SIGGRAPH Asia 2025 Papers with AI Summaries
ACM SIGGRAPH Asia (Transactions on Graphics) · 301 papers
→ SIGGRAPH Asia 2025 papers with code (41)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 301 SIGGRAPH Asia 2025 papers by keyword, author or institution —
free trial on arXivSub.
3D SMoE Splatting for Edge-aware Realtime Radiance Field Rendering
Yi-Hsin Li, Mårten Sjöström
OptimizationComputational EfficiencyMixture of ExpertsNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: Proposed a new 3D SMoE Splatting (3DSMoES) framework, integrating the Steered Mixtures-of-Experts regression method into the CUDA-optimized process of 3D Gaussian Splatting (3DGS) to achieve 3D rendering.
3DPR: Single Image 3D Portrait Relighting with Generative Priors
Pramod Rao (Max Planck Institute for Informatics), Marc Habermann (Max Planck Institute for Informatics)
RestorationGenerationSuper ResolutionConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkImageMesh
🎯 What it does: Developed a method for 3D relighting and view reconstruction from a single portrait image.
4DSloMo: 4D Reconstruction for High Speed Scene with Asynchronous Capture
Yutian Chen, Tianfan Xue
RestorationGenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelGaussian SplattingImageVideo
🎯 What it does: This paper proposes an asynchronous capture scheme combined with a video diffusion model to repair 4D reconstruction from sparse viewpoints, thereby achieving high frame rate and high-quality 4D dynamic reconstruction without the need for high-speed cameras.
A compact stochastic representation for Monte Carlo Path Traced images
Matthias Treder, Roc Ramon Currius
Computational EfficiencyRepresentation LearningDiffusion modelScore-based ModelAuto EncoderContrastive LearningGaussian SplattingImage
🎯 What it does: Propose a learning-based compact representation that captures the complete Monte Carlo sampling distribution of rendered images, enabling fast rendering at arbitrary samples per pixel (SPP) during inference without the need for expensive path tracing.
A Highly-Efficient Hybrid Simulation System for Flight Controller Design and Evaluation of Unmanned Aerial Vehicles
Jiwei Wang, Xiaopei Liu
Autonomous DrivingOptimizationComputational EfficiencyRobotic IntelligenceSimultaneous Localization and MappingWorld ModelOptical FlowPhysics Related
🎯 What it does: A hybrid simulation system combining a far-field adaptive block fluid simulator and a parameterized empirical model at the aircraft boundary, with automatic parameter calibration, was developed for the design and evaluation of drone flight controllers.
A Nonconforming Formulation of Cloth
Elias Gueidon, Maurizio M. Chiaramonte
Physics Related
🎯 What it does: A fabric simulation method based on nonconforming function spaces is proposed, which calculates the surface bending energy by using an internal penalty term to weakly enforce the continuity of tangent basis functions;
A Stack-Free Parallel h-Adaptation Algorithm for Dynamically Balanced Trees on GPUs
Lixin Ren, Enhua Wu
OptimizationComputational Efficiency
🎯 What it does: Developed a stackless parallel h-adaptive algorithm for GPUs for the construction and update of dynamic balanced trees to improve fluid simulation performance
Acoustic Reliefs
Jeremy Chew, Bernd Bickel
GenerationOptimizationDiffusion modelGenerative Adversarial NetworkContrastive LearningImageAudio
🎯 What it does: Designed and generated an acoustic relief that can uniformly scatter sound and visually approximate the image provided by the user, achieving multi-objective optimization through differential simulation and rendering.
AD-GS: Alternating Densification for Sparse-Input 3D Gaussian Splatting
Gurutva Patle (Indian Institute of Science), R. Soundararajan
RestorationGenerationDiffusion modelNeural Radiance FieldContrastive LearningGaussian SplattingImagePoint Cloud
🎯 What it does: Proposes an Alternating Dense Framework (AD-GS), which trains the 3D Gaussian Splatting (3DGS) model under sparse views. It first performs high-density processing to extract details, then low-density processing with clipping and geometric constraints to eliminate floating objects, ultimately achieving high-quality novel view synthesis.
Adaptive Neural Kernels for Gradient-domain Rendering
Matthieu Josse, Adrien Gruson
RestorationConvolutional Neural NetworkDiffusion modelScore-based ModelOptical FlowImage
🎯 What it does: Developed an adaptive neural kernel that dynamically selects the best adjacent pixels to enhance noise suppression in gradient domain rendering
Aerial Path Planning for Urban Geometry and Texture Co-Capture
Weidan Xiong (Shenzhen University), Hui Huang (Shenzhen University)
Autonomous DrivingOptimizationDiffusion modelScore-based ModelNeural Radiance FieldGaussian SplattingOptical FlowImagePoint CloudMesh
🎯 What it does: A drone path planning framework based solely on 2D building outline maps and safe flight altitudes is proposed, which can simultaneously capture high-quality geometric and texture images in a single flight.
AGSwap: Overcoming Category Boundaries in Object Fusion via Adaptive Group Swapping
Zedong Zhang, Jun Li
Image TranslationGenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Propose an adaptive group swapping (AGSwap) method for generating images of creative hybrid objects from cross-category text pairs.
An Adjoint Method for Differentiable Fluid Simulation on Flow Maps
Zhiqi Li (Georgia Institute of Technology), Bo Zhu (Georgia Institute of Technology)
OptimizationComputational EfficiencyOptical FlowImageVideoPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a bidirectional flow graph-based adaptive backward solver, achieving high-precision differentiability in incompressible fluid simulation, and avoiding differentiation of intermediate steps in long-term gradient calculations.
Anchored 4D Gaussian Splatting for Dynamic Novel View Synthesis
Yilong Li, Guoping Wang
GenerationData SynthesisGaussian SplattingImageVideo
🎯 What it does: Propose a 4D Gaussian splatting framework based on anchors for novel view synthesis in dynamic scenes.
AniMaker: Multi-Agent Animated Storytelling with MCTS-Driven Clip Generation
Haoyuan Shi (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
GenerationData SynthesisTransformerLarge Language ModelReinforcement LearningAgentic AIMixture of ExpertsDiffusion modelVideoTextMultimodality
🎯 What it does: Automate the generation of coherent animated videos featuring multiple characters and scenarios from text.
AnimaX: Animating the Inanimate in 3D with Joint Video-Pose Diffusion Models
Zehuan Huang (Beihang University), Lu Sheng (Beihang University)
GenerationData SynthesisPose EstimationTransformerPrompt EngineeringVision-Language-Action ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningVideoTextMeshSequential
🎯 What it does: Achieved text-driven animation generation for 3D meshes with arbitrary skeletal structures through a joint video-pose diffusion model.
Animus3D: Text-driven 3D Animation via Motion Score Distillation
Qi Sun (City University of Hong Kong), Jing Liao (City University of Hong Kong)
GenerationData SynthesisOptimizationKnowledge DistillationTransformerPrompt EngineeringDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderContrastive LearningGaussian SplattingOptical FlowImageVideoTextPoint CloudMeshStochastic Differential Equation
🎯 What it does: This paper proposes the Animus3D framework, which automatically learns motion fields for static 3D assets using text prompts, thereby generating high-quality 3D animation sequences.
AniTex: Light-Geometry Consistent PBR Material Generation for Animatable Objects
Jieting Xu, Yuchi Huo
GenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelImageVideoText
🎯 What it does: Proposes the AniTex generative pipeline, which utilizes diffusion models to synthesize high-quality physically based rendering (PBR) materials for animatable objects based on text prompts.
AnySplat: Feed-forward 3D Gaussian Splatting from Unconstrained Views
Lihan Jiang (University of Science and Technology of China), Bo Dai (University of Hong Kong)
GenerationPose EstimationDepth EstimationComputational EfficiencyKnowledge DistillationTransformerNeural Radiance FieldContrastive LearningGaussian SplattingImage
🎯 What it does: This paper proposes AnySplat, a single forward-pushing 3D Gaussian splatting network that can generate 3D Gaussian primitives and camera poses of a scene directly from arbitrary multi-view images without relying on camera calibration, achieving real-time novel view synthesis.
ART-DECO: Arbitrary Text Guidance for 3D Detailizer Construction
Qimin Chen (Simon Fraser University), Zhiqin Chen (Adobe Research)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelScore-based ModelContrastive LearningImageTextMesh
🎯 What it does: Propose ART-DECO, a refiner capable of instantly refining rough 3D shapes into high-quality detailed models based on arbitrary text prompts.
ARTalk: Speech-Driven 3D Head Animation via Autoregressive Model
Xuangeng Chu (University of Tokyo), Tatsuya Harada (University of Tokyo)
GenerationPose EstimationTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningVideoMeshAudio
🎯 What it does: Proposes ARTalk, a self-recursive framework capable of real-time generating 3D head animations synchronized with speech.
Artifact-Resilient Real-Time Holography
Victor Chu, Felix Heide
GenerationOptimizationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkOptical FlowImagePhysics Related
🎯 What it does: Proposes a new metric for quantifying the anti-artifact performance of holograms and applies it to the optimization of computer-generated holograms (CGH), proposes a simulation method to model the effects of anterior and posterior occlusions of the eye on holographic display, and trains a real-time neural network phase generator to produce artifact-resistant three-dimensional holograms.
ArtiLatent: Realistic Articulated 3D Object Generation via Structured Latents
Honghua Chen (Nanyang Technological University), Xingang Pan (Nanyang Technological University)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelFlow-based ModelAuto EncoderContrastive LearningGaussian SplattingImagePoint CloudMesh
🎯 What it does: Propose ArtiLatent, a generative framework capable of generating artificial articulated 3D objects with high-fidelity geometry, accurate joint motion, and realistic appearance based on a single real image.
ASIA: Adaptive 3D Segmentation using Few Image Annotations
Sai Raj Kishore Perla (Simon Fraser University), Hao Zhang (Simon Fraser University)
SegmentationTransformerPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImagePoint CloudMesh
🎯 What it does: This study proposes an adaptive 3D segmentation framework called ASIA, which can generate part segmentations that conform to user annotations on different shapes.
Assembler: Scalable 3D Part Assembly via Anchor Point Diffusion
Wang Zhao (Tencent ARC Lab), Ying Shan (Tencent ARC Lab)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelAuto EncoderImagePoint CloudMesh
🎯 What it does: Proposes an extensible 3D part assembly framework called Assembler, which can generate complete objects under the condition of a single reference image, using sparse anchor point clouds, and further achieve high-resolution editable 3D models.
Audio Driven Real-Time Facial Animation for Social Telepresence
Jiye Lee (Seoul National University), Shaojie Bai (Meta)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationGraph Neural NetworkTransformerDiffusion modelScore-based ModelGaussian SplattingMultimodalitySequentialAudio
🎯 What it does: Proposes a system capable of generating real-time 3D facial animation driven by speech, balancing high fidelity, generality, and extremely low latency.
Audio Driven Universal Gaussian Head Avatars
Kartik Teotia (Max Planck Institute for Informatics), C. Theobalt
GenerationData SynthesisTransformerDiffusion modelAuto EncoderGaussian SplattingImageVideoMeshAudio
🎯 What it does: Propose a audio-driven universal avatar synthesis framework based on 3D Gaussian Splatting, which can generate realistic 3D head animations from pure audio and support rapid personalization from a very small amount of multi-view or monocular data.
Auto Hair Card Extraction for Smooth Hair with Differentiable Rendering
Zhongtian Zheng (LIGHTSPEED), Kui Wu (LIGHTSPEED)
GenerationOptimizationDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningImageMesh
🎯 What it does: This paper implements a fully automated pipeline that converts high-precision hair strand-based models into a limited number of low-memory hair card models.
AutoBrep : Autoregressive B-Rep Generation with Unified Topology and Geometry
Xiang Xu (Autodesk Research), Peter Meltzer
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelAuto EncoderPoint CloudMesh
🎯 What it does: Proposed and trained a single autoregressive Transformer called AutoBrep, which generates complete B-Reps using a unified discrete token sequence based on face-edge topology, supporting auto-completion.
Automated design of compound lenses with discrete-continuous optimization
Arjun Teh (Carnegie Mellon University), Matthew O'Toole (Carnegie Mellon University)
OptimizationMixture of ExpertsDiffusion modelImagePhysics RelatedStochastic Differential Equation
🎯 What it does: Developed an automated compound lens design method that combines MCMC sampling with gradient optimization, capable of simultaneously optimizing continuous parameters (curvature, thickness, spacing, etc.) and discrete topology (number and type of elements) of the lens.
Automatic Sampling for Discontinuities in Differentiable Shaders
Yash Belhe, Tzu-Mao Li
OptimizationExplainability and InterpretabilityComputational Efficiency
🎯 What it does: A new method for computing the integral gradient of discontinuous functions is proposed, which can directly perform differentiation on shader programs;
Autoregressive Generation of Static and Growing Trees
Hanxiao Wang (CASIA), Peter Wonka (KAUST)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelGenerative Adversarial NetworkContrastive LearningImagePoint CloudGraphAgriculture Related
🎯 What it does: Proposed a self-attention tree structure generation framework based on the Hourglass Transformer, capable of generating static trees, growing trees, and completing tasks such as image-to-tree and point cloud-to-tree under various conditional generation scenarios.
AutoSketch: VLM-assisted Style-Aware Vector Sketch Completion
Hsiao-Yuan Chin (National Taiwan University), Bing-Yu Chen (National Taiwan University)
GenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose AutoSketch, a style-aware vector sketch completion method assisted by visual language models.
BlobCtrl: Taming Controllable Blob for Element-level Image Editing
Yaowei Li (Peking University), Yuexian Zou (Peking University)
Image HarmonizationRestorationGenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelContrastive LearningGaussian SplattingImageBenchmark
🎯 What it does: Propose the BlobCtrl framework, which utilizes probabilistic Blobs as visual primitives to achieve element-level image editing, supporting operations such as adding, translating, scaling, removing, replacing objects, and their combinations.
Bokeh Diffusion: Defocus Blur Control in Text-to-Image Diffusion Models
Armando Fortes (Nanyang Technological University), Xingang Pan (Nanyang Technological University)
RestorationGenerationDepth EstimationTransformerVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the Bokeh Diffusion framework to achieve precise and scene-consistent bokeh blur control in text-to-image diffusion models.
BrepGPT: Autoregressive B-rep Generation with Voronoi Half-Patch
Pu Li (Institute of Automation, Chinese Academy of Sciences), Dongming Yan
GenerationData SynthesisGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringDiffusion modelAuto EncoderGenerative Adversarial NetworkMeshGraphRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented BrepGPT, a self-attention based B-rep generation framework based on Transformer, which utilizes Voronoi Half-Patch to uniformly encode B-rep, achieving complete CAD generation tasks from unconditional to multi-modal conditional generation, automatic completion, and interpolation.
BSP-OT: Sparse transport plans between discrete measures in loglinear time
Baptiste Genest, David Coeurjolly
OptimizationComputational EfficiencyImagePoint CloudMesh
🎯 What it does: Propose a BSP-OT algorithm based on a variant of quicksort, which solves the sparse optimal transport problem between two discrete measures (including uniformly sized same and non-uniformly sized different measures) in logarithmic linear time.
CamCloneMaster: Enabling Reference-based Camera Control for Video Generation
Yawen Luo (Chinese University of Hong Kong), Kun Gai (Kuaishou Technology)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkVideo
🎯 What it does: Propose CamCloneMaster, a general video generation framework that directly clones camera motion from reference videos without requiring camera parameters or fine-tuning during inference.
CameraVDP: Perceptual Display Assessment with Uncertainty Estimation via Camera and Visual Difference Prediction
Yancheng Cai (University of Cambridge), Rafal Mantiuk
Anomaly DetectionOptimizationExplainability and InterpretabilityComputational EfficiencyDiffusion modelScore-based ModelContrastive LearningGaussian SplattingOptical FlowImageVideoPoint Cloud
🎯 What it does: Propose a complete display evaluation system called CameraVDP, which combines a calibrated consumer-grade camera with a visual difference predictor. It can achieve precise photometric, color, and geometric correction on the original images captured by the camera, and assess human-perceptible display defects through the visual difference predictor, while providing uncertainty estimation.
CamPVG: Camera-Controlled Panoramic Video Generation with Epipolar-Aware Diffusion
Chenhao Ji (Tongji University), Cairong Zhao (Tongji University)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelAuto EncoderOptical FlowImageVideo
🎯 What it does: Proposed the CamPVG framework, achieving panoramic video generation based on precise camera trajectories.
Can Any Model Be Fabricated? Inverse Operation Based Planning for Hybrid Additive-Subtractive Manufacturing
Yongxue Chen (University of Manchester), Charlie C. L. Wang
OptimizationComputational EfficiencyRobotic IntelligenceDiffusion modelAuto EncoderContrastive LearningPoint CloudMeshBenchmark
🎯 What it does: Propose an inverse operation (accretion and erosion) planning algorithm that can generate executable hybrid additive-subtractive manufacturing sequences for any complex geometry, ensuring that the final part is completely consistent with the target geometry.
Capturing Non-Linear Human Perspective in Line Drawings
Jinfan Yang (University of British Columbia), Alla Sheffer (University of British Columbia)
Image TranslationGenerationConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageMesh
🎯 What it does: Learn and model the nonlinear deviations of artists in perspective within their sketches to make the rendered sketches more resemble hand-drawn ones.
CFC: Simulating Character-Fluid Coupling using a Two-Level World Model
Zhiyang Dou, Taku Komura
Diffusion modelScore-based ModelWorld ModelPhysics Related
🎯 What it does: Proposed a dual-layer world model for simulating the interaction between characters and fluids through bidirectional coupling;
Chapper: Carvable Hull-and-Pack for Subtractive Manufacturing
Zhenmin Zhang, Haisen Zhao
OptimizationDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingPoint CloudMesh
🎯 What it does: Proposes a complete method for solving the carvable hull-and-pack problem, integrating high-density packing of irregular objects, detachable sequence planning, carving hull generation, and cutting path planning.
CHARM: Control-point-based 3D Anime Hairstyle Auto-Regressive Modeling
Yuze He (Tsinghua University), Wei Yang (Tencent AIPD)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningImagePoint CloudMesh
🎯 What it does: Propose the CHARM framework, which includes a reversible five-parameter control point parameterization and an autoregressive Transformer, enabling the generation of high-quality, editable 3D anime hairstyles under point cloud or image conditions.
Chord: Chain of Rendering Decomposition for PBR Material Estimation from Generated Texture Images
Zhi Ying (Ubisoft La Forge), Maoyuan Xu (Ubisoft La Forge)
GenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelScore-based ModelNeural Radiance FieldImage
🎯 What it does: Propose a two-stage generation-estimation framework, first generating stitchable texture RGB images using a fine-tuned SDXL, and then predicting SVBRDF channels through chained decomposition.
CityGo: Lightweight Urban Modeling and Rendering with Proxy Buildings and Residual Gaussians
Weihang Liu (ShanghaiTech University), Yingliang Zhang (ShanghaiTech University)
Autonomous DrivingComputational EfficiencyNeural Radiance FieldContrastive LearningGaussian SplattingOptical FlowImagePoint CloudMesh
🎯 What it does: This paper proposes the CityGo framework, which utilizes texturized proxy building grids, residual Gaussians, and surrounding environment Gaussians to achieve lightweight, high-quality rendering of large-scale urban scenes;
Closed-form Cauchy Coordinates and Their Derivatives for 2D High-order Cages
Shibo Liu, Xiao-Ming Fu
🎯 What it does: Proposes closed-form Cauchy coordinates and their derivatives applicable to two-dimensional closed high-order input cages, and can convert input polynomial curves into output curves of any desired polynomial order.
Closed-Form Construction of Voronoi Diagrams with Star-Shaped Metrics
Haoyang Zhou, Bernhard Thomaszewski
Optimization
🎯 What it does: Proposes a closed-form, differentiable star-shaped metric Voronoi diagram construction method that can simultaneously optimize the site positions and metric parameters.
Clustered Error Correction with Grouped 4D Gaussian Splatting
Taeho Kang (Seoul National University), Youngki Lee (Seoul National University)
RestorationGenerationOptimizationDiffusion modelGaussian SplattingOptical FlowVideoPoint CloudBenchmark
🎯 What it does: An improved method based on clustering error correction and group dynamic transformation is proposed for 4D Gaussian point cloud rendering. It uses elliptical error clustering to locate missing or occluded errors and dynamically adds new Gaussian points through back-projection or foreground segmentation. At the same time, group-shared temporal transformations are introduced to maintain consistent correspondence between dynamic objects and point clouds.
Compact shape representation utilizing local surface similarities
A. Garifullin, V. Frolov
CompressionAuto EncoderMesh
🎯 What it does: Propose a compact shape representation method called SCom Tree based on the Signed Distance Field for 3D model storage and rendering.
ComposeMe: Attribute-Specific Image Prompts for Controllable Human Image Generation
G. Qian (Snap Inc.), Kfir Aberman (Snap Inc.)
GenerationTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: This paper proposes a model called ComposeMe based on attribute-specific image prompts, for attribute-level controllable generation of human portraits under text prompts.
Computational Design of Shape-Aware Sieves
David Cha, Oded Stein
OptimizationDiffusion modelMesh
🎯 What it does: Construct and solve a mathematical framework that formulates the dice design problem as a two-player game, utilizing differential rendering and particle swarm optimization to achieve manufacturable dice designs that satisfy given shape set allow/block constraints.
Computational Modeling and Design of Capacitive Stretch Sensors
Arvi Gjoka, D. Panozzo
OptimizationDiffusion modelScore-based ModelPhysics Related
🎯 What it does: Proposed an accurate and differentiable capacitive stretch sensor simulator that supports inverse design and significantly reduces calibration requirements
Consecutive Frame Extrapolation with Predictive Sparse Shading
Zhizhen Wu, Yuchi Huo
RestorationGenerationRecurrent Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderOptical FlowVideo
🎯 What it does: A method is proposed that sparsely colors key regions and reuses pixels from low-variation regions to achieve frame extrapolation in neural networks, and a Predictive Error-Flow-eXtrapolation Network (EFXNet) architecture is designed, which can predict extrapolation errors, estimate optical flow, and complete frame extrapolation in one go.
ConsistEdit: Highly Consistent and Precise Training-free Visual Editing
Zixin Yin (Hong Kong University of Science and Technology), Xili Dai (Hong Kong University of Science and Technology)
Image TranslationImage HarmonizationRestorationGenerationTransformerPrompt EngineeringDiffusion modelScore-based ModelImageVideoTextStochastic Differential Equation
🎯 What it does: Propose ConsistEdit, a training-free attention control method for the MM-DiT architecture, which can achieve multi-round, multi-region, and fine-grained image and video editing while maintaining the structural consistency of the source image and the non-edited regions.
ConsiStyle: Style Diversity in Training-Free Consistent T2I Generation
Yohai Mazuz (Tel Aviv University), Lior Wolf (Tel Aviv University)
GenerationTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: Propose an untrained Consistyle method to achieve the unification of character consistency and style diversity in text-to-image (T2I) generation.
Constructing Diffusion Avatar with Learnable Embeddings
Xuan Gao, Juyong Zhang
GenerationData SynthesisTransformerDiffusion modelScore-based ModelGaussian SplattingImageMesh
🎯 What it does: Built a diffusion avatar model that learns through controllable signals and adapts to synthetic data.
Context as Memory: Scene-Consistent Interactive Long Video Generation with Memory Retrieval
Jiwen Yu (University of Hong Kong), Xihui Liu (University of Hong Kong)
GenerationRetrievalTransformerVision-Language-Action ModelDiffusion modelAuto EncoderOptical FlowVideoRetrieval-Augmented Generation
🎯 What it does: Propose the Context-as-Memory framework, which directly uses historically generated frames as memory, achieving long video scene consistency by retrieving relevant frames.
Control Operators for Interactive Character Animation
Ruiyu Gou, Daniel Holden
GenerationPose EstimationReinforcement Learning from Human FeedbackTransformerPrompt EngineeringDiffusion modelFlow-based ModelImageVideo
🎯 What it does: Proposed the Control Operators framework, allowing non-technical designers to specify interactive character controllers through semantic control operations, and verified its feasibility on two advanced models.
CrossGen: Learning and Generating Cross Fields for Quad Meshing
Qiujie Dong (University of Hong Kong), Wenping Wang (Texas A&M University)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelScore-based ModelAuto EncoderPoint CloudMesh
🎯 What it does: Propose an end-to-end framework called CrossGen based on a sparse convolutional autoencoder, which can directly predict high-quality cross fields from point cloud surfaces and generate corresponding SDFs for fast quadrilateral meshing.
Curvature Enthusiasm: Correspondence-Free Interpolation and Matching of Articulated 3D Shapes using Compressed Normal Cycles
Adam Hartshorne, Neill D. F. Campbell
Pose EstimationOptimizationComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerDiffusion modelAuto EncoderPoint CloudMeshStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed an unsupervised framework for physically plausible 3D motion shape interpolation and dense correspondence estimation.
Curve-Based Slicer for Multi-Axis DLP 3D Printing
Chengkai Dai (Centre for Perceptual and Interactive Intelligence (CPII) Limited), Charlie C. L. Wang
OptimizationRobotic IntelligenceDiffusion modelScore-based ModelMesh
🎯 What it does: This paper proposes a curve-based slicer that generates variable layers for multi-axis DLP printing by optimizing the curve trajectory, achieving support structure minimization, elimination of overhanging points, and improvement of surface quality.
Cut2Next: Generating Next Shot via In-Context Tuning
Jingwen He (Chinese University of Hong Kong), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelVideoTextMultimodalityBenchmark
🎯 What it does: Propose a 'Next Shot Generation' (NSG) technique that automatically generates the next shot complying with director's editing rules based on the given previous shot;
DeMapGS: Simultaneous Mesh Deformation and Surface Attribute Mapping via Gaussian Splatting
Shuyi Zhou (University of Tokyo), Takeshi Oishi (University of Tokyo)
GenerationOptimizationComputational EfficiencyGraph Neural NetworkDiffusion modelGaussian SplattingOptical FlowImagePoint CloudMesh
🎯 What it does: Propose the DeMapGS framework, which jointly optimizes deformable meshes and surface-attached Gaussian Splat, and extracts high-quality surface attribute maps such as diffuse, normal, and displacement from the optimization results.
Design for Descent: What Makes a Shape Grammar Easy to Optimize?
Milin Kodnongbua, Adriana Schulz
OptimizationMeshStochastic Differential Equation
🎯 What it does: Proposed and implemented the Stochastic Rewrite Descent (SRD) algorithm, combining structured rewriting with continuous parameter updates, and verified its optimization performance on three shape grammars.
Designing and Fabricating Color BRDFs with Differentiable Wave Optics
Yixin Zeng, Min H. Kim
Diffusion modelOptical FlowPhysics Related
🎯 What it does: Designed and fabricated various full-color BRDFs using a differentiable wave optics framework.
Designing with Tension: Nearly-Developable Patch Layouts
Anna Maria Eggler, Bernd Bickel
Mesh
🎯 What it does: Automatically approximate free-form surfaces using approximately developable patches to form tension structures while ensuring manufacturability and structural feasibility
Detail Enhanced Gaussian Splatting for Large-Scale Volumetric Capture
Julien Philip (Eyeline Labs), Paul E. Debevec
RestorationGenerationTransformerDiffusion modelGaussian SplattingOptical FlowImageVideo
🎯 What it does: Proposed a complete pipeline for large-scale, multi-actor, high-resolution 4D voxel capture and rendering, combining Scene Rig and Face Rig hardware setups, and using dynamic Gaussian splatting (Poly4DGS) to reconstruct dynamic performances, further enhancing facial close-ups with a detail-enhancing diffusion model;
DiffCamera: Arbitrary Refocusing on Images
Yiyang Wang (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
Image TranslationRestorationDepth EstimationTransformerDiffusion modelScore-based ModelRectified FlowAuto EncoderContrastive LearningImageBenchmark
🎯 What it does: Propose the DiffCamera model, which can arbitrarily specify the focus and blur degree on existing images for refocusing, while maintaining scene consistency.
Differentiable Light Transport with Gaussian Surfels via Adapted Radiosity for Efficient Relighting and Geometry Reconstruction
Kaiwen Jiang (University of California, San Diego), Ravi Ramamoorthi (University of California, San Diego)
RestorationGenerationOptimizationComputational EfficiencyDiffusion modelNeural Radiance FieldGaussian SplattingImagePoint Cloud
🎯 What it does: Propose a differentiable illumination transport framework based on Gaussian surfaces for high-quality relighting and geometry reconstruction.
DiffTex: Differentiable Texturing for Architectural Proxy Models
Weidan Xiong (Shenzhen University), Hui Huang (Shenzhen University)
GenerationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingOptical FlowImagePoint CloudMesh
🎯 What it does: This paper proposes an end-to-end method called DiffTex based on differentiable rendering, which generates high-quality, view-consistent, and photorealistic texture maps for building proxy models (simplified polygons) from a set of unordered and calibrated RGB camera images.
Diffusion-Guided Relighting for Single-Image SVBRDF Estimation
Youxin Xing, Beibei Wang
Image TranslationRestorationConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Generate material images under various illumination conditions using diffusion models, and use a shuffle-based background consistency module and a lightweight specular prior encoder to help multi-image SVBRDF estimators recover reflectance maps without specular highlights.
Discovering Folding Lines for Surface Compression
Toshiki Aoki, Mina Konaković Luković
CompressionOptimizationMesh
🎯 What it does: A new origami crease pattern computation method is proposed, which uses an adapted material point method (MPM) to simulate the compression of the target surface and obtain an initial folded configuration. Then, a folding line extraction (FLE) algorithm is used to optimize the smooth surface, resulting in a smooth curved crease line that achieves the target compression while minimizing deformation and stretching outside the crease line.
DreamID: High-Fidelity and Fast diffusion-based Face Swapping via Triplet ID Group Learning
Fulong Ye (ByteDance), Xinglong Wu (ByteDance)
Image TranslationRestorationGenerationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningImage
🎯 What it does: Propose a facial replacement method called DreamID based on diffusion models, which utilizes Triplet ID Group data to achieve explicit supervision, enabling high similarity, attribute preservation, and efficient inference in facial replacement.
DreamO: A Unified Framework for Image Customization
Chong Mou (ByteDance), Xinglong Wu (ByteDance)
Image TranslationGenerationData SynthesisTransformerPrompt EngineeringDiffusion modelFlow-based ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the DreamO unified framework, which supports image customization under various conditions (identity, subject, style, try-on, etc.).
DSCombiner: Double Shrinkage for Combining Biased and Unbiased Monte Carlo Renderings
Chenxi Zhou, Jie Guo
OptimizationDiffusion modelNeural Radiance FieldContrastive LearningImage
🎯 What it does: Proposed DSCombiner, a dual shrinkage estimator that flexibly combines unbiased and biased estimators in image space into a single estimation process.
DvD: Unleashing a Generative Paradigm for Document Dewarping via Coordinates-based Diffusion Model
Weiguang Zhang (Xi'an Jiaotong-Liverpool University), Qiufeng Wang (Xi'an Jiaotong-Liverpool University)
RestorationTransformerDiffusion modelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Propose DvD, a document dewarping method based on coordinate-level diffusion models, which generates dewarping mappings instead of directly generating images, and construct a large-scale multi-domain evaluation benchmark called AnyPhotoDoc6300.
DynVFX: Augmenting Real Videos with Dynamic Content
Danah Yatim (Weizmann Institute of Science), Tali Dekel (Weizmann Institute of Science)
Image HarmonizationGenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelVideoTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a zero-shot, fully automatic text-driven method called DynVFX, which can naturally insert and synchronize dynamic objects or effects in real videos based on short text instructions without requiring additional masks or reference materials.
EBREnv: SVBRDF Estimation in Uncontrolled Environment Lighting via Exemplar-Based Representation
Li Wang, Jiawan Zhang
RestorationConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Propose an environment light exemplar representation based on the sample surface domain, used to estimate spatially varying bidirectional reflectance distribution functions (SVBRDF) in uncontrolled environments.
Echo: Enhancing Conversational Behavior Generation via Hierarchical Semantic Comprehension with Large Language Models
Haiwei Xue, Zhiyong Wu
GenerationTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelTextMultimodalityAudio
🎯 What it does: Propose a system named Echo, which deeply understands dialogue semantics through large language models and combines audio-action features to enhance the generation of conversational behaviors.
Editable Physically-based Reflections in Raytraced Gaussian Radiance Fields
Yohan Poirier-Ginter (Université Laval), G. Drettakis
OptimizationComputational EfficiencyDiffusion modelNeural Radiance FieldGaussian SplattingOptical FlowImagePoint CloudPhysics Related
🎯 What it does: The paper proposes a lighting field reconstruction method based on 3D Gaussian splatting, which can achieve real-time and editable reconstruction of reflections in the scene by separately optimizing diffuse and specular reflections and using path tracing to compute physically accurate multiple bounce specular effects.
Efficient and Scalable Spatial Regularization of Optimal Transport
Lucas Brifault, Mathieu Desbrun
Optimization
🎯 What it does: Propose a new spatial regularization method to solve the optimal transport problem, utilizing forward and backward mean mapping for regularization.
Efficient Object Reconstruction with Differentiable Area Light Shading
Yaoan Gao, Weiwei Xu
OptimizationComputational EfficiencyDiffusion modelNeural Radiance FieldGaussian SplattingPoint CloudMesh
🎯 What it does: Proposed an inverse rendering method using active area illumination to improve the accuracy of material attribute estimation in 3D object reconstruction.
EGG-Fusion: Efficient 3D Reconstruction with Geometry-aware Gaussian Surfel on the Fly
Xiaokun Pan (Zhejiang University), Guofeng Zhang (Zhejiang University)
Depth EstimationAutonomous DrivingOptimizationComputational EfficiencyNeural Radiance FieldGaussian SplattingSimultaneous Localization and MappingOptical FlowImageVideoPoint Cloud
🎯 What it does: Proposed the EGG-Fusion real-time RGB-D SLAM system, which utilizes geometry-aware Gaussian point clouds to achieve high-precision 3D reconstruction.
ELAD: Blind Face Restoration using Expectation-based Likelihood Approximation and Diffusion Prior
Sean Man, Michael Elad
RestorationTransformerDiffusion modelScore-based ModelAuto EncoderImage
🎯 What it does: Propose ELAD, a plug-and-play method for blind face restoration, which explicitly models the likelihood and obtains its first and second moments through an attenuation estimator, thereby enabling Bayesian inference.
Environment-aware Motion Matching
J. L. Pontón, N. Pelechano
Pose EstimationAutonomous DrivingComputational EfficiencyReinforcement Learning from Human FeedbackOptical FlowVideoPoint CloudRetrieval-Augmented Generation
🎯 What it does: Proposed an Environment-aware Motion Matching system that can automatically adjust full-body posture and root trajectory in real-time interactive environments based on dynamic obstacles and other characters.
Evaluating and Sampling Glinty NDFs in Constant Time
Pauli Kemppinen, T. Boubekeur
GenerationComputational EfficiencyDiffusion modelScore-based ModelGaussian SplattingMesh
🎯 What it does: Proposes a rendering method that represents segmented geometry as a 4D point process and uses an implicit multi-scale grid to quickly locate per-pixel glints, aiming to efficiently generate glint effects.
Example-Based Feature Painting on Textures
Andrei-Timotei Ardelean (Friedrich-Alexander-Universität Erlangen-Nürnberg), Tim Weyrich (Friedrich-Alexander-Universität Erlangen-Nürnberg)
RestorationGenerationAnomaly DetectionConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper proposes a complete workflow that automatically extracts and clusters defect features from a small number of unlabeled textured images, trains conditional diffusion models, and supports interactive drawing, editing, and feature transfer on textured images of arbitrary size.
FairyGen: Storied Cartoon Video from a Single Child-Drawn Character
Jiayi Zheng (Great Bay University), Xiaodong Cun (Great Bay University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Generate multi-shot cartoon animation videos featuring a single child's hand-drawn character, while maintaining consistency in the character's style and motion
Fast & Stable Control of Coupled Solid-Fluid Dynamic Systems
Jie Chen, Bo-Ning Ren
Reinforcement Learning
🎯 What it does: Proposed a reinforcement learning control algorithm for coupled solid-fluid dynamics systems, which can achieve more stable and robust control;
Fast Galerkin Multigrid Method for Unstructured Meshes
Jia-Ming Lu, Shi-min Hu
OptimizationComputational EfficiencyMeshPhysics Related
🎯 What it does: A multigrid solver framework for unstructured grids was developed, significantly improving the efficiency of physical simulations.
Fine-Grained Spatially Varying Material Selection in Images
Julia Guerrero-Viu, V. Deschaintre
SegmentationConvolutional Neural NetworkTransformerDiffusion modelContrastive LearningImage
🎯 What it does: Propose a material selection method based on multi-resolution ViT, supporting two-level selection at the texture layer and sub-texture layer, and combining multi-query sampling to enhance training stability;
Fire-X: Extinguishing Fire with Stoichiometric Heat Release
Helge Wrede, Soeren Pirk
Diffusion modelGaussian SplattingPhysics Related
🎯 What it does: A flame simulation framework across solids, liquids, and gases was developed, utilizing multi-component thermodynamics and reaction transport to achieve precise simulation of the combustion process, supporting various scenarios such as jet fires, water spray fire suppression, fuel evaporation and starvation, and achieving interactive heat sources, fire detectors, and flame rendering ranging from blue to orange, from laminar to turbulent flows;
Force-Dual Modes: Subspace Design from Stochastic Forces
Otman Benchekroun (University of Toronto), Philip A. Etter
OptimizationExplainability and InterpretabilityComputational EfficiencyContrastive LearningGaussian SplattingPoint CloudMeshTabularTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper proposes a Force-Dual mode based on constructing a subspace using a random force distribution, addressing the lack of interpretability and generality in the subspace design of traditional ROM.
Fovea Stacking: Imaging with Dynamic Localized Aberration Correction
Shi Mao (King Abdullah University of Science and Technology), Wolfgang Heidrich (King Abdullah University of Science and Technology)
RestorationOptimizationTransformerDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowImageMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Propose a 'retinal stacking' imaging scheme using a variable phase plate, achieving high-quality full-field images by locally correcting aberrations in a simplified optical system.
Frame-Free Representation of Polarized Light for Resolving Stokes Vector Singularities
Shinyoung Yi, Min H. Kim
Physics Related
🎯 What it does: A polarization light Stokes vector representation without a local coordinate system was studied, and its superiority was verified in two application scenarios.
FreeArt3D: Training-Free Articulated Object Generation using 3D Diffusion
Chuhao Chen (University of California San Diego), Minghua Liu (Hillbot Inc)
GenerationData SynthesisPose EstimationTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkImagePoint CloudMesh
🎯 What it does: Proposes FreeArt3D, a training-free 3D articulated object generation framework that can directly recover high-quality textured multi-jointed 3D models from a few images of different poses.
FreeMusco: Motion-Free Learning of Latent Control for Morphology-Adaptive Locomotion in Musculoskeletal Characters
Minkwan Kim (Hanyang University), Yoonsang Lee (Hanyang University)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelAuto EncoderContrastive LearningWorld ModelTime SeriesSequential
🎯 What it does: This paper proposes a 'FreeMusco' framework that does not rely at all on motion capture data. It can jointly learn the latent gait representation and low-level control strategies of musculoskeletal characters, achieving energy-aware, morphology-adaptive diverse gaits, and can further be used for high-level tasks such as target navigation and path following.
From Rigging to Waving: 3D-Guided Diffusion for Natural Animation of Hand-Drawn Characters
Jie Zhou (City University of Hong Kong), Hongbo Fu (Hong Kong University of Science and Technology)
GenerationPose EstimationDomain AdaptationTransformerSupervised Fine-TuningDiffusion modelOptical FlowImageVideoMesh
🎯 What it does: Propose a hybrid system that first generates a coarse animation with geometric consistency using skeletal animation, and then refines the texture and secondary motion using a domain-adapted video diffusion model, enabling natural 3D animation for hand-drawn characters.