arXivSub Start free trial

SIGGRAPH Asia 2025 Papers — Page 3

ACM SIGGRAPH Asia (Transactions on Graphics) · 301 papers

PanoDreamer: Optimization-Based Single Image to 360 3D Scene With Diffusion

Avinash Paliwal (Texas A&M University), N. Kalantari

GenerationDepth EstimationOptimizationTransformerDiffusion modelScore-based ModelGaussian SplattingOptical FlowImagePoint Cloud

🎯 What it does: Based on a single input image, generate a panorama and the corresponding depth map through an optimized diffusion model, and then construct a 360° 3D scene.

PartComposer: Learning and Composing Part-Level Concepts from Single-Image Examples

Junyu Liu (Brown University), Daniel Ritchie (Brown University)

GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelContrastive LearningImageMultimodality

🎯 What it does: Learn and combine fine-grained part concepts from a single image to achieve cross-category controllable generation

PartUV: Part-Based UV Unwrapping of 3D Meshes

Zhaoning Wang (Hillbot Inc.), Minghua Liu (Hillbot Inc.)

Mesh

🎯 What it does: Propose a 3D mesh UV unfolding method called PartUV based on semantic part priors, which can significantly reduce the number of charts while maintaining low distortion.

Performance Analysis of Catch-Up Eye Movements in Visual Tracking

Jenna Kang, Qi Sun

Object TrackingOptical FlowImageVideo

🎯 What it does: Study and measure the catch-up saccade performance in visual tracking, and propose a behavioral analysis model based on predicting user reaction delay using target visibility.

PhysHMR: Learning Humanoid Control Policies from Vision for Physically Plausible Human Motion Reconstruction

Qiao Feng (University of Pennsylvania), Lingjie Liu (University of Pennsylvania)

Pose EstimationOptimizationKnowledge DistillationRobotic IntelligenceConvolutional Neural NetworkTransformerReinforcement LearningVision-Language-Action ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageVideoPoint CloudMesh

🎯 What it does: Propose PhysHMR, which directly learns a vision-to-action control policy from monocular videos, achieving physically plausible human motion reconstruction.

PhySIC: Physically Plausible 3D Human-Scene Interaction and Contact from a Single Image

Pradyumna Yalandur Muralidhar (University of Tübingen), Gerard Pons-Moll (University of Tübingen)

Pose EstimationDepth EstimationOptimizationConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingImagePoint CloudMesh

🎯 What it does: Proposes the PhySIC method, which can simultaneously reconstruct a metric-scale 3D human mesh, complete scene geometry, and vertex-level contact maps from a single RGB image, achieving physically plausible human-scene interactions.

Physics-Based Motion Imitation with Adversarial Differential Discriminators

Ziyu Zhang (Simon Fraser University), Xue Bin Peng (Simon Fraser University)

OptimizationRobotic IntelligenceReinforcement LearningGenerative Adversarial NetworkVideoSequentialPhysics Related

🎯 What it does: A multi-objective optimization method based on an adversarial differential discriminator is proposed for physics-driven character animation motion imitation, eliminating the dependence on manually designed reward functions.

PhysiOpt: Physics-Driven Shape Optimization for 3D Generative Models

Xiao Zhan, Mina Konaković Luković

OptimizationDiffusion modelAuto EncoderGenerative Adversarial NetworkPoint CloudMeshPhysics RelatedStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: Propose a differentiable physics optimizer, PhysiOpt, which directly optimizes 3D shapes in the latent space of a generative model to improve their physical feasibility.

PoissonNet: A Local-Global Approach for Learning on Surfaces

Arman Maesumi (Brown University), Noam Aigerman (Université de Montréal)

ClassificationSegmentationPose EstimationRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningPoint CloudMesh

🎯 What it does: This paper proposes a novel graph learning framework called PoissonNet, which combines local gradient transformation and global Poisson equation solving to achieve efficient learning of surface features on triangular meshes.

Potentially Visible Set Generation with the Disocclusion Buffer

Sebastian Künzel, Dieter Schmalstieg

Depth EstimationComputational EfficiencyGaussian SplattingOptical Flow

🎯 What it does: Proposed a potential visible set (PVS) generation algorithm based on the occlusion buffer and implemented the algorithm.

PowerGS: Display-Rendering Power Co-Optimization for Neural Rendering in Power-Constrained XR Systems

Weikai Lin (University of Rochester), Yuhao Zhu (University of Rochester)

OptimizationComputational EfficiencyNeural Radiance FieldGaussian SplattingImageVideoPoint Cloud

🎯 What it does: Proposes the PowerGS framework, which jointly optimizes the rendering power consumption and display power consumption of the 3D Gaussian Splatting (3DGS) model under the constraint of visual quality, and supports disparity rendering.

Practical Gaussian Process Implicit Surfaces with Sparse Convolutions

Kehan Xu, Wojciech Jarosz

GenerationComputational EfficiencyConvolutional Neural NetworkDiffusion modelScore-based ModelNeural Radiance FieldGaussian Splatting

🎯 What it does: This paper eliminates the expensive solution of linear systems by rewriting Gaussian Process Implicit Surfaces (GPIS) as procedural noise, achieving efficient randomized surface and volume rendering.

PractiLight: Practical Light Control Using Foundational Diffusion Models

Yotam Erel, C. Theobalt

Image TranslationRestorationGenerationTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelImage

🎯 What it does: This paper proposes PractiLight, a method for controllable relighting that utilizes the base diffusion model (Stable Diffusion 1.5) to manipulate lighting and shadows. It trains a lightweight LoRA regressor on the self-attention layer to predict the direct-irradiance map of an image, and during image generation, it employs Classifier Guidance, ControlNet, and self-attention query injection techniques to freely modify the light source position, intensity, and material on the generated image.

Precise Gradient Discontinuities in Neural Fields for Subspace Physics

Mengfei Liu (University of Toronto), E. Grinspun

Diffusion modelNeural Radiance FieldAuto EncoderContrastive LearningPoint CloudMeshGraphPhysics RelatedStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: This paper proposes a neural field construction through input space lifting and smooth truncated distance functions, which can accurately represent physical fields with discontinuous gradients without encoding the interface geometry into the network weights, thereby enabling subspace simulations that are independent of discretization for parameterized shape families and heterogeneous materials.

Prior-Enhanced Gaussian Splatting for Dynamic Scene Reconstruction from Casual Video

Meng-Li Shih (University of Washington), Brian Curless (University of Washington)

RestorationObject DetectionSegmentationGenerationDepth EstimationConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingOptical FlowVideoPoint Cloud

🎯 What it does: A fully automated pipeline is proposed to achieve dynamic scene reconstruction from casually captured monocular RGB videos. Techniques such as video segmentation, EPI error mask, depth refinement, skeleton sampling, and mask-guided trajectory re-identification are used to enhance the 2D prior, and the geometric and motion consistency of dynamic Gaussian Splatting is improved through virtual viewpoint depth loss and structural projection loss.

PriorAvatar: Efficient and Robust Avatar Creation from Monocular Video Using Learned Priors

Tianjian Jiang, Jie Song

GenerationPose EstimationDepth EstimationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelGaussian SplattingVideo

🎯 What it does: Propose the PriorAvatar method, which utilizes learned 3D shape and appearance priors of multiple people, generating renderable 3D Gaussians through a shared U-Net decoder to high-fidelity reconstruct avatars from monocular videos.

Procedural Scene Programs for Open-Universe Scene Generation: LLM-Free Error Correction via Program Search

Maxim Gumin (Brown University), Daniel Ritchie (Brown University)

GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelContrastive LearningImageTextPoint CloudMeshRetrieval-Augmented Generation

🎯 What it does: This paper proposes a programmatic scene description language (PSDL) based on a Python embedded DSL, which directly lays out 3D scenes using imperative programs generated by LLMs, and corrects errors through a local search-based program error correction mechanism without additionally calling LLMs.

Progressive Outfit Assembly and Instantaneous Pose Transfer

Dewen Guo, Huamin Wang

GenerationPose EstimationDiffusion modelAuto EncoderGenerative Adversarial NetworkMesh

🎯 What it does: Propose a progressive clothing assembly and instant pose transfer framework using mid-surface representation, addressing cross-occlusion conflicts in multi-layered clothing and reconstructing complete collision-free geometry.

Proteus-ID: ID-Consistent and Motion-Coherent Video Customization

Guiyu Zhang (Chinese University of Hong Kong Shenzhen), Li Jiang (Chinese University of Hong Kong Shenzhen)

GenerationData SynthesisTransformerVision Language ModelDiffusion modelOptical FlowImageVideoTextMultimodality

🎯 What it does: Developed a diffusion model called Proteus-ID, which can generate videos that maintain identity consistency and exhibit coherent motion based on a single reference image and text prompts;

QMF-Blend: Quantized Matrix Factorization for Efficient Blendshape Compression

R. Fedotov, L. Kavan

CompressionAuto EncoderMesh

🎯 What it does: An efficient blendshape compression algorithm based on quantized matrix decomposition is proposed, significantly reducing storage requirements and computational complexity.

RaRa Clipper: A Clipper for Gaussian Splatting Based on Ray Tracer and Rasterizer

Da Li (King Abdullah University of Science and Technology), Ivan Viola (King Abdullah University of Science and Technology)

Gaussian SplattingPoint CloudBiomedical Data

🎯 What it does: Propose RaRa Clipper, an efficient and high-fidelity Gaussian Splatting clipping framework that combines rasterization with ray tracing

RCTrans: Transparent Object Reconstruction in Natural Scene via Refractive Correspondence Estimation

Fangzhou Gao, Jiawan Zhang

RestorationTransformerNeural Radiance FieldContrastive LearningImage

🎯 What it does: Proposes the RCTrans method, achieving high-quality reconstruction of transparent objects in natural scenes using ray-background intersection constraints and a pre-trained correspondence estimation network;

Realistic Cloth Rendering with a Ray-Wave Hybrid Shading Model

Yunchen Yu, A. Weidlich

GenerationDiffusion modelNeural Radiance FieldGaussian SplattingOptical FlowImagePhysics Related

🎯 What it does: A fabric coloring model based on the hybrid of ray and wave optics is proposed to simultaneously simulate the reflection and transmission of fabrics.

Reconfigurable Hinged Kirigami Tessellations

Aviv Segall, O. Sorkine-Hornung

OptimizationRobotic IntelligenceDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningMeshPhysics Related

🎯 What it does: Proposes a computational framework for designing rotatable and deployable Kirigami structures that can approximate arbitrary free-form surfaces.

ReSTIR PG: Path Guiding with Spatiotemporally Resampled Paths

Zheng Zeng, Daqi Lin

🎯 What it does: Developed a real-time ReSTIR Path Guiding (ReSTIR-PG) method, which utilizes ReSTIR's resampling path extraction to derive a guiding distribution for generating improved initial candidate paths for the next frame.

RibbonSculpt: Voronoi Ball based 3D Sculpting from Sparse VR Ribbons

Anandhu Sureshkumar, Marie-Paule Cani

GenerationOptimizationDiffusion modelGaussian SplattingPoint CloudMesh

🎯 What it does: Real-time interactive freeform shape design is achieved by progressively drawing sparse oriented bands in VR, generating and refining closed surfaces with arbitrary topological genus; volume proxies are constructed using filtered Voronoi spheres, and smoothing is performed through Laplacian energy minimization after mesh extraction; users can add/remove bands to further sculpt the proxy.

Rigidity-Aware 3D Gaussian Deformation from a Single Image

Jinhyeok Kim (UNIST), Kyungdon Joo (UNIST)

RestorationPose EstimationContrastive LearningGaussian SplattingOptical FlowImagePoint Cloud

🎯 What it does: Studies how to deform the already reconstructed 3D Gaussians using a single target image while maintaining the original geometry undistorted.

RL-ACD: Reinforcement Learning-based Approximate Convex Decomposition

Yuzhe Luo, Xifeng Gao

OptimizationReinforcement LearningMesh

🎯 What it does: Proposes a reinforcement learning-based approximate convex decomposition method, RL-ACD, for efficiently and approximately optimally decomposing complex 3D shapes into convex components

Robust Derivative Estimation with Walk on Stars

Zihan Yu, Shuang Zhao

OptimizationPhysics RelatedStochastic Differential Equation

🎯 What it does: Proposes an extension of the walk on stars method for solving PDEs, aimed at high-precision estimation of spatial derivatives.

Robust Single-shot Structured Light 3D Imaging via Neural Feature Decoding

Jiaheng Li, Wenzheng Chen

Data SynthesisDepth EstimationConvolutional Neural NetworkRecurrent Neural NetworkTransformerPrompt EngineeringNeural Radiance FieldAuto EncoderContrastive LearningOptical FlowImagePoint CloudMeshBenchmark

🎯 What it does: Propose a single-frame structured light 3D reconstruction framework called NSL based on neural feature decoding, which can achieve high-precision depth recovery under a single camera + projection mode;

S3 Imagery: Specular Shading from Scratch-Anisotropy

Pengfei Shen, Yifan Peng

Ordinary Differential Equation

🎯 What it does: This paper proposes achieving continuous view-dependent specular shadow effects of 3D virtual objects through scratch-based reflection technology. It optimizes the distribution of scratches on the substrate surface by solving an ordinary differential equation constrained by the bidirectional reflectance distribution function (BRDF), and finally fabricates realistic reflectors on planar and developable curved surfaces using a commercial engraving machine, demonstrating continuous and realistic specular shadows.

Sample Space Partitioning and Spatiotemporal Resampling for Specular Manifold Sampling

P. Hong, Daqi Lin

OptimizationComputational EfficiencyDiffusion modelScore-based Model

🎯 What it does: Improve Specular Manifold Sampling to achieve high-quality holographic illumination rendering in interactive environments by using sample space partitioning and spatiotemporal reuse techniques.

Scattering-Aware Color Calibration for 3D Printers Using a Simple Calibration Target

Tomáš Iser, A. Wilkie

OptimizationDiffusion modelOptical FlowPoint CloudMeshPhysics Related

🎯 What it does: Proposes a 3D printer full-color optical property calibration method based on multi-scattering light transport, using a single printable calibration target.

Self-supervised Underwater Color Restoration via Wavelet-Diffusion Model with Filtered Multi-Scale Feature Distillation

Xin Zhang, Binghua Su

RestorationTransformerDiffusion modelScore-based ModelContrastive LearningImage

🎯 What it does: Proposes a self-supervised underwater color restoration framework based on waveform diffusion models, utilizing filtered multi-scale feature distillation;

SeqTex: Generate Mesh Textures in Video Sequence

Ze Yuan (HKU), Xiaojuan Qi (HKU)

GenerationData SynthesisTransformerSupervised Fine-TuningVision Language ModelDiffusion modelFlow-based ModelImageVideoTextMesh

🎯 What it does: SeqTex proposes an end-to-end framework that directly generates complete UV texture maps using a pre-trained video diffusion model, supporting image or text conditioning.

Shape-aware Inertial Poser: Motion Tracking for Humans with Diverse Shapes Using Sparse Inertial Sensors

Lu Yin (Xiamen University), Shihui Guo (Xiamen University)

Pose EstimationOptimizationRecurrent Neural NetworkTransformerContrastive LearningOptical FlowMeshTime SeriesSequential

🎯 What it does: Proposes a shape-aware sparse inertial capture system, SAIP, which can achieve real-time motion capture for different body types (including children) using only six IMUs, and first realizes human shape estimation in IMU-based capture.

Shape-for-Motion: Precise and Consistent Video Editing With 3D Proxy

Yuhao Liu (City University of Hong Kong), Rynson W. H. Lau

Image TranslationRestorationGenerationData SynthesisTransformerDiffusion modelAuto EncoderContrastive LearningGaussian SplattingVideoPoint CloudMesh

🎯 What it does: Propose the Shape-for-Motion framework to achieve precise and consistent video editing based on 3D proxies.

ShapeGen: Towards High-Quality 3D Shape Synthesis

Yangguang Li (Chinese University of Hong Kong), Yan-Pei Cao (VAST)

GenerationData SynthesisTransformerDiffusion modelScore-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningImagePoint CloudMesh

🎯 What it does: Achieve single-image to high-quality 3D shape synthesis through improved VAE representation, resolution enhancement, hybrid conditional training, linear attention, and inference time scaling.

Shaping Strands with Neural Style Transfer

Beyzanur Coban, Vinicius C. Azevedo

Image TranslationGenerationConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkImageMesh

🎯 What it does: Proposes the first neural style transfer pipeline that supports hair and fur, achieving complex, 3D-consistent, and temporally coherent styling through customized hair/fur representations.

Shoot-Bounce-3D: Single-Shot Occlusion-Aware 3D from Lidar by Decomposing Two-Bounce Light

Tzofi Klinghoffer (Massachusetts Institute of Technology), Rakesh Ranjan (Meta)

Data SynthesisDepth EstimationAutonomous DrivingDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningImagePoint CloudMeshGraph

🎯 What it does: By utilizing multi-point parallel illumination with single-photon LiDAR, the method learns to unmix double-bounce reflections, recovering dense depth, occlusion geometry, and specular objects from a single frame, achieving single-shot 3D reconstruction.

Simplifying Textured Triangle Meshes in the Wild

H. Liu (Roblox), Cem Yuksel (University of Utah)

Diffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningMesh

🎯 What it does: This paper proposes a method for high-quality simplification on triangular meshes that include non-manifolds, multiple connected components, and textures, while preserving both geometry and texture fidelity;

Single-Image 3D Human Reconstruction with 3D-Aware Diffusion Priors and Facial Enhancement

Jie Yang, Lin Gao

GenerationData SynthesisPose EstimationSuper ResolutionTransformerSupervised Fine-TuningDiffusion modelScore-based ModelGaussian SplattingImageVideo

🎯 What it does: Generate high-quality, photorealistic 3D digital humans from a single image, and enhance facial details to improve facial consistency.

Single-Shot Facial Capture using Polarized RGB Sinusoidal Illumination

Arvin Lin, Abhijeet Ghosh

Image TranslationRestorationPose EstimationDepth EstimationDiffusion modelNeural Radiance FieldAuto EncoderOptical FlowImage

🎯 What it does: Propose a single-shot facial capture method that uses linearly polarized RGB sinusoidal illumination and consumer-grade cameras to achieve per-pixel diffuse albedo, specular albedo, specular roughness, and optical normal estimation.

Sketch2PoseNet: Efficient and Generalized Sketch to 3D Human Pose Prediction

Li Wang (Nanjing University), Hao Zhu (Nanjing University)

Data SynthesisPose EstimationConvolutional Neural NetworkGraph Neural NetworkTransformerVision-Language-Action ModelDiffusion modelAuto EncoderContrastive LearningImagePoint CloudMesh

🎯 What it does: Propose an end-to-end Sketch2PoseNet network that can quickly predict 3D human poses from hand-drawn sketches of different styles.

SMF: Template-free and Rig-free Animation Transfer using Kinetic Codes

Sanjeev Muralikrishnan (University College London), Niloy J. Mitra (University College London)

GenerationData SynthesisPose EstimationTransformerAuto EncoderContrastive LearningImageVideoPoint CloudMeshStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: Proposes a self-supervised animation transfer framework called Self-Supervised Motion Fields (SMF), which can map sparse 3D/2D keypoint motions to arbitrary 3D meshes, enabling full-body animation generation without templates, bindings, or skin weights.

Snapping Deployable Toroids for Modular Gridshells

Felix Dellinger, K. Sharifmoghaddam

Physics Related

🎯 What it does: Proposed a polyhedral toroidal structure (PQ-toroids) that can rapidly switch between two stable states: planar and deployed, and realized its inverse generation and assembly through a design process.

Social Agent: Mastering Dyadic Nonverbal Behavior Generation via Conversational LLM Agents

Zeyi Zhang (Peking University), Libin Liu (Peking University)

GenerationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelDiffusion modelVideoMultimodalityRetrieval-Augmented GenerationAudio

🎯 What it does: Propose the Social Agent framework, a two-person dialogue-based nonverbal behavior generation system guided by large language models (LLMs), achieving context-aware synchronized generation of two-person gestures and eye contact through autoregressive diffusion models.

SOF: Sorted Opacity Fields for Fast Unbounded Surface Reconstruction

L. Radl, Markus Steinberger

Computational EfficiencyRepresentation LearningDiffusion modelNeural Radiance FieldGaussian SplattingImagePoint CloudMesh

🎯 What it does: Propose the Sorted Opacity Fields (SOF) method, enabling the fast extraction of high-quality unbounded surface meshes from 3D Gaussian representations.

Solid-Shell Labeling for Discrete Surfaces

Siqi Wang, H. Liu

OptimizationDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningMesh

🎯 What it does: A method is proposed that assigns solid or shell labels to each face of the input surface mesh. The internal space of solid faces is defined, and the volume of shell faces is determined by thresholding the generalized winding number field. Subsequently, shell faces are processed using an offset mesher, and solid faces are processed using a tetrahedral mesher, and finally, the volume mesh is obtained by merging them.

Sparse Cache Updates for Scalable Distributed Effect-Based Rendering

W. Tatzgern, M. Steinberger

Computational EfficiencyGaussian SplattingOptical Flow

🎯 What it does: Proposed an scalable multi-GPU cloud rendering system, which uses OSC (surface cache) for multi-view scenes and decouples shading from perspective to enable reuse of rendering information for multiple users;

Sparse SVBRDF Acquisition via Importance-Aware Illumination Multiplexing

Lianghao Zhang, Jiawan Zhang

OptimizationComputational EfficiencyRepresentation LearningDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImagePhysics Related

🎯 What it does: Proposes an SVBRDF reconstruction method that combines adaptive sampling with illumination multiplexing, targeting sparse images and using planar light sources for sampling.

Spectral Prefiltering of Neural Fields

Mustafa B. Yaldiz (University of California San Diego), Ravi Ramamoorthi (University of California San Diego)

RestorationSuper ResolutionNeural Radiance FieldAuto EncoderImageMesh

🎯 What it does: Propose a method to train neural fields that can perform linear pre-filtering in a single forward pass and support various symmetric low-pass filter kernels (such as Gaussian, Box, Lanczos)

Spectral Reconstruction with Uncertainty Quantification via Differentiable Rendering and Null-Space Sampling

Mengqi Xia, Holly Rushmeier

RestorationSuper ResolutionDiffusion modelNeural Radiance FieldOptical FlowImageMultimodality

🎯 What it does: Proposes a spectral reconstruction method based on differentiable rendering and null space sampling, which can recover spectral information from multispectral images and provide uncertainty quantification.

Spectral-GS: Taming 3D Gaussian Splatting with Spectral Entropy

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

Gaussian SplattingImagePoint Cloud

🎯 What it does: Propose the Spectral-GS method, which improves the rendering of high-frequency details in 3D Gaussian Splatting through spectral entropy analysis, solving the needle-like artifact problem.

SPGen: Spherical Projection as Consistent and Flexible Representation for Single Image 3D Shape Generation

Jingdong Zhang (Texas A&M University), Xin Li (Texas A&M University)

GenerationData SynthesisTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImagePoint CloudMesh

🎯 What it does: Proposes the SPGen framework, which utilizes multi-layer spherical projection (Spherical Projection, SP) as a consistent and flexible 3D shape representation, and performs conditional generation on SP maps through a pre-trained 2D diffusion model, enabling the rapid generation of high-quality meshes from single-view images.

Split4D: Decomposed 4D Scene Reconstruction Without Video Segmentation

Yongzhen Hu, Sida Peng

SegmentationGenerationRepresentation LearningTransformerDiffusion modelContrastive LearningGaussian SplattingOptical FlowImageVideo

🎯 What it does: Propose a method to recover a decomposed 4D scene representation (Split4D) from multi-view videos without requiring video segmentation, and achieve high-quality instance-level 4D segmentation and editing.

SRBTrack: Terrain-Adaptive Tracking of a Single-Rigid-Body Character Using Momentum-Mapped Space-Time Optimization

Hanyang Cao, Tae-Joung Kwon

OptimizationRobotic IntelligencePhysics Related

🎯 What it does: Train a single rigid body controller to generate realistic motion in complex physical environments and support zero-shot adaptation.

SS4D: Native 4D Generative Model via Structured Spacetime Latents

Zhibing Li (Chinese University of Hong Kong), Dahua Lin (Chinese University of Hong Kong)

GenerationData SynthesisTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkOptical FlowImageVideoPoint Cloud

🎯 What it does: Generate high-quality, spatiotemporally consistent dynamic 3D objects (4D content) from monocular videos

StableMotion: Training Motion Cleanup Models with Unpaired Corrupted Data

Yuxuan Mu (Simon Fraser University), Xue Bin Peng (Simon Fraser University)

RestorationPose EstimationTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkVideoSequential

🎯 What it does: Train a motion cleaning model on unpaired corrupted motion capture data, constructing the StableMotion framework that can automatically detect and repair various real-world motion capture noise.

Star-Shaped Distance Voronoi Diagrams for 3D Metamaterial Design

Logan Numerow, Bernhard Thomaszewski

OptimizationPhysics Related

🎯 What it does: A differentiable three-dimensional star-shaped distance Voronoi diagram was constructed and integrated into the optimization process of mechanical metamaterials, demonstrating that the method can generate periodic structures with diverse directional stiffness and stress-strain curves as well as gradually varying heterogeneous structures.

Statistical Error Reduction for Monte Carlo Rendering

Hiroyuki Sakai, David Hahn

OptimizationComputational EfficiencyDiffusion modelScore-based ModelImagePhysics Related

🎯 What it does: A general framework for statistical error reduction is proposed, covering denoising of estimated radiance and variance, and a new variance estimation denoising method and multi-transform denoising scheme are introduced;

StereoFG: Generating Stereo Frames from Centered Feature Stream

Chenyu Zuo, Rui Wang

GenerationSuper ResolutionConvolutional Neural NetworkRecurrent Neural NetworkDiffusion modelGenerative Adversarial NetworkOptical FlowImageVideo

🎯 What it does: Designed a stereo frame generation network based on centralized feature flow, achieving high-quality and temporally stable outputs for binocular views through low-resolution rendering with alternating eyes, generating results at four times the scale.

STGlight: Online Indoor Lighting Estimation via Spatio-Temporal Gaussian Fusion

Shiyuan Shen, Chunxia Xiao

Depth EstimationComputational EfficiencyTransformerNeural Radiance FieldGaussian SplattingVideoPoint Cloud

🎯 What it does: Propose STGlight, a lightweight online indoor lighting estimation method that can process LDR RGB-D video streams, maintain incrementally updated global geometric and lighting representations, and predict HDR environment maps for any position in each frame.

Strands2Cards: Automatic Generation of Hair Cards from Strands

Kenji Tojo, Hao Li

GenerationComputational EfficiencyDiffusion modelContrastive LearningGaussian SplattingOptical FlowMesh

🎯 What it does: Automatically convert fiber-based hair models into efficient mesh representations of hair cards for real-time rendering, while generating polygon strips and semi-transparent textures to maintain the original appearance.

StyleSculptor: Zero-Shot Style-Controllable 3D Asset Generation with Texture-Geometry Dual Guidance

Zefan Qu (City University of Hong Kong), Rynson W.H. Lau (City University of Hong Kong)

GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelScore-based ModelRectified FlowImagePoint CloudMesh

🎯 What it does: Proposes StyleSculptor, a zero-shot, pre-trained 3D generation model-based framework capable of generating 3D assets with dual controllability over texture and geometry from content images and style images.

Sums of Wedges: Conforming Weighted Delaunay Triangulations are Polynomial in Fixed Dimension

Dimitrios Bogiokas, Marc Alexa

🎯 What it does: Propose transforming the d-dimensional triangulation problem under the sub-simple x constraint into a lower envelope computation problem of wedge function sums, and prove the output polynomial complexity. A heuristic method for pruning wedge intersection points is proposed to reduce Steiner points, ultimately achieving an output-sensitive algorithm with near-linear time complexity.

Supra-threshold Contrast Perception in Augmented Reality

Dongyeon Kim, Rafał K. Mantiuk

OptimizationComputational EfficiencyRepresentation LearningContrastive LearningOptical FlowImage

🎯 What it does: Conduct experiments on optically transparent AR displays to investigate two hypotheses: background discount and suprathreshold contrast perception, and verify through controlled experiments which mechanism better explains enhanced contrast perception in AR.

Surface-Aware Distilled 3D Semantic Features

Lukas Uzolas (Delft University of Technology), Petr Kellnhofer (Delft University of Technology)

Knowledge DistillationRepresentation LearningTransformerDiffusion modelAuto EncoderContrastive LearningPoint CloudMesh

🎯 What it does: By performing self-supervised contrastive learning on the base features extracted from pre-trained 2D vision models, a surface-aware 3D semantic feature space is constructed, enabling the distinction of instances within the same semantic category (e.g., left and right hands, left and right feet), and achieving many-to-many 3D shape correspondence within this space.

SymBridge: A Human-in-the-Loop Cyber-Physical Interactive System for Adaptive Human-Robot Symbiosis

Haoran Chen, Changhe Tu

Pose EstimationRobotic IntelligenceReinforcement Learning from Human FeedbackConvolutional Neural NetworkRecurrent Neural NetworkTransformerVision-Language-Action ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowVideoMultimodalityPoint CloudBenchmark

🎯 What it does: This paper proposes SymBridge, a closed-loop human-robot collaboration system that uses AR to enable real humans and virtual robots to interact in physical space in real time, and continuously improves the interaction model through human feedback.

SZ Sequences: Binary-Based (0, 2q)-Sequences

Abdalla G. M. Ahmed, Hui Huang

Diffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: Propose a novel construction of (0,4)-sequences based on binary, and extend it to higher dimensions that are powers of two; achieve the ability to embed and combine low-dimensional sequences; use 2×2 block matrices as symbols to generate larger matrices, constructing (0,s)-sequences with the target base; provide a complete recursive generation formula with full resolution (64 bits); utilize binary generation matrices to achieve efficient bit operations, which can directly replace Sobol matrices;

TC-GS: A Faster Gaussian Splatting Module Utilizing Tensor Cores

Zimu Liao (Shanghai Jiao Tong University), Rong Fu (Shanghai Artificial Intelligence Laboratory)

Computational EfficiencyNeural Radiance FieldGaussian SplattingPoint CloudMesh

🎯 What it does: Propose the TC-GS module, which accelerates the alpha computation of 3D Gaussian splatting using Tensor Cores, achieving seamless integration.

Teamwork: Collaborative Diffusion with Low-rank Coordination and Adaptation

Sam Sartor (College of William & Mary), Pieter Peers (College of William & Mary)

GenerationData SynthesisTransformerDiffusion modelImage

🎯 What it does: Proposes the Teamwork framework, which utilizes multi-instance low-rank adaptation and coordination mechanisms to achieve input/output channel expansion and task transfer without modifying the structure of pre-trained diffusion models.

Temporally Smooth Mesh Extraction for Procedural Scenes with Long-Range Camera Trajectories using Spacetime Octrees

Zeyu Ma (Princeton University), Jia Deng (Princeton University)

GenerationData SynthesisNeural Radiance FieldGaussian SplattingOptical FlowPoint CloudMeshTime Series

🎯 What it does: Propose the BinocMesher method, which extracts temporally smooth 3D meshes from an occupancy function under long-distance camera trajectories using a 4D Binary-Octree.

The Aging Multiverse: Generating Condition-Aware Facial Aging Tree via Training-Free Diffusion

Bang Gong (University of North Carolina Chapel Hill), Roni Sengupta (University of North Carolina Chapel Hill)

Image TranslationGenerationTransformerPrompt EngineeringDiffusion modelRectified FlowImage

🎯 What it does: Generate a diverse aging trajectory tree (aging multiverse) for single facial images, conditionally generating multiple possible aging paths based on external factors such as environment, health, and lifestyle;

The Granule-In-Cell Method for Simulating Sand–Water Mixtures

Yizao Tang (Peking University), Bao Chen

Physics Related

🎯 What it does: Propose the Granule-In-Cell (GIC) method to achieve bidirectional coupling between sand grains and water, and maintain mass and volume conservation of the mixture by combining Discrete Element Method (DEM) with Particle-In-Cell (PIC).

Topology-Aware Optimization of Gaussian Primitives for Human-Centric Volumetric Videos

Yuheng Jiang (Max Planck Institute for Informatics), C. Theobalt

CompressionOptimizationAuto EncoderGaussian SplattingOptical FlowVideoPoint Cloud

🎯 What it does: Propose a topology-aware dynamic Gaussian representation (TaoGS) to achieve high-quality reconstruction and rendering of long-term motion tracking and topology changes (such as undressing and object interaction) in human videos.

TrackerSplat: Exploiting Point Tracking for Fast and Robust Dynamic 3D Gaussians Reconstruction

Daheng Yin (Simon Fraser University), Jiangchuan Liu (Simon Fraser University)

Object TrackingOptimizationComputational EfficiencyNeural Radiance FieldContrastive LearningGaussian SplattingOptical FlowVideoPoint Cloud

🎯 What it does: By combining pixel trajectories from multi-view point tracking with a 3D Gaussian model, TrackerSplat rapidly and robustly reconstructs dynamic 3D scenes in multi-view videos.

Training-Free Instance-Aware 3D Scene Reconstruction and Diffusion-Based View Synthesis from Sparse Images

Jiatong Xia (University of Adelaide), Lingqiao Liu (University of Adelaide)

SegmentationGenerationDepth EstimationTransformerDiffusion modelGaussian SplattingOptical FlowImagePoint Cloud

🎯 What it does: Achieve instance-aware 3D scene reconstruction and view synthesis based on diffusion models from sparse RGB images without training and pose preprocessing, supporting object-level editing;

Transient LASSO: Transient Large-Scale Scene Reconstruction

Dominik Scheuble, Felix Heide

Depth EstimationNeural Radiance FieldImagePoint Cloud

🎯 What it does: Proposes a neural scene reconstruction method based on transient imaging (Transient LASSO), for reconstructing the geometric structure and attributes of scenes in outdoor environments.

Ultrafast and Controllable Online Motion Retargeting for Game Scenarios

Tianze Guo, Xiaogang Jin

Pose EstimationOptimizationComputational EfficiencyOptical FlowMeshGraph

🎯 What it does: Proposed a super-fast, lightweight, controllable online motion retargeting framework that supports joint-level control;

UltraZoom: Generating Gigapixel Images from Regular Photos

Jingwei Ma (University of Washington), Steven M. Seitz (University of Washington)

GenerationData SynthesisSuper ResolutionTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelOptical FlowImageVideo

🎯 What it does: Generate a gigapixel-level ultra-high-resolution image of an object using a conventional-resolution panoramic image and several close-up images;

Underwater Optical Backscatter Communication using Acousto-Optic Beam Steering

A. Agarwal, A. Pediredla

OptimizationFederated LearningComputational EfficiencyRobotic IntelligenceOptical FlowUltrasoundPhysics Related

🎯 What it does: A water under optical backscattering communication technology based on acousto-optic beam steering is proposed.

Uni-Inter: Unifying 3D Human Motion Synthesis Across Diverse Interaction Contexts

Sheng Liu (Nanjing University), Xuelong Li (China Telecom)

GenerationData SynthesisPose EstimationTransformerDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningVideoPoint CloudMesh

🎯 What it does: Propose a unified interactive motion generation framework called Uni-Inter, which can handle human-human, human-object, human-scene, and multi-entity compound interactions under the same model.

Uni3C: Unifying Precisely 3D-Enhanced Camera and Human Motion Controls for Video Generation

Chenjie Cao (Alibaba), Yanwei Fu (Fudan University)

GenerationData SynthesisDepth EstimationTransformerVision-Language-Action ModelDiffusion modelGaussian SplattingOptical FlowImageVideoPoint CloudMeshBenchmark

🎯 What it does: Propose the Uni3C framework, unifying 3D-enhanced camera trajectory control with human motion control, for generating high-quality controllable videos from single views.

Unifying Latent Action and Latent State Pre-training for Policy Learning from Videos

Guangyan Chen, Yufeng Yue

Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelContrastive LearningImageVideoBenchmark

🎯 What it does: Propose the UniMimic method, which unifies the use of video data for pre-training on latent actions and latent states: first train a unified tokenizer to learn latent states from video frames and infer the latent actions between them, then pre-train a policy on videos to predict these latent actions and subsequent latent states, and finally fine-tune the policy on a robot dataset with action labels to achieve precise execution.

UTDesign: A Unified Framework for Stylized Text Editing and Generation in Graphic Design Images

Yiming Zhao (Peking University), Zhouhui Lian (Peking University)

Image TranslationGenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelDiffusion modelRectified FlowAuto EncoderImageTextMultimodalityBenchmark

🎯 What it does: Propose the UTDesign unified framework to achieve high-precision stylized text editing and generation of English labels in design images, and integrate a complete text-to-design (T2D) pipeline.

Variational Neural Surfacing of 3D Sketches

Yutao Zhang, Mikhail Bessmeltsev

OptimizationDiffusion modelScore-based ModelNeural Radiance FieldPoint CloudMesh

🎯 What it does: A new method is proposed to recover the implicit neural surface of 3D line drawings through variational optimization, using a curvature smoothing and input line alignment loss to reconstruct the surface.

Vertex Features for Neural Global Illumination

Rui Su (Peking University), Sheng Li (Peking University)

Computational EfficiencyRepresentation LearningGraph Neural NetworkDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingPoint CloudMesh

🎯 What it does: This paper proposes an encoding method that directly stores learnable features on the vertices of a 3D mesh (Neural Vertex Features), and combines adaptive multi-resolution refinement (LOD) to achieve efficient neural global illumination rendering.

Vertical Binocular Misalignment in AR Impairs Reading Performance

Daniel Gurman, Kevin W. Rio

OptimizationExplainability and InterpretabilityComputational EfficiencyText

🎯 What it does: Measured the time users take to detect, fuse, and process virtual reality content under different VBI and VID conditions by presenting short text on a multi-display haploscope.

VideoFrom3D: 3D Scene Video Generation via Complementary Image and Video Diffusion Models

Geon-Yeong Kim, Sunghyun Cho (POSTECH)

GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelScore-based ModelGaussian SplattingOptical FlowImageVideoPoint CloudMesh

🎯 What it does: Generate high-quality 3D scene videos using rough geometry, camera trajectories, and reference images, providing fast and style-consistent rendering results.

Virtually Being: Customizing Camera-Controllable Video Diffusion Models with Volumetric Performance Captures

Yuancheng Xu, Ning Yu

GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelGaussian SplattingVideoPoint Cloud

🎯 What it does: Propose a framework that achieves multi-view character consistency and 3D camera control in video diffusion models through a customized data pipeline.

Viscous Vortex Dynamics on Surfaces

Cuncheng Zhu, Albert Chern

Optical FlowMeshPhysics Related

🎯 What it does: A vorticity method was proposed on surfaces to simulate incompressible viscous flows, incorporating terms related to Gaussian curvature into the viscous force, thereby improving the evolution of the vorticity equation and harmonic components; meanwhile, an IMEX numerical scheme was introduced;

Voost: A Unified and Scalable Diffusion Transformer for Bidirectional Virtual Try-On and Try-Off

Seungyong Lee (NXN Labs), Jeong-gi Kwak (NXN Labs)

Image TranslationGenerationTransformerSupervised Fine-TuningDiffusion modelRectified FlowImage

🎯 What it does: Propose a unified diffusion transformer called Voost, which can perform virtual try-on and try-off tasks on the same model, achieving high-quality generation of clothing and human images.

Voyager: Long-Range and World-Consistent Video Diffusion for Explorable 3D Scene Generation

Tianyu Huang (Harbin Institute of Technology), Chunchao Guo (Tencent Hunyuan)

GenerationData SynthesisDepth EstimationTransformerDiffusion modelScore-based ModelGaussian SplattingWorld ModelOptical FlowImageVideoPoint Cloud

🎯 What it does: Propose the Voyager framework, which can generate long-range, globally consistent RGB-Depth videos based on a single image and user-defined camera trajectory, and directly apply them to 3D scene reconstruction and exploration.

Waste-to-Value: Reutilized Material Maximization for Additive and Subtractive Hybrid Remanufacturing

Fanchao Zhong, Haisen Zhao

OptimizationGraph Neural NetworkPrompt Engineering

🎯 What it does: Propose a computational framework that strictly considers global and local collision constraints to plan the addition and subtraction of material hybrid remanufacturing processes in order to maximize the reusable volume.

WATER: Watertight Tessellation for Real-Time Pixel-Accurate Rendering of Large-Scale Surfaces

Yajun Zeng, Ligang Liu

Computational EfficiencyMesh

🎯 What it does: Developed WATER, a software framework for real-time pixel-accurate rendering of large surfaces, supporting seamless (watertight) non-uniform subdivision;

Wavelet Fluids

Luan Lyu, Enhua Wu

Diffusion modelOptical FlowPhysics Related

🎯 What it does: A unified boundary condition framework based on wavelets is proposed for single-phase and two-phase fluid simulations.

WorldExplorer: Towards Generating Fully Navigable 3D Scenes

Manuel Schneider (Technical University of Munich), M. Nießner (Technical University of Munich)

GenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelGaussian SplattingImageVideoText

🎯 What it does: A panoramic preview is generated from text, followed by iteratively generating multi-view images through autoregressive video expansion trajectories and scene memory mechanisms, ultimately reconstructing a fully navigable 3D scene that can be rendered in real-time from any viewpoint using 3D Gaussian splats.

X-Actor: Emotional and Expressive Long-Range Portrait Acting from Audio

Chenxu Zhang (Bytedance), Linjie Luo (Bytedance)

GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelScore-based ModelAuto EncoderContrastive LearningImageVideoAudio

🎯 What it does: Propose the X-Actor framework, achieving the generation of actor-level, long-duration emotional portrait animations from a single image and audio;

X-UniMotion: Animating Human Images with Expressive, Unified and Identity-Agnostic Motion Latents

Guoxian Song (ByteDance), Linjie Luo (ByteDance)

GenerationPose EstimationDepth EstimationTransformerDiffusion modelAuto EncoderContrastive LearningImageVideo

🎯 What it does: Propose a unified and expressive implicit motion latent representation, achieving human image animation using four decoupled motion latent codes (face, body, left and right hands).