These 46 SIGGRAPH 2025 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every SIGGRAPH 2025 paper, free trial on arXivSub.
🎯 What it does: Inject the visual style of reference images into 3D geometry using large 3D reconstruction models (e.g., InstantMesh), achieving stylization of 3D objects without the need for training or testing during optimization.
A Divide-and-Conquer Approach for Global Orientation of Non-Watertight Scene-Level Point Clouds with 0-1 Integer Optimization
Zhuodong Li (Chinese Academy of Sciences), Ying He (Nanyang Technological University)
CodeOptimizationPoint Cloud
🎯 What it does: This paper proposes the DACPO (Divide-and-Conquer Point Orientation) framework for achieving global normal orientation in non-closed, scene-level point clouds. First, the point cloud is divided into several mutually connected small blocks, and each block is independently processed with normal initialization and iterative Poisson reconstruction (iPSR) to obtain locally consistent normals. Subsequently, an inter-block graph is built based on visible connected regions (VCR), and a 0-1 integer optimization is applied to determine whether each block needs to be flipped, thereby achieving globally consistent orientation across the entire scene.
A Fully-statistical Wave Scattering Model for Heterogeneous Surfaces
Zhengze Liu, Rui Wang
CodePhysics Related
🎯 What it does: A model is proposed that provides a fully statistical description of heterogeneous surfaces at the microscale, and calculates their BRDF and scattering field covariance based on the Harvey-Shack theory, thereby enabling sampling of light spots and enriching surface appearance without explicitly defining the surface.
A Platform for Interactive AI Character Experiences
Rafael Wampfler (ETH Zurich), Markus Gross (ETH Zurich)
CodeGenerationData SynthesisRecommendation SystemExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Built an scalable digital character interaction platform, demonstrating dialogues and narrative-driven experiences with a virtual Einstein avatar.
🎯 What it does: Propose a two-step UAV path planning method based on prior reconstruction models and change probability statistics, specifically designed for change detection and local updates in urban scenarios.
🎯 What it does: This paper proposes the Anymate dataset and a three-stage learning framework for automatically generating skeletal assemblies and skin weights from static 3D meshes, achieving fully automatic 3D object animation.
Asymptotic analysis and design of linear elastic shell lattice metamaterials
Di Zhang (University of Science and Technology of China), Ligang Liu (University of Science and Technology of China)
CodeOptimizationMeshPhysics Related
🎯 What it does: Based on Ciarlet's thin shell theory, an asymptotic analysis is conducted on shell lattice structures with thickness approaching zero, introducing and strictly defining 'Asymptotic Directional Stiffness (ADS)', and providing its convergence theorem, upper bound, and optimality conditions; subsequently, a discrete and shape optimization framework based on triangular mesh is developed, enabling efficient design of periodic mid-surface structures such as TPMS.
Automated Task Scheduling for Cloth and Deformable Body Simulations in Heterogeneous Computing Environments
Chen He, Huamin Wang
CodeOptimizationComputational EfficiencyDiffusion modelGaussian SplattingOptical FlowMeshPhysics Related
🎯 What it does: Developed an automatic task scheduling framework to optimize the performance of fabric and deformable body simulations in heterogeneous computing environments.
🎯 What it does: Propose a C² continuous, radius-2 compact kernel Material Point Method (CK-MPM), and achieve robust mapping between particles and grids through a dual-grid framework, compatible with APIC and MLS-MPM;
🎯 What it does: Proposed a long-context reference line drawing coloring framework called Cobra, which can efficiently and accurately color comic line drawings and supports more than 200 reference images and color prompts.
🎯 What it does: Propose the DualMS framework, which optimizes the separation surface through dual-channel minimal surface design for free-form heat exchangers, aiming to achieve a balance between maximizing heat transfer efficiency and minimizing pressure drop.
🎯 What it does: Generate complete cello performance actions directly from audio, including hand details and bow movements, forming a full-body interactive performance.
🎯 What it does: A home-based facial appearance capture method based on a Patch-level diffusion model was developed, which reconstructs high-quality illumination, geometry, and reflectance properties using a single smartphone plus flashlight video.
🎯 What it does: Proposed a fast median/percentile filtering algorithm applicable to arbitrary bit depth, arbitrary radius, and arbitrary convex-shaped (especially circular) kernels, with both CPU and GPU implementations.
🎯 What it does: This paper proposes constructing Green coordinates on three-dimensional Bézier surface cages (including tensor product Bézier patches and Bézier triangles), thereby achieving compact and flexible cage deformation.
🎯 What it does: Propose a mesh-free fluid solver based on Gaussian Spatial Representation (GSR), which can continuously and differentiably represent the flow field and achieve the temporal evolution of Navier-Stokes equations through first-order optimization.
🎯 What it does: Transform the video matting task into a conditional generation problem, utilizing the pre-trained Stable Video Diffusion model combined with multi-stage training and synthesized/semi-annotated data to achieve high-quality video matting.
🎯 What it does: Proposes a three-stage learning framework called Hand-Shadow Poser, which is used to inversely estimate the 3D joint poses of both hands from only the binary mask of hand shadows, thereby generating a light projection similar to the target hand shadow.
🎯 What it does: Propose a neural network method called Kernel Predicting Neural Shadow Mapping, which converts hard shadow values into soft shadows by predicting pixel-level local filter weights, and achieves real-time high-quality shadows through the use of dilated filters and a loss function with temporal regularization.
🎯 What it does: Studied the correspondence between eye movements and painting actions during the image-to-image drawing process, and based on this, developed a real-time visual guidance-assisted painting interface.
🎯 What it does: A new geometric representation method called mescher is proposed for handling impossible objects, allowing rendering and relighting of these objects, and performing intrinsic geometric processing operations such as heat diffusion and geodesic distance queries.
🎯 What it does: Propose the Mobius method, which utilizes a pre-trained text-to-video diffusion model, and achieves seamless loop video generation without training by performing cyclic shifts in the latent space during the inference phase.
🎯 What it does: Propose a multi-light importance sampling method based on neural networks, using local illumination information to predict the light source selection distribution for each shading point, and training the network online during rendering.
🎯 What it does: This paper proposes OctGPT, a multi-scale autoregressive model based on serialized octrees, for high-quality, high-resolution 3D shape and scene generation.
PARC: Physics-based Augmentation with Reinforcement Learning for Character Controllers
Michael Xu (Simon Fraser University), Xue Bin Peng (Simon Fraser University)
CodeGenerationData SynthesisRobotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelDiffusion modelScore-based ModelVideoSequentialPhysics Related
🎯 What it does: Propose the PARC framework, starting from a small-scale motion capture dataset, iteratively training a motion generator and a physical motion tracking controller to achieve agile character control on complex terrains.
🎯 What it does: Supports fine-grained, precise, and seamless image editing without the need for manual masks by learning object part-specific text tokens in pre-trained diffusion models.
Piecewise Ruled Approximation for Freeform Mesh Surfaces
Yiling Pan (Tsinghua University), Bailin Deng (Cardiff University)
CodeOptimizationMesh
🎯 What it does: A piecewise supporting surface approximation method for arbitrary free-form triangular meshes is proposed. The target mesh is converted into an approximate supporting surface through sparsification optimization, and the optimization results are used to extract piecewise boundaries, construct initial supporting line segments, and further optimize to improve approximation accuracy.
🎯 What it does: This paper proposes a manifold-based surface orthogonal distribution function (P-NDF) computation method for high-resolution normal map rendering of specular highlights, which preserves micro-details while significantly improving computational speed.
Jiabao Wang, Amir Vaxman (University of Edinburgh)
CodeOptimizationMesh
🎯 What it does: This paper proposes a design method for polar fields based on a piecewise power-linear representation, which can specify singularities with arbitrary exponents at any position on the mesh (faces, edges, points) and generate smooth, non-aliased direction fields;
Sketch3DVE: Sketch-based 3D-Aware Scene Video Editing
Feng-Lin Liu (Institute of Computing Technology, Chinese Academy of Sciences), Lin Gao (Institute of Computing Technology, Chinese Academy of Sciences)
🎯 What it does: Proposed a 3D-aware video editing method based on hand-drawn sketches, named Sketch3DVE, which enables local editing of structural content in scene videos with significant changes in perspective;
CodeImage TranslationImage HarmonizationGenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageMultimodality
🎯 What it does: Propose Stable-Makeup, a novel makeup transfer framework based on diffusion models, which can achieve fine-grained transfer of diverse makeup styles ranging from light to heavy in real-world scenarios.
🎯 What it does: Developed a gradient-optimized implicit surface representation called TetWeave, which jointly optimizes point clouds, directional SDF, and Delaunay tetrahedral mesh, achieving differentiable, self-intersection-free, two-manifold, and watertight mesh reconstruction.
🎯 What it does: Generate high-quality animatable 3D human face avatars based on textual descriptions and align them precisely with the SMPL-X parameterized model.
🎯 What it does: Propose a novel Topological Offset algorithm that can generate an offset surface with the same topology as the original surface, without self-intersections, closed, and maintaining a certain distance from the original surface;
Towards Understanding Depth Perception in Foveated Rendering
Sophie Kergaßner (Università della Svizzera italiana), P. Didyk
CodeDepth EstimationExplainability and InterpretabilityComputational EfficiencyContrastive LearningGaussian SplattingOptical FlowImage
🎯 What it does: Studied the depth perception thresholds of disparity signals affected by blur in the peripheral visual field, and constructed the corresponding perception model.
🎯 What it does: Proposed a dynamic IMU calibration framework based on Transformer, achieving implicit real-time calibration for sparse inertial motion capture;
🎯 What it does: This paper proposes an entropy-based SVBRDF uncertainty assessment method, achieving efficient and fast computation through frequency domain (spherical harmonics) analysis, and utilizes uncertainty for guidance, information sharing, and diffusion model filling.
🎯 What it does: This paper proposes the Virtualized 3D Gaussians (V3DG) system, which enables real-time rendering in scenes composed of large-scale 3D Gaussian assets, and dynamically controls rendering detail through offline construction of multi-level clustering and online Footprint selection.
Zhaofeng Luo (Carnegie Mellon University), Minchen Li (Carnegie Mellon University)
CodeComputational EfficiencyRobotic IntelligenceGaussian SplattingSimultaneous Localization and MappingOptical FlowPoint CloudMesh
🎯 What it does: Developed VR-Doh, a system that realizes real-time physical simulation and 3D model editing through hand interaction in virtual reality.