SIGGRAPH Asia 2025 Papers — Page 2
ACM SIGGRAPH Asia (Transactions on Graphics) · 301 papers
Fuse3D: Generating 3D Assets Controlled by Multi-Image Fusion
Xuancheng Jin (Zhejiang University), Yuchi Huo (Zhejiang University)
GenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderContrastive LearningImagePoint CloudMeshRetrieval-Augmented Generation
🎯 What it does: Fuse3D is a 3D asset generation method based on multi-graph fusion, which allows users to specify desired geometric and texture information in different regions of multiple control images. These regional features are then fused and directly drive 3D generation models to produce complete and controllable 3D models.
G2 Interpolating Spline with Local Maximum Curvature
Bowen Jiang, Renjie Chen
🎯 What it does: Proposes a class of G2 continuous spline curves with local maximum curvature control, capable of accurately interpolating given control points while maintaining compact local support;
GarmageNet: A Multimodal Generative Framework for Sewing Pattern Design and Generic Garment Modeling
Siran Li (Zhejiang Sci-Tech University), Huamin Wang (Style3D Research)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelAuto EncoderImageTextMultimodalityPoint CloudMesh
🎯 What it does: Propose GarmageNet, a unified multimodal generation framework that can automatically generate 2D sewing patterns, panel seam relationships, and 3D garment initializations directly applicable for physical simulation from text, sketches, point clouds, or planar patterns;
Gaussian Integral Linear Operators for Precomputed Graphics
Haolin Lu, T. Hachisuka
Gaussian Splatting
🎯 What it does: Proposed and implemented an integral linear operator method based on Gaussian mixture approximated kernel functions and input functions, which can analytically evaluate integrals and support operator chain composition.
Gaussian See, Gaussian Do: Semantic 3D Motion Transfer from Multiview Video
Yarin Bekor (Technion - Israel Institute of Technology), O. Litany
RestorationGenerationData SynthesisPose EstimationConvolutional Neural NetworkRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingOptical FlowVideoMultimodalityPoint CloudMeshBenchmark
🎯 What it does: Propose a semantic 3D motion transfer method from multi-view videos to a 3D Gaussian point cloud model
Generalized Unbiased Reconstruction for Gradient-Domain Rendering
Difei Yan, Kun Xu
RestorationImage
🎯 What it does: Propose a general unbiased reconstruction framework for gradient domain rendering, achieving unbiased reconstruction by linearly combining pixel colors and gradients.
Generating 360° Video is What You Need For a 3D Scene
Zhaoyang Zhang (Yale University), Yiwei Hu (Adobe Research)
GenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelGenerative Adversarial NetworkGaussian SplattingImageVideoText
🎯 What it does: Proposed a two-stage text-to-3D scene generation pipeline named WorldPrompter, which first generates a 360° panoramic video and then achieves a navigable 3D scene through fast 3D Gaussian Splatting reconstruction.
Generating Detailed Character Motion from Blocking Poses
Purvi Goel (Stanford University), Kayvon Fatahalian (Stanford University)
GenerationPose EstimationTransformerDiffusion model
🎯 What it does: Propose a technique based on diffusion models that can convert sparse, rough blocking poses provided by animators into detailed, natural character animations.
Generating Objects with Part-Articulation from a Single Image
Ruijie Lu, Siyuan Huang
GenerationData SynthesisTransformerVision Language ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkImagePoint CloudMesh
🎯 What it does: Developed a three-stage framework called DreamArt for generating high-fidelity, interactive multi-jointed objects from a single image.
Generating the Past, Present and Future from a Motion-Blurred Image
S. Tedla (York University), David B. Lindell (University of Toronto)
RestorationGenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelAuto EncoderImageVideo
🎯 What it does: This paper proposes a framework based on a large-scale video diffusion model to recover continuous video frames from past, present, to future using blurred images.
Generative Head-Mounted Camera Captures for Photorealistic Avatars
Shaojie Bai (Meta Reality Labs), Hyung Jun Kim (Meta Reality Labs)
GenerationData SynthesisRepresentation LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelContrastive LearningImageVideo
🎯 What it does: Proposed a GenHMC method based on diffusion models to generate head-mounted camera (HMC) captured images, thereby establishing correspondence between HMC and digital avatars without requiring paired data, and using these synthetic data to train a more robust general facial encoder.
GenLit: Reformulating Single-Image Relighting as Video Generation
Shrisha Bharadwaj (Max Planck Institute for Intelligent Systems), Michael J. Black (Max Planck Institute for Intelligent Systems)
Image TranslationRestorationGenerationTransformerSupervised Fine-TuningVision-Language-Action ModelDiffusion modelScore-based ModelImageVideo
🎯 What it does: Convert the local relighting problem of a single image into a video generation task, utilizing the Stable Video Diffusion model to generate illumination changes on a single image by controlling the position and intensity of point light sources.
GigaSLAM: Large-Scale Monocular SLAM with Hierarchical Gaussian Splats
Kai Deng (Nankai University), Jin Xie (Nankai University)
Depth EstimationAutonomous DrivingOptimizationComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkNeural Radiance FieldGaussian SplattingSimultaneous Localization and MappingOptical FlowImageVideoPoint Cloud
🎯 What it does: Proposes GigaSLAM, a monocular RGB odometry and mapping framework that utilizes a hierarchical sparse voxel structure and Gaussian projection rendering, enabling real-time localization and high-quality scene rendering in kilometer-scale, unbounded outdoor environments.
Glare Pattern Depiction: High-Fidelity Physical Computation and Physiologically-Inspired Visual Response
Yuxiang Sun, G. Baranoski
Diffusion modelOptical FlowImagePhysics Related
🎯 What it does: A high-fidelity glare pattern depiction framework based on physical computation and visual perception models is studied and proposed, applicable to both daytime and nighttime scenarios.
Gradient-Weighted Feature Back-Projection: A Fast Alternative to Feature Distillation in 3D Gaussian Splatting
Joji Joseph (Indian Institute of Science), Shalabh Bhatnagar (Indian Institute of Science)
SegmentationComputational EfficiencyKnowledge DistillationContrastive LearningGaussian SplattingImagePoint Cloud
🎯 What it does: This paper proposes a training-free feature projection method that projects 2D features back into pre-trained 3D Gaussians, achieving efficient and fast feature rendering and 3D segmentation.
GS-RoadPatching: Inpainting Gaussians via 3D Searching and Placing for Driving Scenes
Guo Chen (Beijing Normal University), Sheng Yang (Alibaba)
Autonomous DrivingDiffusion modelGaussian SplattingOptical FlowImageVideoPoint Cloud
🎯 What it does: Removing and completing target objects in driving scenes through 3D Gaussian Splatting, i.e., directly completing on the reconstructed 3D Gaussian scene, avoiding traditional 2D priors or additional retraining.
GSWT: Gaussian Splatting Wang Tiles
Yu Zeng, Pedro V. Sander
GenerationData SynthesisGaussian SplattingMesh
🎯 What it does: Proposed a texture framework based on Wang Tiles for generating large-scale or unlimited terrains from a single example, leveraging Gaussian Splatting for efficient rendering;
Harmonic Caching for Walk on Spheres
Zihong Zhou, Wojciech Jarosz
OptimizationComputational EfficiencyPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose a variance reduction technique that solves elliptic PDEs through overlapping harmonic expansions and unbiased Monte Carlo random walks.
Harnessing Diffusion-Yielded Score Priors for Image Restoration
Xinqi Lin (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Chao Dong (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelScore-based ModelGenerative Adversarial NetworkImageText
🎯 What it does: This paper proposes a image restoration framework called HYPIR based on pre-trained diffusion models. It first initializes the restoration network with the weights of the diffusion model, and then performs lightweight adversarial fine-tuning on this basis, eliminating the multi-step sampling process of traditional diffusion models, achieving restoration with a single forward pass.
Hierarchical Neural Semantic Representation for 3D Semantic Correspondence
Keyu Du (University of Electronic Science and Technology of China), Shuaicheng Liu (University of Electronic Science and Technology of China)
Representation LearningGraph Neural NetworkTransformerDiffusion modelScore-based ModelContrastive LearningPoint CloudMesh
🎯 What it does: This paper proposes a training-free hierarchical neural semantic representation (HNSR) that leverages pre-trained 3D generative models to achieve high-precision 3D semantic correspondence, supporting shapes with different topologies, geometries, and categories.
High-Fidelity Dynamic Portrait Animation via Direct Preference Optimization and Temporal Motion Modulation
Jiahao Cui, Siyu Zhu
GenerationData SynthesisReinforcement Learning from Human FeedbackConvolutional Neural NetworkTransformerDiffusion modelImageVideoAudio
🎯 What it does: Proposes a diffusion framework based on human preference alignment for high-dynamic, realistic portrait animation driven by audio and skeletal motion.
HiWave: Training-Free High-Resolution Image Generation via Wavelet-Based Diffusion Sampling
Tobias Vontobel (ETH Zurich), Romann M. Weber
GenerationSuper ResolutionTransformerDiffusion modelScore-based ModelAuto EncoderImage
🎯 What it does: Leverages pre-trained diffusion models to achieve high-resolution (4K and above) image synthesis through an untrained two-stage process (first generating a base image, then performing block-level DDIM inversion and waveform domain detail enhancement).
HOMA: Towards Generic Human-Object Interaction in Multimodal Driven Human Animation with Weak Conditions
Ziyao Huang, Fan Tang
GenerationData SynthesisTransformerPrompt EngineeringVision-Language-Action ModelDiffusion modelVideoMultimodalityAudio
🎯 What it does: Proposed a weakly conditioned, multi-modal driven HOI video generation framework called HOMA, which achieves controllable human-robot interaction animation through sparse decoupled motion guidance.
How Does a Virtual Agent Decide Where to Look? Symbolic Cognitive Reasoning for Embodied Head Rotation
Juyeong Hwang (Korea University), HyeongYeop Kang (Korea University)
Explainability and InterpretabilityRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageVideoText
🎯 What it does: Proposes the SCORE framework, which utilizes symbolic cognitive reasoning and vision-language models to predict the head movements of virtual agents, and combines FastVLM for online verification, achieving natural and interpretable head motion.
HRC-Net: Learning Visual Hypothesis, Representative, and Collaboration for Multi-Domain Image Inpainting
Xin Wang, Ping Li
RestorationConvolutional Neural NetworkDiffusion modelGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Proposes a framework named HRC-Net, which uses multi-domain image inpainting techniques to repair damaged regions in target images through three sub-networks.
HRM^2Avatar: High-Fidelity Real-Time Mobile Avatars from Monocular Phone Scans
Chao Shi (Alibaba Group), Chengfei Lv (Alibaba Group)
GenerationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingOptical FlowImageVideoMesh
🎯 What it does: Proposed an end-to-end single-phone scanning workflow that utilizes two-phase static and dynamic video sequences to construct high-fidelity full-body clothing avatars that can be rendered in real-time on mobile devices;
Hyperspectral Polarimetric BRDFs of Real-world Materials
Yunseong Moon (Pohang University of Science and Technology), Seung-Hwan Baek (Pohang University of Science and Technology)
Data SynthesisComputational EfficiencyDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingOptical FlowImageTabularPhysics Related
🎯 What it does: Collected and constructed a hpBRDF dataset covering the full Mueller matrix from visible to near-infrared light, covering 68 bands for the first time; meanwhile, an efficient acquisition and reconstruction system was proposed.
Imaginarium: Vision-guided High-Quality 3D Scene Layout Generation
Xiaoming Zhu (Tsinghua University), Long Zeng (Tsinghua University)
GenerationPose EstimationDepth EstimationRetrievalTransformerPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelImageTextPoint CloudMeshRetrieval-Augmented Generation
🎯 What it does: A visually driven system is proposed, which can generate high-quality, logically coherent 3D scene layouts from a predefined 3D asset library by combining text prompts, refined image generation models, image parsing, and asset retrieval.
Img2CAD: Reverse Engineering 3D CAD Models from Images through VLM-Assisted Conditional Factorization
Yang You (Stanford University), Leonidas J. Guibas (Stanford University)
Image TranslationGenerationData SynthesisConvolutional Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelFlow-based ModelImagePoint CloudMesh
🎯 What it does: A two-stage framework called Img2CAD is constructed, which first uses a VLM to predict the discrete CAD structure from a single-view image, and then employs a Transformer to predict continuous attributes, thereby achieving reverse engineering from images to 3D CAD.
Imperfect Image-Space Control Variates for Monte Carlo Rendering
Chanu Yang, Bochang Moon
OptimizationImage
🎯 What it does: Propose an image-space based control variate technique, using neighboring pixel estimates as control variates to reduce variance in Monte Carlo rendering.
Implicit Incompressible Porous Flow using SPH
Timna Böttcher (RWTH Aachen University), Jan Bender (RWTH Aachen University)
Optical FlowPhysics Related
🎯 What it does: Propose an implicit incompressible porous flow solver based on SPH;
Implicit Position-Based Fluids
Elier Diaz, Cem Yuksel
OptimizationOptical Flow
🎯 What it does: A new incompressible Smoothed Particle Hydrodynamics (SPH) scheme is proposed, which uses a second-order implicit descent method to optimize a special variational energy to achieve incompressibility.
Improving Curl Noise
J. A. Bærentzen, Sylvain Lefebvre
Diffusion modelScore-based ModelFlow-based ModelImage
🎯 What it does: Propose an n-dimensional divergence-free vector noise and verify its effectiveness in various scenarios such as image distortion and surface texture
In-2-4D: Inbetweening from Two Single-View Images to 4D Generation
Sauradip Nag (Simon Fraser University), A. M. Amiri
GenerationData SynthesisTransformerDiffusion modelScore-based ModelAuto EncoderGaussian SplattingOptical FlowImageVideoPoint CloudBenchmark
🎯 What it does: Interpolate 4D motion from two single-view images to generate continuous 3D+ motion sequences.
In-Context Brush: Zero-shot Customized Subject Insertion with Context-Aware Latent Space Manipulation
Yu Xu (Institute of Computing Technology, Chinese Academy of Sciences), Tong-Yee Lee (National Cheng-Kung University)
Image TranslationImage HarmonizationGenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: Propose the zero-shot subject insertion framework In-Context Brush, which leverages context learning in multi-modal diffusion transformers to seamlessly insert user-specified subject objects into target images, and achieve diverse control through text prompts.
INF-3DP: Implicit Neural Fields for Collision-Free Multi-Axis 3D Printing
Jiasheng Qu (Chinese University of Hong Kong), Guoxin Fang (Chinese University of Hong Kong)
OptimizationRobotic IntelligenceDiffusion modelScore-based ModelNeural Radiance FieldPoint CloudMesh
🎯 What it does: Proposes the INF-3DP framework based on implicit neural fields, unifying the optimization of toolpath generation and global collision-free motion planning for multi-axis 3D printing.
InfiniHuman: Realistic 3D Human Creation with Precise Control
Yuxuan Xue (University of Tübingen), Gerard Pons-Moll (University of Tübingen)
GenerationData SynthesisPose EstimationDepth EstimationSuper ResolutionTransformerPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingImageTextMultimodalityMesh
🎯 What it does: Built a complete framework called InfiniHuman that can infinitely generate diverse and precisely controllable 3D human characters.
Input-Aware Sparse Attention for Real-Time Co-Speech Video Generation
Beijia Lu (Carnegie Mellon University), Jun-Yan Zhu (Carnegie Mellon University)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationTransformerDiffusion modelScore-based ModelContrastive LearningImageVideoAudio
🎯 What it does: Propose a conditional video distillation method that realizes real-time lip-sync video generation using input human pose information, combining input-aware sparse attention and distillation loss.
Inverse Radiative Transport for Infrared Scenes with Gaussian Primitives
Zhenyuan Liu, Bernd Bickel
GenerationData SynthesisGaussian SplattingImagePhysics Related
🎯 What it does: Use Gaussian primitives to represent the scene, estimate surface temperature and material properties, and generate infrared renderings that match the input thermal images.
Inverse Tiling of 2D Finite Domains
Rulin Chen, Peng Song
Optimization
🎯 What it does: Propose the two-dimensional finite field inverse tiling problem, and provide a computational method for gradually constructing and optimizing the prototype shape set
Jackknife Transmittance and MIS Weight Estimation
Christoph Peters
OptimizationComputational EfficiencyPhysics Related
🎯 What it does: Proposed a transmission rate estimation method based on Jackknife, and extended it for estimating MIS weights, improving the bias and variance of traditional methods.
JoruriPuppet: Learning Tempo-Changing Mechanisms Beyond the Beat for Music-to-Motion Generation with Expressive Metrics
Ran Dong, Xi Yang
GenerationDiffusion modelGenerative Adversarial NetworkVideoMultimodalityAudio
🎯 What it does: This paper proposes new rhythm variation features, the JoruriPuppet dataset, and three evaluation metrics, and integrates the rhythm variation features into a neural network-based music-to-motion generation model.
Jump Restore Light Transport
Sascha Holl (Max Planck Institute for Informatics), Hans-Peter Seidel (Max Planck Institute for Informatics)
OptimizationComputational EfficiencyMixture of ExpertsImagePhysics RelatedStochastic Differential Equation
🎯 What it does: In this work, the authors propose a continuous-time MCMC framework that can convert any locally exploring Markov process into a global process, keeping it invariant under the target distribution, and apply this framework to light transport simulation.
JumpingGS: Level-jump 3D Gaussian Representation for Delicate Textures in Aerial Large-scale Scene Rendering
Jiongming Qin, Chunxia Xiao
Computational EfficiencyRepresentation LearningGaussian SplattingPoint CloudMesh
🎯 What it does: Propose the JumpingGS hierarchical Gaussian representation method to improve the rendering quality of fine textures in large-scale aerial scenes
Kinetic Free-Surface Flows and Foams with Sharp Interfaces
Haoxiang Wang, Wei Li
Physics Related
🎯 What it does: Proposes a free surface dynamics solver named HOME-FREE LBM, which utilizes sharp interface techniques to simulate liquid free surface flows, including phenomena such as turbulence, brewing, and foaming, and achieves fast and stable foam generation, splitting, and merging.
KISSColor: Kinetic and Intuitive Stroke Stretching for Vector Drawing Colorization
Yiming Dong, Wenping Wang
Mixture of ExpertsDiffusion modelOptical FlowImageMesh
🎯 What it does: Developed the KISSColor method, which uses intuitive stretching along tangent contour lines of the winding-number field for open strokes, and infers the user's intended closed regions through kinetic stroke stretching and redundant area suppression, achieving automatic coloring of vector drawings;
Lang3D-XL: Language Embedded 3D Gaussians for Large-scale Scenes
Shai Krakovsky (Tel Aviv University), Hadar Averbuch-Elor (Cornell University)
SegmentationRetrievalComputational EfficiencyRepresentation LearningTransformerVision Language ModelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingImageTextMultimodality
🎯 What it does: Propose a method that embeds language into a 3D Gaussian distribution, enabling semantic localization and visualization of large scenes through natural language interaction.
Large-Area Fabrication-aware Computational Diffractive Optics
Kaixuan Wei (King Abdullah University of Science and Technology), Wolfgang Heidrich (King Abdullah University of Science and Technology)
OptimizationConvolutional Neural NetworkDiffusion modelAuto EncoderImagePoint CloudTabularPhysics Related
🎯 What it does: We have designed and implemented a differentiable optical design framework aimed at large-area, mass-producible applications, taking into account process distortions and numerical high-resolution capabilities. It combines grayscale direct-write photolithography plus nanoimprint replication, utilizing a neural network digital twin to achieve process modeling, and employs GSPMD distributed FFT and convolution for large-scale DOE optimization.
LARM: A Large Articulated Object Reconstruction Model
Sylvia Yuan (University of California San Diego), Minghua Liu (Hillbot Inc.)
RestorationSegmentationGenerationPose EstimationDepth EstimationTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImagePoint CloudMesh
🎯 What it does: LARM is a fast, unified feedforward model that can reconstruct three-dimensional objects with joint motion from sparse perspective images.
LayerPeeler: Autoregressive Peeling for Layer-wise Image Vectorization
Rong Wu, Jing Liao (City University Of Hong Kong)
Image TranslationRestorationGenerationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageTextGraph
🎯 What it does: Propose LayerPeeler, an autoregressive hierarchical peeling framework that progressively removes the top non-occluded layers from images and restores the occluded content, ultimately generating a complete and structurally consistent SVG;
Learning Human Motion with Temporally Conditional Mamba
Quang Nguyen (FPT Software AI Center), Anh Nguyen (University of Liverpool)
GenerationData SynthesisPose EstimationTransformerDiffusion modelScore-based ModelVideoMultimodalityTime SeriesAudio
🎯 What it does: Proposed a Mamba-based diffusion framework called Temporally Conditional Mamba, for generating and estimating human motion that is temporally aligned with input signals under temporal conditions.
Learning to Ball: Composing Policies for Long-Horizon Basketball Moves
Pei Xu (Stanford University), C. K. Liu
Robotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningPrompt EngineeringMixture of ExpertsGenerative Adversarial NetworkImageVideoSequentialChain-of-Thought
🎯 What it does: Propose a new strategy integration framework that enables a basketball agent in physical simulation to smoothly perform actions such as dribbling, catching, and shooting in multi-stage, long-horizon tasks.
Learning to Refocus with Video Diffusion Models
S. Tedla, Shumian Xin (Adobe)
RestorationGenerationTransformerDiffusion modelAuto EncoderImageVideo
🎯 What it does: Generate full focal stack from a single defocused image using video diffusion models, enabling post-capture arbitrary focusing.
LEGO®-Maker: Autoregressive Image-Conditioned LEGO® Model Creation
Jiahao Ge, Chi-Wing Fu
GenerationData SynthesisKnowledge DistillationTransformerVision Language ModelDiffusion modelContrastive LearningImageText
🎯 What it does: Propose a learning-based generative model called LEGO®-Maker, which can quickly generate LEGO® models under image conditions.
LegoACE: Autoregressive Construction Engine for Expressive LEGO® Assemblies
Hao Xu, Xiaogang Jin
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextSequential
🎯 What it does: Proposes LegoACE, an autoregressive construction engine for automated LEGO assembly, capable of generating structurally complete LEGO models based on text prompts or multi-view normal maps.
Lifted Surfacing of Generalized Sweep Volumes
Yiwen Ju, Tao Ju
🎯 What it does: Propose a general algorithm for computing the boundary surface of the volume generated by the sweeping of rigid or deformable solids in three-dimensional space, which can ensure the generation of seamless and non-intersecting surfaces.
Light-SQ: Structure-aware Shape Abstraction with Superquadrics for Generated Meshes
Yuhan Wang (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
GenerationData SynthesisDiffusion modelScore-based ModelContrastive LearningGaussian SplattingPoint CloudMesh
🎯 What it does: Propose Light-SQ, a superquadric shape abstraction framework tailored for user-generated content scenarios;
Lightweight, Edge-Aware, and Temporally Consistent Supersampling for Mobile Real-Time Rendering
Sipeng Yang, Xiaogang Jin
Super ResolutionOptical FlowImage
🎯 What it does: A lightweight, edge-aware, and temporally consistent supersampling framework is proposed for real-time rendering on mobile devices, aiming to enhance image reconstruction quality and temporal coherence.
LLM-Primitives: Large Language Model for 3D Reconstruction with Primitives
Kuan Tian, Jun Zhang
GenerationData SynthesisTransformerLarge Language ModelVision-Language-Action ModelTextMultimodalityPoint Cloud
🎯 What it does: High-quality 3D primitives reconstruction is achieved by combining large language models (LLMs) with multimodal conditional inputs.
LookUp3D: Data-Driven 3D Scanning
Giancarlo Pereira (New York University), D. Panozzo
Depth EstimationGaussian SplattingSimultaneous Localization and MappingOptical FlowPoint CloudMesh
🎯 What it does: Propose LookUp3D, a structured light 3D scanning method based on a per-pixel color-to-depth lookup table, capable of acquiring high-resolution 3D geometry at high frame rates.
Low-Rank Adaptation of Neural Fields
Anh Truong (Massachusetts Institute of Technology), Justin Solomon (Massachusetts Institute of Technology)
RestorationCompressionOptimizationComputational EfficiencySupervised Fine-TuningNeural Radiance FieldImageVideoPoint CloudMesh
🎯 What it does: Proposes a parameter-efficient update method for instance-specific neural fields using low-rank adaptation (LoRA), which can encode minor edits (such as image filtering, geometric deformation, video compression, energy minimization) as low-rank weight increments;
LSF-Animation: Label-Free Speech-Driven Facial Animation via Implicit Feature Representation
Xin Lu (University of Chinese Academy of Sciences), Jun Xiao (University of Chinese Academy of Sciences)
GenerationRepresentation LearningTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningMultimodalityMeshAudio
🎯 What it does: A fully unlabeled voice-driven 3D facial animation framework, LSF-Animation, was developed, which can directly extract emotional and identity features from raw audio and neutral facial meshes, achieving unlabeled generation.
LVT: Large-Scale Scene Reconstruction via Local View Transformers
Tooba Imtiaz (Google), John Flynn (Google)
GenerationData SynthesisTransformerNeural Radiance FieldAuto EncoderGaussian SplattingImageVideo
🎯 What it does: Proposed the Local View Transformer (LVT), achieving high-quality 3D Gaussian plenoptic reconstruction and novel view synthesis for large-scale scenes with a single forward pass.
MALeR: Improving Compositional Fidelity in Layout-Guided Generation
Shivanker Saxena, Makarand Tapaswi (IIIT Hyderabad)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningImageTextBenchmark
🎯 What it does: For layout-guided text-to-image generation, MALeR addresses the issues of background semantic leakage, out-of-distribution generation, and attribute binding errors through three training-agnostic regularization mechanisms, achieving more accurate multi-agent multi-attribute scene synthesis.
Marching Neurons: Accurate Surface Extraction for Neural Implicit Shapes
Christian Stippel (TU Wien), P. Hermosilla
GenerationComputational EfficiencyRepresentation LearningSpiking Neural NetworkTransformerDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingPoint CloudMesh
🎯 What it does: Proposes an analytical surface extraction method based on neural implicit functions called Marching Neurons, which can directly obtain precise polygonal meshes from ReLU MLPs, avoiding errors caused by traditional sampling;
MaskedManipulator: Versatile Whole-Body Control for Loco-Manipulation
Chen Tessler, Gal Chechik
Knowledge DistillationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: Propose a full-body control framework that allows users to specify high-level goals (such as object pose or body pose), using MaskedManipulator to generate control policies. These policies are obtained by distilling a tracking controller trained on a large-scale human motion capture dataset, generating object manipulation behaviors that align with the specified goals.
MATStruct: High-quality Medial Mesh Computation via Structure-aware Variational Optimization
Ningna Wang (University of Texas at Dallas), Xiaohu Guo (University of Texas at Dallas)
OptimizationDiffusion modelScore-based ModelMesh
🎯 What it does: Proposed a structure-aware variational optimization framework called MATStruct for high-quality generation of intrinsic axis transform (medial axis transform) and mesh representations of three-dimensional shapes.
Medial Sphere Preconditioning for Knot Untangling and Volume-Filling Curves
Yuta Noma, Karan Singh
OptimizationComputational EfficiencyMeshBiomedical Data
🎯 What it does: A fast, robust, and controllable algorithm for rope untangling and volumetric curve filling is proposed.
MiGumi: Making Tightly Coupled Integral Joints Millable
Aditya Ganeshan (Brown University), M. Larsson
OptimizationDiffusion modelScore-based ModelContrastive LearningMeshBenchmark
🎯 What it does: Propose a process based on CNC machine tools that enables traditional woodworking joints (Kigumi) to be machined using standard flat-end cylindrical cutters, achieving tight fitting of the joints.
MILO: A Lightweight Perceptual Quality Metric for Image and Latent-Space Optimization
Ugur Çogalan, Colin Groth (Max Planck Institute for Informatics)
RestorationOptimizationConvolutional Neural NetworkMixture of ExpertsVision-Language-Action ModelDiffusion modelContrastive LearningImage
🎯 What it does: Proposes a lightweight multi-scale full-reference image quality assessment metric called MILO, which can output global MOS and spatial visibility maps, and supports use as a perceptual loss in both image and latent spaces.
MILo: Mesh-In-the-Loop Gaussian Splatting for Detailed and Efficient Surface Reconstruction
Antoine Guédon (École Polytechnique), M. Ovsjanikov
GenerationData SynthesisOptimizationComputational EfficiencyDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingImagePoint CloudMesh
🎯 What it does: Achieved bidirectional consistency with 3D Gaussian Splatting by dynamically extracting and optimizing the mesh during training, thus completing high-quality, low-vertex full scene reconstruction.
Modeling and Exploiting the Time Course of Chromatic Adaptation for Display Power Optimizations in Virtual Reality
Ethan Chen (University of Rochester), Yuhao Zhu (University of Rochester)
OptimizationComputational EfficiencyDiffusion modelContrastive LearningOptical FlowImageVideoTime Series
🎯 What it does: This paper proposes a method to reduce OLED display power consumption in VR devices by utilizing the color adaptation process. The core idea is to gradually change the scene's illumination color temperature, allowing the visual system to complete adaptation within an acceptable perceptual error range, thus significantly saving energy without significantly degrading the quality of experience.
MODepth: Benchmarking Mobile Multi-frame Monocular Depth Estimation with Optical Image Stabilization
Yu Lu, Guangtao Xue
Depth EstimationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningOptical FlowImageVideoBenchmark
🎯 What it does: Proposed a multi-frame monocular depth estimation system called MODepth, which is based on controlling motion through an optical image stabilization (OIS) module, and designed the MODNet network as well as the principal point offset estimation module and the pose estimation module.
Motion In-Betweening for Densely Interacting Characters
Xiaotang Zhang (Durham University), Hubert P. H. Shum (Durham University)
GenerationPose EstimationGraph Neural NetworkDiffusion modelAuto EncoderGenerative Adversarial NetworkVideoSequential
🎯 What it does: A motion interpolation method for dense interaction roles is studied, which can generate continuous and natural dual-role motion between given keyframes.
Motion2Motion: Cross-topology Motion Transfer with Sparse Correspondence
Lingxin Chen, Lei Zhang (International Digital Economy Academy)
GenerationData SynthesisPose EstimationGraph Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningOptical FlowVideoGraph
🎯 What it does: Propose the Motion2Motion framework, which utilizes a small number of target skeleton animations and sparse skeletal correspondences to achieve cross-topology motion transfer.
Multiphase Particle-Based Simulation of Poro-Elasto-Capillary Effects
Ruolan Li, Xiaokun Wang
Physics Related
🎯 What it does: A multiphase particle-based framework is proposed to simulate the pore-elastic-capillary coupled effects between fluids and porous media as a whole, and a physics-driven model is developed to capture elastic deformation and dynamic pore structure evolution under capillary action. Meanwhile, a saturation-aware pressure Poisson equation is constructed to ensure incompressibility of the fluid both inside and outside the porous media, and a representative elementary volume (REEB) method is proposed to unify the modeling of macroscopic homogeneous porous media and embedded cavity structures.
MV-Performer: Taming Video Diffusion Model for Faithful and Synchronized Multi-view Performer Synthesis
Yihao Zhi (CUHKSZ), Xiaoguang Han (CUHKSZ)
GenerationData SynthesisPose EstimationDepth EstimationTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowImageVideo
🎯 What it does: Proposed MV-Performer, a framework that converts monocular videos into synchronized multi-view 360-degree human novel view synthesis.
MVP4D: Multi-View Portrait Video Diffusion for Animatable 4D Avatars
Felix Taubner, David B. Lindell
GenerationData SynthesisTransformerDiffusion modelAuto EncoderContrastive LearningGaussian SplattingImageVideoMultimodality
🎯 What it does: Generate animatable 360° multi-view videos from a single reference image and distill them into 4D avatars that can be rendered in real-time;
Navigating with Annealing Guidance Scale in Diffusion Space
Shai Yehezkel (Tel Aviv University), Daniel Cohen-Or (Tel Aviv University)
GenerationTransformerPrompt EngineeringDiffusion modelScore-based ModelImageTextMultimodality
🎯 What it does: Developed an adaptive annealing guidance scheduler that dynamically adjusts the guidance scale in diffusion models to improve image quality and text consistency in text-to-image generation.
Neighbor-Aware Data-Driven Relaxation of Stitch Mesh Models for Knits
Yura Hwang, Cem Yuksel
OptimizationData-Centric LearningDiffusion modelScore-based ModelContrastive LearningOptical FlowMesh
🎯 What it does: A neighbor-aware data-driven relaxation method was developed for modeling seam behavior in lightweight fabric mesh models.
NeLiF: Neural Lighting Function Generation for Real-Time Indoor Rendering
Hongtao Sheng, Hujun Bao
GenerationData SynthesisTransformerDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkGaussian SplattingImagePoint Cloud
🎯 What it does: Proposed a two-stage neural illumination function generation model that can synthesize diverse lighting effects for unseen dynamic scenes and complex indoor lighting fixtures in real-time rendering.
Neural Hamiltonian Deformation Fields for Dynamic Scene Rendering
Hai-Long Qin (Beijing University of Posts and Telecommunications), Jincheng Dai (Beijing University of Posts and Telecommunications)
OptimizationComputational EfficiencyRepresentation LearningTransformerDiffusion modelNeural Radiance FieldGaussian SplattingImageVideoBenchmark
🎯 What it does: Propose a neural deformation field based on Hamiltonian dynamics, named NeHaD, to improve the rendering quality and motion coherence of dynamic Gaussian splatting.
Neural Image abstraction using long smoothing B-splines
Daniel Berio (University of London), A. Shamir
GenerationOptimizationDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageMultimodality
🎯 What it does: By linearly mapping uniform B-splines into the DiffVG rendering pipeline, the paper constructs an optimization framework that combines high-order continuity with gradient-differentiable image loss, enabling the generation of long, smooth, and controllable vector curves. Experiments are conducted on multiple tasks including image abstraction, texture inpainting, and text cartoonization.
Neural Kinematic Bases for Fluids
Yibo Liu (University of Victoria), T. Schneider
Diffusion modelAuto EncoderContrastive LearningOptical FlowPoint CloudMeshPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a neural motion foundation based on MLP for real-time, interactive animation generation in 2D and 3D fluid simulations.
Neural Octahedral Field: Octahedral Prior for Simultaneous Smoothing and Sharp Edge Regularization
Ruichen Zheng (Tsinghua University), Ruizhen Hu (Shenzhen University)
RestorationDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderContrastive LearningPoint CloudMesh
🎯 What it does: By introducing the Neural Octahedral Field (NOF) as a prior, simultaneous point cloud denoising and adaptive regularization of smooth and sharp edges are achieved for neural implicit surfaces, thereby recovering clearer and more sharply featured 3D surfaces from noisy point clouds.
Neural Texture Splatting: Expressive 3D Gaussian Splatting for View Synthesis, Geometry, and Dynamic Reconstruction
Yiming Wang (ETH Zurich), Siyu Tang (ETH Zurich)
GenerationData SynthesisComputational EfficiencyConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderGaussian SplattingImageVideoPoint Cloud
🎯 What it does: Propose Neural Texture Splatting (NTS), which adds a learnable RGBA texture field to each Gaussian primitive based on 3D Gaussian Splatting, to enhance local expressiveness.
Neural Visibility of Point Sets
Jun-Hao Wang (Peking University), Peng-Shuai Wang (Peking University)
ClassificationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkGraph Neural NetworkVision Language ModelAuto EncoderContrastive LearningPoint CloudMesh
🎯 What it does: This paper proposes a deep learning-based method for point cloud visibility determination, directly predicting the visibility of each point under a given viewpoint;
NeuralPVS: Learned Estimation of Potentially Visible Sets
Xiangyu Wang, Dieter Schmalstieg
Depth EstimationOptimizationComputational EfficiencyConvolutional Neural NetworkDiffusion modelScore-based ModelAuto EncoderContrastive LearningGaussian SplattingOptical FlowImageVideoPoint CloudMesh
🎯 What it does: Propose NeuralPVS, a real-time depth learning-based method for computing potential visibility sets (PVS) from regions, capable of outputting nearly error-free visible geometry sets at a frame rate of 100 Hz.
NeuVAS: Neural Implicit Surfaces for Variational Shape Modeling
Pengfei Wang, Wenping Wang
GenerationData SynthesisOptimizationDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingPoint CloudMeshGraph
🎯 What it does: Propose NeuVAS, a variational shape modeling framework based on neural implicit surfaces, which can directly generate high-quality, controllable surfaces from sparse curve networks, disconnected curve sketches, and sparse point clouds.
Nonlinear Noise2Noise for Efficient Monte Carlo Denoiser Training
Andrew Tinits (University of Waterloo), Stephen Mann (University of Waterloo)
RestorationConvolutional Neural NetworkScore-based ModelContrastive LearningImagePhysics Related
🎯 What it does: Proposed a theoretical framework that enables the safe use of nonlinear transformations (such as tone mapping) in Noise2Noise training, and applied it to denoising Monte Carlo rendered images
Numerical Homogenization of Sand from Grain-level Simulations
Yi-Lu Chen, C. Wojtan
Physics Related
🎯 What it does: A numerical homogenization method is used to automatically extract continuum material properties from granular rigid body simulations, converting them into elastic parameters and yield criteria, which are then simulated in MPM using an improved backward mapping technique.
NURBS-Based Grid Shell Form Finding on Domains with Topologically Arbitrary Boundaries
Masaaki Miki, Toby Mitchell
OptimizationDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowMeshPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This study proposes a shell shape finding method based on NURBS grids. To address the limitation of the traditional Airy stress function in describing stress distribution in domains with arbitrary topological boundaries, additional stress functions proposed by Schaefer (1953) and Gurtin (1963) are reintroduced. It is proven that combining these additional stress functions with the Airy stress function can fully characterize all possible stress states, thereby enabling the solution of shell shapes on domains with complex topological boundaries.
Object-level Visual Prompts for Compositional Image Generation
Gaurav Parmar (Carnegie Mellon University), Kfir Aberman (Snap Research)
GenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose VisualComposer, a method that can directly inject multiple object-level visual prompts into text-to-image diffusion models, enabling the generation of scenes with multiple objects, diverse layouts, and good identity preservation.
Off-Centered WoS-Type Solvers with Statistical Weighting
Anchang Bao (Tsinghua University), Jianmin Wang (Tsinghua University)
OptimizationComputational EfficiencyMeshTabularPhysics Related
🎯 What it does: A statistically weighted centrifugal Walk on Spheres (WoS) estimator is proposed to reduce bias and variance without losing the mean property, applicable to various PDEs such as Poisson and shielded Poisson.
OmnimatteZero: Fast Training-free Omnimatte with Pre-trained Video Diffusion Models
D. Samuel, Rami Ben-Ari
Image TranslationRestorationSegmentationGenerationTransformerDiffusion modelAuto EncoderOptical FlowImageVideoBenchmark
🎯 What it does: Propose OmnimatteZero, a fully training-free real-time video matting method that achieves object removal, foreground extraction, and layer composition while preserving object effects such as shadows and reflections.
OmniPart: Part-Aware 3D Generation with Semantic Decoupling and Structural Cohesion
Yu-nuo Yang, Xihui Liu (University of Hong Kong)
GenerationData SynthesisTransformerDiffusion modelRectified FlowContrastive LearningImagePoint CloudMesh
🎯 What it does: Propose the OmniPart framework, which can generate three-dimensional objects with editable structures from a single 2D image and controllable 2D part masks.
One String to Pull Them All: Fast Assembly of Curved Structures from Flat Auxetic Linkages
Akib Zaman, Mina Konaković Luković
Optimization
🎯 What it does: Propose a computational method that decomposes free-form structural surfaces into quadrilateral rigid tiles, which can be rapidly assembled into target 3D structures using a single rope from a flat configuration.
One-shot Embroidery Customization via Contrastive LoRA Modulation
Jun Ma (Zhejiang Sci-Tech University), Huamin Wang (Style3D Research)
Image TranslationGenerationData SynthesisTransformerDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: This paper proposes a contrastive learning-based low-rank adaptation (LoRA) modulation framework that achieves one-shot embroidery style customization using a single reference embroidery image;
Overlap Region Extraction of Two NURBS Surfaces
Jieyin Yang, Xiaohong Jia
OptimizationMesh
🎯 What it does: Propose an algorithm based on a two-layer optimization framework to compute the boundary of the overlapping region between two NURBS surfaces within a given error threshold.
PAD3R: Pose-Aware Dynamic 3D Reconstruction from Casual Videos
Ting-Hsuan Liao (University of Maryland College Park), Jia-Bin Huang (University of Maryland College Park)
GenerationPose EstimationDepth EstimationOptimizationTransformerDiffusion modelScore-based ModelContrastive LearningGaussian SplattingVideoMesh
🎯 What it does: Propose PAD3R, which can fully reconstruct the 4D structure of deformable 3D objects from casually captured monocular videos.