SIGGRAPH 2025 Papers — Page 2
ACM SIGGRAPH (Transactions on Graphics) · 306 papers
Fast Isotropic Median Filtering
Ben Weiss (Google Research)
RestorationComputational EfficiencyAuto EncoderContrastive LearningOptical FlowImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed a fast median/percentile filtering algorithm applicable to arbitrary bit depth, arbitrary radius, and arbitrary convex-shaped (especially circular) kernels, with both CPU and GPU implementations.
Fast Physics-Based Modeling of Knots and Ties using Templates
Dewen Guo, Huamin Wang
Diffusion modelGenerative Adversarial NetworkContrastive LearningGaussian SplattingSimultaneous Localization and MappingWorld ModelOptical FlowMeshPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a physics-based fast modeling system that maps fabric segments to controllable and non-penetrating 3D shapes of knots and ties using tubular parameterized templates, and supports interactions with surrounding meshes;
Fast Subspace Fluid Simulation with a Temporally-Aware Basis
Siyuan Chen, Zhecheng Wang
Computational EfficiencyVideoTime SeriesPhysics Related
🎯 What it does: Proposes a spatiotemporal adaptive subspace fluid simulation method based on Dynamic Mode Decomposition (DMD), which significantly reduces computational costs while maintaining high visual quality.
Feature-Aligned Parametrization in Penner Coordinates
Ryan Capouellez, Denis Zorin
OptimizationMesh
🎯 What it does: Extend Penner coordinates to support surfaces with sharp features, and propose a set of algorithms to align parameterization, while providing a two-stage method to satisfy feature constraints or minimize residuals when infeasible.
Feature-Preserving Mesh Repair via Restricted Power Diagram
Huibiao Wen, Changhe Tu
RestorationDiffusion modelContrastive LearningOptical FlowMesh
🎯 What it does: A unified mesh repair framework is proposed, which utilizes constrained power diagrams and wrapped surfaces to repair various mesh defects, achieving closure and manifoldness while preserving feature lines;
Feeling Blue or Seeing Red? Investigating the effect of light color, shadow and realism on the perception of emotion of real and virtual humans
Rachel Mcdonnell, Pisut Wisessing
RecognitionDiffusion modelNeural Radiance FieldAuto EncoderImage
🎯 What it does: By controlling light color, shadows, and illumination temperature, the study investigates their impact on emotion recognition, emotional intensity, and sincerity of photorealistic virtual humans and portrait photographs.
Field Smoothness-Controlled Partition for Quadrangulation
Zhongxuan Liang, Xiao-Ming Fu
Mesh
🎯 What it does: A novel partitioning method is proposed to achieve reliable feature-aligned quadrilateral meshing;
FlexiAct: Towards Flexible Action Control in Heterogeneous Scenarios
Shiyi Zhang (Tsinghua University), Yansong Tang (Tsinghua University)
Image TranslationGenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageVideoText
🎯 What it does: Designed and implemented FlexiAct, an image-to-video (I2V) framework tailored for heterogeneous scenarios, which can accurately transfer actions from reference videos to arbitrary target images without requiring prior pose or structural alignment;
Flexible 3D Cage-based Deformation via Green Coordinates on Bézier Patches
Dong Xiao (University of Science and Technology of China), Renjie Chen (University of Science and Technology of China)
Diffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingOptical FlowMeshStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes constructing Green coordinates on three-dimensional Bézier surface cages (including tensor product Bézier patches and Bézier triangles), thereby achieving compact and flexible cage deformation.
FLoD: Integrating Flexible Level of Detail into 3D Gaussian Splatting for Customizable Rendering
Y. Seo (Yonsei University), Youngjung Uh (Yonsei University)
Computational EfficiencyNeural Radiance FieldGaussian SplattingPoint CloudMesh
🎯 What it does: Propose a multi-layer representation that introduces an adjustable level of detail (LoD) into the 3D Gaussian Splatting (3DGS) framework, allowing users to choose between single-layer or multi-layer rendering under different GPU memory constraints.
Fluid Simulation on Compressible Flow Maps
Duowen Chen, Bo Zhu
Optical FlowPhysics Related
🎯 What it does: A unified compressible flow mapping framework is proposed and validated on three representative classes of compressible flow systems.
Fluid Simulation on Vortex Particle Flow Maps
Sinan Wang (Georgia Institute of Technology), Bo Zhu (Georgia Institute of Technology)
Diffusion modelOptical FlowPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a VPFM method that combines vortex particles with flow diagrams to simulate incompressible fluids and achieve long-term flow diagrams and high-fidelity vortex structures.
ForceGrip: Reference-Free Curriculum Learning for Realistic Grip Force Control in VR Hand Manipulation
Dong-qin Han, HyeongYeop Kang
Robotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningVision-Language-Action ModelDiffusion modelMultimodalityTime SeriesSequential
🎯 What it does: ForceGrip is a VR controller hand simulation system based on deep reinforcement learning, capable of mapping user trigger inputs to physically feasible and realistic grip forces.
GAIA: Generative Animatable Interactive Avatars with Expression-conditioned Gaussians
Zhengming Yu, Shalini De Mello
GenerationData SynthesisTransformerDiffusion modelNeural Radiance FieldGenerative Adversarial NetworkGaussian SplattingPoint CloudMesh
🎯 What it does: Developed the GAIA system, capable of generating high-fidelity 3D head avatars, achieving realistic animation and rendering, by employing Gaussian representations embedded in deformation models, shared UV parameterization, and expression-conditioned generators, and designing a two-branch architecture to achieve decoupling of identity and expression.
GarmentImage: Raster Encoding of Garment Sewing Patterns with Diverse Topologies
Yuki Tatsukawa, Takeo Igarashi
GenerationData SynthesisRepresentation LearningConvolutional Neural NetworkDiffusion modelAuto EncoderImageMesh
🎯 What it does: Proposes GarmentImage, a representation method that encodes garment sewing patterns into multi-channel raster grids, demonstrating its advantages in VAE, text editing, and image prediction tasks.
Gaussian Compression for Precomputed Indirect Illumination
Zhi Zhou, Zhangjin Huang
CompressionAuto EncoderGaussian SplattingImage
🎯 What it does: Propose a light field probe compression method based on low-bit adaptive Gaussian functions, achieving high-density sampling representation of indirect illumination in complex scenes
Gaussian Fluids: A Grid-Free Fluid Solver based on Gaussian Spatial Representation
Jingrui Xing, Bao Chen
OptimizationGaussian SplattingOptical FlowPoint CloudPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose a mesh-free fluid solver based on Gaussian Spatial Representation (GSR), which can continuously and differentiably represent the flow field and achieve the temporal evolution of Navier-Stokes equations through first-order optimization.
Gaussian Wave Splatting for Computer-Generated Holography
Suyeon Choi, G. Wetzstein
GenerationComputational EfficiencyDiffusion modelScore-based ModelRectified FlowNeural Radiance FieldGaussian SplattingOptical FlowImagePoint CloudMeshPhysics Related
🎯 What it does: Convert 2D Gaussian scene representations into optical phase fields to generate phase maps that can be directly used for holographic display.
GaVS: 3D-Grounded Video Stabilization via Temporally-Consistent Local Reconstruction and Rendering
Zinuo You (ETH Zurich), Dengxin Dai (Huawei Research Zurich)
RestorationSupervised Fine-TuningDiffusion modelGaussian SplattingOptical FlowVideo
🎯 What it does: This paper proposes GaVS, a full-frame video stabilization framework based on 3D scene reconstruction, which employs local Gaussian Splatting for reconstruction and rendering, and achieves stable results without cropping, distortion, and temporal inconsistency through fine-tuning during testing, dynamic compensation, extended regularization, and scene extrapolation.
GenAnalysis: Joint Shape Analysis by Learning Man-Made Shape Generators with Deformation Regularizations
Yuezhi Yang (University of Texas at Austin), Qixing Huang (University of Texas at Austin)
SegmentationOptimizationRepresentation LearningDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningPoint CloudMesh
🎯 What it does: Propose the GenAnalysis framework, which utilizes a learned implicit shape generator combined with an as-affine-as-possible (AAAP) deformation regularization to achieve joint shape matching and consistent segmentation of man-made shapes.
Generating Past and Future in Digital Painting Processes
Lvmin Zhang, Maneesh Agrawala
GenerationTransformerDiffusion modelImageVideo
🎯 What it does: A framework is constructed to generate past and future states of the painting process video based on the canvas image uploaded by the user, and the generated states can continue to be used as input to further generate more states.
Generative detail enhancement for physically based materials
Saeed Hadadan (University of Maryland), Matthias Zwicker (University of Maryland)
RestorationGenerationTransformerDiffusion modelScore-based ModelImageMesh
🎯 What it does: Utilize existing diffusion models and inverse rendering techniques to enhance the physical material textures of existing 3D assets, making them appear more aged, weathered, etc.
Generative Neural Materials
Nithin Raghavan, Ravi Ramamoorthi
GenerationData SynthesisConvolutional Neural NetworkRecurrent Neural NetworkTransformerVision Language ModelDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: This paper proposes the first image-conditioned diffusion model tailored for neuromorphic materials, and extends it to text-conditioned generation; meanwhile, it introduces a general MLP variant called NeuMIP, defines a 16-channel feature texture as a universal base, and trains conditional diffusion models on this base to generate neuromorphic materials from flash images, natural images, and text prompts; in addition, a new neuromorphic materials dataset containing 150k samples across 16 categories is constructed.
Generative Video Matting
Yongtao Ge (University of Adelaide), Chunhua Shen (Zhejiang University of Technology)
SegmentationGenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelFlow-based ModelAuto EncoderImageVideo
🎯 What it does: Transform the video matting task into a conditional generation problem, utilizing the pre-trained Stable Video Diffusion model combined with multi-stage training and synthesized/semi-annotated data to achieve high-quality video matting.
Geometric Contact Potential
Zizhou Huang (New York University), Denis Zorin (New York University)
OptimizationRobotic IntelligenceDiffusion modelScore-based ModelOptical FlowMeshStochastic Differential Equation
🎯 What it does: A new geometric contact potential (Geometric Contact Potential) is proposed for safe, differentiable, and pseudo-force-free self-collision detection and handling on continuous surfaces;
GSHeadRelight: Fast Relightability for 3D Gaussian Head Synthesis
Henglei Lv, Lin Gao
GenerationData SynthesisTransformerNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Proposed a fast portrait relighting method based on a 3D Gaussian model
Guided Lens Sampling for Efficient Monte Carlo Circle-of-Confusion Rendering
Jiawei Huang, Jiaping Wang
Computational EfficiencyNeural Radiance FieldGaussian Splatting
🎯 What it does: Propose a guided lens sampling method for efficiently rendering circular blur (CoC).
Guiding-Based Importance Sampling for Walk on Stars
Tianyu Huang (Tsinghua University), Feng Xu (Tsinghua University)
OptimizationComputational EfficiencyMixture of ExpertsDiffusion modelScore-based ModelPoint CloudMeshTabularPhysics Related
🎯 What it does: Propose a guided importance sampling method based on the Walk on Stars (WoSt) recursive term to significantly reduce variance.
Hand-Shadow Poser
Hao Xu (Chinese University of Hong Kong), Chi-Wing Fu (Chinese University of Hong Kong)
GenerationPose EstimationOptimizationConvolutional Neural NetworkTransformerDiffusion modelContrastive LearningOptical FlowImageMesh
🎯 What it does: Proposes a three-stage learning framework called Hand-Shadow Poser, which is used to inversely estimate the 3D joint poses of both hands from only the binary mask of hand shadows, thereby generating a light projection similar to the target hand shadow.
HexHex: Highspeed Extraction of Hexahedral Meshes
Tobias Kohler, D. Bommes
OptimizationComputational EfficiencyMesh
🎯 What it does: Proposed a novel, unconditional, and robust hexahedral mesh extraction algorithm called HexHex, specifically designed for local injective integer grid mapping to achieve high-performance and scalable mesh generation.
High-Fidelity Novel View Synthesis via Splatting-Guided Diffusion
Xiang Zhang, Christopher Schroers
GenerationData SynthesisDepth EstimationConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelAuto EncoderGaussian SplattingOptical FlowImageVideo
🎯 What it does: Propose a video diffusion model called SplatDiff based on pixel projection guidance, which is used to generate high-fidelity, geometrically consistent new views from a single image.
High-performance CPU Cloth Simulation Using Domain-decomposed Projective Dynamics
Zixuan Lu, Yin Yang
OptimizationComputational Efficiency
🎯 What it does: Proposed a CPU algorithm for high-resolution fabric simulation, utilizing domain decomposition and the global-local decomposition of projective dynamics to achieve parallel computing.
Histogram Stratification for Spatio-Temporal Reservoir Sampling
Corentin Salaün, K. Myszkowski
🎯 What it does: A two-step method is proposed, first organizing the sampling candidate points into local histograms, and then using quasi Monte Carlo and antithetic sampling patterns to sample from the histograms, in order to reduce real-time rendering noise and improve visual quality.
HOIGaze: Gaze Estimation During Hand-Object Interactions in Extended Reality Exploiting Eye-Hand-Head Coordination
Zhiming Hu (University of Stuttgart), Andreas Bulling (University of Stuttgart)
Pose EstimationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkGraph Neural NetworkTransformerContrastive LearningVideoPoint CloudGraph
🎯 What it does: This paper proposes a method for eye gaze estimation in hand-object interaction (HOI) scenarios, named HOIGaze.
HoLa: B-Rep Generation using a Holistic Latent Representation
Yilin Liu (Shenzhen University), Hui Huang (Shenzhen University)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextMultimodalityPoint CloudMesh
🎯 What it does: Designed and implemented HoLa (Holistic Latent), a system that learns Boundary Representation (B-Rep) models of CAD in a unified latent space, and trained a latent diffusion model based on this to generate unconditional and multimodal (image, point cloud, sketch, text) conditioned B-Rep models.
HumanRAM: Feed-forward Human Reconstruction and Animation Model using Transformers
Zhiyuan Yu (Hong Kong University of Science and Technology), Xiaowei Zhou (Zhejiang University)
RestorationGenerationPose EstimationTransformerDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingImageMesh
🎯 What it does: This paper proposes a forward-end-to-end full human reconstruction and animation model called HumanRAM, which is designed for sparse/single-frame human portrait images and can simultaneously accomplish novel view reconstruction and novel pose animation;
Hybrid Tours: A Clip-based System for Authoring Long-take Touring Shots
Xinrui Liu, Abe Davis
Neural Radiance FieldAuto EncoderOptical FlowVideo
🎯 What it does: Propose Hybrid Tours, which uses short video capture to represent potential tour segments, and generates long-distance camera trajectories by filtering and combining them through a custom interactive application, thus producing long shooting tour shots.
Hyper-Dimensional Deformation Simulation
Alvin Shi, Theodore Kim
OptimizationComputational EfficiencyDiffusion modelOptical FlowMeshPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes a four-dimensional deformable body simulation method, including 5-cell mesh generation, 4D hyper-elastic energy construction, and 4D collision detection and response.
Image-GS: Content-Adaptive Image Representation via 2D Gaussians
Yunxiang Zhang (New York University), Qi Sun (New York University)
CompressionSupervised Fine-TuningAuto EncoderGaussian SplattingImage
🎯 What it does: Proposed a content-adaptive image representation called Image-GS based on a 2D anisotropic Gaussian distribution.
Image-space Adaptive Sampling for Fast Inverse Rendering
Kai Yan, Shuang Zhao
Computational EfficiencyDiffusion modelNeural Radiance FieldAuto EncoderImage
🎯 What it does: Proposed an image space adaptive sampling framework for accelerating inverse rendering.
Image-Space Collage and Packing with Differentiable Rendering
Zhenyu Wang (Shenzhen University), Min Lu (Shenzhen University)
OptimizationComputational EfficiencyDiffusion modelGenerative Adversarial NetworkContrastive LearningImageTextGraph
🎯 What it does: This paper proposes an image-space collage and packing method based on differentiable rendering, which optimizes geometric elements using a loss function in the image space, achieving efficient shape filling and non-overlapping arrangements.
IMLS-Splatting: Efficient Mesh Reconstruction from Multi-view Images via Point Representation
Kaizhi Yang, Hao Su
OptimizationComputational EfficiencyRepresentation LearningDiffusion modelNeural Radiance FieldAuto EncoderGaussian SplattingImagePoint CloudMesh
🎯 What it does: Propose an end-to-end method for sparse point cloud representation and grid reconstruction called IMLS-Splatting
Improving Global Motion Estimation in Sparse IMU-based Motion Capture with Physics
Xinyu Yi (Tsinghua University), Feng Xu (Tsinghua University)
Pose EstimationOptimizationRecurrent Neural NetworkContrastive LearningSimultaneous Localization and MappingOptical FlowTime SeriesSequentialPhysics Related
🎯 What it does: A physics-driven method is proposed to develop a real-time global motion capture system using only 6 IMUs, capable of recovering full-body posture, global translation, contact points, contact forces, joint forces, and interactions with the environment in 3D space.
In Search of Empty Spheres: 3D Apollonius Diagrams on GPU
Cyprien Plateau–Holleville, Stéphane Mérillou
🎯 What it does: Propose a complete Apollonius graph construction algorithm tailored for GPU, based on core unit topological updates and iterative insertion of new sites, utilizing nearest neighbor queries to achieve parallel construction.
Instance Segmentation of Scene Sketches Using Natural Image Priors
Mia Tang (Stanford University), Maneesh Agrawala (Stanford University)
Object DetectionSegmentationDepth EstimationTransformerSupervised Fine-TuningDiffusion modelContrastive LearningImageTextRetrieval-Augmented Generation
🎯 What it does: A scene sketch instance segmentation method based on natural image priors, called InkLayer, was studied, which can group pixels in sparse sketches into different objects and generate editable hierarchical structures.
InstanceGen: Image Generation with Instance-level Instructions
Etai Sella (Tel Aviv University), Hadar Averbuch-Elor (Cornell University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes InstanceGen, a method that, without requiring additional training, generates high-quality images that conform to complex textual prompts involving multiple objects, attributes, and spatial relationships by combining the fine-grained structural initialization of image generation models and instance-level instructions generated by large language models.
Instant Self-Intersection Repair for 3D Meshes
W. Jang, Seungyong Lee
OptimizationMesh
🎯 What it does: Proposed an instant self-crossing repair framework that transforms local contact handling into global repair through local signed tangent point energy and gradient diffusion.
InstantRestore: Single-Step Personalized Face Restoration with Shared-Image Attention
Howard Zhang (Snap Inc.), Kfir Aberman
RestorationConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Propose InstantRestore, which utilizes a single-step diffusion model and a shared image attention mechanism to achieve personalized facial image restoration with a single forward pass.
Interactive Optimization of Scaffolded Procedural Patterns
D. Sforza, Fabio Pellacini
Optimization
🎯 What it does: Proposes an interactive parameter estimation method based on structured program patterns, which generates the target style hierarchically and optimizes parameters in real time.
Intersection-Free Garment Retargeting
Zizhou Huang, Victor Zordan
OptimizationMesh
🎯 What it does: Proposes a training-free mesh optimization method for transferring artist-designed clothing from a standard model to more general avatars, achieving non-intersecting and well-fitted clothing.
Interspatial Attention for Efficient 4D Human Video Generation
Ruizhi Shao (Tsinghua University), Gordon Wetzstein (Stanford University)
GenerationData SynthesisPose EstimationTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderGenerative Adversarial NetworkImageVideoPoint Cloud
🎯 What it does: The paper proposes a cross-space attention mechanism for high-quality 4D human video generation and builds the ISA-DiT model.
IntrinsicEdit: Precise generative image manipulation in intrinsic space
Linjie Lyu (Max-Planck-Institute for Informatics), Iliyan Georgiev (Adobe Research)
Image TranslationRestorationGenerationConvolutional Neural NetworkTransformerPrompt EngineeringDiffusion modelScore-based ModelAuto EncoderImage
🎯 What it does: Propose a framework for precise generative image editing in the intrinsic image space, supporting tasks such as object insertion/removal, material adjustment, and global relighting.
Inverse Design of Discrete Interlocking Materials with Desired Mechanical Behavior
Pengbin Tang, Bernd Bickel
OptimizationDiffusion modelScore-based ModelPhysics Related
🎯 What it does: Propose a computational method that utilizes an implicit toroidal union representation to model discrete interlocking materials (DIMs), and performs gradient optimization to design materials that achieve desired mechanical behaviors.
Inverse Geometric Locomotion
Quentin Becker, Mark Pauly
OptimizationOptical Flow
🎯 What it does: Propose a computational framework for optimizing shape sequences to achieve user-specified motion goals in scenarios where deformable bodies determine their motion through dynamic shape changes.
IP-Composer: Semantic Composition of Visual Concepts
Sara Dorfman (Tel Aviv University), Daniel Cohen-Or (Tel Aviv University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Propose IP-Composer, a training-free method for multi-image visual concept composition generation, which extracts concept subspaces from text instructions and merges concepts from different images by projecting them, generating new images.
IP-Prompter: Training-Free Theme-Specific Image Generation via Dynamic Visual Prompting
Yuxin Zhang, Changsheng Xu
GenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Proposes IP‑Prompter, a training‑free and model‑modification‑free visual prompting method that generates subject-specific images through Dynamic Visual Prompts (DVP).
James-Stein Gradient Combiner for Inverse Monte Carlo Rendering
Jeongmin Gu, Bochang Moon
OptimizationDiffusion modelScore-based ModelNeural Radiance FieldImagePhysics Related
🎯 What it does: A gradient mixer is proposed, combining unbiased gradients and biased gradients to improve scene parameter inference in inverse Monte Carlo rendering.
JGS2: Near Second-order Converging Jacobi/Gauss-Seidel for GPU Elastodynamics
Zixuan Lu (University of Utah), Yin Yang (University of Utah)
OptimizationComputational EfficiencyPoint CloudMeshGraphPhysics Related
🎯 What it does: Propose a GPU-based elastodynamic simulation algorithm called JGS2, which achieves efficient parallel solving using Jacobi/Gauss–Seidel iterations with near second-order convergence.
Kernel Predicting Neural Shadow Maps
Xuejun Hu, Kun Xu
Image TranslationRestorationComputational EfficiencyConvolutional Neural NetworkDiffusion modelAuto EncoderOptical FlowImage
🎯 What it does: Propose a neural network method called Kernel Predicting Neural Shadow Mapping, which converts hard shadow values into soft shadows by predicting pixel-level local filter weights, and achieves real-time high-quality shadows through the use of dilated filters and a loss function with temporal regularization.
LAM: Large Avatar Model for One-shot Animatable Gaussian Head
Yisheng He (Alibaba Group), Liefeng Bo (Alibaba Group)
GenerationData SynthesisTransformerGaussian SplattingImageVideoPoint Cloud
🎯 What it does: Generate high-quality 3D Gaussian avatars (Gaussian avatar) that can be instantly animated using a single image with a single forward pass, and can be directly real-time rendered in traditional rendering pipelines.
Large-Scale Multi-Character Interaction Synthesis
Ziyi Chang (Durham University), Hubert P. H. Shum (Durham University)
GenerationData SynthesisTransformerReinforcement LearningDiffusion modelVideo
🎯 What it does: Construct a conditional autoregressive generation pipeline to synthesize large-scale multi-agent coordinated interactions in the absence of multi-agent data.
LayerFlow: A Unified Model for Layer-aware Video Generation
Sihui Ji, Hengshuang Zhao
GenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderImageVideoText
🎯 What it does: Propose LayerFlow, a unified layer-aware video generation model that can generate three types of videos based on text prompts for foreground, background, and mixed layers, and supports layer decomposition and recombination;
LayerPano3D: Layered 3D Panorama for Hyper-Immersive Scene Generation
Shuai Yang (Shanghai Jiao Tong University), Dahua Lin (Chinese University of Hong Kong)
GenerationData SynthesisDepth EstimationSuper ResolutionTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelGaussian SplattingImageTextPoint Cloud
🎯 What it does: Generate fully explorable 360°×180° 3D panoramic scenes from a single text prompt.
Leapfrog Flow Maps for Real-Time Fluid Simulation
Yuchen Sun, Bo Zhu
Flow-based ModelOptical FlowPhysics Related
🎯 What it does: Propose Leapfrog Flow Maps (LFM), combining a hybrid velocity-momentum scheme with the leapfrog method to achieve real-time incompressible fluid simulation, and accelerate the projection from momentum to velocity using adaptive multigrid preconditioned conjugate gradient (AMGPCG);
Learning to Assemble with Alternative Plans
Ziqi Wang, Maryam Kamgarpour
Robotic IntelligenceGraph Neural NetworkReinforcement Learning
🎯 What it does: Propose a multi-robot collaboration assembly framework based on reinforcement learning, which can generate backup assembly plans in real-time during construction to adapt to unexpected on-site changes.
Learning to Draw Is Learning to See: Analyzing Eye Tracking Patterns for Assisted Observational Drawing
Feng Liu, Zeyu Wang
Image TranslationGenerationDiffusion modelOptical FlowImageVideo
🎯 What it does: Studied the correspondence between eye movements and painting actions during the image-to-image drawing process, and based on this, developed a real-time visual guidance-assisted painting interface.
Lifting the Winding Number: Precise Discontinuities in Neural Fields for Physics Simulation
Yue Chang (University of Toronto), E. Grinspun
Diffusion modelScore-based ModelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingOptical FlowPoint CloudMeshPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a Wind Lifter neural field method, which constructs a volumetric field that remains continuous during evolutionary cutting by embedding the generalized vortex number of the cutting curve into the input, thereby achieving precise modeling of discontinuities when cutting thin-walled flexible structures.
Light Pipe Holographic Display: Bandwidth-preserved Kaleidoscopic Guiding for AR Glasses
Minseok Chae (Seoul National University), Yoonchan Jeong (Seoul National University)
OptimizationComputational EfficiencyNeural Radiance FieldGaussian SplattingOptical FlowImagePoint CloudPhysics Related
🎯 What it does: Proposed a full-bandwidth preserving optical tiling technology utilizing light pipes to achieve full-bandwidth, full-field-of-view holographic display in AR glasses, and spatially separated the light source driving unit from the image combiner;
LightLab: Controlling Light Sources in Images with Diffusion Models
Nadav Magar (Tel Aviv University), Yedid Hoshen (Hebrew University of Jerusalem)
Image TranslationRestorationGenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelImageRetrieval-Augmented Generation
🎯 What it does: By combining real-world lighting variation paired data with large-scale synthesized rendered data, the diffusion model is fine-tuned to achieve parameterizable control over the intensity, color, and environmental illumination of visible light sources from a single image.
Lightning-fast Boundary Element Method
Jiong Chen, Mathieu Desbrun
OptimizationComputational EfficiencyPhysics Related
🎯 What it does: Constructing a sparse inverse LU decomposition approximation as a preconditioner for GMRES to accelerate the solution of boundary element methods
Linear-Time Transport with Rectified Flows
Khoa Do, Nicolas Bonneel
GenerationOptimizationFlow-based ModelRectified FlowOptical Flow
🎯 What it does: This paper proposes a linear-time algorithm based on Rectified Flows for generating transport plans from a uniform density to a density stored on a grid, and further extends it to match two non-uniform distributions, achieving applications such as shape interpolation and centroid calculation.
MAGNET: Muscle Activation Generation Networks for Diverse Human Movement
Jungnam Park, Jungdam Won
GenerationData SynthesisPose EstimationComputational EfficiencyKnowledge DistillationGraph Neural NetworkTransformerReinforcement LearningDiffusion modelScore-based ModelBiomedical Data
🎯 What it does: Developed MAGNET, an scalable framework for reconstructing full-body muscle activation for diverse human motions.
MASH: Masked Anchored SpHerical Distances for 3D Shape Representation and Generation
Changhao Li (University of Science and Technology of China), Ruizhen Hu (Shenzhen University)
GenerationRepresentation LearningDiffusion modelAuto EncoderContrastive LearningGaussian SplattingPoint CloudMesh
🎯 What it does: Proposes MASH (Masked Anchored Spherical Distances), a multi-view, parameterizable 3D shape representation method, and provides a differentiable optimization algorithm to convert point clouds into continuous surface representations.
MatCLIP: Light- and Shape-Insensitive Assignment of PBR Material Models
Michael Birsak (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
RecognitionRepresentation LearningTransformerVision Language ModelDiffusion modelContrastive LearningImageMesh
🎯 What it does: MatCLIP achieves illumination and geometry adaptive material assignment for 3D models by learning PBR material embeddings that are invariant to shape and illumination.
MaterialPicker: Multi-Modal DiT-Based Material Generation
Xiaohe Ma (Zhejiang University), Yiwei Hu (Adobe Research)
RestorationGenerationTransformerPrompt EngineeringDiffusion modelScore-based ModelAuto EncoderContrastive LearningImageVideoTextMultimodality
🎯 What it does: Developed a multi-modal Diffusion Transformer (DiT) model called MaterialPicker, which can generate high-quality PBR materials (albedo, normal, roughness, height, metallic) based on image crops or text prompts, and automatically correct issues such as perspective, distortion, and occlusion.
Meschers: Geometry Processing of Impossible Objects
Ana Dodik (MIT CSAIL), Justin Solomon (MIT CSAIL)
OptimizationComputational EfficiencyRepresentation LearningDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningMesh
🎯 What it does: A new geometric representation method called mescher is proposed for handling impossible objects, allowing rendering and relighting of these objects, and performing intrinsic geometric processing operations such as heat diffusion and geodesic distance queries.
MGPBD: A Multigrid Accelerated Global XPBD Solver
Chunlei Li (Beihang University), Qinping Zhao (Beihang University)
OptimizationMesh
🎯 What it does: Propose a global XPBD solver MGPBD accelerated by using unsmoothed aggregation (UA) AMG preconditioned conjugate gradient (PCG).
MIND: Microstructure INverse Design with Generative Hybrid Neural Representation
Tianyang Xue (Shandong University), Bernd Bickel (ETH Zurich)
GenerationData SynthesisTransformerMixture of ExpertsDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkGaussian SplattingPoint CloudMeshPhysics Related
🎯 What it does: MIND achieves inverse design of 3D printable microstructures through generative hybrid neural representations (Holoplane), directly generating diverse and geometrically valid microstructures based on target elastic tensors.
MiSo: A DSL for Robust and Efficient Solve and MInimize Problems
Federico Sichetti, D. Panozzo
Optimization
🎯 What it does: Propose a domain-specific language called MiSo and its compiler, specifically designed for low-dimensional optimization and solving problems, used to generate efficient C++ code;
Mobius: Text to Seamless Looping Video Generation via Latent Shift
Xiuli Bi (Chongqing University of Post and Telecommunications), Bin Xiao (Chongqing University of Post and Telecommunications)
GenerationTransformerPrompt EngineeringDiffusion modelAuto EncoderVideoText
🎯 What it does: Propose the Mobius method, which utilizes a pre-trained text-to-video diffusion model, and achieves seamless loop video generation without training by performing cyclic shifts in the latent space during the inference phase.
Model See Model Do: Speech-Driven Facial Animation with Style Control
Yifang Pan (University of Toronto), Luiz Gustavo Hafemann (Ubisoft La Forge)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelScore-based ModelAuto EncoderContrastive LearningVideoMultimodalityAudio
🎯 What it does: Propose a example-based speech-driven 3D facial animation generation framework, which can control the animation's expression style through style reference videos while maintaining precise lip synchronization.
Modeling and Rendering Glow Discharge
Venkataram Edavamadathil Sivaram, Tzu-Mao Li
Neural Radiance FieldOptical FlowPhysics Related
🎯 What it does: Proposed a glow discharge model for neon and gas discharge lamps, and developed an efficient point-wise solver suitable for traditional volume rendering systems;
Moment Bounds are Differentiable: Efficiently Approximating Measures in Inverse Rendering
Markus Worchel, Marc Alexa
Computational EfficiencyDiffusion modelNeural Radiance Field
🎯 What it does: Proposes a moment-based differentiable approximation for light propagation attenuation modeling in inverse rendering
MonetGPT: Solving Puzzles Enhances MLLMs' Image Retouching Skills
Niladri Shekhar Dutt (University College London), Niloy J. Mitra (University College London)
Image TranslationImage HarmonizationRestorationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This work proposes MonetGPT, an image editing system based on a multimodal large language model (MLLM), which trains the model to master image processing operations by solving visual puzzles. Subsequently, the model can identify issues in the original photo, plan a series of operational steps, and generate corresponding parameters, ultimately invoking a predefined procedural editing library to complete high-quality, interpretable image editing.
Monocular Online Reconstruction with Enhanced Detail Preservation
Songyin Wu (University of California Santa Barbara), Zhao Dong (Meta Reality Labs Research)
Depth EstimationAutonomous DrivingOptimizationConvolutional Neural NetworkTransformerContrastive LearningGaussian SplattingSimultaneous Localization and MappingOptical FlowImageVideoPoint Cloud
🎯 What it does: Proposed an online monocular RGB stream 3D high-quality reconstruction framework, utilizing a three-dimensional Gaussian distribution to achieve instant, detail-rich scene reconstruction.
Motion Control via Metric-Aligning Motion Matching
Naoki Agata (University of Tokyo), Takeo Igarashi (University of Tokyo)
GenerationData SynthesisPose EstimationOptimizationTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowVideoMultimodalitySequentialAudio
🎯 What it does: This paper proposes a method called Metric-Aligning Motion Matching (MAMM), which enables motion control by performing metric alignment between original motion sequences and any control sequences (such as sketches, labels, audio, or another motion sequence).
Motion Inversion for Video Customization
Luozhou Wang (Hong Kong University Of Science And Technology (Guangzhou)), Yingcong Chen (Hong Kong University Of Science And Technology (Guangzhou))
GenerationData SynthesisTransformerDiffusion modelContrastive LearningOptical FlowVideoTextMultimodality
🎯 What it does: This paper proposes a new video customization method, which utilizes motion embeddings learned from reference videos to control the temporal attention of text-to-video (T2V) diffusion models, thereby enabling the transfer of motion from source videos to arbitrary target content.
Motion-example-controlled Co-speech Gesture Generation Leveraging Large Language Models
Bohong Chen (Zhejiang University), Kun Zhou (Zhejiang University)
GenerationPose EstimationTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderVideoMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Propose a motion example-controlled co-speech gesture generation framework MECo based on large language models;
MotionCanvas: Cinematic Shot Design with Controllable Image-to-Video Generation
Jinbo Xing (Chinese University of Hong Kong), Feng Liu (Adobe Research)
GenerationData SynthesisDepth EstimationTransformerVision-Language-Action ModelDiffusion modelAuto EncoderContrastive LearningOptical FlowImageVideoMultimodality
🎯 What it does: Designed a MotionCanvas that allows users to control camera and object motion within a 3D scene space, generating high-quality cinematic-level videos.
MoVer: Motion Verification for Motion Graphics Animations
Jiaju Ma, Maneesh Agrawala
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringImageVideoTextRetrieval-Augmented Generation
🎯 What it does: Proposed MoVer DSL for checking spatiotemporal properties of motion graphics animations, and built a synthesis and verification pipeline based on LLM, achieving iterative refinement of animations.
Multi-Dimensional Procedural Wave Noise
Pascal Guehl, J. Dischler
GenerationData SynthesisDiffusion modelScore-based Model
🎯 What it does: Proposes a waveform-based procedural noise model that can be used in any dimension, utilizing spectral domain representation and implemented via inverse Fourier transform, employing precomputed complex-valued hyperplane wave functions to achieve efficient integration.
Multi-Person Interaction Generation from Two-Person Motion Priors
Wenning Xu (University of Glasgow), Edmond S. L. Ho (University of Glasgow)
GenerationData SynthesisPose EstimationGraph Neural NetworkDiffusion modelScore-based ModelContrastive LearningOptical FlowVideoMeshGraph
🎯 What it does: This paper proposes a graph-based interactive sampling method that generates multi-person interactive animations using existing two-person motion diffusion models;
Multiple Importance Reweighting for Path Guiding
Zhimin Fan, Jie Guo
Mixture of ExpertsGaussian Splatting
🎯 What it does: Propose a multi-importance reweighting method, treating path-guided estimates combination as a path-level reweighting process and calculating composite weights that account for spatial-directional variations.
MyTimeMachine: Personalized Facial Age Transformation
Luchao Qi (University of North Carolina at Chapel Hill), Roni Sengupta (University of North Carolina at Chapel Hill)
Image TranslationRestorationGenerationTransformerPrompt EngineeringDiffusion modelGenerative Adversarial NetworkContrastive LearningImageVideo
🎯 What it does: This paper proposes a system called MyTimeMachine, which learns a personalized facial age transformation model by using only 10–50 personal photos spanning 20–40 years and global age priors.
NAM: Neural Adjoint Maps for refining shape correspondences
G. Viganò, S. Melzi
OptimizationRepresentation LearningGraph Neural NetworkContrastive LearningPoint CloudMesh
🎯 What it does: Proposes using neural adjoint maps based on multilayer perceptrons to refine 3D shape correspondences, and designs an iterative algorithm to improve accuracy and robustness.
Nested Attention: Semantic-aware Attention Values for Concept Personalization
Or Patashnik (Tel Aviv University), Daniel Cohen-Or (Tel Aviv University)
Image TranslationGenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposed and implemented the Nested Attention mechanism, injecting highly expressive personalized themes into text-to-image diffusion models with a single text token, thereby enhancing identity fidelity and prompt alignment.
Neural BRDF Importance Sampling by Reparameterization
Liwen Wu (University of California San Diego), R. Ramamoorthi
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageMeshBenchmark
🎯 What it does: Propose a neural BRDF importance sampling method based on reparameterization, which directly learns a single-step trainable MLP to convert the BRDF integral into an easily sampled form and integrates it unbiasedly with MIS.
Neural Co-Optimization of Structural Topology, Manufacturable Layers, and Path Orientations for Fiber-Reinforced Composites
Tao Liu (University of Manchester), Charlie C. L. Wang (University of Manchester)
OptimizationDiffusion modelNeural Radiance FieldAuto EncoderPoint CloudMeshTabularFibre Orientation DistributionBenchmark
🎯 What it does: This study proposes a neural network framework based on implicit neural fields for simultaneously optimizing the structural topology, manufacturable bending layers, and path directions of fiber-reinforced composite materials, achieving synergistic optimization of strength and manufacturability during multi-axis 3D printing processes.
Neural Importance Sampling of Many Lights
P. Figueiredo, Nima Khademi Kalantari
OptimizationComputational EfficiencyDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningImageMesh
🎯 What it does: Propose a multi-light importance sampling method based on neural networks, using local illumination information to predict the light source selection distribution for each shading point, and training the network online during rendering.
NeurCross: A Neural Approach to Computing Cross Fields for Quad Mesh Generation
Qiujie Dong (Shandong University), Wenping Wang (Texas A&M University)
GenerationOptimizationConvolutional Neural NetworkGraph Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningPoint CloudMesh
🎯 What it does: Generate high-quality quadrilateral meshes by simultaneously learning the signed distance function and cross field through neural networks.