (Crysta AI); and Jingyi Yu and Yanjun Zho u (ShanghaiTech University)\n---------------------\nInverse Rendering for Discrete X-Ray Computed Tomography\n\nWe propose a gradient-based discrete tomography method that models each voxel as a probability distribution ov er known materials, minimizing an object-space loss for projection consist ency. Inspired by inverse rendering, it outperforms classical methods, sup ports scattering, and excels in sparse and...\n\n\nLovro Nuic and Ziyi Zha ng (Ecole Polytechnique Fédérale de Lausanne), Korbinian Sager (Carl Zeiss AG), and Wenzel Jakob (Ecole Polytechnique Fédérale de Lausanne)\n------- --------------\nAmbient-robust Inverse Rendering using Robot-assisted RGB- NIR Imaging\n\nThis paper presents an ambient-robust inverse rendering met hod using robot-assisted RGB–NIR imaging. By combining multi-view RGB imag es with NIR flash images, we reconstruct accurate geometry and reflectance under multiple lighting conditions. We also introduce a robot-assisted im aging system ...\n\n\nHoon-Gyu Chung, Jinnyeong Kim, Hyunwoo Kang, and Seu ng-Hwan Baek (POSTECH)\n---------------------\nEAG-PT: Emission-Aware Gaus sians and Path Tracing for Diffuse Indoor Scene Reconstruction and Editing \n\nEAG-PT reconstructs indoor scenes with 2D Gaussians for editable diffu se global illumination. By separating emissive and non-emissive components and combining efficient reconstruction with path tracing, it enables more natural, physically consistent scene editing than radiance-field methods, while a...\n\n\nXijie Yang (Zhejiang University, Shanghai Artificial Inte lligence Laboratory); Mulin Yu (Shanghai Artificial Intelligence Laborator y); Changjian Jiang (Zhejiang University); Kerui Ren (Shanghai Jiao Tong U niversity, Shanghai Artificial Intelligence Laboratory); Tao Lu and Jiangm iao Pang (Shanghai Artificial Intelligence Laboratory); Dahua Lin (The Chi nese University of Hong Kong, Shanghai Artificial Intelligence Laboratory) ; Bo Dai (The University of Hong Kong, Feeling AI); and Linning Xu (The Ch inese University of Hong Kong, Shanghai Artificial Intelligence Laboratory )\n---------------------\nRadiance Caching for Differentiable Path Tracing \n\nWe introduce a spatially blended radiance-caching framework for differ entiable path tracing that improves speed and robustness in material recov ery under unknown lighting. \nBy jointly optimizing cache, materials, and blending, the method produces physically meaningful materials that general ize to ne...\n\n\nZiyi Zhang (Google, École Polytechnique Féderale de Laus anne (EPFL)); Delio Vicini, Sebastian Winberg, and Stephan Garbin (Google) ; and Wenzel Jakob (Ecole Polytechnique Fédérale de Lausanne)\n----------- ----------\nSample Matching for Joint Extinction Gradient Estimation in Di fferentiable Volume Rendering\n\nWe propose sample matching, a variance re duction principle for differentiable volume rendering. By jointly estimati ng scattering and transmittance gradient components at shared sample posit ions, our method reduces gradient variance by up to 80%, significantly imp roving convergence speed and reconstr...\n\n\nRuihan Yu (The University of Tokyo, Tsinghua University); Yu-Chen Wang (University of California Irvin e); Jingwang Ling (University of Illinois Urbana-Champaign); Feng Xu (Tsin ghua University); and Shuang Zhao (University of Illinois Urbana-Champaign )\n---------------------\nRobust Computation of Boundary Path Integrals Us ing Kernel-Density Estimation\n\nWe present a simple, robust, and consiste nt estimator for boundary path integrals in physics-based differentiable r endering. By reformulating boundary integration with kernel-density estima tion, our method avoids fragile measure-zero sampling and costly reparamet erization. It matches finite-differen...\n\n\nPeiyu Xu (University of Illi nois Urbana-Champaign); Lifan Wu and Benedikt Bitterli (NVIDIA); Ravi Rama moorthi (NVIDIA, University of California San Diego); and Shuang Zhao (Uni versity of Illinois Urbana-Champaign)\n---------------------\nLifting Line s and Tone: Image-Space Stylization in Path-Space\n\nWe present a framewor k for lifting image-space stylizations into path-space rendering, and demo nstrate it on line and tone rendering. By constructing geometric mappings between image and path space, our method preserves line and tone structure under complex light transport, enabling coherent, expres...\n\n\nRex West (Aoyama Gakuin University), Sayan Mukherjee (The University of Tokyo), an d Yonghao Yue (Aoyama Gakuin University)\n---------------------\nAbstracti on in Style: Beyond Texture and Color\n\nArtistic styles often involve sha pe reinterpretation, not merely a change of surface appearance. Convention al style transfer methods miss this deeper abstraction behavior. We introd uce Abstraction in Style (AiS), a two-stage generative framework that lear ns and transfers style at shape level, beyond...\n\n\nMin Lu and Yuanfeng He (Shenzhen University); Anthony Chen (Peking University); Jianhuang He a nd Pu Wang (Shenzhen University); Daniel Cohen-Or (Tel Aviv University, Sh enzhen University); and Hui Huang (Shenzhen University)\n----------------- ----\nProgressive Photorealistic Simplification\n\nWe introduce a method f or simplifying photorealistic scenes by progressively removing elements ba sed on semantic importance. Our framework systematically removes elements in a hierarchical order, using generative inpainting to maintain photoreal ism throughout the process. This generates plausible ph...\n\n\nAdi Rosent hal (Reichman University, Google); Dana Berman (Google); Yedid Hoshen (Heb rew University of Jerusalem, Google); and Ariel Shamir (Reichman Universit y, Google)\n---------------------\nGimmBO: Interactive Generative Image Mo del Merging via Bayesian Optimization\n\nGimmBO helps users interactively explore combinations of customized image generation styles to create desir ed results. It replaces manual slider tuning with a tailored Preferential Bayesian Optimization approach, outperforming alternatives in simulations and user studies. It further extends to combi...\n\n\nChenxi Liu and Selen a Ling (University of Toronto) and Alec Jacobson (University of Toronto, V ector Institute)\n---------------------\nSemantic-Structural Alignment for Generative Pictorial Charts\n\nWe introduce a generative framework that t ransforms traditional statistical graphics into memorable, engaging pictor ial charts. It utilizes structural alignment to anchor spatial layouts and semantic alignment to transfer expressive textures. This ensures the resu lting visuals are artistically compe...\n\n\nZhida Sun, Yulin Zhang, Zheng Gu, and Min Lu (Shenzhen University); Bongshin Lee (Yonsei University); D aniel Cohen-Or (Tel Aviv University, Shenzhen University); and Hui Huang ( Shenzhen University)\n---------------------\nGradient Descent in the ALPS: Abstracted Low-Poly Stylization and Fabrication\n\nWe generate 2D low-pol y meshes that abstract images using score distillation sampling. We focus on ensuring the validity of the output mesh, quantized palette colors and semantic similarity to the input. The resulting vector images are easy to edit and fabricated as mosaics, embroidery, crocheting, ...\n\n\nRuben Wie rsma (ETH Zurich, Adobe Research) and Alexandre Binninger, Peizhuo Li, Ann ika Oehri, Aviv Segall, Tanguy Magne, Danielle Luterbacher, Marcel Padilla , Jing Ren, and Olga Sorkine-Hornung (ETH Zurich)\n---------------------\n Divide and Truncate: A Penetration and Inversion Free Framework for Couple d Multi-physics System\n\nDivide and Truncate (DAT) is a unified framework for penetration-free coupling of multi-physics systems, including rigid b odies, volumetric soft bodies, thin shells, rods, and animated objects. DA T eliminates artificial damping and deadlock issues. Material- and solver- agnostic, DAT plugs into any i...\n\n\nAnka H. Chen (NVIDIA), Jerry Hsu (Z eroMatter), Youssef Ayman (The American University in Cairo), and Miles Ma cklin (NVIDIA)\n---------------------\nRobust and Efficient Penetration-Fr ee Elastodynamics without Barriers\n\nWe introduce a barrier-free optimiza tion framework for nonpenetration elastodynamic simulation using a smooth augmented Lagrangian formulation and efficient active-set exploration. Unl ike IPC barriers, our method avoids TOI locking, maintains a compact antic ipatory constraint set, and delivers faste...\n\n\nJuntian Zheng and Zhaof eng Luo (Carnegie Mellon University); Minchen Li (Carnegie Mellon Universi ty, Genesis AI); and Juntian Zheng\n---------------------\nHigh-Order Cont inuous Geometrical Validity\n\nWe propose a conservative algorithm to test the geometrical validity of polynomial finite elements as they deform lin early in time. In elastodynamic simulation, our algorithm guarantees that the system remains physically valid during the entire trajectory, not only at discrete time steps, even when ...\n\n\nFederico Sichetti (Università di Genova); Zizhou Huang (New York University, Roblox); Marco Attene (IMAT I CNR Genova); Denis Zorin (New York University); Enrico Puppo (Università di Genova); Daniele Panozzo (New York University); and Federico Sichetti\ n---------------------\nFloating-Point Robustness in Parametric Surface Co ntinuous Collision Detection: From Algorithm to Benchmarking\n\nWe present solutions to the floating-point–induced decision errors arising in CCD fo r parametric surfaces, including an error-resistant CCD method and a data construction pipeline with known GT for algorithm evaluation under floatin g-point perturbations. Experimental results demonstrate that t...\n\n\nXuw en Chen (SIST, Peking University); Junyu Wang (Southeast University); Chen g Yu (SIST, Peking University); Xingyu Ni (The University of Hong Kong); M eng Zhang (Nanjing University of Science and Technology); Bin Wang (Indepe ndent); Mengyu Chu (Peking University, State Key Laboratory of General Art ificial Intelligence); and Baoquan Chen (Peking University)\n------------- --------\nAGIPC: Adaptive In-Solve Algebraic Coarsening for GPU IPC\n\nWe propose a GPU-friendly adaptive algebraic coarsening method guided by Gree n strain increments and combined with affine embedding to avoid topologica l changes and irregular memory access, yielding the first fully GPU-optimi zed adaptive IPC solver with up to 3× speedup over state-of-the-art GPU IP C...\n\n\nXuan Wang (The University of Hong Kong (HKU)); Zhaofeng Luo (CMU ); Minchen Li (Carnegie Mellon University, Genesis AI); and Taku Komura an d Kemeng Huang (The University of Hong Kong (HKU))\n---------------------\ nYASPS: A Symbolic Framework for Extensible, High-Performance IPC Simulati on\n\nYASPS is a programming system for automatically computing derivative s of your physical simulation energy from symbolic definitions, specializi ng to contact-rich scenarios. YASPS captures the structure of the simulati on data, parameterization, and sparsity, and produces modular and efficien t GPU code...\n\n\nXuan Tang (University of California San Diego); Kemeng Huang (The University of Hong Kong (HKU)); Gilbert Bernstein (University o f Washington); Minchen Li (Carnegie Mellon University, Genesis AI); and Tz umao Li (University of California San Diego)\n---------------------\nConst ant Mean Curvature Surfaces from Discrete Harmonic Maps\n\nWe present a si mple discretization of constant mean curvature surfaces based on the class ical observation that their Gauss maps are harmonic. Our construction is e lementary---requiring only discrete Dirichlet energy minimization and a Po isson solve---yet it exactly mirrors this aspect of the smooth ...\n\n\nYo usuf Soliman (Side Effects Software Inc); Peter Schröder (University of Bo nn, California Institute of Technology); and Ulrich Pinkall (Technical Uni versity of Berlin)\n---------------------\nSynchronizing Fields with Singu larities\n\nWe cast a variety of field-processing problems into the langua ge of synchronization, unlocking a unified algorithm based on semidefinite relaxation. Our method produces better optima, enables new boundary condi tions, and comes with a geometric interpretation in terms of lifting topol ogical defects (...\n\n\nNatalia Pacheco-Tallaj (Massachusetts Institute o f Technology (MIT)); Mattéo Couplet and Edward Chien (Boston University); and David Palmer (Institute of Science and Technology Austria (ISTA), Harv ard University)\n---------------------\nGlobal Discrete Optimization for Q uad-Dominant Mesh Reduction\n\nWe propose a global optimization framework for simplifying quad-dominant meshes used in game and animation production . By formulating edge selection as an Integer Linear Programming problem w ith partial poly-chord segmentation and curvature-aware constraints, our m ethod preserves clean edge flows an...\n\n\nYuzhe Luo (Sate Key Lab of CAD &CG, Zhejiang University; LIGHTSPEED, China); Jingchen Gao (Hangzhou Dianz i University); Zherong Pan (META); Kui Wu (LIGHTSPEED); Xuebo Ji (The Univ ersity of Hong Kong); Xiaogang Jin (State Key Laboratory of CAD & CG, Zhej iang University); and Xifeng Gao (LIGHTSPEED)\n---------------------\nSQua dGen: Generating Simple Quad Layouts via Chart Distance Fields\n\nSQuadGen introduces a diffusion-based approach to simple quad layout generation ba sed on Chart Distance Fields, a continuous representation that encodes dis crete quad structures on surfaces. This formulation avoids direct mesh con nectivity prediction and enables the synthesis of simple, artist-frien...\ n\n\nYou-Kang Kong (Tsinghua University, Microsoft Research Asia); Yang Li u (Microsoft, Microsoft Research Asia); Yue Dong (Microsoft Research Asia) ; Xin Tong (Anuttacon); and Heung-Yeung Shum (International Digital Econom y Academy, Tsinghua University)\n---------------------\nSurface Power Diag rams for Knit Singularity Placement\n\nWe present an algorithm for knit si ngularity placement that leverages a generalization of power diagrams to s urfaces. By optimizing singularity positions in a global fashion, we achie ve faster and more optimal placement, allowing for denser knit graphs. Our framework robustly produces helix-free kni...\n\n\nRahul Mitra, Mattéo Co uplet, and Ruichen Liu (Boston University); Jonathan Ng, Ruza Markov, and William Batara Jeremiah Samosir (VARIANT3D); Megan Hofmann (Northeastern U niversity); and Edward Chien (Boston University)\n---------------------\nL earning Sparse Singularities for Cross Field Design\n\nProblem-aware quad mesh design is challenging due to the choice of neural representation. We propose a two-stage strategy: we first learn sparse singularities, follwed by an analytical step to convert the singularities into cross field and t he final quad. Thanks to the hybrid neural-analytical strat...\n\n\nXiaohu Zhang, Hujun Bao, and Jin Huang (Zhejiang University) and Xiaohu Zhang\n- --------------------\nCo-Optimization of Structure and Manufacturable Semi -Continuous Layers for Laminated Composites\n\nA field-based computational framework for designing fabric-reinforced composites with optimized topol ogy and manufacturable semi-continuous layers, producing laminates up to 4 3.8% stiffer than conventional planar-ply designs.\n\n\nTao Liu, Aoran Lyu , Yongxue Chen, Yu Jiang, and Michael James Petty (The University of Manch ester) and Charlie C.L. Wang (University of Manchester)\n----------------- ----\nDiceplay: A Modular Canvas for Physical Image Composition\n\nDicepla y is a modular, physical, low-resolution canvas that uses identical geomet ric dice to create abstract visual compositions. We propose a gradient-bas ed optimization approach to generate Diceplay configurations from text pro mpts by using a novel shape grammar and relaxing the purely discrete de... \n\n\nMilin Kodnongbua (University of Washington); Zihan Jack Zhang, Shish i Xiao, Vivian Li, and Heather Robertson (Brown University); Rulin Chen (B eijing Normal-Hong Kong Baptist University); and David Laidlaw and Adriana Schulz (Brown University)\n---------------------\nA Unified Homogenizatio n Framework for Straight- and Curved-Crease Origami Materials\n\nWe presen t a unified computational framework for numerical homogenisation of straig ht- and curved-crease origami materials with periodic crease patterns.\n\n \nMingjie Li and Juan Montes Maestre (ETH Zürich); Emilien Ganier (EPFL, C entre National de la Recherche Scientifique - Laboratoire d'informatique d e l'École Polytechnique (LIX)); Klara Mundilova and Mark Pauly (EPFL); and Bernhard Thomaszewski (ETH Zürich)\n---------------------\nmpcGear: Multi -Point Conjugation Gear Mechanisms\n\nThis work presents a new class of ge ar mechanisms, called multi-point conjugation gear mechanisms, as well as computational techniques to modeling these gear mechanisms for exactly gen erating user-specified 3D motions under external loads.\n\n\nKe Chen (Univ ersity of Science and Technology of China, Singapore University of Technol ogy and Design); Joshua John Shi Kai Lee (Singapore University of Technolo gy and Design); Jianmin Zheng (Nanyang Technological University); Ligang L iu (University of Science and Technology of China); and Peng Song (Singapo re University of Technology and Design)\n---------------------\nKinematic Kitbashing\n\nKinematic Kitbashing synthesizes articulated 3D objects by a ssembling reusable parts according to an abstract kinematic graph. It pres erves plausible attachments across motion using kinematics-aware geometric cues and can optimize for functional goals, enabling new articulated desi gns from existing ...\n\n\nMinghao Guo (Massachusetts Institute of Technol ogy (MIT)); Victor Zordan (Clemson University); Sheldon Andrews (École de Technologie Supérieure (ÉTS)); Wojciech Matusik (Massachusetts Institute o f Technology (MIT)); Maneesh Agrawala (Stanford University, Roblox); and H sueh-Ti Derek Liu (Roblox)\n---------------------\nComputational Design of Coordinate-Motion Assemblies\n\nWe present a computational approach for d esigning contact-based coordinate-motion assemblies that meet user-specifi ed target appearance and motion. One key enabler of our approach is that we established a theoretical connection between contact geometry within an assembly and unique coordinate motio...\n\n\nYukun Lu and Ke Chen (Univer sity of Science and Technology of China, Singapore University of Technolog y and Design); Ligang Liu (University of Science and Technology of China); and Peng Song (Singapore University of Technology and Design)\n---------- -----------\nComputational Design of Terrestrial Robots with Anisotropic F riction\n\nThe interplay of morphology, control, and anisotropic friction makes optimal design for terrestrial locomotion challenging. We present a computational pipeline that co-designs friction and controllers across div erse robot morphologies, showing anisotropic friction is critical and achi eving statistic...\n\n\nHang Hu (Tsinghua University, Shanghai Qi Zhi Inst itute); Kangbo Lyu (Shanghai Qi Zhi Institute, Tsinghua University); Chang yu Hu, Zihan Li, and Peiwen Yang (Tsinghua University); Minchen Li (Carneg ie Mellon University, Genesis AI); Shuguang Li (Tsinghua University); and Tao Du (Tsinghua University, Shanghai Qi Zhi Institute)\n----------------- ----\nExact predicates, exact constructions and combinatorics for mesh CSG \n\nThis article introduces an algorithm that exactly constructs the so-ca lled Weiler model (also called a 3D mesh arrangement) and that uses it to implement CSG with arbitrary multi-operand expressions. The main contribut ion is a 2D Constrained Delaunay Triangulation with exact coordinates and symbolic...\n\n\nBruno Levy (Inria Saclay and Laboratoire de Mathématiques d'Orsay Université Paris Saclay) and Bruno Levy\n---------------------\nS urface chamfering for robust tetrahedral meshing\n\nWe describe a new algo rithm for conforming high quality tetrahedral meshing with 100% success ra te. A novel input "chamfering" that eliminates all acute angles guarantees formal convergence of the underlying Delaunay refinement, while novel imp licit Steiner points coupled with indirect geometric pre...\n\n\nLorenzo D iazzi (CNR IMATI), Daniele Panozzo and Jiacheng Dai (NYU), and Marco Atten e (CNR IMATI)\n---------------------\nDJM: Compact Base Meshes for Displac ement Mapping using Triangle Jacobians\n\nA new method for computing base meshes for displacement mapping, exploiting the Jacobian of the displaceme nt function to guide base-mesh computation. DJM outperforms prior art in t erms of reconstruction accuracy-to-size trade off and can be used for both traditional micromesh-based rendering and neu...\n\n\nCongyi Zhang (Unive rsity of Texas Dallas); Nicholas Vining (NVIDIA, The University of British Columbia); Yanhong Lin (TransGP); Alireza Khatami (University of Texas Da llas); Ziyu Sun (The University of British Columbia); Xiaohu Guo (Universi ty of Texas Dallas); Wenping Wang (Texas A&M University); and Alla Sheffer (The University of British Columbia)\n---------------------\nFeature-Pres erving Offset Meshing\n\nThis paper introduces a novel method for generati ng offset meshes from 3D surfaces. It uniquely supports mitered joins and variable offset distances while robustly preserving sharp features. Evalua ted on a broad dataset, our approach outperforms state-of-the-art methods in feature fidelity and eleme...\n\n\nHongyi Cao, Gang Xu, Renshu Gu, and Jinlan Xu (Hangzhou Dianzi University); Xiaoyu Zhang (Beijing Institute of Spacecraft System Engineering); Rabczuk Timon (Bauhaus University Weimar) ; Yuzhe Luo (State Key Laboratory of CAD&CG Zhejiang University); Xifeng G ao (LightSpeed Studios); and Hongyi Cao\n---------------------\nPR-Cage: P rogressive Feasibility Relaxation for Tight Bounding Cage Generation\n\nPR -Cage is an automated framework for generating high-quality cages—simplifi ed mesh proxies that tightly enclose complex geometry. Using a staircase r elaxation of a thickness parameter and constrained QEM optimization, it ac hieves extreme simplification while preserving shape fidelity, enablin...\ n\n\nHuibiao Wen (Shandong University, University of Health and Rehabilita tion Sciences); Kaikai Qin (Hangzhou Dianzi University); Xinxin Su (Shando ng Huayun Technology Co., Ltd.); Jingcheng Mei (HoteamSoft); Shuangmin Che n (Qingdao University of Science and Technology, Shandong Key Laboratory o f Deep Sea Equipment Intelligent Networking); Chongyang Deng (Hangzhou Dia nzi University); Changhe Tu and Shiqing Xin (Shandong University); and Wen ping Wang (Texas A&M University)\n---------------------\nC^0 Generalized C oons Volumes over Arbitrary Polyhedra\n\nIn this paper, we generalize the Coons volume from hexahedral topology to arbitrary polyhedral topologies v ia generalized barycentric coordinates. We prove that the proposed general ized Coons volume possesses several desirable geometric properties and dem onstrate its applications in computer graphic...\n\n\nKaikai Qin and Zeqi Ge (Hangzhou Dianzi University), Péter Salvi (Budapest University of Techn ology and Economics), Chenhao Ying (Hangzhou Dianzi University), Huibiao W en (Shandong University), Kepeng Xu (Xidian University), Shiqing Xin (Shan dong University), and Chongyang Deng (Hangzhou Dianzi University)\n------- --------------\nTexture-Aware Remeshing for Texture-aware Geometry Process ing\n\nThis work demonstrates how to hoist textures to vertex color to all ow for incorporating textures into geometry processing algorithms. The use of vertex colors for texture-aware results is then shown across a number of algorithms, including parameterization, clustering, remeshing, and othe rs.\n\n\nJulian Knodt and Seung-Hwan Baek (POSTECH) and Julian Knodt\n---- -----------------\nBoxCtrl: 3D-Aware Visual Prompting for Geometric Image Editing\n\nBoxCtrl introduces a 3D-aware visual prompting framework for pr ecise image editing. Using RGB 3D bounding boxes to decouple geometry from appearance enables accurate translation, rotation, and scaling. A two-sta ge training of first SFT and then Reinforcement Learning achieves superior precision and ...\n\n\nFeifei Wang and Shiyuan Yang (City University of H ong Kong), Xiaoyu Li (Tencent), and Jing Liao (City University of Hong Kon g)\n---------------------\nViewWeaver: Geometry-Grounded Generative Render ing for 3D-Aware Image Customization\n\nViewWeaver is a geometry-grounded generative framework for 3D-aware image customization. It leverages multi- view references and explicit camera control to synthesize instruction-foll owing images with consistent identity and structure, improving viewpoint a ccuracy and realism through efficient multi-...\n\n\nYaowei Li (Peking Uni versity), Xiaoyu Li and Zhaoyang Zhang (Tencent), Hongxiang Li and Long Ch en (HKUST), Ying Shan (Tencent), and Yuexian Zou (Peking University)\n---- -----------------\nContinuous Control of Editing Models via Adaptive-Origi n Guidance\n\nDiffusion editing models lack smooth control over edit inten sity, and scaling standard Classifier-Free Guidance fails to provide it. W e introduce Adaptive-Origin Guidance (AdaOr), which dynamically adjusts th e guidance origin using an identity-conditioned prediction. This enables c ontinuous, fine-gr...\n\n\nAlon Wolf (Tel Aviv University, Decart.ai); Che n Katzir and Kfir Aberman (Decart.ai); and Or Patashnik (Tel Aviv Universi ty, Snap)\n---------------------\nHeadRouter: A Training-free Image Editin g Framework for MM-DiTs by Adaptively Routing Attention Heads\n\nHeadRoute r is a training-free image editing framework tailored for Multimodal Diffu sion Transformers (MM-DiTs). It addresses diminishing text guidance by ada ptively routing attention heads based on semantic sensitivity and refining token representations. The method improves semantic alignment and st...\n \n\nYu Xu (University of Chinese Academy of Sciences), Fan Tang and Juan C ao (Institute of Computing Technology Chinese Academy of Sciences), Xiaoyu Kong (Beihang University), Yuxin Zhang (University of Chinese Academy of Sciences), Jintao Li (Institute of Computing Technology Chinese Academy of Sciences), Oliver Deussen (University of Konstanz), Tong-Yee Lee (Nationa l Cheng Kung University), and Yu Xu\n---------------------\nLooseRoPE: Con tent-aware Attention Manipulation for Semantic Harmonization\n\nLooseRoPE enables prompt-free image editing by letting users crop and paste objects directly into new scenes. It improves diffusion-based harmonization by mod ulating RoPE with saliency, balancing identity preservation and contextual blending. This yields intuitive, spatially precise edits that maint...\n\ n\nEtai Sella (Tel Aviv University, Snap); Yoav Baron (Tel Aviv University ); Hadar Averbuch-Elor (Cornell Tech); and Daniel Cohen-Or and Or Patashni k (Tel Aviv University, Snap)\n---------------------\nMAOAM: Unified Objec t and Material Selection with Vision-Language Models\n\nMAOAM is a unified selection framework that enables precise object- and material- level sele ction across both text- and click-based interactions. MAOAM leverages a VL M which interprets the user's selection intent, and encodes information to the segmentation head. MAOAM achieves strong performance an...\n\n\nJaden Park (University of Wisconsin-Madison); Valentin Deschaintre and Jason Ku en (Adobe Research); Kangning Liu (Adobe); Iliyan Georgiev and Krishna Kum ar Singh (Adobe Research); Yong Jae Lee (Adobe Research, University of Wis consin-Madison); and Michael Fischer (Adobe Research)\n------------------- --\nProgressing Level-of-Detail Animation for Volumetric Elastodynamics\n\ nWe extend Progressive Dynamics [Zhang et al. 2024, 2025] from cloth and s hells to volumetric finite elements, enabling an efficient level-of-detail (LOD) animation-design pipeline with predictive coarse-resolution preview s for rapid iteration toward a final high-resolution volumetric elastodyna mics ...\n\n\nJiayi Eris Zhang and Doug James (Stanford University) and Da nny Kaufman (Adobe Inc.)\n---------------------\nMixed Material Point Meth ods for Stiff Elastoplasticity\n\nOur variant of the Material Point Method discretizes both the velocity and stress fields over mixed finite element s, avoiding the need for repeated costly particle-to-grid transfers within the implicit elastoviscoplastic solve. Our method supports materials rang ing from sand and snow to elastic soli...\n\n\nGilles Daviet (NVIDIA)\n--- ------------------\nMPM Lite: Linear Kernels and Integration without Parti cles\n\nMPM Lite is a hybrid Lagrangian/Eulerian method that removes parti cle-based quadrature at solve time. It achieves 15.9 times speedup over im plicit MPM, and 1.88 times speedup over explicit MPM.\n\n\nXiang Feng (Uni versity of California Los Angeles, University of California San Diego); Yu nuo Chen and Chang Yu (University of California Los Angeles); Hao Su (Univ ersity of California San Diego); Demetri Terzopoulos (University of Califo rnia Los Angeles); Yin Yang (University of Utah); Joe Masterjohn and Aleja ndro Castro (Toyota Research Institute); and Chenfanfu Jiang (University o f California Los Angeles)\n---------------------\nTube Maps: Fast SPH Boun dary Handling with Tubular Coordinates\n\nWe introduce Tube Maps, a drop-i n replacement for SPH boundary density computation that achieves accuracy comparable to implicit methods while dramatically reducing their computati onal cost. Our key observation is that the boundary density integral is fu lly determined by the local surface geometry n...\n\n\nDaria Nogina and Si lvia Sellán (Columbia University)\n---------------------\nLow-Rank Koopman Deformables with Log-Linear Time Integration\n\nWe present a Koopman oper ator method for fast deformable simulation using Dynamic Mode Decompositio n, enabling efficient long-term prediction without sequential time steppin g. Our discretization-agnostic formulation generalizes across shapes and r esolutions, making Koopman-based reduced models pract...\n\n\nYue Chang (U niversity of Toronto); Peter Yichen Chen (University of British Columbia); Eitan Grinspun (University of Toronto); and Maurizio M. Chiaramonte (Real ity Labs Research, Meta)\n---------------------\nPhysics-Inspired Procedur al Texturing of Extremely Deformable Surfaces\n\nApplying texture maps to dynamic deformable surfaces presents a significant challenge, due to ever- shifting differences in length scales involved. \nWe present two novel wav e-based procedural texturing algorithms which reproduce common physical pr operties like advection and self-similarity, enabling ...\n\n\nAleksei Kal inov, Mickaël Ly, Christian Hafner, and Chris Wojtan (Institute of Science and Technology Austria)\n---------------------\nWoodstock: Interactive Mo deling of Fungal Wood Decay\n\nWoodstock is a novel biophysical simulation framework for fungal wood decay. By tightly coupling rot dynamics, wood s tates, and strand-based mechanics, our method reproduces characteristic de cay phenomena such as fungal colonization, trunk hollowing, wood fracturin g, and eventual structural collapse...\n\n\nZhanyu Yang (Purdue University ), Nikolas Schwarz (Kiel University), Bosheng Li (Samsung), Dominik Michel s (KAUST), Bedrich Benes (Purdue University), Sören Pirk (Kiel University) , and Wojtek Palubicki (Adam Mickiewicz University in Poznań)\n----------- ----------\nGraphical X Splatting (GraphiXS): A Graphical Model for 4D Gau ssian Splatting under Uncertainty\n\nGraphical X Splatting (GraphiXS) is a novel probabilistic framework for systematically modelling data uncertain ty in 4D Gaussian Splatting. It handles various types of spatial and tempo ral sparsity, including missing frames and sparse viewpoints, and generali zes across diverse primitives. Extensive ...\n\n\nDoga Yilmaz (University College London (UCL)); Jialin Zhu (Baidu Research); Deshan Gong (The Unive rsity of Hong Kong); and He Wang (University College London (UCL), Univers ity of Leeds)\n---------------------\nObject-Space Analysis of Local Contr ast Sensitivity for Hierarchical Representations of 3D Gaussians\n\nWe pre sent an object-space contrast sensitivity analysis for 3D Gaussian represe ntations that estimates spatial frequency responses without explicit raste rization using analytical Fourier transforms and the projection-slice theo rem. This enables contrast-sensitivity-based perceptual evaluation and e.. .\n\n\nNaoto Yoshii and Suguru Saito (Institute of Science Tokyo) and Masa taka Sawayama and Yoshinori Dobashi (Hokkaido University, Prometech CG Res earch)\n---------------------\nLearning View-Dependent Splatting Kernels\n \nWe present a differentiable framework to automatically learn view-depend ent 2D kernels in a splatting-based pipeline to improve reconstruction qua lity and representation efficiency for novel 3D view synthesis.\n\n\nHuake ng Ding, Zhanpeng Liu, Fan Pei, Kun Zhou, and Hongzhi Wu (State Key Lab of CAD and CG, Zhejiang University)\n---------------------\nCAdam: Context-A daptive Moment Estimation for 3D Gaussian Densification in Generative Dist illation\n\nCAdam resolves the mismatch between reconstruction-native dens ification and stochastic generative guidance in optimization-based 3DGS. B y reinterpreting densification as statistical signal verification via mome ntum and context-adaptive selection, it reduces Gaussian primitives by 85% –97% acros...\n\n\nSeungJeh Chung (Kyung Hee University), Geonho Park (Kor ea University), Misong Kim (Kyung Hee University), and HyeongYeop Kang (Ko rea University)\n---------------------\nRadiance Fields from Photons\n\nNe ural radiance fields and Gaussian splats often produce artifacts in low li ght, high dynamic range, or with rapid motion. Instead, by training at the granularity of individual photons, as enabled by single-photon cameras, w e can achieve sharp, reliable reconstructions even in these challenging co nd...\n\n\nSacha Jungerman, Aryan Garg, and Mohit Gupta (University of Wis consin-Madison) and Sacha Jungerman\n---------------------\nRef-DGS: Refle ctive Dual Gaussian Splatting\n\nRef-DGS is a dual Gaussian splatting fram ework that decouples surface reconstruction and specular reflection, model ing near- and far-field specular reflections without ray tracing, achievin g efficient, high-quality surface reconstruction and novel view synthesis. \n\n\nNingjing Fan and Yiqun Wang (Chongqing University); Dong-Ming Yan (I nstitute of Automation, Chinese Academy of Sciences); and Peter Wonka (Kin g Abdullah University of Science and Technology (KAUST))\n---------------- -----\nLoBoFit: Flexible Garment Refitting via Local Bone Mapping Blending \n\nLoBoFit brings robust, high-fidelity garment refitting to challenging avatar changes. It preserves original design features, fine wrinkles, and fit style by optimizing garments in blended local bone coordinates, enabli ng stable convergence, efficient refinement, and reliable results across l arge bod...\n\n\nMeng Zhang (Nanjing University of Science and Technology) , Yu Xin (University of Science and Technology of China), Feiya Guo (Nanji ng University of Science and Technology), Kaizhang Kang (King Abdullah Uni versity of Science and Technology (KAUST)), Mengyu Chu (Peking University) , and Ruizhen Hu (Shenzhen University)\n---------------------\nPatternGSL: A Structured Specification Language for Template-Free and Simulation-Read y 3D Garments\n\nWe present PatternGSL, a structured and template-free rep resentation for simulation-ready 3D garments. Given a single garment image , our VLM-based method predicts editable sewing patterns with explicit pan el, edge, and stitch topology. Trained on the large-scale PatternGSLData d ataset, it enables ac...\n\n\nZhenyang Li (The University of Hong Kong (HK U); Tsinghua University, China); Lutao Jiang (The Hong Kong University of Science and Technology (Guangzhou)); Yizhou Zhao (Carnegie Mellon Universi ty); Weikai Chen (LIGHTSPEED); Ying-Cong Chen (The Hong Kong University of Science and Technology (Guangzhou), The Hong Kong University of Science a nd Technology); Xin Wang (LIGHTSPEED); and Yifan Peng (The University of H ong Kong (HKU))\n---------------------\nGarment Particles: A 2D–3D Symmetr ic Garment Representation for Generation and Editing\n\nGarment Particles is a 5D point-cloud representation that jointly encodes 2D sewing patterns and 3D geometry. This representation enables Garment Particles Flow (GPF) and Particles-to-Pattern Flow (PPF), a rectified flow framework for both simulation-ready sewing pattern generation from text and ima...\n\n\nKiyoh iro Nakayama (Stanford University); I-Chao Shen (The University of Tokyo); Ruofan Liu (Institute of Science Tokyo, Stanford University); Yiming Wang (ETH Zürich, Stanford University); Gordon Wetzstein (Stanford University) ; and Takeo Igarashi (The University of Tokyo)\n---------------------\nLea rning Sewing Patterns via Latent Flow Matching of Implicit Fields\n\nWe in troduce an implicit representation for sewing patterns using distance fiel ds. Combined with latent flow matching, the model learns distributions ove r panel geometry and stitching, enabling generation, image-based estimatio n, and optimization. The framework supports flexible topology and improve. ..\n\n\nCong Cao (MBZUAI); Ren Li (MBZUAI, SUSTech); Corentin Dumery (EPFL ); and Hao Li (Mohamed Bin Zayed University of Artificial Intelligence)\n- --------------------\nLearned Universal Interoperable Virtual Try-ON\n\nWe introduce a fully automated 3D virtual try-on system that fits diverse mu ltilayer garments onto arbitrary humanoid characters. Using SMPL as an int ermediate proxy, it handles unrigged, non-manifold assets, supports garmen t resizing, and produces high-quality draping without manual correspondenc e,...\n\n\nCong Cao and Xianhang Cheng (Mohamed Bin Zayed University of Ar tificial Intelligence); Jingyuan Liu (The University of Tokyo, Mohamed Bin Zayed University of Artificial Intelligence); Yujian Zheng and Zhenhui Li n (Mohamed Bin Zayed University of Artificial Intelligence); Ren Li (Moham ed Bin Zayed University of Artificial Intelligence, SUSTech); Meriem Chkir (Mohamed Bin Zayed University of Artificial Intelligence); and Hao Li (Mo hamed Bin Zayed University of Artificial Intelligence, Pinscreen)\n------- --------------\nFIT: A Large-Scale Dataset for Fit-Aware Virtual Try-On\n\ nWe introduce FIT (Fit-Inclusive Try-on), a virtual try-on dataset of over 1M samples covering diverse garment fits, each annotated with precise bod y/garment measurements. Our novel data generation pipeline leverages synth etic garment simulation and photorealistic retexturing. Then, we leverage FIT t...\n\n\nYuanhao Wang and Johanna Karras (University of Washington, G oogle Research); Yingwei Li (Google Research); and Ira Kemelmacher-Shlizer man (University of Washington, Google Research)\n---------------------\nAd apting Quality Metrics to Tone Mapping\n\nThis work provides a recipe for adapting existing image/video quality metrics for evaluating tone-mapping algorithms. We demonstrate across multiple datasets that when tone-mapped images and their HDR reference are represented in a perceptually-uniform s pace, existing metrics such as TOPIQ, DISTS, a...\n\n\nKenneth Chen (New Y ork University); Dongyeon Kim (University of Cambridge); Yuta Asano and Al exandre Chapiro (Reality Labs Research, Meta); Qi Sun (New York University ); and Rafał Mantiuk (University of Cambridge)\n---------------------\nLUC ID: Learning Unified Control for Image Deflaring and Exposure Mastery in N ighttime Photography\n\nMaster the night with LUCID: Learning Unified Cont rol for Image Deflaring and Exposure Mastery in Nighttime Photography. Mer ging flare disentanglement and generative diffusion, our framework conquer s entangled degradations. LUCID unlocks continuous control over global ill umination and light source s...\n\n\nTingyu Yang, Yuan Cheng, and Xiaoyun Yuan (Shanghai Jiao Tong University)\n---------------------\nLucky High Dy namic Range Smartphone Imaging\n\nStage 3 revisions add a DNG-to-sRGB pipe line description and ISP-limitation caveat, promote six supplement figures /tables to the main paper, clarify use of unseen handheld captures for rea l-world evaluation, added a failure case, limitation and future work in th e main paper, and add supplementary mo...\n\n\nBaiang Li, Ruyu Yan, and Et han Tseng (Princeton University); Zhoutong Zhang (Adobe); Adam Finkelstein (Princeton University); Jiawen Chen (Adobe); and Felix Heide (Princeton U niversity)\n---------------------\nSoft Anisotropic Diagrams for Different iable Image Representation\n\nSoft Anisotropic Diagrams (SAD) is a fast, d ifferentiable image representation based on a soft anisotropic additively weighted Voronoi partition (i.e., an Apollonius diagram) with learnable pe r-site temperatures. It reconstructs images with sharp, content-aligned bo undaries, compact storage, fast ra...\n\n\nLaki Iinbor, Zhiyang Dou, and W ojciech Matusik (MIT)\n---------------------\nSingle-shot HDR using conven tional image sensor shutter functions and optical randomization\n\nWe pres ent a single-shot HDR imaging method using a global reset release (GRR) se nsor and optical shuffling via a random fiber bundle. Our prototype enable s HDR recovery with a simple prior and flexible exposure control, demonstr ating a dynamic range increase from 48 dB to 73 dB in validation.\n\n\nXia ng Dai (UC San Diego), Kyrollos Yanny (UC Berkeley), Kristina Monakhova (C ornell University), Nicholas Antipa (UC San Diego), and Xiang Dai\n------- --------------\nFinding Fast Filters\n\nImage and audio processing applica tions want to use large FIR filters while adhering to strict runtime and l atency requirements. Prior methods for fast filter approximation offer lim ited speed versus quality tradeoffs. We unify these approximation techniqu es as primitives within an automatically sea...\n\n\nKarima Ma and Andrew Adams (Adobe Inc), Jonathan Ragan-Kelley (Massachusetts Institute of Techn ology - EECS), and Karima Ma\n---------------------\nR-DMesh: Video-Guided 3D Animation via Rectified Dynamic Mesh Flow\n\nWe present R-DMesh, a nov el feed-forward video-guided mesh animation framework that solves the pose misalignment dilemma between an initial mesh and a condition video throug h jump offset modeling. It generates high-fidelity, motion aligned 4D anim ations and enables downstream applications like pose ...\n\n\nZijie Wu (Hu azhong University of Science and Technology, Tencent Hunyuan); Lixin Xu, P uhua Jiang, Sicong Liu, and Chunchao Guo (Tencent Hunyuan); and Xiang Bai (Huazhong University of Science and Technology)\n---------------------\nTo poCap: Learning Topology-Agnostic Motion Priors for Monocular Video-to-Ani mation\n\nTopoCap is a unified framework for motion capture and retargetin g from monocular video, supporting arbitrary skeletal topologies without t est-time optimization. Leveraging a two-stage generative pipeline, TopoCap achieves motion transfer across diverse creatures, backed by Mobjaverse—a dataset...\n\n\nChengfeng Pu, Jia-Peng Zhang, and Meng-Hao Guo (CS Dept, Tsinghua University); Yan-Pei Cao (VAST); and Shi-Min Hu (CS Dept, Tsinghu a University)\n---------------------\nStylized Text-to-Motion Generation v ia Hypernetwork-Driven Low-Rank Adaptation\n\nWe propose a hypernetwork-dr iven LoRA framework for stylized text-to-motion generation that dynamicall y modulates a pretrained diffusion model using style embeddings extracted from reference motions. By structuring the style latent space with supervi sed contrastive learning and incorporating style-g...\n\n\nJunhyuk Jeon, S eokhyeon Hong, and Junyong Noh (Korea Advanced Institute of Science and Te chnology (KAIST))\n---------------------\nMUSIC: Learning Muscle-Driven De xterous Hand Control\n\nWe present a data-driven framework for physics-bas ed, muscle-driven dexterous control that enables musculoskeletal hands to perform piano playing for novel musical scores. Our framework adopts a hie rarchical architecture, combining high-frequency muscle control with low-f requency coordination. The m...\n\n\nPei Xu and Yufei Ye (Stanford Univers ity); Shuchu Sun (Clemson University); and Yu Ding, Elizabeth Schumann, an d C. Karen Liu (Stanford University)\n---------------------\nMOCHI: Motion Enhancement of Collaborative Human-object Interactions\n\nOur method enha nces noisy multi-human object interaction (MHOI) data by automatically gen erating plausible hand and finger motions and refining noisy body movement s, all while preserving the original interaction semantics.\n\n\nJiye Lee, Yonghun Choi, and Jungdam Won (Seoul National University)\n-------------- -------\nACT: A Unified Framework for Rigging and Animating Characters wit h Arbitrary Topologies\n\nACT is a unified AI framework that effortlessly brings static 3D characters to life. By integrating rigging, skinning, and motion generation into a single diffusion-based pipeline, ACT automatical ly animates characters of any shape, delivering high-quality, physically p lausible motions that overcome...\n\n\nPengyu Long (ShanghaiTech Universit y, ByteDance Games); Weirui Wang (ShanghaiTech University, Deemos Technolo gy); Qingcheng Zhao (University of Toronto, Deemos Technology); Xiaoyang G uo and Xiaoyu Pan (ByteDance Games); Qixuan Zhang (ShanghaiTech University , Deemos Technology); Jiaqing Zhou and Tianlei Hu (ByteDance Games); Wei Y ang (Huazhong University of Science and Technology); and Lan Xu and Jingyi Yu (ShanghaiTech University)\n---------------------\nGeoQuery: Geometry-Q uery Diffusion for Sparse-View Reconstruction\n\nGeoQuery is a geometry-gu ided single-step diffusion framework for mitigating query contamination in existing multi-view self attention mechanism. By retrieving proxy feature s from clean reference views via geometric correspondences, it bypasses co rrupted query features and enables high-fidelity rest...\n\n\nXiao Cao (Un iversity of Electronic Science and Technology of China), Yuze Li (Tianjin University), Youmin Zhang and Jiayu Song (Rawmantic AI), Cheng Yan (Tianji n University), and Wen Li and Lixin Duan (University of Electronic Science and Technology of China)\n---------------------\nVidSplat: Gaussian Splat ting Reconstruction with Geometry-Guided Video Diffusion Priors\n\nWe pres ent VidSplat, a generative reconstruction framework that leverages powerfu l video diffusion priors to synthesize novel views that compensate for mis sing input coverage, and thereby recover complete 3D scenes from sparse in puts. VidSplat performs robustly to sparse input and even a single imag... \n\n\nJimin Tang, Wenyuan Zhang, Junsheng Zhou, and Zian Huang (Tsinghua U niversity); Kanle Shi and Shenkun Xu (Kuaishou Technology); Yu-Shen Liu (T singhua University); and Zhizhong Han (Wayne State University)\n---------- -----------\nImmediate 3D Gaussian Splat Reconstruction of Unordered Input with Global Consistency\n\nWe propose a 3D Gaussian Splatting reconstruct ion method with immediate feedback that handles unordered image captures a nd very large scenes. Through fast place-recognition-driven matching, clus ter-based loop closure with graph-propagated correction, and a progressive Gaussian hierarchy construction...\n\n\nAndreas Meuleman (Inria, Universi té de Rennes); Linus Franke (Inria, Université Côte d'Azur); Boris Zhestia nkin (Inria, Université Côte d'Azur; EPFL); Camille Montemagni (Inria, Uni versité de Rennes); and George Drettakis (Inria Université Côte d'Azur)\n- --------------------\nMoonSplat: Monocular Online Gaussian Splatting with Sim(3) Global Optimization\n\nMoonSplat is a robust and efficient online v oxelized 3DGS reconstruction framework integrated with global Sim(3) optim ization, which enables reliable camera tracking and efficient global loop closure for both camera poses and voxelized 3DGS. MoonSplat is equipped wi th a color residual learning strat...\n\n\nGuo Pu and Yixuan Han (Peking U niversity); Haofeng Li, Yao Zhang, and Hui Zhou (Beijing Hydrogen Intellig ent Tech. Co., Ltd.); and Zhouhui Lian (Peking University)\n-------------- -------\nRaDe-GS: Rasterizing Depth in Gaussian Splatting\n\nGaussian Spla tting enables high-quality, real-time novel-view synthesis, but extracting detailed 3D geometry remains challenging. We introduce a rasterized appro ach to render the depth and normal maps for general 3D Gaussian primitives . Our method significantly enhances shape reconstruction accuracy...\n\n\n Baowen Zhang and Chuan Fang (Hong Kong University of Science and Technolog y), Rakesh Shrestha (Simon Fraser University), Yixun Liang (Hong Kong Univ ersity of Science and Technology), xiaoxiao Long (Nanjing University), Pin g Tan (Hong Kong University of Science and Technology), and Baowen Zhang\n ---------------------\nImplicit Minimal Surfaces for Bijective Corresponde nces\n\nWe introduce an implicit representation of continuous, bijective, orientation-preserving maps between genus zero surfaces with or without bo undary. The distortion of these maps can easily be minimized by optimizing the Ginzburg-Landau functional---a ubiquitous model in physics and differ ential geome...\n\n\nEtienne Corman (CNRS, Inria, LORIA); Yousuf Soliman ( Side Effects Software Inc); Robin Magnet (INRIA, Université Paris Cité); a nd Mark Gillespie (INRIA, University of Utah)\n---------------------\nUnta ngling Surfaces via Shape and Mesh Repulsion\n\nWe present an energy-based framework for removing self-intersections from surface meshes. A shape-le vel Gaussian energy resolves global entanglement while a mesh-level Minkow ski penalty handles local discrete degeneracies. Together, they eliminate all self-intersections across benchmark datasets, ou...\n\n\nJiří Minarčík (Carnegie Mellon University, Resistant AI); Michael Liu (Carnegie Mellon University); Keenan Crane (Carnegie Mellon University, Roblox Research); a nd Minchen Li (Carnegie Mellon University, Genesis AI)\n------------------ ---\nQuasi-Medial Distance Field (Q-MDF): A Robust Method for Approximatin g and Discretizing Neural Medial Axes\n\nWe introduce an implicit method f or medial axis extraction from point clouds and meshes, addressing challen ges in noisy or incomplete data. By connecting the signed distance field a nd the medial field, we reformulate the problem as implicit reconstruction , offering a novel, robust solution for accur...\n\n\nJiayi Kong (S-Lab Na nyang Technological University); Chen Zong (S-Lab Nanyang Technological Un iversity, School of Mathematics Nanjing University of Aeronautics and Astr onautics); Jun Luo (Nanyang Technological University); Shiqing Xin (School of Computer Science and Technology Shandong University); Fei Hou (Institu te of Software, Chinese Academy of Sciences; University of Chinese Academy of Sciences); Hanqing Jiang and Chen Qian (SenseTime Research); Ying He ( Nanyang Technological University); and Jiayi Kong\n---------------------\n Structural MAT: Clean and Scalable Medial Axis Simplification via Explicit Surface Correspondence\n\nWe present Structural MAT, a fast and scalable medial axis simplification framework that explicitly tracks the correspond ence between medial vertices and surface regions throughout simplification . The method produces clean, structure-aware medial axes with accurate fea ture alignment and high mesh qu...\n\n\nPengfei Wang (Shandong University) ; Shuangmin Chen (Qingdao University of Science and Technology); Dongming Yan (Institute of Automation, Chinese Academy of Sciences); Ying He (Nanya ng Technological University, Singapore); Shiqing Xin and Changhe Tu (Shand ong University); and Wenping Wang (Texas A&M University)\n---------------- -----\nPersistence-guided Prescribed Topological Simplification\n\nWe deve loped a new algorithm to simplify the topology of a 3D shape to user-presc ribed numbers of connected components, handles, and voids, while minimizin g geometric changes.\n\n\nLinxuan Rong and Tao Ju (Washington University i n St. Louis)\n---------------------\nPointLLM-R: Enhancing 3D Point Cloud Reasoning via Chain-of-Thought\n\nPointLLM-R introduces chain-of-thought r easoning for 3D point cloud understanding via a data-centric pipeline. It builds a large reasoning dataset (PoCoTI) and trains a multimodal model th at achieves state-of-the-art performance, improving accuracy, interpretabi lity, and robustness across synthetic ...\n\n\nChaoqi Chen, Qile Xu, Wenju n Zhou, and Hui Huang (Shenzhen University)\n---------------------\nRecipr ocal Latent Fields for Precomputed Sound Propagation\n\nWave-coding method s for sound propagation in games are very memory-intensive. We present Rec iprocal Latent Fields leveraging Riemannian metric learning to reduce memo ry usage by orders of magnitude. Our method infers relevant acoustic param eters nearly instantaneously and results in perceptually ide...\n\n\nHugo Seuté (Ubisoft Divertissements Inc.) and Pranai Vasudev, Etienne Richan, a nd Louis-Xavier Buffoni (Audiokinetic Inc.)\n---------------------\nMAViD: A Multimodal Framework for Audio-Visual Dialogue Understanding and Genera tion\n\nMAViD is a multimodal framework for interactive audio-visual dialo gue. Its Conductor-Creator architecture integrates multimodal understandin g and reasoning into response generation. By combining AR and diffusion mo dels, MAViD enables synchronized, long-duration audio-video generation wit h consistent...\n\n\nYouxin Pang (Tsinghua University); Jiajun Liu, Lingfe ng Tan, Yong Zhang, and Feng Gao (meituan); Xiang Deng (Tsinghua Universit y); Zhuoliang Kang and Xiaoming Wei (meituan); and Yebin Liu (Tsinghua Uni versity)\n---------------------\nAudio-Omni: Extending Multi-modal Underst anding to Versatile Audio Generation and Editing\n\nAudio-Omni is the firs t unified framework for multi-modal audio generation and editing across so und, music, and speech. By combining a frozen multimodal language model wi th a trainable diffusion transformer, it achieves state-of-the-art perform ance on multiple benchmarks while exhibiting remarkable ...\n\n\nZeyue Tia n (Hong Kong University of Science and Technology), Binxin Yang (Tencent), Zhaoyang Liu (Hong Kong University of Science and Technology), Jiexuan Zh ang (Peking University), Ruibin Yuan (Hong Kong University of Science and Technology), Hubery Yin (Tencent), Qifeng Chen (Hong Kong University of Sc ience and Technology), Chen Li and Jing Lyu (Tencent), and Wei Xue and Yik e Guo (Hong Kong University of Science and Technology)\n------------------ ---\nJust-Dub-It: Video dubbing via Joint Audio-Visual Diffusion\n\nJUST-D UB-IT is a unified audio-video diffusion model for automatic video dubbing . It jointly synthesizes translated speech and synchronized lip movements via a lightweight LoRA adapter, replacing traditional pipeline approaches. Trained on synthetically generated multilingual data with strategic inp.. .\n\n\nAnthony Chen (Peking University, Lightricks); Naomi Ken Korem, Tavi Halperin, Matan Ben Yosef, Urska Jelercic, and Ofir Bibi (Lightricks); an d Or Patashnik and Daniel Cohen-Or (Tel Aviv University)\n---------------- -----\nVfxDB: A Visual Effects Volume Dataset and Benchmark for VDB-Native Generative Modeling\n\nVfxDB introduces a pioneering 1-million-sample dat aset tailored for VDB-native generation of sparse volumetric VFX sequences , alongside a reproducible benchmark suite. Validated by a scalable diffus ion baseline featuring our Atomic-Continuous prior, this comprehensive pip eline establishes essential ...\n\n\nJunwei Shu (East China Normal Univers ity); Hantang Liu (KunByte); Dawei Miao (Independent Researcher); and Wenz heng Song, Mingyang Yuan, Wenjie Liu, Changgu Chen, Yang Li, and Changbo W ang (East China Normal University)\n\nInterest Area: Arts & Design, Gaming & Interactive, New Technologies, Production & Animation, Research & Educa tion\n\nKeyword: Animation, Art, Artificial Intelligence/Machine Learning, Audio, Augmented Reality, Capture/Scanning, Computer Vision, Digital Twin s, Display, Dynamics, Education, Ethics and Society, Fabrication, Games, G enerative AI, Geometry, Haptics, Hardware, Image Processing, Lighting, Mat h Foundations and Theory, Modeling, Performance, Physical AI, Pipeline Too ls and Work, Real-Time, Rendering, Robotics, Scientific Visualization, Sim ulation, Spatial Computing, Virtual Reality\n\nRegistration Category: Full Conference Supporter, Full Conference, Experience END:VEVENT END:VCALENDAR