Paperid:1
Authors:Lanjiong Li,Guanhua Zhao,Lingting Zhu,Zeyu Cai,Lequan Yu,Jian Zhang,Zeyu Wang
The Hong Kong University of Science and Technology (Guangzhou),School of Electronic and Computer Engineering, Peking University,The University of Hong Kong,The Hong Kong University of Science and Technology (Guangzhou),The University of Hong Kong,School of Electronic and Computer Engineering, Peking University,The Hong Kong University of Science and Technology (Guangzhou)
Title: AssetDropper: Asset Extraction via Diffusion Models with Reward-Driven Optimization
Abstract:
AssetDropper is a novel framework for extracting standardized assets from reference images, addressing challenges such as occlusion and distortion. Leveraging both synthetic and real-world datasets, along with a reward-driven feedback mechanism, it achieves state-of-the-art performance in asset extraction and provides designers with a versatile open-world asset palette.



Paperid:2
Authors:Jiaju Ma,Maneesh Agrawala
Stanford University,Stanford University
Abstract:
Large vision-language models often fail to capture spatio-temporal details in text-to-animation tasks. We introduce MoVer, a verification system using first-order logic to check properties like timing and positioning in motion graphics animations. Integrated into an LLM pipeline, MoVer enables iterative refinement, significantly improving animation generation accuracy from 58.8% to 93.6%.



Paperid:3
Authors:Marcelo Sandoval-Castañeda,Bryan Russell,Josef Sivic,Gregory Shakhnarovich,Fabian David Caba Heilbron
TTIC,Adobe Research,Adobe Research,TTIC,Adobe Research
Abstract:
We automate video nonlinear editing using a multi-agent system. An Editor agent uses tools to create sequences from clips and instructions, while a Critic agent provides feedback in natural language. Our learning-based approach enhances agent communication. Evaluations with an LLM-as-a-judge metric and user studies show our system’s superior performance.



Paperid:4
Authors:Bo Yang,Ying Cao
ShanghaiTech University,ShanghaiTech University
Abstract:
We propose a Generative Order Learner (GOL) that optimizes element ordering for graphic design generation. Our approach learns a content-aware neural order, which can significantly improve graphic generation quality, generalize across different types of generative models and help design generators scale up greatly.



Paperid:5
Authors:Weitao You,Yinyu Lu,Zirui Ma,Nan Li,Mingxu Zhou,Xue Zhao,Pei Chen,Lingyun Sun
Zhejiang University,Zhejiang University,Zhejiang University,Zhejiang University,Zhejiang University,Zhejiang University,Zhejiang University,Zhejiang University
Abstract:
DesignManager is an AI-powered design support system that functions as an interactive copilot throughout the creative workflow. With node-based visualization of design evolution and conversational interaction modes, it helps designers track, modify, and branch their processes while providing context-aware assistance through an innovative agent framework.



Paperid:6
Authors:Xinrui Liu,Longxiulin Deng,Abe Davis
Cornell University,Cornell University,Cornell University
Abstract:
We propose Hybrid Tours, a hybrid approach to creating long-take shots by combining short video clips in a virtual interface. We show that Hybrid Tours makes capturing long-take touring shots much easier, and that clip-based authoring and reconstruction lead to higher-fidelity results at lower compute costs.



Paperid:7
Authors:Pascal Guehl,Rémi Allègre,Guillaume Gilet,Basile Sauvage,Marie-Paule Cani,Jean-Michel Dischler
LIX - Ecole Polytechnique/CNRS,ICube, Université de Strasbourg,Université de Sherbrooke,ICube, Université de Strasbourg,LIX - Ecole Polytechnique/CNRS,ICube, Université de Strasbourg
Abstract:
We introduce a fast, wave-based procedural noise model enabling precise spectral control in any dimension. Using precomputed wave functions and inverse Fourier transforms, it supports Gaussian and non-Gaussian noises—including Gabor, Phasor, and novel recursive cellular patterns—making it ideal for compact, controllable, and animated solid textures in 2D, 3D, and time.



Paperid:8
Authors:Gilles Daviet,Tianchang Shen,Nicholas Sharp,David Levin,Gilles Daviet
NVIDIA Corp,NVIDIA Corp,NVIDIA Corp,NVIDIA Corp,NVIDIA
Abstract:
We train a network to map signed distance fields to the quadrature points and weights of non-conforming numerical integration rule in a Mixed Finite Element formulation, enabling differentiable elastic simulation over evolving domains. We demonstrate applications to image-guided material and topology optimization.



Paperid:9
Authors:Jiong Chen,Florian Schäfer,Mathieu DESBRUN
INRIA Saclay,Georgia Institute of Technology,INRIA Saclay
Abstract:
We introduce an inverse-LU preconditioner to solve for the typical asymmetric and dense matrices generated by boundary element methods (BEM). The computational efficiency and low memory requirements of our approach conspire to scale up to millions of degrees of freedom, with orders of magnitude speedups in solving times.



Paperid:10
Authors:Bailey Miller,Rohan Sawhney,Keenan Crane,Ioannis Gkioulekas
Carnegie Mellon University,NVIDIA,Carnegie Mellon University,Carnegie Mellon University
Abstract:
We solve partial differential equations in domains involving complex microparticle geometry that is impractical, or intractable, to model explicitly. Drawing inspiration from volume rendering, we treat the domain as a participating medium with stochastic microparticle geometry and develop a volumetric variant of the Monte Carlo walk on spheres algorithm.



Paperid:11
Authors:Paul Himmler,Tobias Günther
Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU),Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU)
Abstract:
We present a novel Monte Carlo approach to solve boundary integral equations with Dirichlet boundary conditions in two dimensions. While Walk-on-Spheres uses largest empty circles, which touch the boundary in only one point, we utilize semicircles and circle sectors that share one or two boundary edges resulting in shorter walks.



Paperid:12
Authors:Tianyu Huang,Jingwang Ling,Shuang Zhao,Feng Xu
School of Software and BNRist, Tsinghua University,School of Software and BNRist, Tsinghua University,University of California Irvine,School of Software and BNRist, Tsinghua University
Abstract:
Walk on stars (WoSt) has shown its power in being applied to Monte Carlo methods for solving PDEs but the sampling techniques in WoSt are not satisfactory, leading to high variance. Inspired by Monte Carlo rendering, we propose a guiding-based importance sampling method to reduce the variance of WoSt.



Paperid:13
Authors:Luozhou Wang,Ziyang Mai,Guibao Shen,Yixun Liang,Xin Tao,Pengfei Wan,Di Zhang,Yijun Li,Yingcong Chen
Hong Kong University of Science and Technology, Guangzhou,Hong Kong University of Science and Technology, Guangzhou,Hong Kong University of Science and Technology, Guangzhou,Hong Kong University of Science and Technology,Kuaishou Technology,Kuaishou Technology,Kuaishou Technology,Adobe Research,Hong Kong University of Science and Technology, Guangzhou
Abstract:
We propose Motion Embeddings for video generation, enabling precise motion in video transfer across diverse scenes and objects. These embeddings disentangle motion from appearance, preserving original dynamics while adapting to new prompts. Experiments show that our method achieved high-quality, prompt-aligned video generation across a wide range of scenarios.



Paperid:14
Authors:Zinuo You,Stamatios Georgoulis,Anpei Chen,Siyu Tang,Dengxin Dai
ETH Zürich,Huawei Research Zürich,ETH Zürich,ETH Zürich,Huawei Research Zürich
Abstract:
GaVS: Transform unstable shaky videos into smooth, professional-quality footage. We design novel 3D rednering technology that preserves the motion intent while eliminating shakes and distortions—no cropping, no distortion and workable under dynamics and intense motions. GaVS delivers natural-looking results validated by users as superiority. Capture life steadily!



Paperid:15
Authors:Or Patashnik,Rinon Gal,Daniil Ostashev,Sergey Tulyakov,Kfir Aberman,Daniel Cohen-Or
Tel Aviv University,Tel Aviv University,Snap,Snap,Snap,Tel Aviv University
Abstract:
This paper introduces Nested Attention, a mechanism that improves text-to-image personalization by injecting query-dependent subject features into cross-attention layers, achieving strong identity preservation and prompt alignment. The method maintains the model’s prior, enabling multi-subject generation across diverse domains.



Paperid:16
Authors:Daniel Garibi,Shahar Yadin,Roni Paiss,Omer Tov,Shiran Zada,Ariel Ephrat,Tomer Michaeli,Inbar Mosseri,Tali Dekel
Tel Aviv University,Technion - Israel Institute of Technology,DeepMind,DeepMind,DeepMind,DeepMind,Technion - Israel Institute of Technology,DeepMind,Weizmann Institute of Science
Abstract:
TokenVerse extracts complex visual elements from images by identifying semantic directions in per-token modulation space of DiT models for each word in the image caption. It's capable of combining concepts from multiple sources by adding corresponding directions, enabling flexible generation of new combinations including abstract concepts like lighting and poses.



Paperid:17
Authors:Rameen Abdal,Or Patashnik,Ivan Skorokhodov,Willi Menapace,Aliaksandr Siarohin,Sergey Tulyakov,Daniel Cohen-Or,Kfir Aberman
Snap,Snap,Snap,Snap,Snap,Snap,Snap,Snap
Abstract:
Personalizing text-to-video models is challenging because dynamic concepts require capturing both appearance and motion. We propose Set-and-Sequence, a framework that personalizes DiT-based video models by first learning an identity LoRA basis from unordered frames, then fine-tuning coefficients with motion residuals on full videos, enabling superior editability and compositionality for applications.



Paperid:18
Authors:Shiyi Zhang,Junhao Zhuang,Zhaoyang Zhang,Ying Shan,Yansong Tang
Tsinghua University,Tencent,Tencent,Tencent,Tsinghua University
Abstract:
We propose FlexiAct, an image animation framework that transfers actions from a reference video to any target image, enabling variations in layout, viewpoint, and skeletal structure while maintaining identity consistency.



Paperid:19
Authors:Bernhard Braun,Jan Bender,Nils Thuerey
Technical University Munich,RWTH Aachen University,Technical University Munich
Abstract:
We present an algorithm for simulating large-scale, violently turbulent two-phase flows—such as breaking ocean waves, tsunamis, and asteroid impacts—at extreme resolutions of the coupled water-air velocity field. This is achieved by integrating a new multiphase FLIP variant with highly efficient dual particle–grid adaptivity and a novel adaptive Poisson solver.



Paperid:20
Authors:Siyuan Chen,Yixin Chen,Jonathan Panuelos,Otman Benchekroun,Yue Chang,Eitan Grinspun,Zhecheng Wang
University of Toronto,University of Toronto,University of Toronto,University of Toronto,University of Toronto,University of Toronto,University of Toronto
Abstract:
We introduce a novel reduced-order fluid simulation technique leveraging Dynamic Mode Decomposition (DMD) to enable fast, memory-efficient, and user-controllable subspace simulation. Combining spatial ROM compression with spectral physical insights, our method excels in animation, real-time interaction, artistic control, and time-reversible fluid effects.



Paperid:21
Authors:Jingrui Xing,Bin Wang,Mengyu Chu,Baoquan Chen
School of Intelligence Science and Technology, Peking University,Independent,Peking University,Peking University
Abstract:
We present a grid-free fluid simulator featuring a novel Gaussian spatial representation (GSR) for velocity field. The advantages of GSR over traditional Lagrangian/Eulerian data structures are 4-folded: memory compactness, spatial adaptivity, vorticity preservation and continuous differentiability. Our method also greatly outperforms similar competitors in terms of quality and performance.



Paperid:22
Authors:Fumiya Narita,Nimiko Ochiai,Takashi Kanai,Ryoichi Ando
The University of Tokyo,GAME FREAK Inc.,The University of Tokyo,Unaffiliated
Abstract:
This paper introduces a novel grid structure that extends tall cell methods for efficient deep water simulation. Unlike previous methods, our approach subdivides tall cells horizontally, allowing for more aggressive adaptivity. We demonstrate that this novel form of adaptivity delivers superior performance compared to traditional uniform tall cells.



Paperid:23
Authors:Duowen Chen,Junwei Zhou,Bo Zhu,Duowen Chen
Georgia Institute of Technology,University of Michigan,Georgia Institute of Technology,Georgia Institute of Technology
Abstract:
We propose Neural PLS, a neural particle level-set method for tracking and evolving dynamic neural representations. Oriented particles serve as interface trackers and sampling seeders, enabling efficient evolution on a multi-resolution grid-hash structure. Our approach integrates traditional PLS and implicit neural representations, achieving superior performance in benchmarks and physical simulations.



Paperid:24
Authors:Naoki Agata,Takeo Igarashi
The University of Tokyo,The University of Tokyo
Abstract:
Metric-Aligning Motion Matching (MAMM) is a novel method for controlling motion sequences using sketches, labels, audio, or another motion sequence without requiring training or annotations. By aligning within-domain distances, MAMM provides a flexible and efficient solution for motion control across various control modalities.



Paperid:25
Authors:Bowen Zheng,Ke Chen,Yuxin Yao,Zijiao Zeng,Xinwei Jiang,He Wang,Joan Lasenby,Xiaogang Jin
Zhejiang University,Zhejiang University,University of Cambridge,Tencent Games Digital Content Technology Center,Tencent Games Digital Content Technology Center,UCL Centre for Artificial Intelligence, Department of Computer Science, University College London,University of Cambridge,Zhejiang University
Abstract:
We present AutoKeyframe, a novel framework that simultaneously accepts dense and sparse control signals for motion generation by generating keyframes directly. Our method reduces manual efforts for keyframing while maintaining precise controllability, using an autoregressive diffusion model and a new skeleton-based gradient guidance method for flexible spatial constraints.



Paperid:26
Authors:Arjun Lakshmipathy,Jessica Hodgins,Nancy Pollard,Arjun Lakshmipathy
Carnegie Mellon University,Carnegie Mellon University,Carnegie Mellon University,Carnegie Mellon University
Abstract:
We present a simple, but effective framework for kinematically retargeting contact-rich anthropomorphic hand-object manipulations by exploiting contact areas. We reliably retarget contact area data between diverse hands using a novel non-isometric shape matching process and generate high quality results using the retargeted contacts alongside a straightforward IK-based motion synthesis pipeline.



Paperid:27
Authors:Inbar Gat,Sigal Raab,Guy Tevet,Yuval Reshef,Amit Haim Bermano,Daniel Cohen-Or
Tel Aviv University,Tel Aviv University,Tel Aviv University,Tel Aviv University,Tel Aviv University,Tel Aviv University
Abstract:
AnyTop generates motion for diverse character skeletons using only skeletal structure as input. This diffusion model incorporates topology information and textual joint descriptions to learn semantic correspondences across different skeletons. It generalizes with minimal training examples and supports joint correspondence, temporal segmentation, and motion editing tasks.



Paperid:28
Authors:Kemeng Huang,Xinyu Lu,Huancheng Lin,Taku Komura,Minchen Li,Kemeng Huang
Carnegie Mellon University,TransGP,Carnegie Mellon University,The University of Hong Kong,Carnegie Mellon University,University of Hong Kong, Carnegie Mellon University
Abstract:
We present a GPU-optimized IPC framework achieving up to 10× speedup across soft, stiff, and hybrid simulations. Key innovations include a connectivity-enhanced MAS preconditioner, a parallel-friendly inexact strain limiting energy, and a hash-based two-level reduction strategy for fast Hes-sian assembly and efficient affine-deformable coupling.



Paperid:29
Authors:Lei Lan,Zixuan Lu,Chun Yuan,Weiwei Xu,Hao Su,Huamin Wang,Chenfanfu Jiang,Yin Yang
University of Utah,University of Utah,University of Utah,State Key Lab of CAD&CG, Zhejiang University, China,UCSD,Style3D Research,UCLA,University of Utah
Abstract:
This paper presents a new GPU simulation algorithm, which converges as fast as global Newton's method and as efficient as Jacobi method.



Paperid:30
Authors:Alvin Shi,Haomiao Wu,Theodore Kim
Yale University,Yale University,Yale University
Abstract:
Ever feel like three dimensions isn't quite enough? We performed the analysis necessary to simulate the motion of deformables in four spatial dimensions! Along the way, we developed techniques for generating simulation-ready hyper-meshes, analyzing hyper-dimensional deformation energies, and detecting and responding to collision scenarios for softbodies in any dimension.



Paperid:31
Authors:Chengxu Zuo,Jiawei Huang,Xiao Jiang,Yuan Yao,Xiangren Shi,Rui Cao,Xinyu Yi,Feng Xu,Shihui Guo,Yipeng Qin
Xiamen University,Xiamen University,Xiamen University,Xiamen University,Bournemouth University,Xiamen University,Tsinghua University,Tsinghua University,Xiamen University,Cardiff University
Abstract:
This work proposes a dynamic calibration system for inertial motion capture, which can dynamically remove non-static IMU drift and sensor-body offset during usage, enable user-friendly calibration (without T-pose and IMU heading reset), and ensure long-term robustness.



Paperid:32
Authors:Rafael Wampfler,Chen Yang,Dillon Elste,Nikola Kovacevic,Philine Witzig,Markus Gross
ETH Zürich,ETH Zürich,ETH Zürich,ETH Zürich,ETH Zürich,ETH Zürich
Abstract:
We present a platform for creating believable, conversational digital characters that combine conversational AI, speech, animation, memory, personality, and emotions. Demonstrated through Digital Einstein, our system enables interactive, story-driven experiences and generalizes to any character, making immersive, AI-powered character experiences more accessible than ever.



Paperid:33
Authors:Fengqi LIU,Longji Huang,Zhengyu Huang,Zeyu Wang
The Hong Kong University of Science and Technology (Guangzhou),The Hong Kong University of Science and Technology (Guangzhou),The Hong Kong University of Science and Technology (Guangzhou),The Hong Kong University of Science and Technology (Guangzhou)
Abstract:
We present an image-to-image drawing setup capturing eye tracking and stroke data across 156 drawings from 10 artists. Our findings reveal consistent fixation patterns, strong gaze–stroke correlations, and structured drawing sequences, offering new insights into professional observation strategies and observation-guided assistive drawing system design.



Paperid:34
Authors:Zhiming Hu,Daniel Haeufle,Syn Schmitt,Andreas Bulling
University of Stuttgart,University of Tuebingen,University of Stuttgart,University of Stuttgart
Abstract:
We present HOIGaze – a novel approach for gaze estimation during hand-object interactions in extended reality. HOIGaze features: 1) a novel hierarchical framework that first recognises attended hand and then estimates gaze; 2) a new gaze estimator combining CNN, GCN, and cross-modal Transformers; and 3) a novel eye-head coordination loss.



Paperid:35
Authors:DongHeun Han,Byungmin Kim,RoUn Lee,KyeongMin Kim,Hyoseok Hwang,HyeongYeop Kang
Kyung Hee University,Korea University,Kyung Hee University,Korea University,Kyung Hee University,Korea University
Abstract:
ForceGrip is a reference-free reinforcement learning-based agent for realistic VR hand manipulation. It uses a progressive curriculum (finger positioning, intention adaptation, dynamic stabilization) and physics simulation to convert VR controller inputs into faithful grip forces. In user studies, participants achieved higher realism and precise control than competitors, ensuring immersive interaction.



Paperid:36
Authors:Qingqin Liu,Ziqi Fang,Jiayi Wu,Shaoyu Cai,Jianhui Yan,Tiande Mo,Shuk Ching CHAN,Kening Zhu
School of Creative Media, City University of Hong Kong,School of Creative Media, City University of Hong Kong,School of Creative Media, City University of Hong Kong,National University of Singapore,School of Creative Media, City University of Hong Kong,Hong Kong Productivity Council,Hong Kong Productivity Council,City University of Hong Kong
Abstract:
This paper presents VirCHEW Reality, a face-worn haptic device for virtual food intake in VR. It uses pneumatic actuation to simulate food textures, enhancing the chewing experience. User studies demonstrated its effectiveness in providing distinct kinesthetic feedback and improving virtual eating experiences, with applications in dining, healthcare, and entertainment.



Paperid:37
Authors:Nicolas Violante,Andréas Meuleman,Alban Gauthier,Fredo Durand,Thibault Groueix,George Drettakis
INRIA, Université Côte d'Azur,INRIA, Université Côte d'Azur,INRIA, Université Côte d'Azur,Massachusetts Institute of Technology (MIT),Adobe Research,INRIA, Université Côte d'Azur
Abstract:
We leverage repetitions in 3D scenes to improve reconstruction in low-quality parts due to poor coverage and occlusions. Our methods segments the repetitions, registers them together, and optimizes a shared representation with multi-view information flowing from all repetitions, improving the reconstruction of each individual repetition.



Paperid:38
Authors:Sara Sabour,Lily Goli,George Kopanas,Mark Matthews,Dmitry Lagun,Leonidas Guibas,Alec Jacobson,David Fleet,Andrea Tagliasacchi,Sara Sabour,Lily Goli
Google Inc - Deepmind,University of Toronto,Runway,Google Inc - Deepmind,Google Inc - Deepmind,Google Inc - Deepmind,University of Toronto,Google Inc - Deepmind,Google Inc - Deepmind,Google Inc - Deepmind, University of Toronto,University of Toronto, Waabi AI
Abstract:
3D Gaussian Splatting (3DGS) enables fast 3D reconstruction and rendering but struggles with real-world captures due to transient elements and lighting changes. We introduce SpotLessSplats, which leverages semantic features from foundation models and robust optimization to remove transient effects, achieving state-of-the-art qualitative and quantitative reconstruction quality on casual scene captures.



Paperid:39
Authors:Ziyi Zhang,Nicolas Roussel,Thomas Muller,Tizian Zeltner,Merlin Nimier-David,Fabrice Rousselle,Wenzel Jakob
EPFL,EPFL,NVIDIA Research,NVIDIA Research,NVIDIA Research,NVIDIA Research,EPFL
Abstract:
We present a simple and fast method to reconstruct radiance surfaces by directly supervising the radiance field via image projection.Unlike volumetric approaches, we move alpha blending and ray marching from image formation into loss computation.This simple modification enables high-quality surface reconstruction while preserving baseline efficiency and robustness.



Paperid:40
Authors:Junkai Huang,Saswat Subhajyoti Mallick,Alejandro Amat,Marc Ruiz Olle,Albert Mosella-Montoro,Bernhard Kerbl,Francisco Vicente Carrasco,Fernando De la Torre
Carnegie Mellon University,Carnegie Mellon University,Carnegie Mellon University,Carnegie Mellon University,Carnegie Mellon University,Carnegie Mellon University,Carnegie Mellon University,Carnegie Mellon University
Abstract:
We present a revolutionary method for experiencing live sports in stunning 3D, redefining the way games are seen, through immersive, interactive replays. Alongside, we release a large-scale synthetic dataset built to benchmark realism, motion, and human interaction in dynamic scenes, to fuel the next wave of research in 3D streaming.



Paperid:41
Authors:Akshat Dave,Tianyi Zhang,Aaron Young,Ramesh Raskar,Wolfgang Heidrich,Ashok Veeraraghavan,Akshat Dave
Massachusetts Institute of Technology Media Lab,Rice University,Massachusetts Institute of Technology Media Lab,Massachusetts Institute of Technology Media Lab,King Abdullah University of Science and Technology,Rice University,Massachusetts Institute of Technology (MIT)
Abstract:
NeST enables non-destructive 3D stress analysis of transparent objects using the polarization of light. Traditional 2D methods require destructively slicing the object. Instead, we reconstruct the entire 3D stress field by jointly handling phase unwrapping and tensor tomography using neural implicit representations and inverse rendering, enabling novel 3D stress visualizations.



Paperid:42
Authors:Peng Li,Zeyong Wei,Honghua Chen,Xuefeng Yan,Mingqiang Wei
Nanjing University of Aeronautics and Astronautics,Nanjing University of Aeronautics and Astronautics,Nanjing University of Aeronautics and Astronautics,Nanjing University of Aeronautics and Astronautics,Nanjing University of Aeronautics and Astronautics
Abstract:
To combine deep learning's generalization with traditional methods' interpretability, we propose CustomBF—a hybrid framework that customizes bilateral filter components per point. By addressing key limitations of the classic bilateral filter, CustomBF achieves robust, interpretable, and effective point cloud denoising across diverse scenarios.



Paperid:43
Authors:Nathan King,Haozhe Su,Mridul Aanjaneya,Steven Ruuth,Christopher Batty,Nathan King
University of Waterloo,LightSpeed Studios,Rutgers University,Simon Fraser University,University of Waterloo,University of Waterloo
Abstract:
Geometry processing often requires the solution of PDEs with boundary conditions on the manifold’s interior. However, input manifolds can take many forms, each requiring specialized discretizations. Instead, we develop a unified framework for general manifold representations by extending the closest point method to handle interior boundary conditions.



Paperid:44
Authors:Zhuodong Li,Fei Hou,Wencheng Wang,Xuquan Lu,Ying He
Institute of Software, Chinese Academy of Sciences,Institute of Software, Chinese Academy of Sciences,Institute of Software, Chinese Academy of Sciences,The University of Western Australia,Nanyang Technological University
Abstract:
We propose a divide-and-conquer approach for orienting large-scale, non-watertight point clouds. The scene is first segmented into blocks, and normal orientations are estimated independently within each block. These local orientations are then globally unified through a graph-based formulation, solved via 0-1 integer optimization. Experiments demonstrate the robustness of our method.



Paperid:45
Authors:Daniel Scrivener,Daniel Cui,Ellis Coldren,Mazdak Abulnaga,Mikhail Bessmeltsev,Edward Chien
Boston University,Boston University,Boston University,Massachusetts Institute of Technology (MIT),Universite de Montreal,Boston University
Abstract:
We propose a novel method for normal estimation of unoriented point clouds and VR ribbon sketches that leverages a modeling of the Faraday cage effect. Our method is uniquely robust to the presence of interior structures and artifacts, producing superior surfacing output when combined with Poisson Surface Reconstruction.



Paperid:46
Authors:Max Kohlbrenner,Marc Alexa
Technical University of Berlin,Technical University of Berlin
Abstract:
Spheres that are disjoint from a given union of spheres can be computing by solving a convex hull problem. This can be exploited for contouring discretely sampled signed distance functions.



Paperid:47
Authors:Yucheol Jung,Hyomin Kim,Hyejeong Yoon,Yoonha Hwang,Seungyong Lee
POSTECH,POSTECH,POSTECH,POSTECH,POSTECH
Abstract:
We propose a novel ICP framework that jointly optimizes a shared template and instance-wise deformations. Our approach automatically captures common shape features from input shapes, achieving state-of-the-art accuracy and consistency while eliminating the need to carefully select a preset template mesh.



Paperid:48
Authors:Diyang Zhang,Zhendong Wang,Zegao Liu,Xinming Pei,Weiwei Xu,Huamin Wang
Style3D Research,Style3D Research,Style3D Research,State Key Laboratory of CAD & CG, Zhejiang University,State Key Laboratory of CAD & CG, Zhejiang University,Style3D Research
Abstract:
We propose a method to estimate optimal cloth mesh resolution based on material stiffness and boundary conditions like shirring or stitching, and dynamic wrinkles from motion-induced collisions. To ensure smooth resolution transitions, we calculate transition distances and generate a mesh sizing map, enhancing realism, efficiency, and versatility for garment design.



Paperid:49
Authors:Chengzhu He,Zhendong Wang,Zhaorui Meng,Junfeng Yao,Shihui Guo,Huamin Wang
Xiamen University,Style3D Research,Xiamen University,Xiamen University,Xiamen University,Style3D Research
Abstract:
This paper introduces an automated scheduling framework to optimize cloth and deformable simulations across heterogeneous computing devices. Using an enhanced HEFT algorithm and asynchronous iteration methods, our approach minimizes communication delays and maximizes parallelism. our experiments demonstrate superior frame rates over single-unit solutions for real-time and resource-constrained environments.



Paperid:50
Authors:Zixuan Lu,Ziheng Liu,Lei Lan,Huamin Wang,Yuko Ishiwaka,Chenfanfu Jiang,Kui Wu,Yin Yang
University of Utah,University of Utah,University of Utah,Style3D Research,SoftBank,UCLA,LightSpeed Studios,University of Utah
Abstract:
This paper presents a CPU-based cloth simulation algorithm that partitions garment models into domains that can be processed by each individual CPU core. Using projective dynamics with domain-level parallelization, this method achieves high performance comparable to GPU methods and runs about an order faster than existing CPU approaches.



Paperid:51
Authors:Tao Huang,Haoyang Shi,Mengdi Wang,Yuxing Qiu,Yin Yang,Kui Wu
LightSpeed Studios,University of Utah,LightSpeed Studios,LightSpeed Studios,University of Utah,LightSpeed Studios
Abstract:
In this work, we introduce the first real-time framework that integrates yarn-level simulation with fiber-level rendering. The whole system provides real-time performance and has been evaluated through various application scenarios, including knit simulation for small patches and full garments and yarn-level relaxation in the design pipeline.



Paperid:52
Authors:Jiayi Eris Zhang,Doug James,Danny Kaufman
Stanford University,Stanford University,Adobe
Abstract:
We propose a general framework, Progressive Dynamics++, for constructing a family of progressive dynamics integration methods that advance physical simulation states forward in both time and spatial resolution. We analyze requirements for stable, continuous, and consistent level-of-detail animation and introduce a novel, stable method that significantly improves temporal continuity.



Paperid:53
Authors:Yue Chang,Mengfei Liu,Zhecheng Wang,Peter Yichen Chen,Eitan Grinspun
University of Toronto,University of Toronto,University of Toronto,MIT CSAIL,University of Toronto
Abstract:
We designed a neural field capable of capturing a diverse family of discontinuities, enabling the simulation of cuts in thin-walled deformable structures. By lifting input coordinates using generalized winding numbers, our approach models discontinuities explicitly and flexibly, supporting accurate, real-time simulations with dynamic cut updates and user-interactive cut shape design.



Paperid:54
Authors:Minseok Chae,Chun Chen,Seung-Woo Nam,Yoonchan Jeong
Seoul National University,Seoul National University,Seoul National University,Seoul National University
Abstract:
We present and analyze a holographic augmented reality display with the bandwidth-preserved guiding method using a light pipe. We propose the use of light pipe to spatially relocate the light engine from the image combiner at the front-module, enabling enhanced weight distribution and obstruction-free view while preserving the wavefront bandwidth.



Paperid:55
Authors:Florian Schiffers,Grace Kuo,Nathan Matsuda,Douglas Lanman,Oliver Cossairt,Florian Schiffers,Oliver Cossairt
Amazon Prime Video,Meta Reality Labs,Meta Reality Labs,Meta Reality Labs,Meta Reality Labs,Northwestern University; Reality Labs Research, Meta,Northwestern University, Facebook
Abstract:
HoloChrome introduces a novel holographic display method by multiplexing multiple wavelengths and two spatial light modulators to enhance image quality. By moving beyond standard three-color primary systems, it significantly reduces speckle noise while preserving natural depth cues while achieving more accurate color reproduction.



Paperid:56
Authors:Peter Michael,Zekun Hao,Serge Belongie,Abe Davis,Peter Michael
Cornell University,Cornell Tech,University of Copenhagen,Cornell University,Cornell University
Abstract:
Video forensics, which focuses on identifying fake or manipulated video, is becoming increasingly difficult with the development of more advanced video editing techniques. We show how coding near-imperceptible, noise-like modulations into the illumination of a scene can create information asymmetry that favors forensic verification of video captured from that scene.



Paperid:57
Authors:Jipeng Sun,Kaixuan Wei,Thomas Eboli,Congli Wang,Cheng Zheng,Zhihao Zhou,Arka Majumdar,Wolfgang Heidrich,Felix Heide
Princeton University,King Abdullah University of Science and Technology (KAUST),Université Paris-Saclay,Princeton University,Princeton University,University of Washington,University of Washington,King Abdullah University of Science and Technology (KAUST),Princeton University
Abstract:
We introduce a collaborative metalens array comprising over 100-million nanopillars for broadband imaging. The proposed array camera is only a few millimeters flat and employs a non-generative reconstruction method, which performs favorably and without hallucinations, irrespective of the scene illumination spectrum.



Paperid:58
Authors:Juhyeon Kim,Craig Benko,Magnus Wrenninge,Ryusuke Villemin,Zeb Barber,Wojciech Jarosz,Adithya Pediredla
Dartmouth College,Aurora Innovation,Aurora Innovation,Aurora Innovation,Aurora Innovation,Dartmouth College,Dartmouth College
Abstract:
We present a general spectral-domain simulation framework for optical heterodyne detection (OHD), extending path integral rendering to capture power spectral density of OHD. Unlike existing domain-specific tools, our approach supports diverse scenes and applications. We validate it against real-world data from FMCW lidar, blood flow velocimetry, and wind Doppler lidar.



Paperid:59
Authors:Suyeon Choi,Brian Chao,Jacqueline Yang,Manu Gopakumar,Gordon Wetzstein
Stanford University,Stanford University,Stanford University,Stanford University,Stanford University
Abstract:
We develop novel and efficient computer-generated holography algorithms, dubbed Gaussian Wave Splatting, that transform Gaussian-based scene representations into holograms. We derive a closed-form 2D Gaussian-to-hologram transform supporting occlusions and alpha blending, along with an efficient, easily parallelizable Fourier-domain approximation of this process, implemented with custom CUDA kernels.



Paperid:60
Authors:Xinyu Ma,Hengyu Meng,Ziwei Wu,Zeyu Wang,Clea von Chamier-Waite
The Hong Kong University of Science and Technology (Guangzhou),The Hong Kong University of Science and Technology (Guangzhou),The Hong Kong University of Science and Technology,The Hong Kong University of Science and Technology (Guangzhou),The Hong Kong University of Science and Technology (Guangzhou)
Abstract:
"Becoming Space" is an installation that explores the agency of AI, discourse, and material intersections through AI-generated forms and 3D printing. Inspired by Ovid's Metamorphoses, it explores human-animal transformations using CLIP-guided diffusion models and stereolithography. The installation reveals limitations in AI's physical form interpretation---which is dominated by discourse---while demonstrating "intra-action" between language, algorithms, machines, and materials. This work tries to discuss authorship and material agency through the entanglement of matters.



Paperid:61
Authors:Troy TianYu LIN,Boyan Zheng,Haichuan Lin,Wen You,Kang Zhang,Chen Liang
Hong Kong University of Science and Technology, Guangzhou,Independent Researcher,Hong Kong University of Science and Technology, Guangzhou,Hong Kong University of Science and Technology, Guangzhou,Hong Kong University of Science and Technology, Guangzhou,Hong Kong University of Science and Technology, Guangzhou
Abstract:
Transforming 2D Chinese calligraphy into 3D forms deepens how traditional art is understood and experienced, combining cultural heritage with modern technology. This approach adds spatial depth, opening new possibilities for digital art, preservation, and interactive design.By merging computational modeling with artistic expression, the work explores how technology can reinterpret historical artforms, encouraging cross-disciplinary dialogue and inspiring new ways to preserve and evolve intangible cultural heritage.



Paperid:62
Authors:Tatuki Hayama,Kotaro Uchibe
Keio University,Hosoo Co.,Ltd
Abstract:
This paper presents a procedural data generation method for Jacquard weaving that uses matrix computations to create textiles with complex shaded patterns and a triple-layer structure. Employing this method, the authors creatively applied noise functions to weave design and produced a textile artwork in collaboration with a traditional craft technique.



Paperid:63
Authors:Yuxuan Han,Junfeng Lyu,Kuan Sheng,Minghao Que,Qixuan Zhang,Lan Xu,Feng Xu
Tsinghua University,Tsinghua University,ShanghaiTech University,Tsinghua University,ShanghaiTech University,ShanghaiTech University,Tsinghua University
Abstract:
Given a single co-located smartphone video captured in a dim room as the input, our method can reconstruct high-quality facial assets within the distribution modeled by a diffusion prior trained on Light Stage scans, which can be exported to common graphics engines like Blender for photo-realistic rendering.



Paperid:64
Authors:Yiqian Wu,Malte Prinzler,Xiaogang Jin,Siyu Tang
ETH Zürich,ETH Zürich,State Key Lab of CAD and CG, Zhejiang University,ETH Zürich
Abstract:
AnimPortrait3D is a novel method for text-based, realistic, animatable 3DGS avatar generation with morphable model alignment. To address ambiguities in diffusion predictions during 3D distillation, we introduce key strategies: initializing a 3D avatar with robust appearance and geometry, and leveraging a ControlNet to ensure accurate alignment with the underlying model.



Paperid:65
Authors:Yisheng He,Xiaodong Gu,Xiaodan Ye,Chao Xu,Zhengyi Zhao,Yuan Dong,Weihao Yuan,Zilong Dong,Liefeng Bo
Alibaba Group,Alibaba Group,Alibaba Group,Alibaba Group,Alibaba Group,Alibaba Group,Alibaba Group,Alibaba Group,Alibaba Group
Abstract:
LAM is an innovative Large Avatar Model for animatable Gaussian head reconstruction from a single image in seconds. Our Gaussian heads are immediately animatable and renderable without additional networks or post-processing. This allows seamless integration into existing rendering pipelines, ensuring real-time animation and rendering across various platforms, including mobile phones.



Paperid:66
Authors:Tingting Liao,Yujian Zheng,Yuliang Xiu,Adilbek Karmanov,Liwen Hu,Leyang Jin,Hao Li
Mohamed bin Zayed University of Artificial Intelligence,Mohamed bin Zayed University of Artificial Intelligence,Westlake University,Mohamed bin Zayed University of Artificial Intelligence,Pinscreen,Mohamed bin Zayed University of Artificial Intelligence,Mohamed bin Zayed University of Artificial Intelligence
Abstract:
SOAP awakens the 3D princess from 2D stylized photos. Unlike other works that directly drive the 2D photos, SOAP reconstructs well-rigged 3D avatars, with detailed geometry and all-around texture, from just a single stylized picture.



Paperid:67
Authors:Chengan He,Junxuan Li,Tobias Kirschstein,Artem Sevastopolskiy,Shunsuke Saito,Qingyang Tan,Javier Romero,Chen Cao,Holly Rushmeier,Giljoo Nam
Yale University,Meta Codec Avatars Lab,Technical University of Munich,Technical University of Munich,Meta Codec Avatars Lab,Meta Codec Avatars Lab,Meta Codec Avatars Lab,Meta Codec Avatars Lab,Yale University,Meta Codec Avatars Lab
Abstract:
We present 3DGH, a generative model that creates realistic 3D human heads with composable hair and face components. By modeling both the separation and correlation between hair and face in a generative paradigm, it enables high-quality, full-head synthesis and flexible 3D hairstyle editing with strong visual consistency and realism.



Paperid:68
Authors:Youyang Du,Lu Wang,Beibei Wang
Shandong University,Shandong University,Nanjing University
Abstract:
Our framework can efficiently synthesize facial microstructure from an unconstrained facial image via differentiable optimization. We propose neural wrinkle simulation for differentiable microstructure parameterization, and direction distribution similarity to align features with blurry image patches. Our framework is also compatible with existing facial reconstruction methods for detail enhancement.



Paperid:69
Authors:Giulio Viganò,Maks Ovsjanikov,Simone Melzi
Università di Milano Bicocca,Centre National de la Recherche Scientifique - Laboratoire d'informatique de l'École Polytechnique (LIX),Università di Milano Bicocca
Abstract:
We introduce Neural Adjoint Maps, a novel representation for correspondences between 3D shapes. Built on and extending the functional map framework, our approach enables accurate, non-linear refinement of shape matching across meshes and point clouds, setting a new standard in diverse scenarios and applications like graphics and medical imaging.



Paperid:70
Authors:Yuezhi Yang,Haitao Yang,George Kiyohiro Nakayama,Xiangru Huang,Leonidas Guibas,Qixing Huang
University of Texas at Austin,University of Texas at Austin,Stanford University,Westlake University,Stanford University,University of Texas at Austin
Abstract:
We present GenAnalysis, an implicit shape generation framework enabling joint shape matching and consistent segmentation by enforcing as-affine-as-possible (AAAP) deformations via regularization loss in latent space. It enables shape analysis via extracting and analysing shape variations in the tangent space. Experiments on ShapeNet demonstrate improved performance over existing methods.



Paperid:71
Authors:Changhao Li,Yu Xin,Xiaowei Zhou,Ariel Shamir,Hao Zhang,Ligang Liu,Ruizhen Hu
University of Science and Technology of China,University of Science and Technology of China,State Key Laboratory of CAD & CG, Zhejiang University,Reichman University,Simon Fraser University,University of Science and Technology of China,Shenzhen University
Abstract:
We introduce Masked Anchored SpHerical Distances (MASH), a novel multi-view and parametrized representation of 3D shapes. MASH is versatilefor multiple applications including surface reconstruction, shape generation, completion, and blending, achieving superior performance thanks to its unique representation encompassing both implicit and explicit features.



Paperid:72
Authors:Chunyi Sun,Junlin Han,Runjia Li,Weijian Deng,Dylan Campbell,Stephen Gould
Australian National University,University of Oxford,University of Oxford,Australian National University,Australian National University,Australian National University
Abstract:
3D2EP transforms 3D shapes into expressive, editable primitives by extruding 2D profiles along 3D curves. This approach creates compact, interpretable representations that support intuitive editing and flexible redesign. It delivers high fidelity and efficiency, outperforming existing methods across digital design, asset creation, and customization workflows.



Paperid:73
Authors:Si-Tong Wei,Rui-Huan Wang,Chuan-Zhi Zhou,Baoquan Chen,Peng-Shuai Wang
Peking University,Peking University,Peking University,Peking University,Peking University
Abstract:
OctGPT is a novel multiscale autoregressive model for 3D shape generation. It introduces hierarchical serialized octree representation, octree-based transformer with 3D RoPE and token-parallel generation schemes. OctGPT significantly accelerates convergence, achieves performance rivaling or surpassing state-of-the-art diffusion models, and supports text/sketch/image-conditioned generation and scene-level synthesis.



Paperid:74
Authors:Longwen Zhang,Qixuan Zhang,Haoran Jiang,Yinuo Bai,Wei Yang,Lan Xu,Jingyi Yu
ShanghaiTech University,ShanghaiTech University,ShanghaiTech University,ShanghaiTech University,Huazhong University of Science and Technology,ShanghaiTech University,ShanghaiTech University
Abstract:
BANG introduces Generative Exploded Dynamics, a novel method that dynamically decomposes 3D objects into meaningful, volumetric parts through smooth, controllable exploded views. Bridging intuitive human understanding and generative AI, it enables precise part-level manipulation, semantic comprehension, and versatile applications in 3D creation, visualization, and printing workflows.



Paperid:75
Authors:Karen Furuta,Jingjing Li,Tatsuki Fushimi,Yoichi Ochiai
University of Tsukuba,University of Tsukuba,University of Tsukuba,University of Tsukuba
Abstract:
This work explores “qi” in kendo through mixed reality and autoethnography, blending tradition and technology. By animating digital humans with “qi”, it frames martial arts as art. The project invites reflection on selfhood and offers fresh insights at the intersection of culture, embodiment, and digital experience.



Paperid:76
Authors:Néill O'Dwyer,Enda Bates,Nicholas Johnson
Trinity College Dublin,Trinity College Dublin,Trinity College Dublin
Abstract:
This practice-based project reimagines Beckett’s Not I in virtual reality, marrying minimalist theatre with immersive technology. A lone, disembodied Metahuman mouth exploits VR’s intense presence while subverting customary embodiment and audience agency. Integrating performing avatars, the work probes authenticity, identity, and authorship, demonstrating how “subtractive dramaturgy” thrives in an additive medium. Findings advance performance studies, XR design, and digital humanities by showing how technology reshapes creativity, embodiment, and storytelling.



Paperid:77
Authors:Yihua Li,Hongyue Chen,Yiqing Li,Yetong Xin
New York University,English Literature and Literary Theory at the University of Freiburg,New York University,Harvard University
Abstract:
PoeSpin is a human-AI cocreating system. By transforming pole dance movements into poetry through AI, we challenge both traditional prejudices against this art form and conventional approaches to human-AI creativity. This work demonstrates how computational systems can preserve the deeply human aspects of artistic expression while creating new possibilities for cross-modal artistic collaboration, suggesting pathways for more inclusive and expressive forms of human-AI co-creation.



Paperid:78
Authors:Kei Iwasaki,Yoshinori Dobashi
Saitama University,Hokkaido University
Abstract:
We introduce spherical harmonics Hessian and solid spherical harmonics, a variant of spherical harmonics, to compute the spherical harmonics Hessian efficiently and accurately to the computer graphics community. These mathematical tools are used to develop an analytical representation of the Hessian matrix of spherical harmonics coefficients for spherical lights.



Paperid:79
Authors:Jiawei Huang,Shaokun Zheng,Kun Xu,Yoshifumi Kitamura,Jiaping Wang
International Digital Economy Academy,Tsinghua University,Tsinghua University,Tohoku University,International Digital Economy Academy
Abstract:
A guided lens sampling technique that improves Monte Carlo rendering of depth-of-field by projecting a global 3D radiance field into lens space via bipolar-cone projection. This method efficiently targets high-contribution regions, significantly reducing noise and improving convergence for circle-of-confusion effects in production rendering.



Paperid:80
Authors:Zhimin Fan,Chen Wang,Yiming Wang,Boxuan Li,Yuxuan Guo,Ling-Qi Yan,Yanwen Guo,Jie Guo
Nanjing University,Nanjing University,Nanjing University,Nanjing University,Nanjing University,University of California Santa Barbara,Nanjing University,Nanjing University
Abstract:
We derive vertex position and irradiance bounds for each triangle tuple, introducing a bounding property of rational functions on the Bernstein basis, to significantly reduce the search domain when systematically simulating specular light transport.



Paperid:81
Authors:Jeffrey Liu,Daqi Lin,Markus Kettunen,Chris Wyman,Ravi Ramamoorthi
University of Illinois Urbana-Champaign,NVIDIA,NVIDIA,NVIDIA,NVIDIA
Abstract:
We introduce reservoir splatting, a technique preserving exact primary hits during temporal ReSTIR. This approach makes temporal path resampling more robust under motion, especially for regions with high-frequency detail. We further demonstrate how reservoir splatting naturally enables ReSTIR support for both motion blur and depth of field.



Paperid:82
Authors:Xuejun Hu,Jinfan Lu,Kun Xu
Tsinghua University,Tsinghua University,Tsinghua University
Abstract:
We present a novel shadow method named kernel predicting neural shadow mapping. By modeling soft shadow values as pixelwise local filtering from basic hard shadow values, we trained a neural network to predict local filter weights, achieving accurate and temporally-stable soft shadows with good generalizability.



Paperid:83
Authors:Naoto Shirashima,Hideki Todo,Yuki Yamaoka,Shizuo Kaji,Kunihiko Kobayashi,Haruna Shimotahira,Yonghao Yue
AGU,Takushoku University,AGU,Kyushu University,AGU,AGU,AGU
Abstract:
We present a stroke-based method for transforming dynamic 3D scenes with smoke, fire, or clouds into painterly animations. Learning from user-provided exemplars, our system transfers stroke styles—color, width, length, and orientation—while preserving motion and structure. This enables expressive and coherent renderings of complex volumetric media.



Paperid:84
Authors:Venkataram Edavamadathil Sivaram,Ravi Ramamoorthi,Tzu-Mao Li
University of California San Diego,University of California San Diego,University of California San Diego
Abstract:
This work presents a physically-based model for simulating and rendering glow discharge, a luminous plasma effect seen in neon lights and gas discharge lamps. The model captures particle interactions and emission dynamics, integrates into volume rendering systems, and enables realistic, interactive visualizations of complex light phenomena.



Paperid:85
Authors:Karran Pandey,Anita Hu,Clement Fuji Tsang,Or Perel,Karan Singh,Maria Shugrina
University of Toronto,NVIDIA,NVIDIA,NVIDIA,University of Toronto,NVIDIA
Abstract:
We present the first interactive system for painting with 3D Gaussian splat brushes. With our tool, artists can sample volumetric fragments from real-world Gaussian splat captures and paint with them in real time. Our tool seamlessly deforms sampled splats along painted strokes, introducing realistic transitions between seams with diffusion inpainting.



Paperid:86
Authors:Kenneth Chen,Nathan Matsuda,Jon McElvain,Yang Zhao,Thomas Wan,Qi Sun,Alexandre Chapiro
New York University,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,New York University,Reality Labs Research, Meta
Abstract:
We studied preferences for different contrasts and peak luminances in HDR. To do this, we collected a new HDR video dataset, developed tone mappers, and built an HDR haploscope that can reproduce high luminance and contrast. Data was fit to a model which is used for applications like display design.



Paperid:87
Authors:Avinab Saha,Yu-Chih Chen,Jean-Charles Bazin,Christian Häne,Ioannis Katsavounidis,Alexandre Chapiro,Alan Bovik
University of Texas Austin,University of Texas Austin,Reality Labs, Meta,Reality Labs, Meta,Reality Labs, Meta,Reality Labs, Meta,University of Texas Austin
Abstract:
We introduce FaceExpressions-70k, a large-scale dataset comprising 70,500 crowdsourced comparisons of facial expressions collected from over 1,000 participants. It supports the training of perceptual models for expression differences and helps guide decisions on acceptable latency and sampling rates for facial expressions when driving a face avatar.



Paperid:88
Authors:Seonghyeon Kim,Chang Wook Seo,Kwanggyoon Seo,Seung Han Song,Junyong Noh
KAIST, Visual Media Lab,Anigma Technologies,KAIST, Visual Media Lab,Chungnam National University Hospital,KAIST, Visual Media Lab
Abstract:
A novel deep learning system enhances realistic virtual oculoplastic surgery simulations.



Paperid:89
Authors:Sophie Kergaßner,Taimoor Tariq,Piotr Didyk
Università della Svizzera italiana,Università della Svizzera italiana,Università della Svizzera italiana
Abstract:
We demonstrate that stereoacuity is remarkably resilient to foveated rendering and remains unaffected with up to 2× stronger foveation than commonly used. To this end, we design a psychovisual experiment and derive a simple perceptual model that determines the amount of foveation that does not affect stereoacuity.



Paperid:90
Authors:Edward Lu,Anthony Rowe
Carnegie Mellon University,Carnegie Mellon University
Abstract:
This paper introduces an improved quad-based geometry streaming method for remote rendering that reduces bandwidth demands through temporal compression and supports QoE-driven adaptation. It achieves high-quality visuals, captures disocclusion events, uses 15× less data than SOTA, and reduces bandwidth down to 100 Mbps, enabling real-time, low-latency rendering on lightweight headsets.



Paperid:91
Authors:Zhaofeng Luo,Zhitong Cui,Shijian Luo,Mengyu Chu,Minchen Li
Carnegie Mellon University,Carnegie Mellon University,Zhejiang University,Peking University,Carnegie Mellon University
Abstract:
VR-Doh, an intuitive VR-based 3D modeling system that lets you sculpt and manipulate soft objects and edit 3D Gaussian Splatting scenes in real time. Combining physics-based simulation and expressive interaction, VR-Doh empowers both novices and experts to create rich, deformable, simulation-ready models with natural hand-based input.



Paperid:92
Authors:Yue Chang,Otman Benchekroun,Maurizio M. Chiaramonte,Peter Yichen Chen,Eitan Grinspun
University of Toronto,University of Toronto,Meta Reality Labs Research,MIT CSAIL,University of Toronto
Abstract:
We introduce shape-space eigenanalysis to compute eigenfunctions across continuously-parameterized shape families. These eigenfunctions are obtained by minimizing a variational principle. To handle eigenvalue dominance swaps at points of multiplicity, we incorporate dynamic reordering during optimization. The method is discretization-agnostic and differentiable, enabling applications in sound synthesis, locomotion, and elastodynamic simulation.



Paperid:93
Authors:Theo Braune,Mark Gillespie,Yiying Tong,Mathieu Desbrun
Centre National de la Recherche Scientifique - Laboratoire d'informatique de l'École Polytechnique (LIX),Inria Saclay,Michigan State University,Inria Saclay
Abstract:
Although discrete connections are ubiquitous in vector field design, their torsion remains unstudied. We extend the existing toolbox to control the torsion of discrete connections: we introduce a new discrete Levi-Civita connection and define torsion as a measure of deviation from this reference, so torsion becomes a simple linear constraint.



Paperid:94
Authors:Haikuan Zhu,Hongbo Li,Hsueh-Ti Derek Liu,Wenping Wang,Jing Hua,Zichun Zhong
Wayne State University,Wayne State University,Roblox,Texas A&M University,Wayne State University,Wayne State University
Abstract:
Our method proposes a novel computational design framework for designing anisotropic tensor fields. It enables flexible control over scalings without requiring users to specify orientations explicitly. We apply these anisotropic tensor fields to various applications, such as anisotropic meshing, structural mechanics, and fabrication.



Paperid:95
Authors:Yiling Pan,Zhixin Xu,Bin Wang,Bailin Deng
Tsinghua University,Tsinghua University,Tsinghua University,Cardiff University
Abstract:
We propose a method to approximate arbitrary freeform surface meshes with piecewise ruled surfaces. Our approach optimizes mesh shape and ruling direction field simultaneously, extracts patch topology, and refines ruling positions and orientations. The technique effectively approximates diverse freeform shapes and has potential applications in architecture and engineering.



Paperid:96
Authors:Ana Dodik,Isabella Yu,Kartik Chandra,Jonathan Ragan-Kelley,Joshua Tenenbaum,Vincent Sitzmann,Justin Solomon
Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT),Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT),Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT),Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT),Massachusetts Institute of Technology (MIT),Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT),Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT)
Abstract:
Meschers are a mesh representation for Escheresque geometry. They allow us to solve partial differential equations on the surface of an impossible object, meaning that we can find impossible shortest paths, perform mescher smoothing, and even inverse render a mescher from an image.



Paperid:97
Authors:Yuou Sun,Bailin Deng,Juyong Zhang,Yuou Sun
University of Science and Technology of China,Cardiff University,University of Science and Technology of China,University of Science and Technology of China
Abstract:
Designing freeform surfaces to reflect or refract light to achieve target light distributions is a challenging inverse problem. We propose an end-to-end optimization strategy using a novel differentiable rendering model driven by image errors, combined with face-based optimal transport initialization and geometric constraints, to achieve high-quality final physical results.



Paperid:98
Authors:Guying Lin,Lei Yang,Congyi Zhang,Hao Pan,Yuhan Ping,Guodong Wei,Taku Komura,John Keyser,Wenping Wang,Guying Lin
Carnegie Mellon University,The University of Hong Kong,The University of British Columbia,Tsinghua University,The University of Hong Kong,The University of Hong Kong,The University of Hong Kong,Texas A&M University,Texas A&M University,Carnegie Mellon University
Abstract:
We introduce Patch-Grid, a unified neural implicit representation that efficiently represents complex shapes, preserves sharp features, and handles open boundaries and thin geometric details. By decomposing shapes into patches encapsulated by adaptive feature grids and merging them through localized CSG operations, Patch-Grid demonstrates superior robustness, efficiency, and accuracy.



Paperid:99
Authors:Dewen Guo,Zhendong Wang,Zegao Liu,Sheng Li,Guoping Wang,Yin Yang,Huamin Wang
Peking University,Style3D Research,Style3D Research,Peking University,Peking University,University of Utah,Style3D Research
Abstract:
We propose a physics-based modeling system for knots and ties using pipe-like parametric templates, defined by Bézier curves and adaptive radii for flexible, intersection-free shapes. Our method maps cloth regions from UV space into 3D knot forms via a penetration-free initialization and supports quasistatic simulation with efficient collision handling.



Paperid:100
Authors:Zizhou Huang,Chrystiano Araújo,Andrew Kunz,Denis Zorin,Daniele Panozzo,Victor Zordan
New York University,Roblox,Roblox,New York University,New York University,Roblox
Abstract:
We introduce an automatic tool to retarget artist-designed garments on a standard mannequin to possibly non-human avatars with unrealistic characteristics, which widely appear in games and animations. We preserve the geometrical features in the original design, guarantee intersection-free, and fit the garment adaptively to the avatars.



Paperid:101
Authors:Anran Qi,Nico Pietroni,Maria Korosteleva,Olga Sorkine-Hornung,Adrien Bousseau
INRIA, Université Côte d'Azur,University of Technology Sydney,ETH Zurich,ETH Zurich,INRIA, Université Côte d'Azur
Abstract:
We present the first algorithm to automatically compute sewing patterns forupcycling existing garments into new designs. Our algorithm takes as inputtwo garment designs along with their corresponding sewing patterns anddetermines how to cut one of them to match the other by following garmentreuse principles.



Paperid:102
Authors:Yuki Tatsukawa,Anran Qi,I-Chao Shen,Takeo Igarashi
The University of Tokyo,INRIA, Université Côte d'Azur,The University of Tokyo,The University of Tokyo
Abstract:
Garment sewing patterns often rely on vector formats, which struggle with discontinuities and unseen topologies. GarmentImage instead encodes geometry, topology, and placement into multi-channel grids, enabling smooth transitions and better generalization. Using simple CNNs, it works well in pattern exploration, prompt-to-pattern generation, and image-to-pattern prediction.



Paperid:103
Authors:Xuan Li,Chang Yu,Wenxin Du,Ying Jiang,Tianyi Xie,Yunuo Chen,Yin Yang,Chenfanfu Jiang
University of California Los Angeles,University of California Los Angeles,University of California Los Angeles,University of California Los Angeles,University of California Los Angeles,University of California Los Angeles,University of Utah,University of California Los Angeles
Abstract:
We introduce Dress-1-to-3 to reconstruct physics-plausible, simulation-ready separated garments from an in-the-wild image. Starting with the image, our approach combines a pre-trained image-to-sewing pattern generation model with a pre-trained multi-view diffusion model. The sewing pattern is refined using a differentiable garment simulator based on the generated multi-view images.



Paperid:104
Authors:Ren Li,Cong Cao,Corentin Dumery,Yingxuan You,Hao Li,Pascal Fua
EPFL,MBZUAI,EPFL,EPFL,MBZUAI,EPFL
Abstract:
We introduce a novel method for accurate 3D garment reconstruction from single-view images, bridging 2D and 3D representations. Our mapping model creates connections among image pixels, UV coordinates, and 3D geometry, resulting in realistic garments with intricate details and enabling downstream applications like garment retargeting and texture editing.



Paperid:105
Authors:Alon Feldman,Mirela Ben-Chen
Technion – Israel Institute of Technology,Technion – Israel Institute of Technology
Abstract:
Logarithmic metric blending enables smooth interpolation between planar shapes while bounding both conformal and area distortions. By blending symmetric positive definite metrics in the log domain, our method geometrically interpolates distortions. This leads to natural transitions that outperform existing techniques in applications such as shape morphing and animation.



Paperid:106
Authors:Shibo Liu,Ligang Liu,Xiao-Ming Fu
University of Science and Technology of China,University of Science and Technology of China,University of Science and Technology of China
Abstract:
We derive closed-form expressions for GWNs of rational parametric curves for robust containment queries.Our closed-form expression enables efficient computation of GWN, even if the query points are located on the rational curve. We also derive the derivatives of GWN for other applications.



Paperid:107
Authors:Elie Michel,Alec Jacobson,Siddhartha Chaudhuri,Jean-Marc Thiery
Adobe Research,Adobe Research,Adobe Research,Adobe Research
Abstract:
We present analytical formulas for evaluating Green and biharmonic 2D coordinates and their gradients and Hessians, for 2D cages made of polynomial arcs.We present results of 2D image deformations by direct interaction with the cage and through variational solvers.We demonstrate the flexibility



Paperid:108
Authors:Shibo Liu,Tielin Dai,Ligang Liu,Xiao-Ming Fu
University of Science and Technology of China,University of Science and Technology of China,University of Science and Technology of China,University of Science and Technology of China
Abstract:
We propose polynomial 2D biharmonic coordinates for closed high-order cages containing polynomial curves of any order by extending the classical2D biharmonic coordinates using high-order BEM. When applying our coordinateto 2D cage-based deformation, users manipulate the \Beziercontrol points to quickly generate the desired conformal deformation.



Paperid:109
Authors:Dong Xiao,Renjie Chen
University of Science and Technology of China,University of Science and Technology of China
Abstract:
This work constructs Green coordinates for cages composed of Bézier patches, which enables flexible deformations with curved boundaries. The high-order structure also allows us to create a compact curved cage for the input models. Additionally, this work proposes a global projection technique for precise linear reproduction.



Paperid:110
Authors:Michal Edelstein,Hsueh-Ti Derek Liu,Mirela Ben-Chen
Technion – Israel Institute of Technology,Roblox,Technion - Israel Institute of Technology
Abstract:
We propose a framework for learning on in-the-wild meshes containing non-manifold elements, multiple components, and interior structures. Our approach uses cages and generalized barycentric coordinates to parametrize and learn volumetric functions, demonstrated by segmentation and skinning weights, achieving state-of-the-art results on wild meshes.



Paperid:111
Authors:Isaac Joseph Clarke,Raul Masu,Theo Papatheodorou
Hong Kong University of Science and Technology (Guangzhou),Hong Kong University of Science and Technology, Guangzhou,Hong Kong University of Science and Technology, Guangzhou
Abstract:
Instead of pursuing the concern of AI displacing artists, we emphasise a role for artists in reshaping technology and branching it in new directions. A role that places us less as a user of AI technology, waiting for its creative outputs, but as a maker of what AI can be, perhaps leading us towards an AI that is as unnatural, occult, and esoteric, as it is practical.



Paperid:112
Authors:Jia SUN,Zheng WEI,Pan HUI
The Hong Kong University of Science and Technology (Guangzhou),The Hong Kong University of Science and Technology,The Hong Kong University of Science and Technology (Guangzhou)
Abstract:
Algorithmic Miner uses VR to reveal the hidden labor behind AI systems. By immersing participants in data annotation tasks, it critically reflects on exploitation, automation, and techno-capitalism, prompting new discussions on ethical, human-centered design in interactive systems.



Paperid:113
Authors:Ana María Zapata Guzmán,Ludovica Schaerf,Darío Negueruela del Castillo,Iacopo Neri
University of Zurich,University of Zurich,University of Zurich,University of Zurich
Abstract:
Both a critique and celebration of digital representation, this project offers multiple perspectives beyond technological homogenization. Through exploring digital f(r)ictions and multiplicities, we reject singular viewpoints in favor of interconnected truths. Our work with AI and Colombian art raises questions about bias, agency, and authenticity in cultural production, prompting reflection on AI's influence on collective imaginaries.



Paperid:114
Authors:Chong Zeng,Yue Dong,Pieter Peers,Hongzhi Wu,Xin Tong
State Key Lab of CAD and CG, Zhejiang University,Microsoft Research Asia,College of William & Mary,State Key Lab of CAD and CG, Zhejiang University,Microsoft Research Asia
Abstract:
We present RenderFormer, a neural rendering pipeline that directly renders an image from a triangle-based representation of scene with full global illumination effects, and that does not require per-scene training or finetuning.



Paperid:115
Authors:Junke Zhu,Zehan Wu,Qixing Zhang,Cheng Liao,Zhangjin Huang
University of Science and Technology of China,Tencent Technology,University of Science and Technology of China,Tencent Technology,University of Science and Technology of China
Abstract:
Our work is a lightweight static global illumination baking solution that achieves competitive lighting effects while using only approximately 5% of the memory required by mainstream industry techniques. By adopting a vertex-probe structure, we ensure excellent runtime performance, making it suitable for low-end devices.



Paperid:116
Authors:Shaohua Mo,Chuankun Zheng,Zihao Lin,Dianbing Xi,Qi Ye,Rui Wang,Hujun Bao,Yuchi Huo
State Key Lab of CAD&CG, Zhejiang University,State Key Lab of CAD&CG, Zhejiang University,State Key Lab of CAD&CG, Zhejiang University,State Key Lab of CAD&CG, Zhejiang University,Zhejiang University,State Key Lab of CAD&CG, Zhejiang University,State Key Lab of CAD&CG, Zhejiang University,State Key Lab of CAD&CG, Zhejiang University
Abstract:
We present a neural global illumination method capable of capturing multi-frequency reflections in dynamic scenes by leveraging object-centric feature grids and a novel dual-band fusion module. Our approach produces high-quality, realistic rendering effects and outperforms state-of-the-art techniques in both visual quality and computational efficiency.



Paperid:117
Authors:Zhi Zhou,Chao Li,Zhenyuan Zhang,Mingcong Tang,Zibin Li,Shuhang Luan,Zhangjin Huang
Tencent,Tencent,Tencent,Tencent,Tencent,Tencent,University of Science and Technology of China
Abstract:
We propose a Gaussian fitting compression method for light field probes, reducing storage and memory demands in large scenes. Using low-bit adaptive Gaussians and GPU-accelerated decompression, our technique replaces traditional PCA-based approaches, achieving 1:50 compression ratios. Real-time cascaded light field textures eliminate redundant baking, preserving visual quality and rendering speed.



Paperid:118
Authors:Wenyou Wang,Rex West,Toshiya Hachisuka
University of Waterloo,Aoyama Gakuin University,University of Waterloo
Abstract:
We introduce a novel segment-based framework for light transport simulation, efficiently assembling paths from disconnected segments. Our method includes innovative segment sampling techniques and corresponding estimation strategies. To demonstrate its strengths, we propose a robust bidirectional path filtering prototype, achieving superior rendering quality and faster convergence than state-of-the-art methods.



Paperid:119
Authors:Pedro Figueiredo,Qihao He,Steve Bako,Nima Khademi Kalantari
Texas A&M University,Texas A&M University,Aurora Innovation,Texas A&M University
Abstract:
Neural approach for estimating spatially varying light selection distributions to improve importance sampling in Monte Carlo rendering. To efficiently manage hundreds or thousands of lights, we integrate our neural approach with light hierarchy techniques, where the network predicts cluster-level distributions and existing methods sample lights within clusters.



Paperid:120
Authors:Sarah Taylor,Salvador Medina,Jonathan Windle,Erica Alcusa Sáez,Iain Matthews
Epic Games,Epic Games,Epic Games,Epic Games,Epic Games
Abstract:
We introduce xADA, a generative model for creating expressive and realistic animation of the face, tongue, and head directly from speech audio.The animation maps directly onto MetaHuman compatible rig controls enabling integration into industry-standard content creation pipelines.xADA generalizes across languages, and voice styles, and can animate non-verbal sounds.



Paperid:121
Authors:Anindita Ghosh,Bing Zhou,Rishabh Dabral,Jian Wang,Vladislav Golyanik,Christian Theobalt,Philipp Slusallek,Chuan Guo
DFKI,Snap Inc.,Max Planck Institute for Informatics,Snap Inc.,Max Planck Institute for Informatics,Max Planck Institute for Informatics,DFKI,Snap Inc.
Abstract:
We present a framework for generating music-driven synchronized two-person dance animations with close interactions. Our system represents the two-person motion sequence as a cohesive entity, performs hierarchical encoding of the motion sequence into discrete tokens, and utilizes dual generative masked transformers to generate realistic and coordinated dance motions.



Paperid:122
Authors:Zhiping Qiu,Yitong Jin,Yuan Wang,Yi Shi,Chao Tan,Chongwu Wang,Xiaobing Li,Feng Yu,Tao Yu,Qionghai Dai
Central Conservatory of Music,Central Conservatory of Music,Central Conservatory of Music,Central Conservatory of Music,Weilan Tech,Central Conservatory of Music,Central Conservatory of Music,Central Conservatory of Music,Tsinghua University,Tsinghua University
Abstract:
Generating string instrument performances with intricate movements and complex interactions poses significant challenges. To address these, we present ELGAR—the first diffusion-based framework for whole-body instrument performance motion generation solely from audio. We further contribute innovative losses, metrics, and dataset, marking a novel attempt with promising results for this emerging task.



Paperid:123
Authors:Bohong Chen,Yumeng Li,Youyi Zheng,Yao-Xiang Ding,Kun Zhou
State Key Laboratory of CAD & CG, Zhejiang University,State Key Lab of CAD and CG, Zhejiang University,State Key Lab of CAD and CG, Zhejiang University,State Key Lab of CAD and CG, Zhejiang University,State Key Lab of CAD and CG, Zhejiang University
Abstract:
We present a framework to utilize Large Language Models (LLMs) for co-speech gesture generation with motion examples as direct conditions. It enables multi-modal controls over co-speech gesture generation, such as motion clips, a single pose, human video, or even text prompts.



Paperid:124
Authors:Linjun Wu,Xiangjun Tang,Jingyuan Cong,He Wang,Bo Hu,Xu Gong,Songnan Li,Yuchen Liao,Yiqian Wu,Chen Liu,Xiaogang Jin
Zhejiang University,Zhejiang University,University of California San Diego,UCL Centre for Artificial Intelligence, Department of Computer Science, University College London (UCL),Tencent Technology (Shenzhen) Co., Ltd.,Tencent Technology (Shenzhen) Co., Ltd.,Tencent Technology (Shenzhen) Co., Ltd.,Tencent Technology (Shenzhen) Co., Ltd.,Zhejiang University,State Key Lab of CAD and CG, Zhejiang University,Zhejiang University
Abstract:
We introduce a novel method that integrates unsupervised style from arbitrary references into a text-driven diffusion model to generate semantically consistent stylized human motion. We leverage text as a mediator to capture the temporal correspondences between motion and style, enabling the seamless integration of temporally dynamic style into motion features.



Paperid:125
Authors:Lei Zhong,Chuan Guo,Yiming Xie,Jiawei Wang,Changjian Li
University of Edinburgh,Snap Inc.,Northeastern University,University of Edinburgh,University of Edinburgh
Abstract:
This paper presents a novel and first approach - Sketch2Anim, to automatically translate 2D storyboard sketches into high-quality 3D animations through multi-conditional motion generation.



Paperid:126
Authors:Qinghe Wang,Yawen Luo,Xiaoyu Shi,Xu Jia,Huchuan Lu,Tianfan Xue,Xintao Wang,Pengfei Wan,Di Zhang,Kun Gai
Dalian University of Technology,The Chinese University of Hong Kong,Kuaishou Technology,Dalian University of Technology,Dalian University of Technology,The Chinese University of Hong Kong,Kuaishou Technology,Kuaishou Technology,Kuaishou Technology,Kuaishou Technology
Abstract:
A 3D-aware and controllable text-to-video generation method allows users to manipulate objects and camera jointly in 3D space for high-quality cinematic video creation.



Paperid:127
Authors:Jinbo Xing,Long Mai,Cusuh Ham,Jiahui Huang,Aniruddha Mahapatra,Chi-Wing Fu,Tien-Tsin Wong,Feng Liu
The Chinese University of Hong Kong,Adobe Research,Adobe Research,Adobe Research,Adobe Research,The Chinese University of Hong Kong,Monash University,Adobe Research
Abstract:
MotionCanvas enables intuitive cinematic shot design in image-to-video generation by letting users control both camera movements and object motions in a 3D-aware scene. Combining classical graphics with modern diffusion models, it translates motion intentions into spatiotemporal signals—without costly 3D data—empowering creative video synthesis for diverse editing workflows.



Paperid:128
Authors:Zekai Gu,Rui Yan,Jiahao Lu,Peng Li,Zhiyang Dou,Chenyang Si,Zhen Dong,Qifeng Liu,Cheng Lin,Ziwei Liu,Wenping Wang,Yuan Liu
Hong Kong University of Science and Technology,Zhejiang University,Hong Kong University of Science and Technology,Hong Kong University of Science and Technology,University of Hong Kong,Nanyang Technological University,Wuhan University,Hong Kong University of Science and Technology,University of Hong Kong,Nanyang Technological University,Texas A&M University,Hong Kong University of Science and Technology
Abstract:
Diffusion as Shader (DaS) is a unified approach for controlled video generation that uses 3D tracking videos to enable versatile editing, including animating mesh-to-video, camera control, motion transfer, and object manipulation, while improving temporal consistency.



Paperid:129
Authors:Xiuli Bi,Jianfei Yuan,Bo Liu,Yong Zhang,Xiaodong Cun,Chi Man Pan,Bin Xiao
Chongqing University of Post and Telecommunications,Chongqing University of Post and Telecommunications,Chongqing University of Post and Telecommunications,Meituan,Great Bay University,University of Macau,Chongqing University of Post and Telecommunications
Abstract:
Mobius is a novel method to generate seamlessly looping videos from text descriptions directly without any user annotations, thereby creating new visual materials for the multi-media presentation.



Paperid:130
Authors:Sihui Ji,Hao Luo,Xi Chen,Yuanpeng Tu,Yiyang Wang,Hengshuang Zhao
The University of Hong Kong,DAMO Academy, Alibaba Group,The University of Hong Kong,The University of Hong Kong,The University of Hong Kong,The University of Hong Kong
Abstract:
We propose LayerFlow, a unified framework for layer-aware video generation, enabling seamless creation of transparent foregrounds, clean backgrounds, and blended scenes. With multi-stage training and LoRA techniques improving layer-wise video quality with limited data, it also supports variants like video layer decomposition, generating backgrounds for given foregrounds and vice versa.



Paperid:131
Authors:Ruizhi Shao,Yinghao Xu,Yujun Shen,Ceyuan Yang,Yang Zheng,Changan Chen,Yebin Liu,Gordon Wetzstein
Tsinghua University,Stanford University,Alibaba Group,ByteDance Inc.,Stanford University,Stanford University,Tsinghua University,Stanford University
Abstract:
We introduce a novel interspatial attention (ISA) for diffusion transformers, which maintains identity and ensures motion consistency while allowing precise control of camera and body poses. Combined with a custom video variation autoencoder, our model achieves state-of-the-art performance for photorealistic 4D human video generation.



Paperid:132
Authors:Yongtao Ge,Kangyang Xie,Guangkai Xu,Mingyu Liu,Li Ke,Longtao Huang,Hui Xue,Hao Chen,Chunhua Shen
The University of Adelaide,Zhejiang University,Zhejiang University,Zhejiang University,Alibaba Group,Alibaba Group,Alibaba Group,Zhejiang University,Zhejiang University of Technology
Abstract:
Limited high-quality ground-truth data hinders traditional video matting's real-world application. This work tackles this by advocating for large-scale training with diverse synthetic segmentation and matting data. A novel generative pipeline is also introduced to predict temporally consistent alpha masks with fine-grained details.



Paperid:133
Authors:Kai Yan,Cheng Zhang,Sébastien Speierer,Guangyan Cai,Yufeng Zhu,Zhao Dong,Shuang Zhao
University of California Irvine,Reality Labs Research, Meta,Reality Labs, Meta,University of California Irvine,Reality Labs, Meta,Reality Labs, Meta,University of California Irvine
Abstract:
Our goal is to accelerate inverse rendering by reducing the sampling budget without sacrificing overall performance. We introduce a novel image-space adaptive sampling framework to accelerate inverse rendering by dynamically adjusting pixel sampling probabilities based on gradient variance and contribution to the loss function.



Paperid:134
Authors:Jeongmin Gu,Bochang Moon
Gwangju Institute of Science and Technology,Gwangju Institute of Science and Technology
Abstract:
This paper introduces a gradient combiner that blends unbiased and biased gradients in parameter space using the James-Stein estimator to infer scene parameters (BSDFs and volumes) from images. This approach enhances optimization accuracy compared to relying solely on either unbiased or biased gradients.



Paperid:135
Authors:Lifan Wu,Nathan Morrical,Sai Praveen Bangaru,Rohan Sawhney,Shuang Zhao,Chris Wyman,Ravi Ramamoorthi,Aaron Lefohn
NVIDIA,NVIDIA,NVIDIA,NVIDIA,NVIDIA,NVIDIA,NVIDIA,NVIDIA
Abstract:
Warped-area reparameterization is a powerful technique to compute differential visibility. The key is constructing a velocity field that is continuous in the domain interior and agrees with defined velocities on boundaries. We present a robust and efficient unbiased estimator for differential visibility, using a fixed-step walk-on-spheres and closest silhouette queries.



Paperid:136
Authors:Markus Worchel,Marc Alexa
TU Berlin,TU Berlin
Abstract:
Measures can be compactly represented and approximated using the theory of moments. This work proves that such moment-based representations are differentiable, leading to principled and efficient approaches for approximating transmittance and visibility in differentiable rendering.



Paperid:137
Authors:Mariia Soroka,Christoph Peters,Steve Marschner
Cornell University,Delft University of Technology,Cornell University
Abstract:
Physically based differentiable rendering computes gradients of the rendering equation. The task is made difficult by discontinuities in the integrand at object silhouettes. To address this challenge, we propose a novel edge sampling approach that outperforms the state-of-the-art among unidirectional differentiable renderers.



Paperid:138
Authors:Ugo Finnendahl,Markus Worchel,Tobias Jüterbock,Daniel Wujecki,Fabian Brinkmann,Stefan Weinzierl,Marc Alexa
TU Berlin,TU Berlin,TU Berlin,TU Berlin,TU Berlin,TU Berlin,TU Berlin
Abstract:
Introducing differentiable path tracing for geometric acoustics with an efficient gradient algorithm based on path replay backpropagation. The system computes derivatives of output spectrograms with respect to arbitrary scene parameters (materials, geometry, emitters, microphones) within the framework of acoustic ray tracing, with applications demonstrated in various geometric scenarios.



Paperid:139
Authors:Yuxuan Zhang,Yirui Yuan,Yiren Song,Jiaming Liu
Shanghai Jiao Tong University,Shanghai Tech University,National University of Singapore,Tiamat AI
Abstract:
Stable-Makeup is a diffusion-based makeup transfer method. It leverages a Detail-Preserving makeup encoder, and content-structure control modules to preserve facial content and structure during transfer. Extensive experiments show that Stable-Makeup outperforms existing methods, offering robust, generalizable performance.



Paperid:140
Authors:Sihui Ji,Yiyang Wang,Xi Chen,Xiaogang Xu,Hao Luo,Hengshuang Zhao
The University of Hong Kong,The University of Hong Kong,The University of Hong Kong,The Chinese University of Hong Kong,DAMO Academy, Alibaba Group,The University of Hong Kong
Abstract:
FashionComposer is a flexible model for compositional fashion image generation, with a universal framework that handles diverse input modalities such as text, human models, and garment images. It personalizes appearance, pose, and human figure, using subject-binding attention to integrate reference features, enabling applications like virtual try-ons and human album generation.



Paperid:141
Authors:Liyuan Zhu,Shengqu Cai,Shengyu Huang,Gordon Wetzstein,Naji Khosravan,Iro Armeni
Stanford University,Stanford University,NVIDIA Research,Stanford University,Zillow,Stanford University
Abstract:
Redesign spaces effortlessly-ReStyle3D transforms indoor scenes by transferring object-specific styles from a single reference image, preserving 3D coherence. Combining semantic-aware diffusion and depth guidance, it enables photo-realistic virtual staging—faithfully redecorating furniture, textures, and decor. Ideal for interior design, our method outperforms existing approaches in realism, detail fidelity, and cross-view consistency.



Paperid:142
Authors:Ipek Oztas,Duygu Ceylan,Aysegul Dundar
Bilkent University,Adobe Research,Bilkent University
Abstract:
Given a 3D object representing the source content and a reference style image, our method performs 3D stylization with a large pre-trained reconstruction model. This is achieved in a zero-shot manner, with no training or test time optimization required, while delivering superior visual fidelity and efficiency compared to existing approaches.



Paperid:143
Authors:Peiying Zhang,Nanxuan Zhao,Jing Liao
City University of Hong Kong,Adobe Research,City University of Hong Kong
Abstract:
We propose a novel text-to-vector pipeline with style customization that disentangles content and style in SVG generation. Our method represents the first feed-forward text-to-vector diffusion model capable of generating SVGs in custom styles.



Paperid:144
Authors:Junhao Zhuang,Lingen Li,Xuan Ju,Zhaoyang Zhang,Chun Yuan,Ying Shan
Tsinghua University,Chinese University of Hong Kong,Chinese University of Hong Kong,Chinese University of Hong Kong,Tsinghua University,Tencent
Abstract:
Cobra is a novel efficient long-context fine-grained ID preservation framework for line art colorization, achieving high precision, efficiency, and flexible usability for comic colorization. By effectively integrating extensive contextual references, it transforms black-and-white line art into vibrant illustrations.



Paperid:145
Authors:Kaizhi Yang,Liu Dai,Isabella Liu,Xiaoshuai Zhang,Xiaoyan Sun,Xuejin Chen,Zexiang Xu,Hao Su
University of Science and Technology of China,University of California San Diego,University of California San Diego,Hillbot Inc.,University of Science and Technology of China,University of Science and Technology of China,Hillbot Inc.,University of California San Diego
Abstract:
We propose IMLS-Splatting, an end-to-end multi-view mesh optimization method that leverages point clouds for surface representation. By introducing a splatting-based differentiable IMLS algorithm, our approach efficiently converts point clouds into SDF and texture field, enabling multi-view mesh optimization in approximately 11 minutes and achieving state-of-the-art reconstruction performance.



Paperid:146
Authors:Mingyang Song,Yang Zhang,Marko Mihajlovic,Siyu Tang,Markus Gross,Tunc Ozan Aydin
Disney Research Studios,Disney Research Studios,ETH Zürich,ETH Zürich,ETH Zürich,Disney Research Studios
Abstract:
We combine splines, a classical tool from applied mathematics, with implicit Coordinate Neural Networks to model deformation fields, achieving strong performance across multiple datasets. The explicit regularization from spline interpolation enhances spatial coherency in challenging scenarios. We further introduce a metric based on Moran’s I to quantitatively evaluate spatial coherence.



Paperid:147
Authors:Selena Ling,Merlin Nimier-David,Alec Jacobson,Nicholas Sharp
NVIDIA Research,NVIDIA Research,University of Toronto,NVIDIA Research
Abstract:
Stochastic preconditioning adds spatial noise to query locations during neural field optimization; it can be formalized as a stochastic estimate for a blur operator. This simple technique eases optimization and significantly improves quality for neural fields optimization, matching or outperforming custom-designed policies and coarse-to-fine schemes.



Paperid:148
Authors:Weizhou Liu,Jiaze Li,Xuhui Chen,Fei Hou,Shiqing Xin,Xingce Wang,Zhongke Wu,Chen Qian,Ying He,Ying He
Beijing Normal University,Nanyang Technological University,Institute of Software, Chinese Academy of Sciences,Institute of Software, Chinese Academy of Sciences,Shandong University,Beijing Normal University,Beijing Normal University,SenseTime Group,Nanyang Technological University College of Computing and Data Science,Nanyang Technological University, School of Computer Science and Engineering
Abstract:
Diffusing Winding Gradients (DWG) efficiently reconstructs watertight 3D surfaces from unoriented point clouds. Unlike conventional methods, DWG avoids solving linear systems or optimizing objective functions, enabling simple implementation and parallel execution. Our CUDA implementation on an NVIDIA GTX 4090 GPU runs 30–120x faster than iPSR on large-scale models (10–20 million points).



Paperid:149
Authors:Jianjun Xia,Tao Ju
Washington University in St. Louis,Washington University in St. Louis
Abstract:
We introduced a new surface reconstruction method from points without normals. The method robustly handles undersampled regions and scales to large input sizes.



Paperid:150
Authors:Kaixin Yao,Longwen Zhang,Xinhao Yan,Yan Zeng,Qixuan Zhang,Jiayuan Gu,Wei Yang,Lan Xu,Jingyi Yu
ShanghaiTech University,ShanghaiTech University,ShanghaiTech University,ShanghaiTech University,ShanghaiTech University,ShanghaiTech University,Huazhong University of Science and Technology,ShanghaiTech University,ShanghaiTech University
Abstract:
We introduce CAST, an innovative method for reconstructing high-quality 3D scenes from a single RGB image. Supporting open-vocabulary reconstruction, CAST excels in managing occlusions, aligning objects accurately, and ensuring physical consistency with the input, unlocking new possibilities in virtual content creation and robotics.



Paperid:151
Authors:Huanyu Chen,Jiahao Wen,Jernej Barbič
University of Southern California,University of Southern California,University of Southern California
Abstract:
We derive a nonlinear elastic rod energy, starting from a general 3D volumetric isotropic material. Validated against FEM, we accurately capture rod stretching, bending and twisting, under finite deformations. We also propose how to separately control linear/nonlinear stretchability/bendability/twistability, supporting rod material design for application in computer graphics.



Paperid:152
Authors:Jerry Hsu,Tongtong Wang,Kui Wu,Cem Yuksel
University of Utah,LightSpeed Studios,LightSpeed Studios,University of Utah
Abstract:
Cosserat rods have become increasingly popular for simulating complex thin elastic rods. However, traditional approaches often encounter significant challenges in robustly and efficiently solving for valid quaternion orientations. We introduce Stable Cosserat rods, which can achieve high accuracy with high stiffness levels and maintain stability under large time steps.



Paperid:153
Authors:Ziqiu Zeng,Siyuan Luo,Fan Shi,Zhongkai Zhang
University of Strasbourg,National University of Singapore,National University of Singapore,Centre for Artificial Intelligence and Robotics, Hong Kong, CAS
Abstract:
This paper presents a GPU-friendly framework for real-time implicit simulation of hyperelastic materials with frictional contacts. Utilizing a novel splitting strategy and efficient solver, the approach achieves robust, high-performance simulation across various stiffness materials, handling large deformations and precise friction interactions with remarkable efficiency, accuracy, and generality.



Paperid:154
Authors:Jiahao Wen,Jernej Barbic,Danny Kaufman
University of Southern California,University of Southern California,Adobe Research
Abstract:
We propose a coupled mesh-adaptation model and physical simulation algorithm to jointly generate, per timestep, optimal adaptive remeshings and implicit solutions for the simulation of frictionally contacting, large-deformation elastica.



Paperid:155
Authors:Leticia Mattos Da Silva,Silvia Sellán,Natalia Pacheco-Tallaj,Justin Solomon
Massachusetts Institute of Technology (MIT),Massachusetts Institute of Technology (MIT),Massachusetts Institute of Technology (MIT),Massachusetts Institute of Technology (MIT)
Abstract:
This paper shows how to express variational time integration for a large class of elastic energies as an optimization problem with a “hidden” convex substructure. Our integrator improves the performance of elastic simulation tasks, while conserving physical invariants up to tolerance/numerical precision.



Paperid:156
Authors:Chris Giles,Elie Diaz,Cem Yuksel
Roblox,University of Utah,University of Utah
Abstract:
We extend the Vertex Block Descent method for fast and unconditionally stable physics-based simulation using an Augmented Lagrangian formulation to enable simulating hard constraints with infinite stiffness and systems with high stiffness ratios. This allows simulating complex contact scenarios involving rigid bodies with stacking and friction, and articulated joint constraints.



Paperid:157
Authors:John Desnoyers-Stewart,Noah Miller,Bernhard Riecke
Univeristy of British Columbia,Simon Fraser University,Simon Fraser University
Abstract:
Synedelica challenges traditional approaches to mixed reality by transforming physical environments through a synesthetic experience. This artwork emphasizes the potential for immersive technology to mediate reality itself, fostering social interaction and shared experiences. By reimagining how we perceive and interact with our surroundings, Synedelica opens new perspectives at the intersection between virtual and physical. Our approach encourages the SIGGRAPH community to explore the innovative capacity of intuitive and serendipitous design.



Paperid:158
Authors:Yuqian Sun,Chenhang Cheng,Chuyan Xu,Chang Hee Lee,Ali Asadipour
Computer Science Research Centre, Royal College of Art,Individual,Individual,Korea Advanced Institute of Science and Technology (KAIST),Computer Science Research Centre, Royal College of Art
Abstract:
Hyborg Agency proposes an artistic perspective on AI agents: We can design AI agents that maintain their distinct non-human nature while meaningfully participating in human social contexts.Presenting AI agents as mechanical deer nurtured by community conversations, this computational ecosystem demonstrates how defamiliarized AI agents can enrich human social experiences while promoting transparency about their artificial nature, contributing to more sustainable and ethical human-AI symbiotic relationship.



Paperid:159
Authors:Xueqi Ma,Yilin Liu,Tianlong Gao,Qirui Huang,Hui Huang
Shenzhen University,Shenzhen University,Shenzhen University,Shenzhen University,Shenzhen University
Abstract:
CLR-Wire is a unified generative framework for 3D curve-based wireframes, jointly modeling geometry and topology in a continuous latent space. Using attention-driven VAEs and flow matching, it enables high-quality, diverse generation from noise, images, or point clouds—advancing CAD design, shape reconstruction, and 3D content creation.



Paperid:160
Authors:Jionghao Wang,Cheng Lin,Yuan Liu,Rui Xu,Zhiyang Dou,Xiaoxiao Long,Haoxiang Guo,Taku Komura,Xin Li,Wenping Wang
Texas A&M University,University of Hong Kong,HKUST,University of Hong Kong,University of Hong Kong,Nanjing University,Skywork AI, Kunlun Inc.,University of Hong Kong,Texas A&M University,Texas A&M University
Abstract:
PDT is a novel framework that uses diffusion models to transform unstructured point clouds into semantically meaningful and structured distributions, such as keypoints, joints, and feature lines. Exploring complex point distribution transformation, PDT captures fine-grained geometry and semantics, offering a versatile tool for diverse tasks.



Paperid:161
Authors:Xiao-Lei Li,Hao-Xiang Chen,Yanni Zhang,Kai Ma,Alan Zhao,Tai-Jiang Mu,Haoxiang Guo,Ran Zhang
Tsinghua University,Tsinghua University,Tencent Video AI Center,Tencent PCG,Tencent Video AI Center,Tsinghua University,Skywork AI, Kunlun Inc.,Tencent Video AI Center
Abstract:
The alignment of text,images,and 3D is very challenging,yet it is crucial and beneficial for many tasks.We explore and reveal the characteristics of the native 3D latent space for 3D generation,make it decomposable and low-rank,thereby enabling efficient learning for multimodal local alignment,achieving precise local enhancement and part-level editing of 3D geometry.



Paperid:162
Authors:Yansong Qu,Dian Chen,Xinyang Li,Xiaofan Li,Shengchuan Zhang,Liujuan Cao,Rongrong Ji
Xiamen University,Xiamen University,Xiamen University,Baidu Inc.,Xiamen University,Xiamen University,Xiamen University
Abstract:
DYG is a 3D drag-based scene editing method for Gaussian Splatting that enables precise, multi-view consistent geometric edits using 3D masks and control points. It combines implicit triplane representation and a drag-based diffusion model for high-quality, fine-grained results. Visit our project page at \url{https://drag-your-gaussian.github.io/}.



Paperid:163
Authors:Peng Li,Suizhi Ma,Jialiang Chen,Yuan Liu,Congyi Zhang,Wei Xue,Wenhan Luo,Alla Sheffer,Wenping Wang,Yike Guo
Hong Kong University of Science and Technology,Johns Hopkins University,Hong Kong University of Science and Technology,Hong Kong University of Science and Technology,Univeristy of British Columbia,Hong Kong University of Science and Technology,Hong Kong University of Science and Technology,Univeristy of British Columbia,Texas A&M University,Hong Kong University of Science and Technology
Abstract:
CMD revolutionizes 3D generation by enabling flexible local editing of 3D models from a single rendering, as well as progressive, interactive creation of complex 3D scenes. At its core, CMD leverages a conditional multiview diffusion model to seamlessly modify/add new components—enhancing control, quality, and efficiency in 3D content creation.



Paperid:164
Authors:Ellie Arar,Yarden Frenkel,Daniel Cohen-Or,Ariel Shamir,Yael Vinker
Tel Aviv University,Tel Aviv University,Tel Aviv University,Reichman University,Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT)
Abstract:
SwiftSketch, a diffusion-based model with a transformer-decoder, generates high-quality vector sketches from images in under a second. It progressively denoises stroke coordinates sampled from a Gaussian distribution, effectively generalizing across various object classes. Training uses the ControlSketch Dataset, a new synthetic high-quality image-sketch pairs created by our ControlSketch optimization method.



Paperid:165
Authors:Sinan Wang,Junwei Zhou,Fan Feng,Zhiqi Li,Yuchen Sun,Duowen Chen,Greg Turk,Bo Zhu
Georgia Institute of Technology,University of Michigan Ann Arbor,Dartmouth College,Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology
Abstract:
We present the Vortex Particle Flow Map (VPFM) method, which revitalizes the traditional Vortex-In-Cell approach for computer graphics. By evolving vorticity and higher-order quantities along particle flow maps, our method achieves significantly improved long-term stability and vorticity preservation, enabling high-fidelity simulation of complex vortical fluid motions in 2D and 3D.



Paperid:166
Authors:Zhiqi Li,Candong Lin,Duowen Chen,Xinyi Zhou,Shiying Xiong,Bo Zhu
Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Zhejiang University,Georgia Institute of Technology
Abstract:
We present a Clebsch PFM fluid solver that accurately transports wave functions using particle flow maps. Key innovations include a new gauge transformation, improved velocity reconstruction on coarse grids, and better fine-scale structure preservation. Benchmarks show superior performance over impulse- or vortex-based methods, especially for small-scale flow features.



Paperid:167
Authors:Mengdi Wang,Fan Feng,Junlin Li,Bo Zhu
Georgia Institute of Technology,Dartmouth College,Georgia Institute of Technology,Georgia Institute of Technology
Abstract:
We introduce an adaptive octree-based GPU simulator for large-scale fluid simulation. Our hybrid particle-grid flow map advection scheme effectively preserves vortex details, enabling high-resolution and high-quality results. The source code has been made publicly available at: https://wang-mengdi.github.io/proj/25-cirrus/.



Paperid:168
Authors:Yuchen Sun,Junlin Li,Ruicheng Wang,Sinan Wang,Zhiqi Li,Bart G. van Bloemen Waanders,Bo Zhu
Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Sandia National Laboratories,Georgia Institute of Technology
Abstract:
We present Leapfrog Flow Maps (LFM), a fast hybrid velocity-impulse scheme with leapfrog integration. The computations are further accelerated by a matrix-free AMGPCG solver optimized for GPUs. As a result, LFM achieves high performance and fidelity across diverse examples, including fireballs and wingtip vortices.



Paperid:169
Authors:Duowen Chen,Zhiqi Li,Taiyuan Zhang,Jinjin He,Junwei Zhou,Bart G. van Bloemen Waanders,Bo Zhu
Georgia Institute of Technology,Georgia Institute of Technology,Dartmouth College,Dartmouth College,University of Michigan,Sandia National Laboratories,Georgia Institute of Technology
Abstract:
We present a unified compressible flow map framework based on Lagrangian path integrals, enabling conservative density-energy transport and flexible pressure treatments. Validated on diverse systems—from shocks to shallow water—it captures complex flow features like vortices and wave interactions, broadening flow-map applicability across compressibility regimes and fluid morphologies.



Paperid:170
Authors:Zhiqi Li,Ruicheng Wang,Junlin Li,Duowen Chen,Sinan Wang,Bo Zhu
Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology,Georgia Institute of Technology
Abstract:
This paper presents Epsilon Difference Gradient Evolution (EDGE), a novel method for accurate flow-map computation on grids without velocity buffers. EDGE enables large-scale, efficient and high-fidelity fluid simulations that capture and preserve complex vorticity structures while significantly reducing memory usage.



Paperid:171
Authors:Nicolas Bonneel,David Coeurjolly,Jean-Claude Iehl,Victor Ostromoukhov
CNRS - LIRIS,CNRS - LIRIS,Université Claude Bernard Lyon 1,Université Claude Bernard Lyon 1
Abstract:
In the context of quasi-Monte Carlo rendering, we introduce a new Sobol' construction and demonstrate that particular pairs of polynomials of the form p and p^2+p+1 in Sobol'-based sampling lead to (1, 2)-sequences. They can be combined to form high-dimensional low discrepancy sequences with good 2D projections.



Paperid:172
Authors:Qingqin Hua,Pascal Grittmann,Philipp Slusallek
Saarland University,Saarland University,Saarland University
Abstract:
Multiple importance sampling (MIS) is vital to most rendering algorithms. MIS computes a weighted sum of samples from different techniques to handle diverse scene types and lighting effects.We propose a practical weight correction scheme that yields better equal-time performance on bidirectional algorithms and resampled importance sampling for direct illumination.



Paperid:173
Authors:Corentin Salaun,Martin Balint,Laurent Belcour,Eric Heitz,Gurprit Singh,Karol Myszkowski
Max Planck Institute for Informatics,Max Planck Institute for Informatics,Intel,Intel,Max Planck Institute for Informatics,Max Planck Institute for Informatics
Abstract:
This paper introduces stratification into resampled importance sampling (RIS) technique for real-time photorealistic rendering. It organizes sample candidates into local histograms and then employs Quasi Monte Carlo and antithetic patterns for efficient sampling. This low-overhead approach significantly reduces rendering noise, improving visual quality compared to existing methods.



Paperid:174
Authors:Zhimin Fan,Yiming Wang,Chenxi Zhou,Ling-Qi Yan,Yanwen Guo,Jie Guo
Nanjing University,Nanjing University,Nanjing University,University of California Santa Barbara,Nanjing University,Nanjing University
Abstract:
We combine the estimates generated in each guiding iteration, leveraging the importance distributions from multiple guiding iterations. We demonstrate that our path-level reweighting makes guiding algorithms less sensitive to noise and overfitting in distributions.



Paperid:175
Authors:Haolin Lu,Delio Vicini,Wesley Chang,Tzu-Mao Li
UC San Diego,Google Inc.,UC San Diego,UC San Diego
Abstract:
Variance reduction techniques are widely used to reduce the noise of Monte Carlo integration. However, these techniques are typically designed with the assumption that the integrand is scalar-valued. To address this, we introduce ratio control variations, an estimator that leverages a ratio-based approach instead of the conventional difference-based control variates.



Paperid:176
Authors:Xiaochun Tong,Toshiya Hachisuka
University of Waterloo,University of Waterloo
Abstract:
We present a practical method for rendering scenes with complex, recursive nonlinear stylization applied to physically based rendering. Our approach introduces nonlinear path filtering(NL-PF) and nonlinear neural radiance caching(NL-NRC), which reduce the exponential sampling cost of stylized rendering to polynomial, enabling rendering of nonlinear stylization with significantly improved efficiency.



Paperid:177
Authors:Ziqi Wang,Wenjun Liu,Jinwen Wang,Gabriel Vallat,Fan Shi,Stefana Parascho,Maryam Kamgarpour
HKUST,HKUST,EPFL,EPFL,National University of Singapore,EPFL,EPFL
Abstract:
We introduce a reinforcement learning framework for assembling structures composed of rigid parts. A pre-trained policy generates alternative assembly plans, enabling rapid adaptation to unexpected disruptions. Our approach supports efficient and robust planning for multi-robot assembly tasks.



Paperid:178
Authors:Lucas N. Alegre,Agon Serifi,Ruben Grandia,David Müller,Espen Knoop,Moritz Bächer
Instituto de Informática - Universidade Federal do Rio Grande do Sul,Disney Research,Disney Research,Disney Research,Disney Research,Disney Research
Abstract:
Presenting AMOR, a policy conditioned on context and a linear combination of reward weights, trained using multi-objective reinforcement learning. Once trained, AMOR allows for on-the-fly adjustments of reward weights, unlocking new possibilities in physics-based and robotic character control.



Paperid:179
Authors:Hewen Xiao,Xiuping Liu,Hang Zhao,Jian Liu,Kai Xu
Dalian University of Technology,Dalian University of Technology,Wuhan University,Shenyang University of Technology,National University of Defense Technology (NUDT)
Abstract:
We introduce a pin-pression gripper featuring parallel-jaw fingers with 2D arrays of independently extendable pins, allowing instant shape adaptation to target object geometry and dynamic in-hand re-orientation for enhanced grasp stability. Reinforcement learning with curriculum-based training enables flexible, robust grasping and grasp-while-lift mode, validated by sim-to-real experiments with superior performance.



Paperid:180
Authors:Sean Memery,Kevin Denamganaï,Jiaxin Zhang,Zehai Tu,Yiwen Guo,Kartic Subr
University of Edinburgh,University of Edinburgh,Lightspeed Studios,Lightspeed Studios,Independent,University of Edinburgh
Abstract:
CueTip is an interactive and explainable automated coaching assistant for a variant of pool/billiards. CueTip has a natural-language interface, the ability to perform contextual, physics-aware reasoning, and its explanations are rooted in a set of predetermined guidelines developed by domain experts. CueTip matches SOTA performance, with grounded and reliable explanations.



Paperid:181
Authors:Wenbin Song,Heng Zhang,Yang Wang,Xiaopei Liu
ShanghaiTech University,ShanghaiTech University,ShanghaiTech University,ShanghaiTech University
Abstract:
We introduce a novel local-domain fluid-solid interaction simulator grounded in a lattice Boltzmann solver. By leveraging an MPC-based domain-tracking approach and an improved convective boundary condition, it offers enhanced stability and efficiency for deriving control policies of virtual agents, holding great promise for applications in both computer animation and robotics.



Paperid:182
Authors:Mingfeng Tang,Ningna Wang,Ziyuan Xie,Jianwei Hu,Ke Xie,Xiaohu Guo,Hui Huang
Shenzhen University,University of Texas at Dallas,Shenzhen University,QiYuan Lab,Shenzhen University,University of Texas at Dallas,Shenzhen University
Abstract:
We present the first scene-update aerial path planning algorithm specifically designed for detecting and updating change areas in urban environments, which paves the way for efficient, scalable, and adaptive UAV-based scene updates in complex urban environments.



Paperid:183
Authors:Yifang Pan,Karan Singh,Luiz Gustavo Hafemann
University of Toronto,University of Toronto,Ubisoft
Abstract:
ModelSeeModelDo presents a speech-driven 3D facial animation method using a latent diffusion model conditioned on a reference clip to capture nuanced performance styles. A novel "style basis" mechanism extracts key poses to guide generation, achieving expressive, temporally coherent animations with accurate lip-sync and strong stylistic fidelity across diverse speech inputs.



Paperid:184
Authors:Zhen Han,Mattias Teye,Derek Yadgaroff,Judith Bütepage
Electronic Arts,Electronic Arts,Electronic Arts,Electronic Arts
Abstract:
The goal of this work is to train lip sync animation models that can run in real-time and on-device. We design a two-stage knowledge distillation framework to distill large, high-quality models. Our results show that we can train small models with low latency and a comparatively small loss in quality.



Paperid:185
Authors:Luchuan Song,Yang Zhou,Zhan Xu,Yi Zhou,Deepali Aneja,Chenliang Xu
University of Rochester,Adobe Research,Adobe Research,Adobe Research,Adobe Research,University of Rochester
Abstract:
The StreamME takes live stream video as input to enable rapid 3D head avatar reconstruction. It achieves impressive speed, capturing the basic facial appearance within 10-seconds and reaching high-quality fidelity within 5-minutes. StreamME reconstructs facial features through on-the-fly training, allowing simultaneous recording and modeling without the need for pre-cached data.



Paperid:186
Authors:Shivangi Aneja,Sebastian Weiss,Irene Baeza,Prashanth Chandran,Gaspard Zoss,Matthias Niessner,Derek Bradley
Technical University of Munich,DisneyResearch|Studios,DisneyResearch|Studios,DisneyResearch|Studios,DisneyResearch|Studios,Technical University Munich,DisneyResearch|Studios
Abstract:
ScaffoldAvatar presents a novel approach for reconstructing ultra-high fidelity animatable head avatars, which can be rendered in real-time. Our method operates on patch-based local expression features and synthesizes 3D Gaussians dynamically by leveraging tiny scaffold MLPs. We employ color-based densification and progressive training to obtain high-quality results and fast convergence.



Paperid:187
Authors:Forrest Iandola,Stanislav Pidhorskyi,Igor Santesteban,Divam Gupta,Anuj Pahuja,Nemanja Bartolovic,Frank Yu,Emanuel Garbin,Tomas Simon,Shunsuke Saito
Meta,Meta,Meta,Meta,Meta,Meta,Meta,Meta,Meta,Meta
Abstract:
Existing Gaussian Splatting avatars require desktop GPUs, limiting mobile device use. SqueezeMe converts these avatars into a lightweight representation, enabling real-time animation and rendering on mobile devices. By distilling the corrective decoder into an efficient linear model, SqueezeMe achieves 72 FPS on a Meta Quest 3 VR headset.



Paperid:188
Authors:Wesley Chang,Andrew Russell,Stephane Grabli,Matt Chiang,Christophe Hery,Doug Roble,Ravi Ramamoorthi,Tzu-Mao Li,Olivier Maury
University of California San Diego,University of California San Diego,Meta,Meta,Meta,Meta,University of California San Diego,University of California San Diego,Meta
Abstract:
Recent methods have been developed to reconstruct 3D hair strand geometry from images. We introduce an inverse hair grooming pipeline to transform these unstructured hair strands into procedural hair grooms controlled by a small set of guide strands and artist-friendly grooming operators, enabling easy editing of hair shape and style.



Paperid:189
Authors:Omer Dahary,Yehonathan Cohen,Or Patashnik,Kfir Aberman,Daniel Cohen-Or
Tel Aviv University,Tel Aviv University,Tel Aviv University,Snap Research,Tel Aviv University
Abstract:
Text-to-image diffusion models struggle with multi-subject generation due to subject leakage. Prior methods impose external layouts that conflict with the model’s prior, harming alignment and natural composition. We introduce a method that leverages the layout encoded in the initial noise, promoting alignment and natural compositions while preserving the model’s diversity.



Paperid:190
Authors:Amirhossein Alimohammadi,Aryan Mikaeili,Sauradip Nag,Negar Hassanpour,Andrea Tagliasacchi,Ali Mahdavi-Amiri
Simon Fraser University,Simon Fraser University,Simon Fraser University,Huawei Canada,Simon Fraser University,Simon Fraser University
Abstract:
Cora is a novel diffusion-based image editing method that achieves complex edits, such as object insertion, background changes, and non-rigid transformations, in only four diffusion steps. By leveraging pixel-wise semantic correspondences between source and target, it preserves key elements of the original image’s structure and appearance while introducing new content.



Paperid:191
Authors:Linjie Lyu,Valentin Deschaintre,Yannick Hold-Geoffroy,Milos Hasan,Jae Shin Yoon,Thomas Leimkuehler,Christian Theobalt,Iliyan Georgiev
Max-Planck-Institute for Informatics & Saarland Informatics Campus,Adobe Research,Adobe Research,Adobe Research,Adobe Research,Max-Planck-Institute for Informatics & Saarland Informatics Campus,Max-Planck-Institute for Informatics & Saarland Informatics Campus,Adobe Research
Abstract:
A generative workflow for precise image editing using an intrinsic-image latent space. Built on RGB-X diffusion, it enables diverse edits—like relighting, color changes, and object manipulation—while preserving identity and ameliorating intrinsic-channel entanglement. All this is done without extra data or fine-tuning, achieving state-of-the-art results.



Paperid:192
Authors:Yen-Chi Cheng,Krishna Kumar Singh,Jae Shin Yoon,Alexander Schwing,Liang-Yan Gui,Matheus Gadelha,Paul Guerrero,Nanxuan Zhao
University of Illinois Urbana-Champaign,Adobe Research,Adobe Research,University of Illinois Urbana-Champaign,University of Illinois Urbana-Champaign,Adobe Research,Adobe Research,Adobe Research
Abstract:
3D-Fixup enables realistic 3D-aware photo editing by leveraging 3D priors and a novel data pipeline that extracts training pairs from real-world videos. Its feed-forward architecture supports efficient, high-quality edits involving complex 3D transformations while preserving object identity, outperforming prior methods in both edit accuracy and user control.



Paperid:193
Authors:Niladri Shekhar Dutt,Duygu Ceylan,Niloy Mitra
University College London (UCL),Adobe,University College London (UCL)
Abstract:
MonetGPT explores using multimodal large language models (MLLMs) for photo retouching by injecting domain knowledge via visual puzzles. These puzzles help MLLMs understand individual operations, visual aesthetics, and generate expert plans. Our procedural pipeline enables explainable edits with detailed reasoning for the plan and individual operations.



Paperid:194
Authors:Aleksandar Cvejic,Abdelrahman Eldesokey,Peter Wonka
King Abdullah University of Science and Technology (KAUST),King Abdullah University of Science and Technology (KAUST),King Abdullah University of Science and Technology (KAUST)
Abstract:
We present PartEdit, a novel diffusion-based system enabling precise, text-based edits of object parts without retraining or manual masks. Optimizing part-aware tokens generates localized non-binary attention maps to guide seamless edits. Our novel blending strategy delivers high-quality visual results and outperforms prior techniques in both synthetic and real-world scenarios.



Paperid:195
Authors:Mia Tang,Yael Vinker,Chuan Yan,Lvmin Zhang,Maneesh Agrawala
Stanford University,Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT),Stanford University,Stanford University,Stanford University
Abstract:
INKi enables instance segmentation for scene sketches by adapting image segmentation models with class-agnostic tuning and depth-based refinement. We introduce a new dataset INK-scene with diverse styles and demonstrate layered sketch organization for advanced editing, including inpainting occluded instances—paving the way for robust, editable sketch understanding.



Paperid:196
Authors:Jiabao Brad Wang,Amir Vaxman
University of Edinburgh,University of Edinburgh
Abstract:
We present a method for designing smooth directional fields on triangle meshes with precise control over singularities. Our approach uses a power-linear polar representation, allowing singularities of any index to be placed anywhere on the mesh. The resulting fields are smooth, robust to mesh quality, and support N-fold symmetry.



Paperid:197
Authors:Qiujie Dong,Huibiao Wen,Rui Xu,Shuangmin Chen,Jiaran Zhou,Shiqing Xin,Changhe Tu,Taku Komura,Wenping Wang
Shandong University,Shandong University,The University of Hong Kong,Qingdao University of Science and Technology,Ocean University of China,Shandong University,Shandong University,The University of Hong Kong,Texas A&M University
Abstract:
We propose NeurCross, a self-supervised framework for quadrilateral mesh generation that jointly optimizes principal curvature direction field and cross field by employing an optimizable neural SDF to approximate the input surface. NeurCross outperforms state-of-the-art methods in terms of singular point placement, robustness to noise and geometric variations, and approximation accuracy.



Paperid:198
Authors:Zhongxuan Liang,Wei Du,Xiao-Ming Fu
University of Science and Technology of China,University of Science and Technology of China,University of Science and Technology of China
Abstract:
Our approach proposes a novel partition method for reliable feature-aligned quadrangulation. The core insight is that singularity-distant smooth streamlines are more suitable as patch boundaries. The key implementation confines patch boundaries to high field smoothness regions.Validated on large-scale datasets, our method generates high-quality quad meshes while preserving reliability.



Paperid:199
Authors:Ryan Capouellez,Rodrigo Singh,Martin Heistermann,David Bommes,Denis Zorin
New York University,New York University,University of Bern,University of Bern,New York University
Abstract:
We extend Penner-coordinate-based methods for seamless parametrizations to surfaces with sharp features to which the parametrization needs to be aligned. We describe a two-phase method to efficiently minimize feature constraint residual errors. We demonstrate that the resulting algorithm works robustly on the Thingi10k dataset, completing the quad mesh generation pipeline.



Paperid:200
Authors:Etienne Corman,Keenan Crane
CNRS,Carnegie Mellon University
Abstract:
This paper presents a method for mapping curved surfaces to the plane without shear, enabling rectangular parameterizations. It introduces a novel approach for computing integrable, orthogonal frame fields. The method improves mesh quality, supports rich user control, and outperforms existing techniques in simulation, modeling, retopology, and digital fabrication tasks.



Paperid:201
Authors:Yuan-Yuan Cheng,Qing Fang,Ligang Liu,Xiao-Ming Fu
University of Science and Technology of China,University of Science and Technology of China,University of Science and Technology of China,University of Science and Technology of China
Abstract:
The paper proposes a construction algorithm based on a divide-and-conquer strategy to map a disk-topology triangular mesh onto any convex polygon., which supports arbitrary numerical precision and exact arithmetic. Under exact arithmetic, it strictly guarantees a bijection for any mesh and convex polygon.



Paperid:202
Authors:Lei Lan,Tianjia Shao,Zixuan Lu,Yu Zhang,Chenfanfu Jiang,Yin Yang
University of Utah,Zhejiang University,University of Utah,University of Utah,UCLA,University of Utah
Abstract:
This paper introduces a nearly second-order convergent training algorithm for 3D Gaussian Splatting that exploits independent kernel attributes and sparse coupling across images. By constructing and solving small Newton systems for parameter groups, it achieves about an-order faster training while maintaining or exceeding SGD-based reconstruction quality.



Paperid:203
Authors:Yunji Seo,Young Sun Choi,HyunSeung Son,Youngjung Uh
Yonsei University,Yonsei University,Yonsei University,Yonsei University
Abstract:
Flexible Level of Detail (FLoD) integrates the concept of LoD into 3DGS using a multi-level representation built with 3D Gaussian scale constraints and level-by-level training strategy. FLoD enables flexible rendering through single-level or selective rendering for optimal image quality under varying GPU VRAM constraints.



Paperid:204
Authors:Xijie Yang,Linning Xu,Lihan Jiang,Dahua Lin,Bo Dai
Zhejiang University,The Chinese University of Hong Kong,University of Science and Technology of China,The Chinese University of Hong Kong,University of Hong Kong
Abstract:
V3DG achieves real-time rendering of massive 3D Gaussians in large, composed scenes through a novel LOD approach.Inspired by Nanite, V3DG processes detailed 3D assets into clusters at various granularities offline, and selectively renders 3D Gaussians at runtime—flexibly balancing rendering speed and visual fidelity based on user-defined tolerances.



Paperid:205
Authors:Jorge Condor,Sébastien Speierer,Lukas Bode,Božič Aljaž,Simon Green,Piotr Didyk,Adrián Jarabo,Jorge Condor
Universita della Svizzera Italiana,Meta Reality Labs,Meta Reality Labs,Meta Reality Labs,Meta Reality Labs,Universita della Svizzera Italiana,Meta Reality Labs,Università della Svizzera Italiana
Abstract:
We formalize the path-tracing of volumes composed of anisotropic kernel mixture models. Our work enables computing physically-based light transport on complex volumetric assets efficiently, on tiny memory budgets. We further introduce Epanechnikov kernels as an efficient alternative in kernel-based rendering, and showcase our method in different forward and inverse volume rendering applications including radiance fields.



Paperid:206
Authors:Keyang Ye,Tianjia Shao,Kun Zhou
Zhejiang University,Zhejiang University,Zhejiang University
Abstract:
We introduce Gaussian-enhanced Surfels (GESs), a bi-scale representation combining opaque surfels and Gaussians for high-fidelity radiance field rendering. GES is entirely sorting free, enabling high-fidelity view-consistent rendering with ultra fast speeds.



Paperid:207
Authors:Rong Liu,Dylan Sun,Meida Chen,Yue Wang,Andrew Feng
USC Institute for Creative Technologies (ICT),University of Southern California,USC Institute for Creative Technologies (ICT),University of Southern California,USC Institute for Creative Technologies (ICT)
Abstract:
Deformable Beta Splatting (DBS) is a novel approach for real-time radiance field rendering that leverages deformable Beta Kernels with adaptive frequency control for both geometry and color encoding. DBS captures complex geometries and lighting with state-of-the-art fidelity, while only using 45% fewer parameters and rendering 1.5x faster than 3DGS-MCMC.



Paperid:208
Authors:Yunxiang Zhang,Bingxuan Li,Alexandr Kuznetsov,Akshay Jindal,Stavros Diolatzis,Kenneth Chen,Anton Sochenov,Anton Kaplanyan,Qi Sun
New York University,New York University,Advanced Micro Devices (AMD),Intel Corporation,Intel Corporation,New York University,Intel Corporation,Intel Corporation,New York University
Abstract:
We introduce Image-GS, a content-adaptive image representation based on colored 2D Gaussians. Image-GS achieves remarkable rate-distortion performance across diverse images and textures while supporting hardware-friendly fast random access and flexible quality control through a smooth level-of-detail hierarchy. We demonstrate its versatility with two applications: semantics-aware compression and image restoration.



Paperid:209
Authors:Mingi Lee,Dongsu Zhang,Clément Jambon,Young Min Kim
Seoul National University,Seoul National University,Massachusetts Institute of Technology (MIT),Seoul National University
Abstract:
We present BrepDiff, a simple, single-stage diffusion model for generating Boundary Representations (B-reps). Our approach generates B-reps by denoising point-based face samples with a dedicated noise schedule. Unlike multi-stage methods, BrepDiff enables intuitive, editable geometry creation, including completion, merging, and interpolation, while achieving competitive performance on unconditional generation.



Paperid:210
Authors:Yingyu Yang,Xiaohong Jia,Bolun Wang,Jieyin Yang,Shiqing Xin,Dong-Ming Yan
State Key Laboratory of Mathematical Sciences, Academy of Mathematics and Systems Science, Chinese Academy of Sciences,State Key Laboratory of Mathematical Sciences, Academy of Mathematics and Systems Science, Chinese Academy of Sciences,Visual Computing Institute, RWTH Aachen University,State Key Laboratory of Mathematical Sciences, Academy of Mathematics and Systems Science, Chinese Academy of Sciences,Shandong University,MAIS, Institute of Automation, Chinese Academy of Sciences
Abstract:
We propose a novel algorithm for efficient and accurate Boolean operations on B-Rep models by mapping them bijectively to controllable-error triangle meshes. Using conservative intersection detection on the mesh to locate all surface intersection curves and carefully handling degeneration and topology errors ensure that the results are watertight and correct.



Paperid:211
Authors:Bingchen Yang,Haiyong Jiang,Hao Pan,Guosheng Lin,Jun Xiao,Peter Wonka,Bingchen Yang
School of Artificial Intelligence, University of Chinese Academy of Sciences,School of Artificial Intelligence, University of Chinese Academy of Sciences,Tsinghua University,Nanyang Technological University,School of Artificial Intelligence, University of Chinese Academy of Sciences,KAUST,School of Artificial Intelligence, University of Chinese Academy of Sciences; Nanyang Technological University
Abstract:
We propose an iterative prompt-and-select architecture to progressively reconstruct the CAD modeling sequence of a target point cloud. We propose the concept of local geometric guidance and come up with three ways to integrate this guidance into iterative reconstruction. Experiments demonstrate the superiority over the current state of the art.



Paperid:212
Authors:Pu Li,Wenhao Zhang,Jinglu Chen,Dongming Yan
Institute of Automation, Chinese Academy Of Sciences,Institute of Automation, Chinese Academy of Sciences,Institute of Automation, Chinese Academy of Sciences,Institute of Automation, Chinese Academy of Sciences
Abstract:
Stitch-A-Shape introduces a novel framework for generating B-Rep models by directly addressing both topology and geometry. Using a sequential stitching approach, it assembles 3D shapes from vertices through curves to faces, effectively managing topological and geometric complexities. The framework demonstrates superior performance in shape generation, class-conditional generation, and autocompletion tasks.



Paperid:213
Authors:Jing-En Jiang,Hanxiao Wang,Mingyang Zhao,Dong-Ming Yan,Shuangmin Chen,Shiqing Xin,Changhe Tu,Wenping Wang
School of Computer Science and Technology, Shandong University,Institute of Automation, Chinese Academy of Sciences, Beijing, China The School of Artificial Intelligence, University of Chinese Academy of Sciences,Academy of Mathematics and Systems Science, Chinese Academy of Sciences, the University of Chinese Academy of Sciences,Institute of Automation, Chinese Academy of Sciences,School of Information and Technology, Qingdao University of Science and Technology,School of Computer Science and Technology, Shandong University,School of Computer Science and Technology, Shandong University,Computer Science & Engineering, Texas A&M University
Abstract:
DeFillet, the reverse of CAD filleting, is vital for CAE and redesign but challenging with polygon CAD models. Our algorithm uses Voronoi vertices as rolling-ball center candidates to efficiently identify fillets. Sharp features are then reconstructed via quadratic optimization, validated on diverse models.



Paperid:214
Authors:Yilin Liu,Duoteng Xu,Xingyao Yu,Xiang Xu,Daniel Cohen-Or,Hao Zhang,Hui Huang
Shenzhen University,Shenzhen University,Shenzhen University,Simon Fraser University,Tel Aviv University,Simon Fraser University,Shenzhen University
Abstract:
We introduce a novel representation for learning and generating Computer-Aided Design (CAD) models in the form of boundary representations (BReps). Our representation unifies the continuous geometric properties of BRep primitives in different orders (e.g., surfaces and curves) and theirdiscrete topological relations in a holistic latent (HoLa) space.



Paperid:215
Authors:Chris Careaga,Yağız Aksoy
Simon Fraser University,Simon Fraser University
Abstract:
We present a photograph relighting method that enables explicit control over light sources akin to CG pipelines. We achieve this in a pipeline involving mid-level computer vision, physically-based rendering, and neural rendering. We introduce a self-supervised training methodology to train our neural renderer using real-world photograph collections.



Paperid:216
Authors:Nadav Magar,Amir Hertz,Eric Tabellion,Yael Pritch,Alex Rav-Acha,Ariel Shamir,Yedid Hoshen
Tel Aviv University,Google,Google,Google,Google,Reichman University,Hebrew University of Jerusalem
Abstract:
LightLab is a diffusion-based method for parametric control over light sources in an image. Leveraging the linearity of light we create a dataset of controled illumniation changes from a small set of real image pairs and synthetic renders, which is used to fine-tune a model to enable physically plausible edits.



Paperid:217
Authors:Mutian Tong,Rundi Wu,Changxi Zheng
Columbia University,Columbia University,Columbia University
Abstract:
We propose a method for estimating spatiotemporally varying indoor lighting from videos using a continuous light field represented as an MLP. By leveraging 2D diffusion priors fine-tuned to predict lighting jointly at multiple locations, our approach achieves superior performance and zero-shot generalization to in-the-wild scenes.



Paperid:218
Authors:Henglei Lv,Bailin Deng,Jianzhu Guo,Xiaoqiang Liu,Pengfei Wan,Di Zhang,Lin Gao
Institute of Computing Technology, Chinese Academy of Sciences,Cardiff University,Kuaishou Technology,Kuaishou Technology,Kuaishou Technology,Kuaishou Technology,Institute of Computing Technology, Chinese Academy of Sciences
Abstract:
GSHeadRelight enables fast, high-quality relightability for 3D Gaussian head synthesis. A linear light model based on learnable radiance transfer is integrated into the native 3DGS rasterization process and supports colored illumination. Without requiring expensive light stage data, our method achieves 240+ FPS rendering speed and offers state-of-the-art relighting results.



Paperid:219
Authors:Philippe Weier,Jérémy Riviere,Ruslan Guseinov,Stephan Garbin,Philipp Slusallek,Bernd Bickel,Thabo Beeler,Delio Vicini
Saarland University,Google,Google,Google,Saarland University,Google,Google,Google
Abstract:
This work addresses recovering textured materials using inverse rendering. Our Laplacian mipmapping improves the reconstruction of high-resolution textures. We also propose a novel gradient computation that enables efficiently reconstructing textured, path-traced subsurface scattering. The methods are applied to challenging scenes, including reconstructing realistic human face appearance from sparse captures.



Paperid:220
Authors:Shuai Yang,Jing Tan,Mengchen Zhang,Tong Wu,Gordon Wetzstein,Ziwei Liu,Dahua Lin
Shanghai Jiao Tong University,The Chinese University of Hong Kong,Zhejiang University,The Chinese University of Hong Kong,Stanford University,Nanyang Technological University,The Chinese University of Hong Kong
Abstract:
LayerPano3D is a novel framework that generates hyper-immersive 3D panoramic scenes from a single text prompt. By decomposing panoramas into multiple layers and optimizing them as 3D Gaussians, it enables full 360°×180° exploration with consistent visual quality, unlocking new possibilities for virtual reality and scene generation.



Paperid:221
Authors:Khoa Do,David Coeurjolly,Pooran Memari,Nicolas Bonneel
University of Michigan,CNRS - LIRIS,CNRS - LIX,CNRS - LIRIS
Abstract:
We propose a simple, parallelizable algorithm inspired by rectified flows to match probability distributions. With linear-time complexity, it approximates optimal transport by employing summed-area tables and direct particle advection. We illustrate our applications in stippling, mesh parameterization, and shape interpolation in 2D, 3D.



Paperid:222
Authors:Navid Ansari,HANS-PETER SEIDEL,Vahid Babaei
Max Planck Institute for Informatics,Max Planck Institute for Informatics,Max Planck Institute for Informatics
Abstract:
This paper proposes a scalable framework using Bayesian Neural Networks and a novel 2mD acquisition function to efficiently discover gamut boundaries in performance space. Combining NSGA-II's diversity and Bayesian Optimization's efficiency, the method enables large-batch, parallel optimization, outperforming traditional approaches in real-world engineering and fabrication tasks.



Paperid:223
Authors:Behrooz Zarebavani,Danny M. Kaufman,David I. W. Levin,Maryam Mehri Dehnavi
University of Toronto,Adobe Research,University of Toronto,University of Toronto
Abstract:
Parth delivers adaptive fill-reducing ordering to accelerate Cholesky solvers in simulations with dynamic sparsity patterns, such as contact modelling, achieving up to 255× ordering speedups. With seamless, three-line integration into popular solvers like MKL and Accelerate, Parth ensures reliable, high-performance computations for applications in computer graphics and scientific computing.



Paperid:224
Authors:Federico Sichetti,Enrico Puppo,Zizhou Huang,Marco Attene,Denis Zorin,Daniele Panozzo
Università di Genova,Università di Genova,New York University,CNR IMATI,New York University,New York University
Abstract:
Many problems in graphics can be formulated as a non-linearly constrained global minimization (MINIMIZE), or solution of a system of non-linear constraints (SOLVE). We introduce MiSo, a domain-specific language and compiler for generating efficient code for low-dimensional MINIMIZE and SOLVE problems, using interval methods to guarantee conservative results.



Paperid:225
Authors:Chunlei Li,Peng Yu,Tiantian Liu,Siyuan Yu,Yuting Xiao,Shuai Li,Aimin Hao,Yang Gao,Qinping Zhao
Beihang University,Beihang University,Taichi Graphics,Zenustech,Beihang University,Beihang University,Beihang University,Beihang University,Beihang University
Abstract:
In high-stiffness, high-resolution simulations, while primal space methods typically fail, the dual-space XPBD method produces unphysical softening artifacts due to convergence stall. We design an innovative Algebraic Multigrid method to enhance XPBD, utilizing lazy-update prolongators and near-kernel optimization. Our approach ensures stability, efficiency, and scalability for high-stiffness, high-resolution deformable models.



Paperid:226
Authors:Isabella Liu,Zhan Xu,Yifan Wang,Hao Tan,Zexiang Xu,Xiaolong Wang,Hao Su,Zifan Shi
University of California San Diego,Adobe Research,Adobe Research,Adobe Research,Hillbot Inc.,University of California San Diego,University of California San Diego,Adobe Research
Abstract:
RigAnything is a transformer-based model that autoregressively generates 3D rigging without templates. It sequentially predicts joints and skeleton topology while assigning skinning weights, working on objects in any pose. It’s 20× faster than existing methods, completing rigging in under 2 seconds with state-of-the-art quality across diverse object types.



Paperid:227
Authors:Jia-Peng Zhang,Cheng-Feng Pu,Meng-Hao Guo,Yan-Pei Cao,Shi-Min Hu
CS Dept, Tsinghua University,CS Dept, Tsinghua University,CS Dept, Tsinghua University,VAST,CS Dept, Tsinghua University
Abstract:
Manual 3D rigging is slow. UniRig introduces a unified learning framework for automatic skeletal rigging. Trained on our large, diverse Rig-XL dataset, it uses an autoregressive model and cross-attention to accurately rig various characters and objects, significantly outperforming prior methods and speeding up animation pipelines.



Paperid:228
Authors:Yufan Deng,Yuhao Zhang,Chen Geng,Shangzhe Wu,Jiajun Wu
Stanford University,Stanford University,Stanford University,Stanford University,Stanford University
Abstract:
We present the Anymate Dataset, a large-scale dataset of 230K 3D assets paired with expert-crafted rigging and skinning information---70 times larger than existing datasets. Using this dataset, we propose a learning-based auto-rigging framework with three sequential modules for joint, connectivity, and skinning weight prediction.



Paperid:229
Authors:Wenning Xu,Shiyu Fan,Paul Henderson,Edmond S. L. Ho
University of Glasgow,University of Glasgow,University of Glasgow,University of Glasgow
Abstract:
Generate exciting multi-character interactions, such as team fights, with our training-free method! Multi-character interactions can be decomposed into multiple two-person interactions using a directed graph, which enables repurposing large pre-trained two-character motion synthesis models without any multi-character data. You can compose and vary multi-character interactions spatially and temporally!



Paperid:230
Authors:Ziyi Chang,He Wang,George Koulieris,Hubert Shum
Durham University,UCL Centre for Artificial Intelligence, Department of Computer Science, University College London (UCL),Durham University,Durham University
Abstract:
This work introduces a conditional generative framework for large-scale multi-character interaction synthesis by facilitating natural interactive motions and transitions where characters are coordinated for new interactive partners, proposing a coordinatable multi-character interaction space for interaction synthesis and a transition planning network to plan transitions to achieve scalable, transferable multi-character animations.



Paperid:231
Authors:Runyi Yu,Yinhuai Wang,Qihan Zhao,Hok Wai Tsui,Jingbo Wang,Ping Tan,Qifeng Chen
Hong Kong University of Science and Technology,Hong Kong University of Science and Technology,Hong Kong University of Science and Technology,Hong Kong University of Science and Technology,Shanghai Aritificial Intelligence Laboratory,Hong Kong University of Science and Technology,Hong Kong University of Science and Technology
Abstract:
This work addresses the challenge of learning robust interaction skills from limited demonstrations. By introducing novel data augmentation techniques for skill transitions and recovery patterns, combined with enhanced reinforcement imitation learning methods, we achieve superior performance in learning interaction skills, demonstrating improved generalization and recovery capabilities across diverse manipulation tasks.



Paperid:232
Authors:Minghao Yin,Yukang Cao,Songyou Peng,Kai Han
University of Hong Kong,Nanyang Technological University, Singapore,Google DeepMind,University of Hong Kong
Abstract:
Splat4D generates high-fidelity 4D content from monocular videos by integrating multi-view rendering, inconsistency identification, a video diffusion model, and asymmetric U-Net refinement. Our framework maintains spatial-temporal consistency while preserving details and following user guidance, achieving state-of-the-art benchmark performance. Applications include text/image-conditioned generation, 4D human modeling, and text-guided content editing.



Paperid:233
Authors:Pinxuan Dai,Peiquan Zhang,Zheng Dong,Ke Xu,Yifan Peng,Dandan Ding,Yujun Shen,Yin Yang,Xinguo Liu,Rynson W.H. Lau,Weiwei Xu
Zhejiang University,Zhejiang University,Zhejiang University,City University of Hong Kong,The University of Hong Kong,Hangzhou Normal University,Ant Group,The University of Utah,Zhejiang University,City University of Hong Kong,State Key Lab CAD&CG, Zhejiang University
Abstract:
We present 4D Gaussian Video (4DGV) for high-quality, low-storage volumetric video reconstruction and real-time streaming. Our method effectively handles complex motion and enables effective motion compression, achieving superior performance in both reconstruction quality and storage efficiency.



Paperid:234
Authors:Zhaoyang Lv,Maurizio Monge,Ka Chen,Yufeng Zhu,Michael Goesele,Jakob Engel,Zhao Dong,Richard Newcombe
Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta
Abstract:
This paper investigates photorealistic scene reconstruction using videos captured from an egocentric device in high dynamic range. It presents a novel system utilizing visual-inertial bundle adjustment and a physical image formation model that handles camera motion artifacts. The experiments using Project Aria and Quest3 show substantial improvements in visual quality.



Paperid:235
Authors:Andreas Meuleman,Ishaan Shah,Alexandre Lanvin,Bernhard Kerbl,George Drettakis
INRIA, Université Côte d'Azur,INRIA, Université Côte d'Azur,INRIA, Université Côte d'Azur,TU Wien,INRIA, Université Côte d'Azur
Abstract:
We propose a fast, on-the-fly 3D Gaussian Splatting method that jointly estimates poses and reconstructs scenes. Through fast pose initialization, direct primitive sampling, and scalable clustering and merging, it efficiently handles diverse ordered image sequences of arbitrary length.



Paperid:236
Authors:Songyin Wu,Zhaoyang Lv,Yufeng Zhu,Duncan Frost,Zhengqin Li,Ling-Qi Yan,Carl Ren,Richard Newcombe,Zhao Dong
Meta Reality Labs Research,Meta Reality Labs Research,Meta Reality Labs Research,Meta Reality Labs Research,Meta Reality Labs Research,University of California Santa Barbara,Meta Reality Labs Research,Meta Reality Labs Research,Meta Reality Labs Research
Abstract:
We propose a high-quality online reconstruction pipeline for monocular input streams, reconstructing environments with detail across multiple levels while maintaining high speed.



Paperid:237
Authors:Letian Huang,Dongwei Ye,Jialin Dan,Chengzhi Tao,Huiwen Liu,Kun Zhou,Bo Ren,Yuanqi Li,Yanwen Guo,Jie Guo
State Key Lab for Novel Software Technology, Nanjing University,State Key Lab for Novel Software Technology, Nanjing University,State Key Lab for Novel Software Technology, Nanjing University,State Key Lab for Novel Software Technology, Nanjing University,TMCC, College of Computer Science, Nankai University,State Key Lab of CAD&CG, Zhejiang University,TMCC, College of Computer Science, Nankai University,State Key Lab for Novel Software Technology, Nanjing University,State Key Lab for Novel Software Technology, Nanjing University,State Key Lab for Novel Software Technology, Nanjing University
Abstract:
We propose TransparentGS, a fast inverse rendering pipeline for transparent objects based on 3D-GS. The main contributions are three-fold: efficient transparent Gaussian primitives for specular refraction, GaussProbe to encode ambient light and nearby contents, and the IterQuery algorithm to reduce parallax errors in our probe-based framework.



Paperid:238
Authors:Elad Richardson,Yuval Alaluf,Ali Mahdavi-Amiri,Daniel Cohen-Or
Tel Aviv University,Tel Aviv University,Simon Fraser University,Tel Aviv University
Abstract:
pOps is a framework for learning semantic manipulations in CLIP’s image embedding space. Built on a Diffusion Prior model, it enables concept manipulation by training operators directly on image embeddings. This approach enhances semantic control and integrates easily with diffusion models for image generation.



Paperid:239
Authors:Lvmin Zhang,Chuan Yan,Yuwei Guo,Jinbo Xing,Maneesh Agrawala
Stanford University,Stanford University,CUHK,CUHK,Stanford University
Abstract:
A framework to generate past and future processes for drawing process videos.



Paperid:240
Authors:Eric Chen,Žiga Kovačič,Madhav Aggarwal,Abe Davis
Cornell University,Cornell University,Cornell University,Cornell University
Abstract:
Pocket Time-Lapse is a system to record, explore and visualize long-term changes in the environment, based on data that a user can capture with the phone they carry. Our contributions include a process to conveniently capture a scene, and novel techniques for registering and visualizing panoramic time-lapse data.



Paperid:241
Authors:Etai Sella,Yanir Kleiman,Hadar Averbuch-Elor
Tel Aviv University,Meta,Cornell Tech
Abstract:
We propose InstanceGen - a new technique for improving Text-to-Image models ability to generate images for prompts describing multiple objects, attributes and spatial relationships. InstanceGen requires no training or additional user inputs and achieves state-of-the art results in terms of both accuracy and visual quality on these highly challenging prompts.



Paperid:242
Authors:Yuxin Zhang,Minyan Luo,Weiming Dong,Xiao Yang,Haibin Huang,Chongyang Ma,Oliver Deussen,Tong-Yee Lee,Changsheng Xu
MAIS, Institute of Automation, Chinese Academy of Sciences,MAIS, Institute of Automation, Chinese Academy of Sciences,MAIS, Institute of Automation, Chinese Academy of Sciences,ByteDance Inc.,ByteDance Inc.,ByteDance Inc.,University of Konstanz,National Cheng-Kung University,MAIS, Institute of Automation, Chinese Academy of Sciences
Abstract:
This paper presents T-Prompter, a method for visually prompting generative models to enable continuous image generation for specific themes, characters, and scenes. It introduces Dynamic Visual Prompting to enhance generation accuracy and quality, outperforming existing methods in maintaining character identity, style consistency, and text alignment.



Paperid:243
Authors:Yuanpeng Tu,Xi Chen,Ser-Nam Lim,Hengshuang Zhao
The University of Hong Kong,The University of Hong Kong,UCF,The University of Hong Kong
Abstract:
To address a lack of generalization to novel classes, we propose DreamMask, which systematically explores data generation in the open-vocabulary setting, and how to train the model with both real and synthetic data. It significantly simplifies the collection of large-scale training data, serving as a plug-and-play enhancement for existing methods.



Paperid:244
Authors:Sara Dorfman,Dana Cohen-Bar,Rinon Gal,Daniel Cohen-Or
Tel Aviv University,Tel Aviv University,NVIDIA Research,Tel Aviv University
Abstract:
IP-Composer is a novel, training-free method for compositional image generation from multiple reference images. Extending IP-Adapter, it uses natural language to identify concept-specific subspaces in CLIP, projects input images into these subspaces to extract targeted concepts, and fuses them into composite embeddings—enabling fine-grained, controllable generation across diverse visual concepts.



Paperid:245
Authors:Tianyang Xue,Longdu Liu,Lin Lu,Paul Henderson,Pengbin Tang,Haochen Li,Jikai Liu,Haisen Zhao,Hao Peng,Bernd Bickel
Shandong University,Shandong University,Shandong University,University of Glasgow,ETH Zürich,Shandong University,Shandong University,Shandong University,CrownCAD,ETH Zürich
Abstract:
We introduce MIND, a novel generative framework for inverse-designing diverse, tileable 3D microstructures. Leveraging latent diffusion and our hybrid neural representation, MIND precisely achieves targeted physical properties, ensures geometric validity, and enables seamless boundary compatibility—opening new avenues for advanced metamaterial design and manufacturing applications.



Paperid:246
Authors:Tao Liu,Tianyu Zhang,Yongxue Chen,Weiming Wang,Yu Jiang,Yuming Huang,Charlie C.L. Wang
University of Manchester,University of Manchester,University of Manchester,University of Manchester,University of Manchester,University of Manchester,University of Manchester
Abstract:
We present a computational framework that co-optimizes structural topology, curved layers, and fiber orientations for manufacturable, high-strength composites. Using implicit neural fields, our method integrates design and fabrication objectives into a unified optimization process, achieving up to 33.1% improvement in failure load for multi-axis 3D printed fiber-reinforced thermoplastics.



Paperid:247
Authors:Maxine Perroni-Scharf,Zachary Ferguson,Thomas Butruille,Carlos Portela,Mina Konaković Luković
Massachusetts Institute of Technology (MIT),Massachusetts Institute of Technology (MIT),Massachusetts Institute of Technology (MIT),Massachusetts Institute of Technology (MIT),Massachusetts Institute of Technology (MIT)
Abstract:
We introduce a method for discovering novel microscale TPMS structures with high-energy dissipation. By combining a parametric design space, empirical testing, and uncertainty-aware deep ensembles with Bayesian optimization, we efficiently explore and discover structures with extreme energy dissipation capabilities.



Paperid:248
Authors:Pengbin Tang,Bernhard Thomaszewski,Stelian Coros,Bernd Bickel
ETH Zürich,ETH Zürich,ETH Zürich,ETH Zürich
Abstract:
In this paper, we present a computational approach for designing Discrete Interlocking Materials (DIM) with desired mechanical properties. We demonstrate the effectiveness of our method by designing discrete interlocking materials with diverse limit profiles for in- and out-of-plane deformation and validate our method on fabricated physical prototypes.



Paperid:249
Authors:Di Zhang,Ligang Liu
University of Science and Technology of China,University of Science and Technology of China
Abstract:
This paper introduces a novel asymptotic directional stiffness (ADS) metric to analyze the contribution of middle surface geometry on the stiffness of shell lattice metamaterials, focusing on Triply Periodic Minimal Surfaces (TPMS). It provides a theoretical framework and optimization techniques, advancing the understanding of TPMS shell lattices.



Paperid:250
Authors:Aviv Segall,Jing Ren,Martin Schwarz,Olga Sorkine-Hornung
ETH Zurich,ETH Zurich,University of Basel,ETH Zurich
Abstract:
Gothic microarchitecture—a prevalent feature of late medieval art—comprises sculptural works that replicate monumental Gothic forms, though its original construction techniques remain historically undocumented. Leveraging insights from 15th-century Basel goldsmith drawings, we present an interactive framework for reconstructing these intricate designs from 2D projections into accurate 3D forms.



Paperid:251
Authors:András Simon,Danwu Chen,Philipp Urban,Vincent Duveiller,Henning Lübbe
Fraunhofer IGD,Fraunhofer IGD,Fraunhofer IGD,VITA Zahnfabrik H. Rauter GmbH & Co. KG,VITA Zahnfabrik H. Rauter GmbH & Co. KG
Abstract:
We propose a practical method for dental layer biomimicry and multi-spot shade matching using multi-material 3D printing. It integrates seamlessly into workflows combining dental CAD tools and industrial multi-material slicers.We validated it by printing multiple dentures and teeth with varying inner structures and translucencies to match VITA classical shades.



Paperid:252
Authors:Michele Rocca,Sune Darkner,Kenny Erleben,Sheldon Andrews,Michele Rocca
University of Copenhagen,University of Copenhagen,University of Copenhagen,École de technologie supérieure,University of Copenhagen
Abstract:
We present a new perspective on physics-based character animation. Assuming policies for similar motions should have similar weights, we introduce regularization during RL training to preserve weight similarity. By modeling the weights’ manifold with a diffusion model, we generate a continuum of policies adapting to novel character morphologies and tasks.



Paperid:253
Authors:Jungnam Park,Euikyun Jung,Jehee Lee,Jungdam Won
Seoul National University,Seoul National University,Seoul National University,Seoul National University
Abstract:
We introduce MAGNET (Muscle Activation Generation Networks), a scalable framework for reconstructing full-body muscle activations across diverse human movements, which also includes distilled models for solving downstream tasks or generating real-time muscle activations—even on edge devices. The efficacy is demonstrated through examples of daily life and challenging behaviors.



Paperid:254
Authors:Michael Xu,Yi Shi,KangKang Yin,Xue Bin Peng
Simon Fraser University,Simon Fraser University,Simon Fraser University,Simon Fraser University
Abstract:
PARC is a framework that enhances terrain traversal with machine learning and physics-based simulation. By iteratively training a kinematic motion generator and simulated motion tracker, PARC produces a character controller capable of traversing complex environments using highly agile motor skills, overcoming the challenges of limited motion capture data.



Paperid:255
Authors:Minsu Kim,Eunho Jung,Yoonsang Lee
Hanyang University,Hanyang University,Hanyang University
Abstract:
PhysicsFC introduces a breakthrough in interactive football simulation—enabling real-time control of physically simulated players that perform complex skills with smooth transitions. It combines skill-specific learning, physics-informed rewards, latent-guided training, and transition-aware state initialization, achieving agile, lifelike football behaviors in scenarios ranging from 1v1 play to full 11v11 matches.



Paperid:256
Authors:Minseok Kim,Wonjeong Seo,Sung-Hee Lee,Jungdam Won
Seoul National University,Seoul National University,Korea Advanced Institute of Science and Technology (KAIST),Seoul National University
Abstract:
We introduce ViSA (Virtual Stunt Actors), an interactive animation system using deep reinforcement learning to generate realistic ballistic stunt actions. It efficiently produces dynamic scenes commonly seen in films and TV dramas, such as traffic accidents and stairway falls. A novel action space design enables scene generation within minutes.



Paperid:257
Authors:Jinseok Bae,Younghwan Lee,Donggeun Lim,Young Min Kim
Seoul National University,Seoul National University,Seoul National University,Seoul National University
Abstract:
We introduce a physically-based character animation framework that exploits part-wise latent tokens. The novel structured decomposition enables dynamic exploration to stably adapt to diverse unseen scenarios. Additional refinement networks improve overall motion quality. We show superior performance on multi-body tracking, motion adaptation, and locomotion with damaged body parts.



Paperid:258
Authors:Xiaoyu Huang,Takara Truong,Yunbo Zhang,Fangzhou Yu,Jean Pierre Sleiman,Jessica Hodgins,Koushil Sreenath,Farbod Farshidian
University of California Berkeley,Stanford University,Robotics and AI Institute,Robotics and AI Institute,Robotics and AI Institute,Robotics and AI Institute,Robotics and AI Institute,Robotics and AI Institute
Abstract:
Meet Diffuse-CLoC—a powerful unification of intuitive steering in kinematic motion generation and physics-based character control. By guiding diffusion over joint state-action spaces, it enables agile, steerable, and physically realistic motions across diverse downstream tasks—from obstacle avoidance to task-space control and motion in-betweening—all from a single model, with no fine-tuning required.



Paperid:259
Authors:Xiaohe Ma,Valentin Deschaintre,Milos Hasan,Fujun Luan,Kun Zhou,Hongzhi Wu,Yiwei Hu
State Key Lab of CAD&CG, Zhejiang University,Adobe Research,Adobe Research,Adobe Research,State Key Lab of CAD&CG, Zhejiang University,State Key Lab of CAD&CG, Zhejiang University,Adobe Research
Abstract:
MaterialPicker is a multi-modal material generation model that creates high-quality material maps from images and/or text by fine-tuning a video diffusion model. It robustly extracts materials from real-world photos, even with distortion or occlusion, enhancing fidelity, diversity, and efficiency in material synthesis.



Paperid:260
Authors:Michael Birsak,John Femiani,Biao Zhang,Peter Wonka
King Abdullah University of Science and Technology (KAUST),Miami University,King Abdullah University of Science and Technology (KAUST),King Abdullah University of Science and Technology (KAUST)
Abstract:
MatCLIP assigns realistic PBR materials to 3D models using shape- and lighting-invariant descriptors derived from images, including LDM outputs and photos. It outperforms prior methods by over 15%, enabling consistent material predictions across varied geometry and lighting, with applications to large-scale 3D datasets like ShapeNet and Objaverse.



Paperid:261
Authors:Mengqi Xia,Zhaoyang Zhang,Sumit Chaturvedi,Yutong Yi,Rundong Wu,Holly Rushmeier,Julie Dorsey
Yale University,Yale University,Yale University,Yale University,ByteDance Inc.,Yale University,Yale University
Abstract:
This paper presents a novel pipeline to digitize physical threads and predict fabric appearance before fabricating cloth samples, addressing a real need in the fashion industry. It enables designers to make more informed material choices, thereby promoting sustainable production, reducing costs, and fostering innovation in fabric design.



Paperid:262
Authors:Liwen Wu,Fujun Luan,Miloš Hašan,Ravi Ramamoorthi
University of California San Diego,Adobe Research,Adobe Research,University of California San Diego
Abstract:
Accurate modeling of normal distribution functions (NDF) over a high-resolution normal map enables intriguing glinty appearance but is inefficient. We present a manifold-based glint formulation, transferring the glint NDF computation to mesh intersections. This framework accelerates glint rendering, as well as providing a closed-form shadow-masking derivation for normal-mapped diffuse surfaces.



Paperid:263
Authors:Laurent Belcour,Alban Fichet,Pascal Barla
Intel Labs,Intel Labs,Inria - LaBRI
Abstract:
We introduce a material model for diffuse fluorescence that is compatible with RGB and spectral rendering. This models builds on an analytical integrable Gaussian-based model of the spectral reradiation that is efficient enough to permits real-time rendering and editing of such appearance.



Paperid:264
Authors:Zhengze Liu,Yuchi Huo,Yifan Peng,Rui Wang
State Key Lab of CAD & CG, Zhejiang University,State Key Lab of CAD & CG, Zhejiang University,University of Hong Kong,State Key Lab of CAD & CG, Zhejiang University
Abstract:
This work presents a statistical wave-scattering model for surfaces with nanoscale mixtures in geometry and material. It predicts average appearance (BRDF) and draws realistic speckles directly from surface statistics, without explicit definitions. The proposed model demonstrates various applications including corrosion (natural), particle deposition (man-made) and height-correlated mixture (artistic).



Paperid:265
Authors:Louis Sugy
NVIDIA
Abstract:
This paper introduces a novel median filtering algorithm, using hierarchical tiling to reduce redundant computations and achieve better complexity than prior sorting-based methods. The paper discusses two implementations, for both small and larger kernel sizes, that outperform the state of the art by up to 5x on modern GPUs.



Paperid:266
Authors:Ben Weiss
Google Research
Abstract:
The median filter is a staple of computational image processing. Existing efficient methods share a common flaw, which is that they use a square kernel, producing visual artifacts. Our method overcomes this limitation, enabling fast and high-quality circular-kernel median filtering, across multiple platforms and image types.



Paperid:267
Authors:Xiang Zhang,Yang Zhang,Lukas Mehl,Markus Gross,Christopher Schroers
ETH Zürich,DisneyResearch|Studios,DisneyResearch|Studios,ETH Zürich,DisneyResearch|Studios
Abstract:
We present SplatDiff, a pixel-splatting-guided diffusion model for single-image novel view synthesis (NVS). Leveraging pixel splatting and video diffusion, SplatDiff generates high-quality novel views with consistent geometry and high-fidelity details. SplatDiff achieves state-of-the-art results in single-view NVS and demonstrates remarkable zero-shot performance on sparse-view NVS and stereo video conversion.



Paperid:268
Authors:Youngsik Yun,Jeongmin Bae,Hyunseung Son,Seoha Kim,Hahyun Lee,Gun Bang,Youngjung Uh
Yonsei University,Yonsei University,Yonsei University,Electronics and Telecommunications Research Institute,Electronics and Telecommunications Research Institute,Electronics and Telecommunications Research Institute,Yonsei University
Abstract:
We reveal that existing online reconstruction of dynamic scenes with 3D Gaussian Splatting produces temporally inconsistent results, led by inevitable noise in real-world recordings. To address this, we decompose the rendered images into the ideal signal and the errors during optimization, achieving temporally consistent results across various baselines.



Paperid:269
Authors:Zhe Kong,Le Li,Yong Zhang,Feng Gao,Shaoshu Yang,Tao Wang,Kaihao Zhang,Zhuoliang Kang,Xiaoming Wei,Guanying Chen,Wenhan Luo
Sun Yat-sen University,Tianjin University,Meituan,Meituan,School of Artificial Intelligence, University of Chinese Academy of Sciences,Nanjing University,Harbin Institute of Technology,Meituan,Meituan,Sun Yat-sen University,The Hong Kong University of Science and Technology
Abstract:
In this work, we propose DAM-VSR, an appearance and motion disentanglement framework for video super-resolution. Appearance enhancement is achieved through reference image super-resolution, while motion control is achieved through video ControlNet. Additionally, we propose a motion-aligned bidirectional sampling strategy to support the generation of long videos.



Paperid:270
Authors:Janghyeok Han,Gyujin Sim,Geonung Kim,Hyun-Seung Lee,Kyuha Choi,Youngseok Han,Sunghyun Cho
POSTECH,POSTECH,POSTECH,Samsung Electronics,Samsung Electronics,Samsung Electronics,POSTECH
Abstract:
We propose DC-VSR, a novel video super-resolution approach based on a video diffusion prior. DC-VSR leverages Spatial and Temporal Attention Propagation (SAP and TAP) to ensure spatio-temporally consistent results and Detail-Suppression Self-Attention Guidance (DSSAG) to enhance high-frequency details. DC-VSR restores videos with realistic textures while maintaining spatial and temporal coherence.



Paperid:271
Authors:Ahmed H. Mahmoud,Serban D. Porumbescu,John D. Owens
Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT),University of California, Davis,University of California, Davis
Abstract:
Introducing the first GPU-based system for dynamic triangle mesh processing, delivering order-of-magnitude speedups over CPU solutions across diverse applications. Our system uses patch-based data structure, speculative conflict handling, and a novel programming model, enabling robust, high-performance, and fully dynamic mesh operations directly on the GPU.



Paperid:272
Authors:Abhishek Madan,Nicholas Sharp,Francis Williams,Ken Museth,David I.W. Levin
University of Toronto,NVIDIA,NVIDIA,NVIDIA,University of Toronto
Abstract:
We present a novel stochastic version of the Barnes-Hut approximation. Regarding the level-of-detail (LOD) family of approximations as control variates, we construct an unbiased estimator of the kernel sum being approximated. Through several examples in graphics, we demonstrate that our method outperforms a GPU-optimized implementation of the deterministic Barnes-Hut approximation.



Paperid:273
Authors:Kai Li,Xiaohong Jia,Falai Chen,Kai Li
State Key Laboratory of Mathematical Sciences, Academy of Mathematics and Systems Science, Chinese Academy of Sciences,State Key Laboratory of Mathematical Sciences, Academy of Mathematics and Systems Science, Chinese Academy of Sciences,University of Science and Technology of China,Key Laboratory of Mathematics Mechanization, Academy of Mathematics and Systems Science, Chinese Academy of Sciences; University of Chinese Academy of Sciences
Abstract:
Detecting surface self-intersections is crucial for CAD modeling to prevent issues in simulation and manufacturing. This paper presents an algebraic signature-based algorithm for fast determining self-intersections of NURBS surfaces. This signature is then recursively cross-used to compute the self-intersection locus, guaranteeing robustness in critical cases including tangency and small loops.



Paperid:274
Authors:Hugo Schott,Théo Thonat,Thibaud Lambert,Eric Guérin,Eric Galin,Axel Paris
INSA, Lyon,Adobe,Adobe,INSA, Lyon,Université Claude Bernard Lyon 1,Adobe
Abstract:
We introduce Sphere Carving, a method for automatically computing bounding volumes for conservative implicit surface. SDF queries define a set of spheres, from which we extract intersection points, used to compute a bounding volume with guarantees. Sphere Carving is conceptually simple and independent of the function representation.



Paperid:275
Authors:Cyprien Plateau--Holleville,Benjamin Stamm,Vincent Nivoliers,Maxime Maria,Stéphane Mérillou
Université de Limoges,Universität Stuttgart,Université Claude Bernard Lyon 1,Université de Limoges,Université de Limoges
Abstract:
We introduce a novel construction algorithm of 3D Apollonius diagrams designed for GPUs. Our method features a fast execution while allowing a comprehensive computation. This is made possible thanks to a light data structure, a cell update procedure and a spacial exploration strategy all designed to support the diagram properties.



Paperid:276
Authors:Huadong Zhang,Lizhou Cao,Chao Peng
Rochester Institute of Technology,Rochester Institute of Technology,Rochester Institute of Technology
Abstract:
This paper presents UltraMeshRenderer, a GPU out-of-core method for real-time rendering of 3D scenes with billions of vertices and triangles. It features a balanced hierarchical mesh, coherence-based LOD selection, and parallel in-place GPU memory management, achieving efficient data transfer and memory use with significant improvements over existing out-of-core techniques.



Paperid:277
Authors:Zhengming Yu,Tianye Li,Jingxiang Sun,Omer Shapira,Seonwook Park,Michael Stengel,Matthew Chan,Xin Li,Wenping Wang,Koki Nagano,Shalini De Mello
Texas A&M University,NVIDIA,Tsinghua University,NVIDIA,NVIDIA,NVIDIA,NVIDIA,Texas A&M University,Texas A&M University,NVIDIA,NVIDIA
Abstract:
We present GAIA (Generative Animatable Interactive Avatars) for high-fidelity 3D head avatar generation. GAIA learns dynamic details with expression-conditioned Gaussians, while being animatable consistently with an underlying morphable model. With a novel two-branch architecture, GAIA disentangles identity and expression. GAIA achieves state-of-the-art realism and supports interactive rendering and animation.



Paperid:278
Authors:Gengyan S. Li,Paulo Gotardo,Timo Bolkart,Stephan Garbin,Kripasindhu Sarkar,Abhimitra Meka,Alexandros Lattas,Thabo Beeler
ETH Zürich,Google,Google,Google,Google,Google,Google,Google
Abstract:
By combining a continuous, UVD tangent space 3DGS model with a UNet deformation network while maintaining adaptive densification, we present a novel high-detail 3D head avatar model that preserves even finer detail like pores and eyelashes at 4K resolution.



Paperid:279
Authors:Luchao Qi,Jiaye Wu,Bang Gong,Annie Wang,David Jacobs,Roni Sengupta
University of North Carolina at Chapel Hill (UNC),University of Maryland College Park,University of North Carolina Chapel Hill,University of North Carolina at Chapel Hill (UNC),University of Maryland College Park,University of North Carolina at Chapel Hill (UNC)
Abstract:
We personalize a pre-trained global aging prior using 50 personal selfies, allowing age regression (de-aging) and age progression (aging) with high fidelity and identity preservation.



Paperid:280
Authors:Howard Zhang,Yuval Alaluf,Sizhuo Ma,Achuta Kadambi,Jian Wang,Kfir Aberman
Snap,Tel Aviv University,Snap,University of California Los Angeles,Snap,Snap
Abstract:
InstantRestore is a fast, personalized face restoration framework that uses a single-step diffusion model with an extended self-attention mechanism to match low-quality image patches to high-quality reference patches. Leveraging implicit correspondences in the denoising network, we efficiently transfer identity details in one pass, enabling real-time, identity-preserving restoration without per-identity tuning.



Paperid:281
Authors:Shaofei Wang,Tomas Simon,Igor Santesteban,Timur Bagautdinov,Junxuan Li,Vasu Agrawal,Fabian Prada,Shoou-I Yu,Pace Nalbone,Matt Gramlich,Roman Lubachersky,Chenglei Wu,Javier Romero,Jason Saragih,Michael Zollhoefer,Andreas Geiger,Siyu Tang,Shunsuke Saito
ETH Zürich,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,Reality Labs Research, Meta,University of Tübingen,ETH Zürich,Reality Labs Research, Meta
Abstract:
We present the first drivable full-body avatar model that reconstructs perceptually realistic relightable appearance.



Paperid:282
Authors:Hendrik Junkawitsch,Guoxing Sun,Heming Zhu,Christian Theobalt,Marc Habermann
Max Planck Institute for Informatics,Max Planck Institute for Informatics,Max Planck Institute for Informatics,Max Planck Institute for Informatics,Max Planck Institute for Informatics
Abstract:
In this work, we introduce Expressive Virtual Avatars (EVA), an actor-specific, fully controllable and expressive human avatar framework that achieves high-fidelity, lifelike renderings in real-time, while enabling independent control of facial expressions, body movements, and hand gestures.



Paperid:283
Authors:Siyuan Shen,Tianjia Shao,Kun Zhou,Chenfanfu Jiang,Sheldon Andrews,Victor Zordan,Yin Yang
Zhejiang University,Zhejiang University,Zhejiang University,UCLA,École de technologie supérieure (ÉTS),Roblox,University of Utah
Abstract:
Our framework enables realistic and interesting elastic body locomotion by determining optimal muscle activations to achieve desired movements. It combines interior-point method for contact modeling with a novel mixed second-order differentiation algorithm that merges analytic and numerical approaches, allowing Newton's method optimization to create diverse soft body animations.



Paperid:284
Authors:Quentin Becker,Oliver Gross,Mark Pauly
EPFL,University of California San Diego,EPFL
Abstract:
We present a computational framework for optimizing shape sequences to achieve user-defined motion objectives in deformable bodies undergoing geometric locomotion. Through a reduced spatiotemporal parameterization of the shape sequences, our method is able to efficiently capture the complex coupling between shape changes and motion in different environments.



Paperid:285
Authors:Xinyu Yi,Shaohua Pan,Feng Xu
Tsinghua University,Tsinghua University,Tsinghua University
Abstract:
We propose a physics-driven approach to IMU-based motion capture, improving global motion estimation with 3D contact modeling and gravity awareness. Our method estimates world-aligned 3D motion, contact points, contact forces, joint torques, and proxy surface interactions using only six IMUs in real time.



Paperid:286
Authors:Hao Wang,Taogang Hou,Tianhui Liu,Jiaxin Li,Tianmiao Wang,Hao Wang,Taogang Hou
Beihang University,Beijing Jiaotong University,Beijing Jiaotong University,Beijing Jiaotong University,Beihang University,Biehang University, Beihang University,Beijing Jiaotong University
Abstract:
Marker-based optical motion capture (MoCap) is critical for virtual production and movement sciences. We propose a novel framework for MoCap auto-labeling and matching using uniquely coded clusters of reflective markers (AEMCs). Compared to commercial software, our method achieves higher labeling accuracy for heterogeneous targets and unknown marker layouts.



Paperid:287
Authors:KyeongMin Kim,SeungWon Seo,DongHeun Han,HyeongYeop Kang,KyeongMin Kim
Korea University,Korea University,KyungHee University,Korea University,Korea University
Abstract:
This paper introduces DAMO, a Deep solver for Arbitrary Marker configuration in Optical motion capture. DAMO directly infers the relationship between each raw marker point and 3D model joint, without using predefined marker labels and configuration information.



Paperid:288
Authors:Zhiyuan Yu,Zhe Li,Hujun Bao,Can Yang,Xiaowei Zhou
Department of Mathematics, Hong Kong University of Science and Technology,Huawei,State Key Laboratory of CAD&CG, Zhejiang University,Department of Mathematics, Hong Kong University of Science and Technology,State Key Laboratory of CAD&CG, Zhejiang Univerisity
Abstract:
Existing avatar methods typically require sophisticated dense-view capture and/or time-consuming per-subject optimization processes. HumanRAM proposes a feed-forward approach for generalizable human reconstruction and animation from monocular or sparse human images. Experiments show that HumanRAM achieves state-of-the-art results in terms of reconstruction accuracy, animation fidelity, and generalization performance.



Paperid:289
Authors:Wonjong Jang,Yucheol Jung,Gyeongmin Lee,Seungyong Lee
POSTECH,POSTECH,POSTECH,POSTECH
Abstract:
We present a novel framework that instantly (< 1 sec) repairs self-intersections in static surface meshes, which commonly occur during the 3D modeling process.



Paperid:290
Authors:Huibiao Wen,Guilong He,Rui Xu,Shuangmin Chen,Shiqing Xin,Zhenyu Shu,Taku Komura,Jieqing Feng,Wenping Wang,Changhe Tu
Shandong University,Shandong University,University of Hong Kong,Qingdao University of Science and Technology,Shandong University,NingboTech University,University of Hong Kong,State Key Laboratory of CAD & CG, Zhejiang University,Texas A&M University,Shandong University
Abstract:
We present a unified mesh repair framework using a manifold wrap surface to fix diverse imperfections while preserving sharp features. By optimizing projected samples and leveraging adaptive weighting, our method ensures watertightness, manifoldness, and high geometric fidelity, outperforming existing approaches in both topology correction and feature preservation.



Paperid:291
Authors:Daniel Zint,Zhouyuan Chen,Yifei Zhu,Denis Zorin,Teseo Schneider,Daniele Panozzo
New York University/ Courant,New York University/ Courant,New York University/ Courant,New York University/ Courant,University of Victoria,New York University/ Courant
Abstract:
Topological Offsets is a method for generating offset surfaces that are topologically equivalent to an offset infinitesimally close to the surface. By construction, the offsets are manifold, watertight, self-intersection-free, and strictly enclose the input. Tested on Thingi10k, it supports applications like manifold extraction, layered offsets, and robust finite offset computation.



Paperid:292
Authors:Julian Knodt
LightSpeed Studios
Abstract:
A simple and robust modification to triangle mesh reduction bridges the gap for what artists want in quad-dominant mesh reduction, preserving symmetry, topology, and joints without sacrificing geometric quality, allowing for high-quality level-of-detail meshes at no cost compared to what was done before.



Paperid:293
Authors:Alexandre Binninger,Ruben Wiersma,Philipp Herholz,Olga Sorkine-Hornung
ETH Zurich,ETH Zurich,Independent Contributor,ETH Zurich
Abstract:
TetWeave is a novel isosurface representation that jointly optimizes a tetrahedral grid and directional distances for gradient-based mesh processing like multi-view 3D reconstruction. It dynamically builds adaptive grids via Delaunay triangulation, ensuring watertight, manifold meshes. By resampling high-error regions and promoting fairness, it achieves high-quality results with minimal memory requirements.



Paperid:294
Authors:Tobias Kohler,Martin Heistermann,David Bommes
University of Bern,University of Bern,University of Bern
Abstract:
We present HexHex, which extracts a hexahedral mesh from a locally injective integer-grid map. Key contributions include a conservative rasterization technique and a novel mesh data structure called propeller. Our algorithm is significantly faster and uses less memory than the previous state-of-the-art method, especially for large hex-to-tet ratios.



Paperid:295
Authors:Timo Probst,Matthias Teschner,Timo Probst
University of Freiburg,University of Freiburg,University of Freiburg
Abstract:
We present a new SPH approach to replicate the behavior of droplets and other smaller scale fluid bodies. For this, we develop a new implicit surface tension formulation and implement a Coulomb friction force at the fluid-solid interface. A strong coupling between both forces and pressure is achieved through a unified solving mechanism.



Paperid:296
Authors:Mengyun Liu,Kai Bai,Xiaopei Liu
ShanghaiTech University,ShanghaiTech University,ShanghaiTech University
Abstract:
We present an innovative hybrid near-wall model for the multi-resolution lattice Boltzmann solver to effectively enable simulations of high Reynolds number turbulent boundary layer flows. For the first time, it strikes an excellent balance between the precision demanded by industrial computational design and the efficiency required for various visual animations.



Paperid:297
Authors:Yijie Liu,Taiyuan Zhang,Xiaoxiao Yan,Nuoming Liu,Bo Ren
TMCC, College of Computer Science, Nankai University,Dartmouth College,TMCC, College of Computer Science, Nankai University,TMCC, College of Computer Science, Nankai University,TMCC, College of Computer Science, Nankai University
Abstract:
We present a physics-based method for simulating intricate freezing dynamics on thin films. Our novel Phase Map method integrated with MELP particles reproduces Marangoni freezing dynamics and the "Snow-Globe Effect". The framework captures soap bubble freezing dynamics while ensuring stability in complex scenarios and enabling precise pattern control.



Paperid:298
Authors:Filipe Nascimento,Fabricio S. Sousa,Afonso Paiva
Universidade de São Paulo - USP,Universidade de São Paulo - USP,Universidade de São Paulo - USP
Abstract:
Powder-snow avalanches are natural phenomena that result from an instability in the snow cover on a mountain relief. This paper introduces a physically-based framework to simulate powder-snow avalanches under complex terrains, allowing us to animate the turbulent snow cloud dynamics within the avalanche in a visually realistic way.



Paperid:299
Authors:Zhanyu Yang,Aryamaan Jain,Guillaume Cordonnier,Marie-Paule Cani,Zhaopeng Wang,Bedrich Benes
Purdue University,Inria, Université Côte d'Azur,Inria, Université Côte d'Azur,Centre National de la Recherche Scientifique - Laboratoire d'informatique de l'École Polytechnique (LIX),Purdue University,Purdue University
Abstract:
Arenite is a novel, physics-based simulation method for generating realistic sandstone structures. It combines fabric interlocking, multi-factor erosion, and particle-based deposition. Our GPU-based implementation produces detailed 3D shapes such as arches, alcoves, hoodoos, and buttes in minutes and provides real-time control.



Paperid:300
Authors:Michael Liu,Xinlei Wang,Minchen Li
Carnegie Mellon University,NetEase Games Messiah Engine,Carnegie Mellon University
Abstract:
We introduce a compact, C2-continuous kernel for MPM that reduces numerical diffusion and improves efficiency—without sacrificing stability. Built on a dual-grid framework and compatible with APIC and MLS, our method enables high-fidelity, large-scale simulations, further pushing the limits of MPM.



Paperid:301
Authors:Yuanpeng Tu,Luo Hao,Chen Xi,Sihui Ji,Xiang Bai,Zhao Hengshuang
The University of Hong Kong,DAMO Academy, Alibaba Group,The University of Hong Kong,The University of Hong Kong,Huazhong University of Science and Technology,The University of Hong Kong
Abstract:
We propose VideoAnydoor, a zero-shot video object insertion framework with high-fidelity detail preservation and precise motion control, where a pixel warper and a image-video mix-training strategy are designed to warp the pixel details according to the trajectories. VideoAnydoor demonstrates significant superiority over existing methods and naturally supports various downstream applications.



Paperid:302
Authors:Feng-Lin Liu,Shi-Yang Li,Yan-Pei Cao,Hongbo Fu,Lin Gao
Institute of Computing Technology, Chinese Academy of Sciences,Institute of Computing Technology, Chinese Academy of Sciences,VAST,Hong Kong University of Science and Technology,Institute of Computing Technology, Chinese Academy of Sciences
Abstract:
We propose Sketch3DVE, a sketch-based 3D-aware video editing method to enable detailed local manipulation of videos with significant viewpoint changes. Our approach leverages detailed analysis and editing of underlying 3D scene representations, combined with a diffusion model to synthesize realistic and temporally coherent edited videos.



Paperid:303
Authors:Yuxuan Bian,Zhaoyang Zhang,Xuan Ju,Mingdeng Cao,Liangbin Xie,Ying Shan,Qiang Xu
The Chinese University of Hong Kong,Tencent,The Chinese University of Hong Kong,The University of Tokyo,University of Macau,Tencent,The Chinese University of Hong Kong
Abstract:
VideoPainter introduces a dual-branch framework for video inpainting with a lightweight context encoder that integrates with pre-trained diffusion transformers. Its ID resampling strategy maintains identity consistency across any-length videos, while VPData and VPBench provide the largest segmentation-mask dataset with captions. The system achieves state-of-the-art performance in video inpainting and editing.



Paperid:304
Authors:Hongbo Zhao,Jiaxing Li,Peiyi Zhang,Peng Xiao,Jianxin Lin,Yijun Wang
Hunan University,Hunan University,Hunan University,Hunan University,Hunan University,Hunan University
Abstract:
We propose ColorSurge, a lightweight dual-branch network for end-to-end video colorization. It delivers vivid, accurate, and real-time results from grayscale input, and is easily extensible for high-quality performance at low computational cost.



Paperid:305
Authors:Karlis Martins Briedis,Abdelaziz Djelouah,Raphaël Ortiz,Markus Gross,Christopher Schroers
DisneyResearch|Studios,DisneyResearch|Studios,DisneyResearch|Studios,DisneyResearch|Studios,DisneyResearch|Studios
Abstract:
We present a tracking-based video frame interpolation method, optionally guided by user inputs. It utilizes sparse point tracks, first estimated using existing point tracking methods and then optionally refined by the user. Without any user input, it already achieves state-of-the-art results, with further significant improvements possible through user interactions.



Paperid:306
Authors:Manuel Kansy,Jacek Naruniec,Christopher Schroers,Markus Gross,Romann Weber
ETH Zürich,Disney Research Studios,Disney Research Studios,ETH Zürich,Disney Research Studios
Abstract:
Reenact Anything introduces a unified framework for semantic motion transfer, covering applications from full-body and face reenactment to controlling the motion of inanimate objects and the camera. Thereby, motions are represented using text/image embeddings of an image-to-video diffusion model and are optimized based on a given motion reference video.



Paperid:307
Authors:Bosheng Li,Nikolas Schwarz,Wojtek Palubicki,Sören Pirk,Dominik L. Michels,Bedrich Benes
Purdue University,Kiel University,AMU,Kiel University,King Abdullah University of Science and Technology (KAUST),Purdue University
Abstract:
A novel approach for the computational modeling of lignified tissues, such as those found in tree branches and timber, extends strand-based representation to describe biophysical processes at short and long time scales. The computationally fast simulation leverages Cosserat rod physics and enables the interactive exploration of branches and wood breaking.



Paperid:308
Authors:Hao Xu,Yinqiao Wang,Niloy Mitra,Shuaicheng Liu,Pheng Ann Heng,Chi-Wing Fu
The Chinese University of Hong Kong,The Chinese University of Hong Kong,University College London (UCL),University of Electronic Science and Technology of China,Chinese University of Hong Kong,The Chinese University of Hong Kong
Abstract:
We solve an inverse hand-shadow problem: finding poses of left and right hands that together produce a shadow resembling the target 2D input, e.g., animals, letters, and everyday objects. Our three-stage pipeline decouples the anatomical constraints and semantic constraints, and our benchmark provides 210 diverse shadow shapes of varying complexity.



Paperid:309
Authors:Rahul Mitra,Mattéo Couplet,Tongtong Wang,Megan Hoffman,Kui Wu,Edward Chien
Boston University,Boston University,LightSpeed Studios,Northeastern University,LightSpeed Studios,Boston University
Abstract:
We present a method for the automatic placement of knit singularities based on curl quantization. Our method generates knit graphs that maintain all structural manufacturing constraints as well as any additional user constraints. This approach allows for simulation-free previews of rendered knits and also extends to the popular cut-and-sew setting.



Paperid:310
Authors:Klara Mundilova,Michele Vidulis,Quentin Becker,Florin Isvoranu,Mark Pauly
EPFL,EPFL,EPFL,EPFL,EPFL
Abstract:
C-tubes are 3D tubular structures made of developable strips. We introduce an algorithm to construct C-tubes while guaranteeing exact surface developability and an optimization method for design exploration. Applications span architecture, engineering, and product design. We present prototypes showcasing cost-effective fabrication of complex geometries using different materials.



Paperid:311
Authors:Fanchao Zhong,Yang Wang,Peng-Shuai Wang,Lin Lu,Haisen Zhao
Shandong University,Shandong University,Peking University,Shandong University,Shandong University
Abstract:
The proposed neural network, DeepMill, can efficiently predict inaccessible and occlusion regions in subtractive manufacturing. By utilizing a cutter-aware dual-head octree-based convolutional architecture, it overcomes the computational inefficiency of traditional geometric methods and is capable of real-time prediction of inaccessible and occlusion regions during the 3D shape design phase.



Paperid:312
Authors:Weizheng Zhang,Hao Pan,Lin Lu,Xiaowei Duan,Xin Yan,Ruonan Wang,Qiang Du
Shandong University,School of Software, Tsinghua University,Shandong University,Shandong University,Shandong University,Institute of Engineering Thermophysics, Chinese Academy of Sciences,Institute of Engineering Thermophysics, Chinese Academy of Sciences
Abstract:
DualMS is a novel framework for designing high-performance heat exchangers by directly optimizing the separation surface of two fluids using dual skeleton optimization and neural implicit functions. It offers greater topological flexibility than TPMS and achieves superior thermal performance with lower pressure drop while maintaining comparable heat exchange rates.



Paperid:313
Authors:Zilin Xu,Xiang Chen,Chen Liu,Beibei Wang,Lu Wang,Zahra Montazeri,Ling-Qi Yan
University of California Santa Barbara,Shandong University,Zhejiang Lingdi Digital Technology Co.,Ltd,Nanjing University,Shandong University,University of Manchester,University of California Santa Barbara
Abstract:
We challenge the comprehensive neural material representation by thoroughly considering the essential aspects of the complete appearance. We introduce an int8-quantized model that keeps high fidelity while achieving an order of magnitude speedup compared to previous methods, and a controllable structure-preserving synthesis strategy, along with accurate displacement effects.



Paperid:314
Authors:Nithin Raghavan,Krishna Mullia,Alexander Trevithick,Fujun Luan,Miloš Hašan,Ravi Ramamoorthi
University of California San Diego,Adobe Research,University of California San Diego,Adobe Research,Adobe Research,University of California San Diego
Abstract:
We present the first generative model for neural BTFs, enabling single-shot generation from arbitrary text or image prompts. To achieve this, we introduce a universal neural material basis and train a conditional diffusion model to generate materials in this basis from flash images, natural images and text prompts.



Paperid:315
Authors:Liwen Wu,Sai Bi,Zexiang Xu,Hao Tan,Kai Zhang,Fujun Luan,Haolin Lu,Ravi Ramamoorthi
University of California San Diego,Adobe Research,Hillbot,Adobe Research,Adobe Research,Adobe Research,Max Planck Institute for Informatics,University of California San Diego
Abstract:
We introduce a reparameterization-based formulation of neural BRDF importance sampling. Comparing to previous methods that construct a probability transform to the BRDF through multi-step invertible neural networks, our BRDF sampling is in single step without needing network invertibility, achieving higher inference speed with the best variance reduction.



Paperid:316
Authors:Krishna Mullia,Fujun Luan,Xin Sun,Miloš Hašan,Krishna Mullia
Adobe Research,Adobe Research,Adobe Research,Adobe Research,Adobe Research
Abstract:
We propose a neural representation for 3D assets with complex shading. We precompute shading and scattering on ground-truth geometry, enabling high-fidelity rendering with full relightability, eliminating complex shading models and multiple scattering paths, offering significant speed-ups and seamless integration into existing rendering pipelines.



Paperid:317
Authors:Saeed Hadadan,Benedikt Bitterli,Tizian Zeltner,Jan Novák,Fabrice Rousselle,Jacob Munkberg,Jon Hasselgren,Bartlomiej Wronski,Matthias Zwicker
University of Maryland College Park,NVIDIA Research,NVIDIA Research,NVIDIA Research,NVIDIA Research,NVIDIA Research,NVIDIA Research,NVIDIA Research,University of Maryland College Park
Abstract:
We present a tool for enhancing the detail of physically based materials using an off-the-shelf diffusion model and inverse rendering. Our goal is to enhance the visual fidelity of materials with detail that is often tedious to author, by adding signs of wear, aging, weathering, etc.



Paperid:318
Authors:Yang Zhou,Tao Huang,Ravi Ramamoorthi,Pradeep Sen,Ling-Qi Yan,Ling-Qi Yan
University of California Santa Barbara,University of California Santa Barbara,University of California San Diego,University of California Santa Barbara,University of California Santa Barbara,University of California Santa Barbara
Abstract:
We present a novel volumetric representation for the aggregated appearance of complex scenes and a pipeline for level-of-detail generation and rendering. Our representation preserves accurate far-field appearance and spatial correlation from scene geometry. Our method faithfully reproduces appearance and achieves higher quality than existing scene filtering methods.



Paperid:319
Authors:Hossein Baktash,Nicholas Sharp,Qingnan Zhou,Alec Jacobson,Keenan Crane
Carnegie Mellon University,NVIDIA Research,Adobe Research,Adobe Research,Carnegie Mellon University
Abstract:
We identify stable orientations of any rigid shape, and the probability that it will rest at these orientations if randomly dropped on the ground. We use a differentiable inverse version of our method to design and fabricate shapes with target resting behavior, such as dice with target, nonuniform probabilities.



Paperid:320
Authors:Moritz Bächer,Ruben Grandia,Espen Knoop,Guirec Maloisel,Christian Schumacher
Disney Research,Disney Research,Disney Research,Disney Research,Disney Research
Abstract:
We present an implicitly-integrated, quaternion-based constrained Rigid Body Dynamics (RBD) that guarantees satisfaction of kinematic constraints, unifying the solution strategy for complex mechanical systems with arbitrary kinematic structures, by navigating subspaces spanned by constraint forces and torques for systems with redundant constraints, over actuation, and passive degrees of freedom.



Paperid:321
Authors:Magí Romanyà,Miguel A. Otaduy
Universidad Rey Juan Carlos,Universidad Rey Juan Carlos
Abstract:
This work introduces a forward and differentiable rigid-body dynamics framework using Lie-algebra rotation derivatives. The approach offers simplified, compact derivatives, improved conditioning, and higher efficiency compared to traditional methods. Applications include fundamental rigid-body problems and Cosserat rods, showcasing its potential for multi-rigid-body dynamics and incremental-potential formulations.



Paperid:322
Authors:Zizhou Huang,Maxwell Paik,Zachary Ferguson,Daniele Panozzo,Denis Zorin
New York University,New York University,Massachusetts Institute of Technology,New York University,New York University
Abstract:
We present a systematic derivation of a continuum potential defined for smooth and piecewise smooth surfaces, by identifying a set of natural requirements for contact potentials. Our potential is formulated independently of surface discretization and addresses the shortcomings of existing potential-based methods while retaining their advantages.



Paperid:323
Authors:Xiaodi Yuan,Fanbo Xiang,Yin Yang,Hao Su
University of California San Diego,Hillbot Inc.,University of Utah,University of California San Diego
Abstract:
We propose a fast, single-threaded continuous collision detection (CCD) algorithm for convex shapes under affine motion. By combining conservative advancement with a cone-casting approach, it avoids primitive-level overhead and enables efficient integration into intersection-free simulation methods such as ABD.



Paperid:324
Authors:Anka H. Chen,Jerry Hsu,Ziheng Liu,Miles Macklin,Yin Yang,Cem Yuksel
University of Utah,University of Utah,University of Utah,NVIDIA,University of Utah,University of Utah
Abstract:
We introduce Offset Geometric Contact (OGC), a groundbreaking method offering "penetration-free for free" simulations of codimensional objects. OGC efficiently constructs offset volumetric shapes to ensure stable, artifact-free collisions. Leveraging parallel GPU computations, it delivers real-time simulations at speeds over 100× faster than previous methods, eliminating costly collision detection and global-synchronization.



Paperid:325
Authors:Nuri Ryu,Jiyun Won,Jooeun Son,Minsu Gong,Joo-Haeng Lee,Sunghyun Cho
POSTECH,POSTECH,POSTECH,POSTECH,Pebblous,POSTECH
Abstract:
Elevate3D transforms low-quality 3D models into high-quality assets through iterative texture and geometry refinement. At its core, HFS-SDEdit refines textures generatively while preserving the input’s identity leveraging high-frequency guidance. The resulting texture then guides geometry refinement, allowing Elevate3D to deliver high-quality results with well-aligned texture and geometry.



Paperid:326
Authors:Julian Knodt,Xifeng Gao,Julian Knodt
LightSpeed Studios,LightSpeed Studios,Lightspeed Studios, Tencent America
Abstract:
We develop a method to compress textures and UVs for meshes in a content-aware way. We combine this with overlapping and folding symmetric UV charts, and demonstrate our approach on a dataset from Sketchfab. We outperform prior work in visual similarity to the original mesh.



Paperid:327
Authors:Yuqing Zhang,Hao Xu,Yiqian Wu,Sirui Chen,Sirui Lin,Xiang Li,Xifeng Gao,Xiaogang Jin
Zhejiang University,Zhejiang University,Zhejiang University,Zhejiang University,Zhejiang University,Shenzhen University,Lightspeed Studios, Tencent America,Zhejiang University
Abstract:
AlignTex is a novel framework for generating high-quality textures from 3D meshes and multi-view artwork. It improves texture generation by ensuring both appearance detail and geometric consistency, outpacing traditional methods in quality and efficiency, making it a valuable tool for 3D asset creation in gaming and film production.



Paperid:328
Authors:Maria Larsson,Hodaka Yamaguchi,Ehsan Pajouheshgar,I-Chao Shen,Kenji Tojo,Chia-Ming Chang,Lars Hansson,Olof Broman,Takashi Ijiri,Ariel Shamir,Wenzel Jakob,Takeo Igarashi
The University of Tokyo,Gifu Prefecture Research Institute for Human Life Technology,École Polytechnique Féderale de Lausanne (EPFL),The University of Tokyo,The University of Tokyo,National Taiwan University of Arts,Luleå University of Technology,Luleå University of Technology,Shibaura Institute of Technology,Reichman University,The University of Tokyo,The University of Tokyo
Abstract:
We present the Mokume dataset for solid wood texturing, comprising nearly 190 samples from various species. Using this dataset, we propose an inverse modeling pipeline to infer volumetric wood textures from surface photographs, employing inverse procedural texturing and neural cellular automata (NCA).



Paperid:329
Authors:Crane He Chen,Vladimir Kim
Industrial Light & Magic,Adobe
Abstract:
A real-time deformation method for Escher tiles --- interlocking organic forms that seamlessly tessellate the plane following symmetry rules. Rather than treating tiles as mere boundaries, we consider them as textured shapes, ensuring that both the boundary and interior deform simultaneously. The deformation is achieved via a closed-form solution.



Paperid:330
Authors:Marco Maida,Alberto Crescini,Marco Perronet,Elena Camuffo
Independent Researcher,Independent Researcher,Independent Researcher,Independent Researcher
Abstract:
We introduce a novel scannable 2D code where the payload is stored in the topology of nested color regions, abandoning traditional matrix-based approaches (e.g., QRCodes). Claycodes can be largely deformed, styled, and animated. We present a mapping between bits and topologies, shape-constrained rendering, and a robust real-time decoding pipeline.



Paperid:331
Authors:Cédric Zanni,Cédric Zanni
Université de Lorraine CNRS, Inria, LORIA,Université de Lorraine, CNRS, Inria, LORIA, F-54000 Nancy, France; LORIA
Abstract:
The paper presents a tile-based rendering pipeline for modeling with implicit volumes, using blobtrees and smooth CSG operators. It requires no preprocessing when updating primitives and ensures efficient ray processing with sphere tracing. The method uses a low-resolution A-buffer and bottom-up tree traversal for scalable performance.



Paperid:332
Authors:Kechun Wang,Renjie Chen
University of Science and Technology of China,University of Science and Technology of China
Abstract:
Higher-order surfaces enable compact, smooth geometry but require efficient rendering. We introduce PaRas, a GPU-based rasterizer that directly renders parametric surfaces, avoiding costly tessellation. It integrates seamlessly into existing pipelines, outperforming traditional methods for quartic triangular and bicubic rational Bézier patches. Experimental results confirm its superior efficiency and accuracy.



Paperid:333
Authors:Pengfei Zhu,Jie Guo,Yifan Liu,Qi Sun,Yanxiang Wang,Keheng Xu,Ligang Liu,Yanwen Guo
Nanjing University,Nanjing University,Nanjing University,Nanjing University,Nanjing University,Nanjing University,University of Science and Technology of China,Nanjing University
Abstract:
This paper introduces an appearance-aware adaptive sampling method using deep reinforcement learning to optimize the reconstruction of spatially-varying BRDFs from minimal images. By modeling the sampling as a sequential decision-making problem, the method identifies the next best view-lighting pair, outperforming heuristic sampling strategies for heterogeneous materials.



Paperid:334
Authors:Ruben Wiersma,Julien Philip,Miloš Hašan,Krishna Mullia,Fujun Luan,Elmar Eisemann,Valentin Deschaintre
ETH Zürich,Netflix Eyeline Studios,Adobe Research,Adobe Research,Adobe Research,Delft University of Technology,Adobe Research
Abstract:
We quantify uncertainty for SVBRDF acquisition from multi-view captures using entropy. The otherwise heavy computation is accelerated in the frequency domain, yielding a practical, efficient method. We apply uncertainty to improve SVBRDF capture by guiding camera placement, inpainting uncertain regions, and sharing information from certain regions on the object.



Paperid:335
Authors:Hangming Fan,Yuchi Huo,Chuankun Zheng,Chonghao Hu,Yazhen Yuan,Rui Wang
State Key Laboratory of CAD & CG, Zhejiang University,State Key Laboratory of CAD & CG, Zhejiang University,State Key Laboratory of CAD & CG, Zhejiang University,State Key Laboratory of CAD & CG, Zhejiang University,Game Engine Department, CROS, Tencent, China,State Key Laboratory of CAD & CG, Zhejiang University
Abstract:
To reduce the high rendering costs and transmission bandwidth requirements of path tracing-based cloud rendering, we propose a novel streaming-aware rendering framework that is able to learn a joint optimal model integrating two path-tracing acceleration (adaptive sampling and denoising) and video compression technique with client side G-buffer collaboration.



Paperid:336
Authors:Rachel McDonnell,Bharat Vyas,Uros Sikimic,Pisut Wisessing
Trinity College Dublin,Trinity College Dublin,Epic Games,CMKL University
Abstract:
This study explores how light color influences the perception of emotion of virtual characters. By analyzing various lighting conditions, including red and blue hues, we reveal how light affects emotion intensity, recognition, and genuineness. Findings show that lighting, realism, and shadows are key factors in enhancing the emotional impact.



Paperid:337
Authors:Jingwen Ye,Yuze He,Yanning Zhou,Yiqin Zhu,Kaiwen Xiao,Yong-Jin Liu,Wei Yang,Xiao Han
Tencent AIPD,Tencent AIPD,Tencent AIPD,Tencent AIPD,Tencent AIPD,Tsinghua University,Tencent AIPD,Tencent AIPD
Abstract:
We present PrimitiveAnything, a novel framework that reformulates shape primitive abstraction as a primitive assembly generation task. PrimitiveAnything can generate 3D high-quality primitive assemblies that better align with human perception while maintaining geometric fidelity across diverse shape categories, which benefits various 3D applications.



Paperid:338
Authors:Zhenyu Wang,Min Lu
Shenzhen University,Shenzhen University
Abstract:
This work introduces an efficient image-space collage technique that optimizes geometric layouts using a differential renderer and hierarchical resolution strategy. Our approach simplifies complex shape handling in image-space optimization, offering fixed computational complexity. Experiments show our method is an order of magnitude faster than state-of-the-art while supporting diverse visual expressions.



Paperid:339
Authors:Junming Huang,Chi Wang,Letian Li,Changxin Huang,Qiang Dai,Weiwei Xu
State Key Laboratory of CAD & CG, Zhejiang University,State Key Laboratory of CAD & CG, Zhejiang University,State Key Laboratory of CAD & CG, Zhejiang University,State Key Laboratory of CAD & CG, Zhejiang University,LIGHTSPEED,State Key Lab CAD&CG, Zhejiang University, ZJU-Tencent Game and Intelligent Graphics Innovation Technology Joint Lab
Abstract:
We propose BuildingBlock, a hybrid approach integrating generative models, PCG, and LLMs for diverse and structured 3D building generation. A Transformer-based diffusion model generates layouts, which LLMs refine into hierarchical designs. PCG then constructs high-quality buildings, achieving state-of-the-art results and enabling scalable architectural workflows.



Paperid:340
Authors:Jiepeng Wang,Hao Pan,Yang Liu,Xin Tong,Taku Komura,Wenping Wang,Jiepeng Wang
The University of Hong Kong,Microsoft Research Asia,Microsoft Research Asia,Microsoft Research Asia,The University of Hong Kong,Texas A&M University,University of Hong Kong, Microsoft Research Asia
Abstract:
The paper presents StructRe, a structure rewriting system for 3D shape modeling. It uses an iterative process to rewrite objects, either upwards to more concise structures or downwards to more detailed ones, generating hierarchies. This localized rewriting approach enables probabilistic modeling of ambiguous structures and robust generalization across object categories.



Paperid:341
Authors:Davide Sforza,Marzia Riso,Filippo Muzzini,Nicola Capodieci,Fabio Pellacini
Sapienza University of Rome,Sapienza University of Rome,University of Modena and Reggio Emilia,University of Modena and Reggio Emilia,University of Modena and Reggio Emilia
Abstract:
We introduce a method for interactive design of procedural patterns, allowing users to sketch content incrementally in a level-by-level fashion. Each level, or scaffold, builds on the previous one, making optimization more responsive and controllable. A comprehensive validation demonstrates improved editing experience compared to conventional techniques.