arXivSub Start free trial

SIGGRAPH 2025 Papers — Page 3

ACM SIGGRAPH (Transactions on Graphics) · 306 papers

OctGPT: Octree-based Multiscale Autoregressive Models for 3D Shape Generation

Si-Tong Wei (Peking University), Peng-Shuai Wang (Peking University)

GenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingImageTextMultimodalityPoint CloudMeshRetrieval-Augmented Generation

🎯 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.

Offset Geometric Contact

A. Chen, C. Yuksel

OptimizationComputational EfficiencyRobotic IntelligenceDiffusion modelOptical FlowMeshPhysics RelatedStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: A new contact model called Offset Geometric Contact (OGC) is proposed to ensure real-time simulation of encoded-dimensional objects without penetration.

On Planar Shape Interpolation With Logarithmic Metric Blending

Alon Feldman, M. Ben-Chen

Mesh

🎯 What it does: Proposes a planar shape interpolation method based on logarithmic metric blending

On-the-fly Reconstruction for Large-Scale Novel View Synthesis from Unposed Images

Andréas Meuleman, G. Drettakis

Data SynthesisPose EstimationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelScore-based ModelNeural Radiance FieldContrastive LearningGaussian SplattingSimultaneous Localization and MappingOptical FlowImageVideo

🎯 What it does: This paper proposes a real-time, large-scale novel view synthesis method under unposed image sequences, which can obtain camera poses and complete 3D Gaussian radiance field representations immediately after capture.

One Model to Rig Them All: Diverse Skeleton Rigging with UniRig

Jia-Peng Zhang (Tsinghua University), Shi-Min Hu (Tsinghua University)

GenerationPose EstimationTransformerDiffusion modelAuto EncoderContrastive LearningPoint CloudMesh

🎯 What it does: Proposed UniRig, a unified framework that automatically predicts skeletal constraints and weights for multi-class 3D models by leveraging autoregressive models and bone point cross-attention.

Optimal r-Adaptive In-Timestep Remeshing for Elastodynamics

Jiahao Wen, Danny M. Kaufman

OptimizationMeshPhysics Related

🎯 What it does: Proposed a coupled grid adaptive model with physical simulation algorithm, which can simultaneously generate optimal adaptive remeshing and implicit solution results within each time step, used for frictional contact elastic dynamics simulation.

Order Matters: Learning Element Ordering for Graphic Design Generation

Bo Yang, Ying Cao

GenerationTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkImage

🎯 What it does: This paper proposes a Generative Order Learner (GOL) model, which jointly trains an autoregressive generator and a sorting network during the generation of graphic designs, learning content-adaptive 'neural order' to improve generation quality and convergence speed.

Painless Differentiable Rotation Dynamics

Magí Romanyà-Serrasolsas, M. Otaduy

OptimizationComputational EfficiencyRobotic IntelligenceStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: Proposes a forward differentiable rigid body dynamics formulation utilizing the Lie algebra rotation derivative, and applies it to the incremental potential energy form, as well as introduces a new differentiable dynamics adjoint definition.

Painting with 3D Gaussian Splat Brushes

Karran Pandey, Maria Shugrina

GenerationDiffusion modelGaussian SplattingPoint Cloud

🎯 What it does: Explores the use of 3D Gaussian splatter brushes for interactive painting on 3D Gaussian splatter scenes and other surfaces, proposing algorithms such as selecting a subset of Gaussians, stamp-based painting, and using diffusion models for seamless stamp seam handling, and providing a diverse set of artistic brush parameter combinations.

PaRas: A Rasterizer for Large-Scale Parametric Surfaces

Kechun Wang, Renjie Chen

OptimizationComputational EfficiencyMesh

🎯 What it does: Proposed and implemented PaRas, a high-performance real-time rasterizer for large-scale parametric surfaces

PARC: Physics-based Augmentation with Reinforcement Learning for Character Controllers

Michael Xu (Simon Fraser University), Xue Bin Peng (Simon Fraser University)

GenerationData 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.

PartEdit: Fine-Grained Image Editing using Pre-Trained Diffusion Models

Aleksandar Cvejic (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)

Image TranslationImage HarmonizationRestorationGenerationTransformerPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningImageTextBenchmark

🎯 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.

PDT: Point Distribution Transformation with Diffusion Models

Jionghao Wang (Texas A&M University), Xin Li (Texas A&M University)

GenerationData SynthesisPose EstimationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelContrastive LearningPoint CloudMeshGraph

🎯 What it does: Proposes the PDT framework, which utilizes diffusion models to transform arbitrary point cloud distributions into semantically meaningful point distributions (such as mesh key points, skeletal joints, and garment feature lines).

Photoreal Scene Reconstruction from an Egocentric Device

Zhaoyang Lv (Meta), Richard A. Newcombe

GenerationDepth EstimationAutonomous DrivingOptimizationConvolutional Neural NetworkGraph Neural NetworkDiffusion modelNeural Radiance FieldGaussian SplattingSimultaneous Localization and MappingOptical FlowImageVideoPoint Cloud

🎯 What it does: Developed a video input system tailored for first-person perspective cameras, combining visual-inertial bundle adjustment with high-frequency rolling shutter modeling, and utilizing a high dynamic range physical imaging model to achieve lighting-realistic 3D Gaussian Splatting reconstruction.

Physically Controllable Relighting of Photographs

Chris Careaga (Simon Fraser University), Yağız Aksoy (Simon Fraser University)

Image TranslationRestorationGenerationData SynthesisDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningImageMeshPhysics Related

🎯 What it does: A self-supervised physically controllable photo relighting method is proposed, which converts a single image into a 3D mesh that can be customized with light sources in a PBR engine, and outputs realistic results using a neural renderer.

Physics-inspired Estimation of Optimal Cloth Mesh Resolution

Diyang Zhang, Huamin Wang

OptimizationMeshPhysics Related

🎯 What it does: Propose a physics-inspired method based on material stiffness and boundary conditions to estimate the optimal cloth mesh resolution and generate the corresponding triangular mesh.

PhysicsFC: Learning User-Controlled Skills for a Physics-Based Football Player Controller

Minsu Kim (Hanyang University), Yoonsang Lee (Hanyang University)

Robotic IntelligenceReinforcement LearningVideoPhysics Related

🎯 What it does: This paper proposes the PhysicsFC method, which enables a football agent to perform and seamlessly switch between multiple skills such as hitting, controlling, stealing, and kicking the ball under user input through physics simulation.

Piecewise Ruled Approximation for Freeform Mesh Surfaces

Yiling Pan (Tsinghua University), Bailin Deng (Cardiff University)

OptimizationMesh

🎯 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.

PLT: Part-Wise Latent Tokens as Adaptable Motion Priors for Physically Simulated Characters

Jinseok Bae, Young Min Kim

GenerationTransformerReinforcement LearningDiffusion modelAuto EncoderGenerative Adversarial NetworkVideoSequential

🎯 What it does: Propose a strategy architecture for learning distributed motor skills, utilizing component-specific codebooks and refinement networks, enabling natural full-body motion in physically simulated characters and achieving adaptation to unseen scenarios through the combination of components.

Pocket Time-Lapse

Eric Ming Chen, Abe Davis

RestorationGenerationCompressionOptimizationComputational EfficiencyDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGaussian SplattingOptical FlowImageTime Series

🎯 What it does: Capture and process a large number of unstructured panoramic images quickly and conveniently using a smartphone, recording, exploring, and visualizing long-term environmental changes.

Polynomial 2D Biharmonic Coordinates for High-order Cages

Shibo Liu (University of Science and Technology of China), Xiao-Ming Fu (University of Science and Technology of China)

OptimizationDiffusion modelScore-based ModelOptical FlowMeshBenchmarkStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: Proposed biharmonic coordinates for high-order 2D cages (polynomial curves) to achieve smooth deformation between cages of arbitrary polynomial orders.

pOps: Photo-Inspired Diffusion Operators

Elad Richardson, Daniel Cohen-Or

GenerationTransformerVision Language ModelDiffusion modelContrastive LearningImageText

🎯 What it does: Train specific semantic operators (pOps), embedding them into the CLIP image embedding space to form image-embedding-based diffusion operators.

Position-Normal Manifold for Efficient Glint Rendering on High-Resolution Normal Maps

Liwen Wu (University of California San Diego), R. Ramamoorthi

Computational EfficiencyGaussian SplattingImageMesh

🎯 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.

Power-Linear Polar Directional Fields

Jiabao Wang, Amir Vaxman (University of Edinburgh)

OptimizationMesh

🎯 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;

Practical Inverse Rendering of Textured and Translucent Appearance

Philippe Weier, Delio Vicini

RestorationDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkImage

🎯 What it does: Propose an inverse rendering framework for reconstructing high-resolution surface textures and spatially varying subsurface scattering parameters from images.

Practical Stylized Nonlinear Monte Carlo Rendering

Xiaochun Tong, T. Hachisuka

GenerationData SynthesisComputational EfficiencyDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningImage

🎯 What it does: Proposes two technologies: Nonlinear Path Filtering (NL-PF) and Nonlinear Radiance Caching (NL-NRC), to achieve stylized rendering equation (SRE) under controllable costs, significantly reducing the computational and storage complexity of traditional branch estimators.

Predicting Fabric Appearance Through Thread Scattering and Inversion

Mengqi Xia, Julie Dorsey

GenerationData SynthesisDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkImageMeshPhysics Related

🎯 What it does: Proposed a complete workflow from capturing physical yarns to digitizing them and predicting fabric appearance based on weave patterns

PrimitiveAnything: Human-Crafted 3D Primitive Assembly Generation with Auto-Regressive transformer

Jingwen Ye (Tencent AIPD), Xiao Han (Tencent AIPD)

GenerationData SynthesisTransformerDiffusion modelAuto EncoderContrastive LearningPoint CloudMeshSequential

🎯 What it does: Propose the PrimitiveAnything framework, which transforms the abstraction process of 3D shapes into a task of generating variable-length sequences of primitive geometries, and utilizes autoregressive Transformers to achieve high-quality primitive assembly.

Progressive Dynamics++: A Framework for Stable, Continuous, and Consistent Animation Across Resolution and Time

J. Zhang, Danny M. Kaufman

GenerationDiffusion modelOptical FlowVideoStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: A framework named Progressive Dynamics++ was established to achieve stable, continuous, and consistent animations across resolutions and time.

Putting Rigid Bodies to Rest

H. Baktash, Keenan Crane

Physics Related

🎯 What it does: Studies the analysis and design of static configurations of rigid bodies without physical simulation, identifying all possible static points and their dwell probabilities, and providing forward automatic orientation and backward differentiable inverse design methods.

Quadric-Based Silhouette Sampling for Differentiable Rendering

M. Soroka, Steve Marschner

OptimizationComputational EfficiencyDiffusion modelNeural Radiance FieldMesh

🎯 What it does: Proposed an improved rejection test for importance sampling of contour edges in differentiable rendering to reduce variance and improve sampling efficiency.

Quadtree Tall Cells for Eulerian Liquid Simulation

Fumiya Narita, R. Ando

Optical FlowPhysics RelatedOrdinary Differential Equation

🎯 What it does: Proposed a high-resolution cell structure based on a quadtree for efficient Eulerian liquid simulation;

QUASAR: Quad-based Adaptive Streaming And Rendering

E. Lu, Anthony Rowe

CompressionComputational EfficiencyGaussian SplattingOptical FlowMesh

🎯 What it does: Proposed a quadrilateral-based geometric streaming method to achieve a lightweight client-server architecture for AR/VR devices, supporting real-time rendering of new perspectives over variable bandwidth networks.

Radiance Surfaces: Optimizing Surface Representations with a 5D Radiance Field Loss

Ziyi Zhang (École Polytechnique Fédérale de Lausanne), Wenzel Jakob (École Polytechnique Fédérale de Lausanne)

OptimizationComputational EfficiencyRepresentation LearningDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingImagePoint CloudMesh

🎯 What it does: Propose a method that optimizes surface representation through a 5D radiance field loss, replacing the traditional NeRF volumetric rendering with direct supervision of the light field in the scene, thereby achieving an explicit surface model while maintaining fast convergence.

Rags2Riches: Computational Garment Reuse

Anran Qi, Adrien Bousseau

Optimization

🎯 What it does: Proposes the first algorithm for automatically calculating sewing patterns to realize the reuse of existing clothing into new designs.

Real-Time Knit Deformation and Rendering

Tao Huang, Kui Wu

OptimizationComputational EfficiencyDiffusion modelNeural Radiance FieldGaussian SplattingOptical FlowMesh

🎯 What it does: Propose a real-time framework that, after inputting animated mesh, achieves physical simulation at the yarn level and rendering at the fiber level through a node-based representation.

Rectangular Surface Parameterization

E. Corman, Keenan Crane

OptimizationMesh

🎯 What it does: A method is proposed for computing surface parameterization, which maps an infinitesimal axis-aligned square in the plane to an infinitesimal rectangle on the surface, achieving lower distortion and better preservation of target directions by constructing a perfectly orthogonal and exactly integrable frame field; the method supports user-defined distortion metrics, sharp feature alignment, predefined or automatic singularities, and direct control over boundary behavior; high-quality anisotropic quadrilateral meshes are obtained by quantizing and contouring this mapping.

Reenact Anything: Semantic Video Motion Transfer Using Motion-Textual Inversion

Manuel Kansy (ETH Zürich), Romann M. Weber (DisneyResearch|Studios)

Image TranslationGenerationData SynthesisPose EstimationTransformerPrompt EngineeringDiffusion modelContrastive LearningOptical FlowImageVideoText

🎯 What it does: This work proposes a general semantic video motion transfer method, which achieves motion semantic transfer from a reference video to any target image by performing motion-text reverse optimization on the reference video, obtaining a motion-text embedding that preserves the appearance of the target image.

RELATE3D: REfocusing Latent Adapter for Targeted local Enhancement and Editing in 3D Generation

Xiao-Lei Li, Ran Zhang

GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelAuto EncoderPoint CloudMesh

🎯 What it does: Proposed the RELATE3D method to achieve local enhancement and editing of 3D models

Relightable Full-Body Gaussian Codec Avatars

Shaofei Wang (ETH Zürich), Shunsuke Saito (Meta)

GenerationCompressionDiffusion modelNeural Radiance FieldGaussian SplattingImageVideoPoint CloudMesh

🎯 What it does: Propose a method that utilizes 3D Gaussian Splatting to construct a full-body relightable avatar, capable of reconstructing, animating, and rendering the entire body, face, and hands under arbitrary lighting conditions;

RenderFormer: Transformer-based Neural Rendering of Triangle Meshes with Global Illumination

Chong Zeng (Zhejiang University), Xin Tong (Microsoft Research Asia)

GenerationComputational EfficiencyTransformerDiffusion modelNeural Radiance FieldMesh

🎯 What it does: This paper proposes RenderFormer, a neural rendering pipeline based on Transformer, which can directly render globally illuminated images from scenes described by triangular meshes, without requiring training or fine-tuning for each scene.

Reservoir Splatting for Temporal Path Resampling and Motion Blur

Jeffrey Liu, Ravi Ramamoorthi

Gaussian SplattingOptical FlowVideo

🎯 What it does: Propose and implement the memory splatting method for temporal path resampling, reducing the failure of historical paths to contribute to future frames, improving resampling quality, and supporting motion blur and depth of field.

Revisiting Tradition and Beyond: A Customized Bilateral Filtering Framework for Point Cloud Denoising

Peng Li, Mingqiang Wei

RestorationGraph Neural NetworkPoint Cloud

🎯 What it does: For point cloud denoising, the CustomBF framework is proposed, which customizes the classical bilateral filter at the point level.

RigAnything: Template-Free Autoregressive Rigging for Diverse 3D Assets

Isabella Liu (University of California San Diego), Zifan Shi (Adobe Research)

GenerationPose EstimationTransformerDiffusion modelPoint CloudMesh

🎯 What it does: Proposed a template-free automatic rigging method based on autoregressive transformers called RigAnything, which can generate skeletons and skinning weights for diverse 3D assets in arbitrary poses.

ScaffoldAvatar: High-Fidelity Gaussian Avatars with Patch Expressions

Shivangi Aneja (Technical University of Munich), Derek Bradley (DisneyResearch|Studios)

GenerationData SynthesisDiffusion modelNeural Radiance FieldGaussian SplattingImagePoint CloudMesh

🎯 What it does: Propose ScaffoldAvatar, a high-fidelity 3D Gaussian avatar generation method based on local patch expressions, enabling real-time rendering and detailed representation

Scene-Level Appearance Transfer with Semantic Correspondences

Liyuan Zhu, Iro Armeni

Image TranslationRestorationGenerationDepth EstimationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkImage

🎯 What it does: Proposed a scene-level appearance transfer framework called ReStyle3D that achieves multi-view realistic scene appearance transfer from a single style image

Segment-based Light Transport Simulation

Wenyou Wang, T. Hachisuka

OptimizationComputational EfficiencyOptical FlowPhysics Related

🎯 What it does: Proposed a segment-based light transport framework along with related segment sampling techniques and recursive estimators

Semantically Consistent Text-to-Motion with Unsupervised Styles

Linjun Wu, Xiaogang Jin

GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelVideoText

🎯 What it does: Propose a method to integrate unsupervised style into text-driven diffusion models for generating semantically consistent and stylized human motion.

Shape Space Spectra

Yue Chang (University of Toronto), E. Grinspun

OptimizationComputational EfficiencyRepresentation LearningDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningPoint CloudMeshGraph

🎯 What it does: A method for eigenvalue decomposition in a continuous parameterized shape space is studied, constructing differentiable neural field representations that enable solving the eigenmodes of PDE operators for arbitrary shapes without mesh reconstruction and allow optimization over shape parameters.

Single Edge Collapse Quad-Dominant Mesh Reduction

Julian Knodt (LightSpeed Studios)

Mesh

🎯 What it does: A method for quadrilateral-dominant mesh simplification based on unilateral folding is proposed, using edge-weighted four-dimensional quaternions and local sorting to preserve the input mesh's quadrilateral topology and geometric quality.

Single View Garment Reconstruction Using Diffusion Mapping Via Pattern Coordinates

Ren Li (École Polytechnique Fédérale de Lausanne), Pascal Fua (École Polytechnique Fédérale de Lausanne)

RestorationGenerationConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImagePoint CloudMesh

🎯 What it does: Proposes a method for reconstructing high-precision 3D clothing models from a single clothing photograph.

Sketch2Anim: Towards Transferring Sketch Storyboards into 3D Animation

Lei Zhong (University of Edinburgh), Changjian Li (University of Edinburgh)

Image TranslationGenerationPose EstimationConvolutional Neural NetworkTransformerPrompt EngineeringDiffusion modelScore-based ModelContrastive LearningImageVideoTextMultimodality

🎯 What it does: Propose Sketch2Anim, which can directly convert 2D storyboard sketches into high-quality 3D animations

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)

Image TranslationRestorationGenerationTransformerDiffusion modelAuto EncoderGaussian SplattingOptical FlowImageVideoPoint Cloud

🎯 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;

SkillMimic-V2: Learning Robust and Generalizable Interaction Skills from Sparse and Noisy Demonstrations

Runyi Yu (Hong Kong University Of Science And Technology), Qifeng Chen (Hong Kong University Of Science And Technology)

Robotic IntelligenceTransformerReinforcement LearningPrompt EngineeringAuto EncoderContrastive LearningGraph

🎯 What it does: This paper proposes the SkillMimic-V2 framework, which can learn robust and generalizable interactive skills from sparse and noisy demonstrations.

SOAP: Style-Omniscient Animatable Portraits

Tingting Liao (Mohamed bin Zayed University of Artificial Intelligence), Hao Li (Mohamed bin Zayed University of Artificial Intelligence)

GenerationData SynthesisConvolutional Neural NetworkTransformerPrompt EngineeringDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkImagePoint CloudMesh

🎯 What it does: Generate animatable, style-agnostic 3D head models from a single portrait image.

Sobol' Sequences with Guaranteed-Quality 2D Projections

Nicolas Bonneel, V. Ostromoukhov

🎯 What it does: Proposed and constructed multiple sets of 2D Sobol' sequences, ensuring t=1, and capable of combining to generate high-dimensional low-distortion sequences, while providing directly usable initialization tables and generating matrices for constructing 692-dimensional Sobol' sequences; meanwhile, ensuring that all adjacent dimension pairs possess (1,2)-sequence properties, and guaranteeing that 4D projections have t≤4 within the first 215 points.

Solving partial differential equations in participating media

Bailey Miller, Ioannis Gkioulekas

Point CloudTabularBiomedical DataPhysics RelatedStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: Proposed a new Monte Carlo method for solving partial differential equations (PDEs) with complex micro-particle geometries, called volume-based 'walk on spheres' (VWoS) and 'walk on stars' (VWoSt);

Spatiotemporally Consistent Indoor Lighting Estimation with Diffusion Priors

Mutian Tong (Columbia University), Changxi Zheng (Columbia University)

RestorationGenerationOptimizationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningImageVideo

🎯 What it does: Propose a light field inference method that utilizes 2D diffusion models and MLPs to estimate spatial-temporally consistent high dynamic range illumination from indoor videos.

Sphere Carving: Bounding Volumes for Signed Distance Fields

Hugo Schott, A. Paris

Diffusion modelScore-based ModelMesh

🎯 What it does: An automated method named Sphere Carving is proposed to compute bounding volumes that are tightly tangent to procedural implicit surfaces; this method starts with an initial bounding volume from a distance and iteratively approaches the surface using a signed distance function, constructing the final bounding volume by extracting intersections based on empty sphere queries.

Spherical Lighting with Spherical Harmonics Hessian

Kei Iwasaki, Yoshinori Dobashi

OptimizationComputational EfficiencyNeural Radiance FieldMeshPhysics Related

🎯 What it does: Proposed and applied the second derivative of spherical harmonics (spherical harmonics Hessian) and solid spherical harmonics, developed an analytical representation of the Hessian matrix of spherical harmonic coefficients, and applied it to mesh-based SH lighting rendering for multi-sphere light sources, combining Hessian-based error metrics to achieve adaptive mesh refinement.

Splat and Replace: 3D Reconstruction with Repetitive Elements

Nicolás Violante (Inria & Université Côte d'Azur), G. Drettakis

RestorationGenerationPose EstimationDepth EstimationSuper ResolutionTransformerDiffusion modelNeural Radiance FieldContrastive LearningGaussian SplattingOptical FlowImagePoint CloudMesh

🎯 What it does: Leverages redundant elements in 3D Gaussian Splatting results to enhance the quality of 3D reconstruction and novel view synthesis in uncovered regions through instance segmentation, registration, and shared representation.

Splat4D: Diffusion-Enhanced 4D Gaussian Splatting for Temporally and Spatially Consistent Content Creation

Minghao Yin (University of Hong Kong), Kai Han (University of Hong Kong)

GenerationData SynthesisConvolutional Neural NetworkRecurrent Neural NetworkTransformerPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelContrastive LearningGaussian SplattingImageVideoTextMultimodality

🎯 What it does: Propose the Splat4D framework, which utilizes multi-view video diffusion models, image enhancers, hetero-directional U-Net, and uncertainty masks to generate high-quality, temporally and spatially consistent 4D Gaussian splat representations from monocular videos, images, or text, and supports applications such as text/image conditional generation, human animation, and text editing.

Spline Deformation Field

Mingyang Song (Disney Research Studios), T. Aydın

RestorationGenerationOptimizationComputational EfficiencyTransformerDiffusion modelNeural Radiance FieldContrastive LearningGaussian SplattingOptical FlowImageVideoPoint CloudTime Series

🎯 What it does: Propose a trajectory representation based on cubic Hermite splines, and combine it with coordinate neural networks to achieve continuous deformation field inference in dynamic scenes.

SqueezeMe: Mobile-Ready Distillation of Gaussian Full-Body Avatars

Forrest Iandola (Meta), Shunsuke Saito (Meta)

Computational EfficiencyKnowledge DistillationRepresentation LearningAuto EncoderGaussian SplattingOptical FlowVideoPoint CloudMesh

🎯 What it does: Developed the SqueezeMe framework, which distills high-precision 3D Gaussian full-body avatars into a lightweight representation, enabling real-time animation and rendering of multiple avatars on mobile devices.

Stable Cosserat Rods

J. Hsu, C. Yuksel

OptimizationComputational EfficiencyPhysics Related

🎯 What it does: A stable Cosserat rod solver is proposed, improving numerical stability and computational efficiency, supporting large time steps and parallel computing, and verifying its performance in various applications.

StableMakeup: When Real-World Makeup Transfer Meets Diffusion Model

Yuxuan Zhang (Shanghai Jiao Tong University), Haibo Zhao

Image 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.

Stitch-A-Shape: Bottom-up Learning for B-Rep Generation

Pu Li, Dong-ming Yan

GenerationData SynthesisTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkMesh

🎯 What it does: Propose the Stitch-A-Shape framework, which directly and simultaneously models the topology and geometry of B-Rep.

Stochastic Barnes-Hut Approximation for Fast Summation on the GPU

A. Madan, David I.W. Levin

OptimizationComputational EfficiencyPoint CloudMesh

🎯 What it does: Proposed a Barnes-Hut approximation based on random paths and control variables for fast unbiased summation on the GPU.

Stochastic Preconditioning for Neural Field Optimization

Selena Ling (University of Toronto), Nicholas J. Sharp (NVIDIA)

OptimizationNeural Radiance FieldGaussian SplattingImagePoint CloudMesh

🎯 What it does: This paper proposes a 'random preprocessing' technique that injects random noise into query coordinates during the training of neural fields (such as SDF, NeRF, etc.), improving the optimization process by equivalently performing Gaussian blurring on the field in expectation, reducing the risk of falling into local optima, and enhancing reconstruction quality;

Streaming-Aware Neural Monte Carlo Rendering Framework with Unified Denoising-Compression and Client Collaboration

Hangming Fan, Rui Wang

CompressionDiffusion modelNeural Radiance FieldAuto EncoderVideo

🎯 What it does: Propose a flow-aware rendering framework that jointly learns adaptive sampling, denoising, and video compression techniques, and utilizes client-side rendering G-buffers to assist frame decoding.

StreamME: Simplify 3D Gaussian Avatar within Live Stream

Luchuan Song, Chenliang Xu

GenerationData SynthesisGaussian SplattingVideoPoint Cloud

🎯 What it does: Propose the StreamME method, which utilizes real-time video streams to quickly synchronize and record, and reconstruct 3D avatars of the head, supporting instant training and application without pre-cached data;

Stressful Tree Modeling: Breaking Branches with Strands

Bosheng Li, B. Benes

Diffusion modelScore-based ModelPhysics RelatedStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: We propose a computational modeling method for woody tissues such as branches and wood, utilizing advanced strand-based representations and extending their description of short-term and long-term biomechanical processes to achieve fast interactive simulations of wood fracture.

Stroke Transfer for Participating Media

Naoto Shirashima, Yonghao Yue

GenerationData SynthesisDiffusion modelAuto EncoderGenerative Adversarial NetworkImage

🎯 What it does: Generate brushstroke paintings of participating media such as smoke, flame, and clouds by transferring brush attributes (color, width, length, and direction) from examples;

Style Customization of Text-to-Vector Generation with Image Diffusion Priors

Peiying Zhang, Jing Liao

GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelAuto EncoderContrastive LearningImageTextMesh

🎯 What it does: The study is based on text-to-vector graphic (SVG) generation using diffusion models, and proposes a two-stage style customization process that utilizes path-level representations to achieve high-quality, structurally consistent SVGs.

SwiftSketch: A Diffusion Model for Image-to-Vector Sketch Generation

E. Arar, Yael Vinker (MIT)

GenerationData SynthesisTransformerVision Language ModelDiffusion modelScore-based ModelContrastive LearningImageMultimodality

🎯 What it does: Propose SwiftSketch, a real-time image-to-vector sketch generation method based on diffusion models.

TeGA: Texture Space Gaussian Avatars for High-Resolution Dynamic Head Modeling

Gengyan Li (ETH Zurich), T. Beeler

GenerationSuper ResolutionConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingImagePoint CloudMesh

🎯 What it does: Propose a 3D high-resolution head avatar generation method called TeGA based on texture space, which can achieve 4K-level realistic rendering under controllable expressions and viewpoints.

TetWeave: Isosurface Extraction using On-The-Fly Delaunay Tetrahedral Grids for Gradient-Based Mesh Optimization

Alexandre Binninger (ETH Zurich), O. Sorkine-Hornung

GenerationOptimizationComputational EfficiencyDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingImagePoint CloudMesh

🎯 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.

Text-based Animatable 3D Avatars with Morphable Model Alignment

Yiqian Wu, Siyu Tang

GenerationData SynthesisTransformerDiffusion modelScore-based ModelGaussian SplattingImageVideoTextPoint CloudMeshStochastic Differential Equation

🎯 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.

The Mokume Dataset and Inverse Modeling of Solid Wood Textures

M. Larsson, Takeo Igarashi

GenerationData SynthesisPrompt EngineeringDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningImagePoint CloudMeshComputed Tomography

🎯 What it does: Proposed the Mokume dataset and designed a three-stage inverse modeling pipeline to infer solid wood grain patterns from external photographs.

Tiny is not small enough: High quality, low-resource facial animation models through hybrid knowledge distillation

Zhen Han (Electronic Arts), Judith Bütepage (Electronic Arts)

Computational EfficiencyKnowledge DistillationConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningVideoMultimodalityAudio

🎯 What it does: This paper proposes a hybrid knowledge distillation (Hybrid Knowledge Distillation) framework, which trains a small, low-latency audio-driven 3D facial animation model using pseudo-labels, achieving high-quality real-time animation.

TokenVerse: Versatile Multi-concept Personalization in Token Modulation Space

Daniel Garibi (Google DeepMind), Tali Dekel (Google DeepMind)

GenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelImageMultimodality

🎯 What it does: Propose the TokenVerse method to achieve personalized learning for multi-graph multi-concept and free combination without masks;

Topological Offsets

Daniel Zint (New York University), D. Panozzo

Diffusion modelMesh

🎯 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 Comprehensive Neural Materials: Dynamic Structure-Preserving Synthesis with Accurate Silhouette at Instant Inference Speed

Zilin Xu, Ling-Qi Yan

GenerationData SynthesisConvolutional Neural NetworkDiffusion modelAuto EncoderGenerative Adversarial NetworkOptical FlowImage

🎯 What it does: Propose an int8 quantized neural network that achieves real-time inference speed while maintaining high fidelity, and combines controllable structure-preserving synthesis strategies with dynamic two-step displacement tracking techniques to accurately handle displacement and silhouette effects.

Towards Understanding Depth Perception in Foveated Rendering

Sophie Kergaßner (Università della Svizzera italiana), P. Didyk

Depth 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.

Transformer IMU Calibrator: Dynamic On-body IMU Calibration for Inertial Motion Capture

Chengxu Zuo, Yipeng Qin

Data SynthesisPose EstimationTransformerTime SeriesSequential

🎯 What it does: Proposed a dynamic IMU calibration framework based on Transformer, achieving implicit real-time calibration for sparse inertial motion capture;

Transforming Unstructured Hair Strands into Procedural Hair Grooms

Wesley Chang, Olivier Maury

GenerationOptimizationDiffusion modelImageMesh

🎯 What it does: Propose an inverse hair styling pipeline that converts unstructured 3D hair strands recovered from images into procedural hairstyles composed of a small number of guiding hair strands and styling operations.

TransparentGS: Fast Inverse Rendering of Transparent Objects with Gaussians

Letian Huang (Nanjing University), Jie Guo (Nanjing University)

GenerationOptimizationComputational EfficiencyDiffusion modelNeural Radiance FieldGaussian SplattingImagePoint CloudMesh

🎯 What it does: Propose TransparentGS, a fast inverse rendering pipeline based on 3D Gaussian Splatting, which can reconstruct transparent objects and achieve real-time novel view synthesis within one hour;

UltraMeshRenderer: Efficient Structure and Management of GPU Out-of-core Memory for Real-time Rendering of Gigantic 3D Meshes

Huadong Zhang, Chao Peng

Computational EfficiencyGaussian SplattingOptical FlowMesh

🎯 What it does: Propose a new GPU out-of-core rendering method that combines LOD selection considering memory and inter-frame consistency with a parallel in-place GPU memory management algorithm.

Unbiased Differential Visibility Using Fixed-Step Walk-on-Spherical-Caps And Closest Silhouettes

Lifan Wu, Aaron E. Lefohn

Optical FlowPhysics RelatedStochastic Differential Equation

🎯 What it does: Propose a differential visibility estimation method based on fixed-step walk-on-spheres, utilizing spherical cap random walks and cone queries to find the nearest contour points on the boundary.

Uncertainty for SVBRDF Acquisition using Frequency Analysis

R. Wiersma (ETH Zurich), V. Deschaintre (Adobe Research)

OptimizationComputational EfficiencyDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingImageMeshPhysics Related

🎯 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.

Unsupervised Decomposition of 3D Shapes into Expressive and Editable Extruded Profile Primitives

Chunyi Sun, Stephen Gould

GenerationRepresentation LearningTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningMesh

🎯 What it does: Propose a representation method that decomposes 3D shapes into differentiable, parameterized primitive profile extrusions.

Variable Shared Template for Consistent Non-rigid ICP

Yucheol Jung, Seungyong Lee

Pose EstimationOptimizationOptical FlowPoint Cloud

🎯 What it does: Proposes a non-rigid ICP framework that jointly optimizes the shared template and instance deformation

Variational Elastodynamic Simulation

Leticia Mattos Da Silva, Justin Solomon

OptimizationPhysics Related

🎯 What it does: This paper proposes an optimization framework that leverages hidden convex substructures, transforming variational elastic dynamics integration into a solvable problem combining ADMM with proximal operations.

Variational Green and Biharmonic Coordinates for 2D Polynomial Cages

Élie Michel, Jean-Marc Thiery

🎯 What it does: Derive closed-form expressions for Green and biharmonic coordinates under polynomial curve 2D cages, and provide their first and second-order derivatives, supporting deformation using variational solvers at any point.

Variational Surface Reconstruction Using Natural Neighbors

Jianjun Xia, Tao Ju

RestorationComputational EfficiencyPoint Cloud

🎯 What it does: Propose a local reconstruction method based on natural neighbors, which can reconstruct implicit surfaces from sparse point clouds that cannot be vectorized

Vector-Valued Monte Carlo Integration Using Ratio Control Variates

Haolin Lu, Tzu-Mao Li

Optimization

🎯 What it does: This paper proposes a ratio control variate estimator for vector-valued Monte Carlo integration;

VideoAnydoor: High-fidelity Video Object Insertion with Precise Motion Control

Yuanpeng Tu (University of Hong Kong), Hengshuang Zhao (HUST)

GenerationData SynthesisConvolutional Neural NetworkTransformerPrompt EngineeringDiffusion modelContrastive LearningOptical FlowImageVideo

🎯 What it does: Proposed an end-to-end zero-shot video object insertion framework called VideoAnydoor, which can achieve precise motion control through user-specified boxes or trajectories while maintaining high-fidelity details.

VideoPainter: Any-length Video Inpainting and Editing with Plug-and-Play Context Control

Yuxuan Bian (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)

RestorationGenerationData SynthesisTransformerPrompt EngineeringDiffusion modelAuto EncoderImageVideoText

🎯 What it does: Propose VideoPainter, a dual-branch video inpainting framework that supports filling and editing of videos of arbitrary length, and can be controlled by text prompts for background and foreground in a plug-and-play manner.

VirCHEW Reality: On-Face Kinesthetic Feedback for Enhancing Food-Intake Experience in Virtual Reality

Qingqin Liu, Kening Zhu

Robotic IntelligenceReinforcement Learning from Human Feedback

🎯 What it does: Designed a facial wearable pneumatic tactile device called VirCHEW Reality, and validated its effectiveness in enhancing VR food chewing experience through three user studies.

Virtualized 3D Gaussians: Flexible Cluster-based Level-of-Detail System for Real-Time Rendering of Composed Scenes

Xijie Yang (Zhejiang University), Bo Dai (University of Hong Kong)

Computational EfficiencyGaussian SplattingPoint Cloud

🎯 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.