arXivSub Start free trial

SIGGRAPH Asia 2024 Papers with AI Summaries

ACM SIGGRAPH Asia (Transactions on Graphics) · 265 papers

360-degree Human Video Generation with 4D Diffusion Transformer

Ruizhi Shao (Tsinghua University), Yebin Liu (Tsinghua University)

GenerationData SynthesisConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelImageVideoMultimodalityPoint CloudMesh

🎯 What it does: Propose Human4DiT, a 360° human video generation framework based on 4D diffusion Transformer, which can generate high-quality, spatiotemporally consistent multi-view and 360° dynamic videos from a single reference image.

3D Gaussian Ray Tracing: Fast Tracing of Particle Scenes

Nicolas Moenne-Loccoz, Zan Gojcic

Computational EfficiencyNeural Radiance FieldGaussian SplattingOptical FlowPoint CloudPhysics Related

🎯 What it does: Developed a 3D Gaussian particle field renderer based on GPU ray tracing, using BVH structures and per-pixel ray casting

3D Reconstruction with Fast Dipole Sums

Hanyu Chen (Carnegie Mellon University), Ioannis Gkioulekas (Carnegie Mellon University)

OptimizationComputational EfficiencyRepresentation LearningNeural Radiance FieldGaussian SplattingPoint CloudMesh

🎯 What it does: Using dense point clouds obtained from structured light cameras, a regularized dipole sum representation of geometry and radiance fields is constructed. Inverse rendering optimization is achieved by combining ray tracing with Barnes-Hut fast summation, enabling high-quality multi-view 3D reconstruction.

3DGSR: Implicit Surface Reconstruction with 3D Gaussian Splatting

Xiaoyang Lyu (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)

GenerationComputational EfficiencyDiffusion modelScore-based ModelNeural Radiance FieldContrastive LearningGaussian SplattingPoint CloudMesh

🎯 What it does: Propose the 3DGSR method, combining implicit SDF with 3D Gaussian Splatting to achieve a unified framework for high-quality surface reconstruction and real-time rendering.

A class of new tuned primal subdivision schemes with high-quality limit surface in extraordinary regions

Xu Wang, Weiyin Ma

Diffusion modelScore-based ModelMesh

🎯 What it does: Proposes a unified optimization framework for constructing TPS (Tuned Primal Subdivision) schemes for unstructured quadrilateral meshes with arbitrary topology, achieving high-quality surfaces through local refinement, topological splitting, and multiple local smoothing.

A Cubic Barrier with Elasticity-Inclusive Dynamic Stiffness

R. Ando

Physics Related

🎯 What it does: A new cubic barrier function is proposed, combined with the dynamic stiffness of elastic containment, for achieving penetration-free contact solving and strain limitations.

A Dynamic By-example BTF Synthesis Scheme

Zilin Xu (University of California, Santa Barbara), Lingqi Yan

GenerationData SynthesisTransformerDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkImageMesh

🎯 What it does: This paper proposes a dynamic example-based BTF synthesis scheme that can generate an infinitely large, non-repeating complete 6D BTF in real-time during rendering, using only small-sized BTF samples.

A Flexible Mold for Facade Panel Fabrication

Florian Rist, D. Michels

OptimizationDiffusion modelScore-based ModelSimultaneous Localization and MappingOptical FlowMeshPhysics Related

🎯 What it does: Propose a machine that utilizes a deflatable membrane as a flexible mold to efficiently manufacture building panels, along with an algorithm that optimizes the position of customizable boundary elements to match the finished panel with the input geometry;

A Generalized Ray Formulation For Wave-Optical Light Transport

Shlomi Steinberg, Matt Pharr

Computational EfficiencyNeural Radiance FieldOptical FlowPhysics Related

🎯 What it does: Propose a new wave optics light transmission model based on general rays, using the statistical characteristics of aggregated light waves by sensors to represent the light field;

A Mesh-based Simulation Framework using Automatic Code Generation

Philipp Herholz, L. Kavan

Symbolic ComputationOptimizationComputational EfficiencyMesh

🎯 What it does: Proposed a grid computing symbolic code generation framework for parallel devices

A Plentoptic 3D Vision System

Agastya Kalra, Kartik Venkataraman

Depth EstimationRobotic IntelligenceNeural Radiance FieldContrastive LearningGaussian SplattingOptical FlowImageMultimodalityPoint Cloud

🎯 What it does: Propose a multi-camera, multi-modal visual system for industrial robot applications, used to generate high-quality 3D point clouds and improve the completeness of collision avoidance while reducing hallucinations.

A Progressive Embedding Approach to Bijective Tetrahedral Maps driven by Cluster Mesh Topology

V. Z. Nigolian, D. Bommes

Mesh

🎯 What it does: Propose an algorithm that maps spherical topology tetrahedral meshes to star-shaped domains while ensuring the bijectivity of the mapping.

A Simple Approach to Differentiable Rendering of SDFs

Zichen Wang (Cornell University), Steve Marschner (Cornell University)

OptimizationComputational EfficiencyRepresentation LearningSupervised Fine-TuningDiffusion modelScore-based ModelRectified FlowNeural Radiance FieldAuto EncoderImagePoint CloudMesh

🎯 What it does: Propose a differentiable rendering method that utilizes SDF, employing boundary relaxation to approximate visibility discontinuities as thin bands for sampling;

A Statistical Approach to Monte Carlo Denoising

Hiroyuki Sakai, M. Wimmer

RestorationImage

🎯 What it does: A statistical method is proposed to denoise noise images generated by Monte Carlo rendering, analyzing the distribution of random variables for each pixel and designing a fast pairwise statistical test to screen neighboring pixels.

A Time-Dependent Inclusion-Based Method for Continuous Collision Detection between Parametric Surfaces

Xuwen Chen, Bao Chen

OptimizationComputational EfficiencyOptical FlowMesh

🎯 What it does: A new time-varying inclusion function-based continuous collision detection framework is proposed, eliminating the need for time subdivision and significantly improving performance;

Accelerate Neural Subspace-Based Reduced-Order Solver of Deformable Simulation by Lipschitz Optimization

Aoran Lyu (South China University of Technology), Guoxin Fang (Chinese University of Hong Kong)

OptimizationComputational EfficiencyDiffusion modelScore-based ModelAuto EncoderContrastive LearningPoint CloudMeshTabularTime SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: A neural subspace reduced-order solver based on Lipschitz optimization is studied, aiming to accelerate physics simulation of deformable objects by improving the landscape of the objective function within the subspace.

Actuators A La Mode: Modal Actuations for Soft Body Locomotion

Otman Benchekroun (University of Toronto), Victor Zordan (Roblox Research)

OptimizationRobotic IntelligenceDiffusion modelContrastive LearningOptical FlowMesh

🎯 What it does: Propose a motion control framework for arbitrary soft-body characters, which constructs a low-dimensional spatiotemporal excitation subspace using the characters' own vibration modes, and couples it with a simplified skin-based soft-body simulation. Feasible crawling, running, and jumping motions are obtained through CMA-ES optimization.

AdR-Gaussian: Accelerating Gaussian Splatting with Adaptive Radius

Xinzhe Wang (Shanghai Jiao Tong University), Lizhuang Ma (Shanghai Jiao Tong University)

Computational EfficiencyGaussian SplattingPoint Cloud

🎯 What it does: Propose AdR-Gaussian to accelerate the rendering of 3D Gaussian splatting.

Alignable Lamella Gridshells

Davide Pellis

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

🎯 What it does: Studied the inverse design problem of foldable, flat strip-like aligned lamina emergent shells, and explained their geometric constraints based on differential geometry and the Cartan moving frame theory.

All you need is rotation: Construction of developable strips

Takashi Maekawa, Felix Scholz

Physics Related

🎯 What it does: A new method is proposed to generate developable strips along spatial curves by using rotation angles between the Frenet frame and the Darboux frame;

An Eulerian Vortex Method on Flow Maps

Sinan Wang (Georgia Institute of Technology), Bo Zhu (Georgia Institute of Technology)

Diffusion modelContrastive LearningGaussian SplattingOptical FlowImageVideoPhysics RelatedStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: This paper proposes an Euler vortex method based on a flow graph, which utilizes vorticity as line elements for BFECC advancement with error compensation on a bidirectional flow graph, and achieves incompressible fluid simulation without viscosity or pressure through a new velocity-vorticity coupled Poisson solver.

An Impulse Ghost Fluid Method for Simulating Two-Phase Flows

Yuchen Sun, Bo Zhu

Optical FlowPhysics Related

🎯 What it does: A two-phase interface fluid model based on impulse variables is proposed, and the bidirectional flow graph theory is utilized to improve the accuracy of vortex and interface transport. The impulse ghost fluid method is employed to address the interface dynamics of two-phase incompressible fluids. A path integral formula utilizing spatiotemporal buffering is developed, converting history-dependent jump conditions into geometry-dependent jump conditions. The effectiveness of this method in simulating interface-vortex interactions (such as interface vortices, vortex ring reflections, and jumping bubble rings) is demonstrated.

Anim-Director: A Large Multimodal Model Powered Agent for Controllable Animation Video Generation

Yunxin Li (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)

GenerationData SynthesisTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelImageVideoTextMultimodalityRetrieval-Augmented Generation

🎯 What it does: Developed Anim-Director, an automated animation generation agent based on a large-scale multimodal model, capable of automatically generating coherent and plot-complete long animation videos from short narratives.

Appearance Modeling of Iridescent Feathers with Diverse Nanostructures

Yunchen Yu, Steve Marschner

GenerationData SynthesisDiffusion modelAuto EncoderGenerative Adversarial NetworkImagePhysics Related

🎯 What it does: Taking bird feathers as a case study, we constructed appearance models for various structural color feathers, proposed an approximate wave simulation method that utilizes quasi-periodic structures and effectively simulates natural structural irregularities, and condensed the simulation results into a BRDF distribution based on noise functions to achieve spatially varying iridescent appearances.

Approximation by Meshes with Spherical Faces

A. Ramos-Cisneros, Christian Müller

OptimizationMesh

🎯 What it does: This paper proposes a method to approximate a given surface using a mesh constructed from spherical patches and circular edges, taking into account visual appearance, approximation quality, and the geometric properties of cross circles between adjacent faces, while ensuring a torque-free supporting structure composed of planar circular edges.

AR-DAVID: Augmented Reality Display Artifact Video Dataset

Alexandre Chapiro, Rafał K. Mantiuk

Data SynthesisVideoBenchmark

🎯 What it does: Created the first subjective quality dataset specifically for augmented reality displays, and investigated the impact of ambient light on the visibility of display artifacts (blurriness, color distortion).

Architectural Co-LOD Generation

Runze Zhang, Hui Huang

GenerationDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkMesh

🎯 What it does: Proposes a new method called Co-LOD for level of detail (LOD) management in building models, which can achieve precise and semantically consistent detailing both in individual models and model collections;

Autonomous Character-Scene Interaction Synthesis from Text Instruction

Nan Jiang (Institute for AI, Peking University), Yixin Zhu (Institute for AI, Peking University)

GenerationData SynthesisRobotic IntelligenceTransformerVision Language ModelVision-Language-Action ModelDiffusion modelVideoTextMultimodality

🎯 What it does: Proposes a framework for automatically generating multi-stage, scene-aware human-robot interaction motions from a single textual instruction and a target location, capable of seamlessly integrating actions such as walking, hand reaching, and object interaction in 3D environments.

Barrier-Augmented Lagrangian for GPU-based Elastodynamic Contact

Dewen Guo (Peking University), Guoping Wang (Peking University)

OptimizationComputational EfficiencyMeshBenchmarkPhysics Related

🎯 What it does: Proposed a GPU-based barrier-enhanced Lagrangian method for efficient and robust elastic dynamic contact simulation.

Bijective Volumetric Mapping via Star Decomposition

Steffen Hinderink, M. Campen

Flow-based ModelMesh

🎯 What it does: Proposes a method for constructing bijective volumetric mappings between 3D shapes.

BlobGEN-3D: Compositional 3D-Consistent Freeview Image Generation with 3D Blobs

Chao Liu, Arash Vahdat

GenerationData SynthesisTransformerVision Language ModelDiffusion modelGaussian SplattingImageText

🎯 What it does: Propose BlobGEN-3D, which uses object-level 3D blobs and text descriptions for scene representation, and generates free-viewpoint images through a pre-trained Blob-grounded 2D text-to-image diffusion model.

Body Gesture Generation for Multimodal Conversational Agents

Sunwoo Kim, Jehee Lee

GenerationPose EstimationDiffusion modelGenerative Adversarial NetworkMultimodality

🎯 What it does: Propose a body posture generation system that combines a motion matching framework with a learning-based method for multimodal real-time interactive conversational agents.

Boosting 3D object generation through PBR materials

Yitong Wang (Fudan University), Bo Dai (Shanghai Artificial Intelligence Laboratory)

GenerationData SynthesisConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelAuto EncoderGenerative Adversarial NetworkImageTextPoint CloudMeshRetrieval-Augmented Generation

🎯 What it does: Provide high-quality physically-based rendering (PBR) materials for 3D objects generated from single RGB images, with enhanced geometric details through iterative normal refinement.

BSDF importance sampling using a diffusion model

Ziyang Fu, Tzu-Mao Li

GenerationData SynthesisDiffusion modelScore-based ModelPhysics Related

🎯 What it does: Use diffusion models for importance sampling of BSDF, and propose two variants: one that learns the distribution on the disk, suitable for most reflective materials; the other that learns the distribution on the sphere, suitable for extreme specular reflective materials and full BSDF

C^0 Generalized Coons Patches for High-order Cage-based Deformation

Kaikai Qin, Chongyang Deng

🎯 What it does: A C^0 generalized Coons patch is proposed, enabling high-order cage deformation that simultaneously accommodates spatial deformation and high-order deformation of embedded objects.

Cafca: High-quality Novel View Synthesis of Expressive Faces from Casual Few-shot Captures

Marcel C. Bühler, Kripasindhu Sarkar (Google)

Image TranslationRestorationGenerationConvolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelNeural Radiance FieldAuto EncoderContrastive LearningImageMesh

🎯 What it does: Construct a personalized 3D facial model from only three casually taken facial photos, and generate high-quality, detailed new perspective images.

Camera Settings as Tokens: Modeling Photography on Latent Diffusion Models

I-Sheng Fang, Jun-Cheng Chen

GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelImage

🎯 What it does: Identified the expression bottleneck of camera settings in text-to-image models, constructed the CameraSettings20k dataset, and proposed a generation method that utilizes LoRA adapters to embed camera settings as tokens into LDM for numerical control.

Casper DPM: Cascaded Perceptual Dynamic Projection Mapping onto Hands

Yotam Erel (Tel Aviv University), Amit H. Bermano (Tel Aviv University)

Pose EstimationOptimizationComputational EfficiencyRobotic IntelligenceDiffusion modelAuto EncoderContrastive LearningGaussian SplattingOptical FlowVideoPoint CloudMesh

🎯 What it does: Designed and implemented Casper DPM — a real-time dynamic projection mapping framework based on a single projector, capable of projecting three-dimensional content onto the surface of the human hand, combined with a cascaded perception correction to reduce errors and perception latency.

CBIL: Collective Behavior Imitation Learning for Fish from Real Videos

Yifan Wu (University of Hong Kong), Taku Komura (University of Hong Kong)

Data SynthesisTransformerReinforcement LearningAuto EncoderGenerative Adversarial NetworkContrastive LearningVideo

🎯 What it does: Proposes the CBIL framework, which can directly learn the collective motion of fish schools from real videos without requiring three-dimensional motion trajectories.

Chebyshev Parameterization for Woven Fabric Modeling

Annika Öhri, O. Sorkine-Hornung

OptimizationOptical FlowMesh

🎯 What it does: Proposed a Chebyshev net-based fabric distortion energy for fabric modeling.

Coherent Optical Modems for Full-Wavefield Lidar

Parsa Mirdehghan (University of Toronto and Vector Institute), David B. Lindell (University of Toronto and Vector Institute)

Depth EstimationOptimizationComputational EfficiencyDiffusion modelScore-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowImagePoint CloudPhysics Related

🎯 What it does: This paper reconfigures a commercial coherent optical modulator into a LiDAR, utilizing amplitude and phase modulation under two polarization states to achieve full-wavefield LiDAR, which can simultaneously estimate the depth, radial velocity, and polarization information of targets within microseconds of exposure time.

Colorful Diffuse Intrinsic Image Decomposition in the Wild

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

RestorationConvolutional Neural NetworkDiffusion modelAuto EncoderContrastive LearningImage

🎯 What it does: Propose a multi-stage network framework that decomposes a single image into diffuse albedo, colored diffuse shadows, and non-diffuse residuals.

COMFI: A Calibrated Observer Metameric Failure Index for Color Critical Tasks

R. Wanat, Sally Hattori

Optimization

🎯 What it does: Proposed the CALIBRATED OBSERVER METEAMERIC FAILURE INDEX (COMFI) and experimentally evaluated viewer color matching inconsistencies to drive its parameterization

Compositional Neural Textures

Peihan Tu (University of Maryland), Matthias Zwicker (University of Maryland)

GenerationRepresentation LearningTransformerAuto EncoderContrastive LearningGaussian SplattingImage

🎯 What it does: Proposes a fully unsupervised texture representation method that decomposes texture into editable 2D Gaussian textons, enabling various editing operations such as texture reconstruction, transfer, interpolation, and animation.

Computational Biomimetics of Winged Seeds

Qiqin Le, Tao Du

OptimizationDiffusion modelPoint CloudMesh

🎯 What it does: A computational pipeline was developed that uses 3D scans of natural wings to construct a bio-inspired design space, and generates novel wings that outperform natural wings through gradient-free optimization.

Computational Design of a Kit of Parts for Bending Active Structures

Quentin Becker, Mark Pauly

Optimization

🎯 What it does: Proposes a computational framework that simplifies curved active structures into a sparse set of parts, enabling multiple input designs to be realized with the same set of parts

Consolidating Attention Features for Multi-view Image Editing

Or Patashnik, Fernando de la Torre (Carnegie Mellon University)

Image TranslationRestorationGenerationTransformerDiffusion modelNeural Radiance FieldAuto EncoderImage

🎯 What it does: Propose a multi-view image editing framework based on QNeRF, achieving geometric deformation editing of multi-view images by imposing 3D consistency constraints on the self-attention queries of diffusion models.

Content-aware Tile Generation using Exterior Boundary Inpainting

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

GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelAuto EncoderImageText

🎯 What it does: Proposes a content-aware tile generation method based on exterior boundary inpainting, which can generate diverse and seamlessly mosaicked tile sets from text prompts or example images, and supports multiple tile types (self-mosaic tiles, random self-mosaic tiles, Escher tiles, Wang Ou tiles, Dual Wang tiles).

Controllable Shape Modeling with Neural Generalized Cylinder

Xiangyu Zhu, Xiaoguang Han

GenerationData SynthesisDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderPoint CloudMesh

🎯 What it does: A shape modeling framework based on Neural Generic Cylinder (NGC) is proposed, which enables editable and blendable shapes by controlling the central curve geometry and its neural features.

Controlled Spectral Uplifting for Indirect-Light-Metamerism

M. van de Ruit, E. Eisemann

GenerationData SynthesisComputational EfficiencyDiffusion modelAuto EncoderGenerative Adversarial NetworkImage

🎯 What it does: Propose a new spectral boosting technique that can finely control spectral appearance under direct and indirect light, support placing spectral constraints in specific scenes, provide a flexible creative workflow, and efficiently solve spectral results, while introducing a compact spectral texture format to reduce memory usage during rendering.

Correlation-aware Encoder-Decoder with Adapters for SVBRDF Acquisition

Di Luo, Beibei Wang

RestorationConvolutional Neural NetworkGraph Neural NetworkAuto EncoderContrastive LearningImage

🎯 What it does: Propose a relevance-aware encoder-decoder network combined with graph convolutional networks for recovering high-quality SVBRDF maps from a few images, and introduce a dedicated adapter for each map in the decoder.

CPoser: An Optimization-after-Parsing Approach for Text-to-Pose Generation Using Large Language Models

Yumeng Li, Kun Zhou

GenerationPose EstimationOptimizationTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: Leveraging large language models (LLMs) for text-to-pose generation, first by parsing the text into a pose intermediate representation (Pose-IR) during the parsing phase, followed by robust optimization within a quantized pose prior space to generate expressive poses and gestures, and further refining to enhance naturalness and facial expressions.

Curly-Cue: Geometric Methods for Highly Coiled Hair

Haomiao Wu, Theodore Kim

GenerationDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningOptical FlowMesh

🎯 What it does: A geometric method for generating highly curled hair shapes is proposed, combining three techniques;

Customizing Text-to-Image Diffusion with Object Viewpoint Control

Nupur Kumari (Adobe Research), Jun-Yan Zhu (Carnegie Mellon University)

Image TranslationGenerationPose EstimationTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelNeural Radiance FieldContrastive LearningImageTextMultimodality

🎯 What it does: This paper proposes a custom method for incorporating object viewpoint control into text-to-image diffusion models, enabling the generation of new scenes and images with arbitrary viewpoints and text prompts while maintaining object identity.

Customizing Text-to-Image Models with a Single Image Pair

Maxwell Jones (Carnegie Mellon University), Jun-Yan Zhu (Carnegie Mellon University)

Image TranslationGenerationTransformerDiffusion modelImageText

🎯 What it does: Propose a text-to-image model customization method called Pair Customization based on single image pairs, which can learn and apply new styles from a pair of images with style differences while maintaining structure and color tone;

Dance-to-Music Generation with Encoder-based Textual Inversion

Sifei Li (Institute of Automation, Chinese Academy of Sciences), Changsheng Xu (Institute of Automation, Chinese Academy of Sciences)

GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelAuto EncoderVideoTextAudio

🎯 What it does: Propose an encoder-based text inversion method that encodes the rhythm and genre information of dance videos into pluggable pseudo-words, thereby enhancing pre-trained text-music models to achieve dance-synchronized music generation.

DARTS: Diffusion Approximated Residual Time Sampling for Time-of-flight Rendering in Homogeneous Scattering Media

Qianyue He (Tsinghua University), Xin Jin (Tsinghua University)

OptimizationComputational EfficiencyDiffusion modelScore-based ModelNeural Radiance FieldImageMeshPhysics RelatedStochastic Differential Equation

🎯 What it does: This paper proposes a rendering algorithm called DARTS, which efficiently generates time-resolved paths in uniform scattering media, achieving precise control over the flight time of light.

Decoupling Contact for Fine-Grained Motion Style Transfer

Xiangjun Tang, Xiaogang Jin

GenerationData SynthesisPose EstimationTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowVideoSequential

🎯 What it does: Propose a motion style transfer method that decouples and finely controls contact through hip velocity.

Deformation Recovery: Localized Learning for Detail-Preserving Deformations

Ramanathan Sundararaman (Ecole Polytechnique), M. Ovsjanikov (Ecole Polytechnique)

RestorationGraph Neural NetworkContrastive LearningPoint CloudMesh

🎯 What it does: Propose a Jacobian network based on local rough input to achieve high-quality, detail-preserving shape deformation, and apply it to registration refinement, unsupervised correspondence, and interactive editing.

Dense Server Design for Immersion Cooling

Milin Kodnongbua, Adriana Schulz

OptimizationPhysics Related

🎯 What it does: Designed and implemented a three-dimensional server density optimization computation framework for liquid immersion cooling, and verified its feasibility through prototypes and experiments.

Designing triangle meshes with controlled roughness

Victor Ceballos Inza, H. Pottmann

OptimizationDiffusion modelAuto EncoderGenerative Adversarial NetworkMesh

🎯 What it does: This paper proposes a computational design framework for controlling the roughness of triangular meshes, focusing on small-scale roughness, and achieving the regulation of roughness or smoothness by utilizing the local positioning of mesh edges and faces relative to the reference surface curvature; by constructing a two-dimensional dual graph and generating it in isometric geometry, the analysis of the interaction between curvature and roughness is simplified, and a visualization and interactive local design tool is provided; the framework includes components such as appearance-oriented remeshing, optimization-based automatic roughening, and dihedral angle control.

DiffCSG: Differentiable CSG via Rasterization

Haocheng Yuan (University of Edinburgh), Changjian Li (University of Edinburgh)

OptimizationComputational EfficiencyData-Centric LearningDiffusion modelAuto EncoderOptical FlowMeshBenchmark

🎯 What it does: Propose DiffCSG, which achieves differentiable rendering of Constructive Solid Geometry (CSG) models through differentiable rasterization;

Differentiable Modeling of Material Spreading in Inkjet Printing for Appearance Prediction

Emiliano Luci, Vahid Babaei

GenerationData SynthesisDiffusion modelNeural Radiance FieldAuto EncoderImagePhysics Related

🎯 What it does: A fully differentiable method is proposed for modeling material diffusion behavior in inkjet printing, generating a general material diffusion transformation based on a single calibration image, resulting in an effective material grid.

Differentiable Owen Scrambling

Bastien Doignies, V. Ostromoukhov

OptimizationAuto Encoder

🎯 What it does: Proposes a differentiable Owen scrambling method for gradient optimization of low-discrepancy sequences, while keeping the original t-value unchanged.

Differentiable Photon Mapping using Generalized Path Gradients

J.-G. Xing, Kun Xu

Neural Radiance FieldOptical FlowPhysics RelatedStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: Proposed the first differentiable photon mapping method

Differential Walk on Spheres

Bailey Miller (Carnegie Mellon University), Ioannis Gkioulekas (Carnegie Mellon University)

OptimizationPoint CloudMeshPhysics RelatedStochastic Differential Equation

🎯 What it does: A differentiable Walk on Spheres (DiffWoS) algorithm is proposed, which directly computes the derivative of the solution of partial differential equations (taking the shielded Poisson equation as an example) with respect to shape or boundary parameters, without requiring global solving or meshing, thereby enabling inverse problems and shape optimization.

Differentiating Variance for Variance-Aware Inverse Rendering

Kai Yan, Shuang Zhao

Optimization

🎯 What it does: Derive the mathematical form of the derivative of rendering variance with respect to scene parameters and sampling probabilities, and propose an unbiased Monte Carlo estimator to achieve variance-aware inverse rendering.

DiffH2O: Diffusion-Based Synthesis of Hand-Object Interactions from Textual Descriptions

S. Christen, Bugra Tekin

GenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelImageTextMesh

🎯 What it does: Propose DiffH2O, a framework based on diffusion models, which generates realistic and controllable hand-object interaction actions from natural language text, and can generate them on unseen objects.

DifFRelight: Diffusion-Based Facial Performance Relighting

Mingming He (Netflix Eyeline Studios), P. Debevec

Image TranslationGenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelGaussian SplattingImageVideo

🎯 What it does: Achieve free-viewpoint relighting of multi-view planar illumination facial performances using diffusion models.

DiffUHaul: A Training-Free Method for Object Dragging in Images

Omri Avrahami (Hebrew University of Jerusalem), Weili Nie (NVIDIA Research)

Image TranslationImage HarmonizationGenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelImageTextMultimodality

🎯 What it does: Propose an untrained object dragging method called DiffUHaul, which utilizes the spatial understanding of the local text-to-image model BlobGEN to achieve smooth movement of objects in images.

Direct Manipulation of Procedural Implicit Surfaces

Marzia Riso, Fabio Pellacini

Diffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderMesh

🎯 What it does: A method for directly manipulating procedural implicit surfaces within the viewport is proposed, achieving synchronization between user interaction and shape parameters through the design of specific co-parameterization and automatic differentiation.

DirectL: Efficient Radiance Fields Rendering for 3D Light Field Displays

Zongyuan Yang, Xunbo Yu

Computational EfficiencyNeural Radiance FieldGaussian SplattingOptical FlowImage

🎯 What it does: For self-illuminating optical displays, the DirectL rendering paradigm is proposed, which directly renders light field images, significantly reducing the number of pixels rendered and improving real-time rendering speed.

DIScene: Object Decoupling and Interaction Modeling for Complex Scene Generation

Xiao-Lei Li, Shi-Min Hu

GenerationData SynthesisKnowledge DistillationGraph Neural NetworkTransformerDiffusion modelGaussian SplattingMeshGraph

🎯 What it does: Leveraging the knowledge of pre-trained 2D diffusion models, we distill knowledge for 3D asset generation and propose the DIScene method, which adopts a learnable structured scene graph to represent complete 3D scenes. Nodes explicitly model object appearance, text descriptions, transformations, and geometry (mesh + surface-aligned Gaussians), while edges model object interactions. Objects are optimized in a canonical space, and object-aware rendering is used to avoid erroneous backpropagation.

DrawingSpinUp: 3D Animation from Single Character Drawings

Jie Zhou, Hongbo Fu

Image TranslationGenerationPose EstimationConvolutional Neural NetworkTransformerPrompt EngineeringDiffusion modelAuto EncoderGaussian SplattingOptical FlowImageVideoPoint CloudMesh

🎯 What it does: Propose the DrawingSpinUp system, which can generate characters with 3D animation effects from a single hand-drawn character image, achieving realistic poses and textures through contour removal and geometric refinement.

DreamUDF: Generating Unsigned Distance Fields from A Single Image

Yu-Tao Liu, Lin Gao

GenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldAuto EncoderImage

🎯 What it does: Proposed the DreamUDF framework for generating high-quality 3D objects with arbitrary topology from a single image.

Dynamic Gaussian Marbles for Novel View Synthesis of Casual Monocular Videos

Colton Stearns (Stanford University), Leonidas J. Guibas (Stanford University)

GenerationData SynthesisOptimizationNeural Radiance FieldContrastive LearningGaussian SplattingOptical FlowImageVideo

🎯 What it does: This paper proposes the Dynamic Gaussian Marbles method, extending dynamic scene Gaussian Splatting to arbitrary monocular videos, achieving high-quality novel view synthesis.

Dynamic Neural Radiosity with Multi-grid Decomposition

Rui Su, Sheng Li

Neural Radiance FieldAuto Encoder

🎯 What it does: A variable scene neural radiance decomposition method is proposed, which can efficiently model dynamic global illumination.

Dynamic Skeletonization via Variational Medial Axis Sampling

Qijia Huang, D. Bechmann

OptimizationPoint CloudMesh

🎯 What it does: Propose a method to compute the discrete skeleton of point clouds or triangular meshes through variational medial sampling, optimizing shape partitioning and minimizing plane-sphere and point-sphere distance errors;

Efficient GPU Cloth Simulation with Non-distance Barriers and Subspace Reuse

Lei Lan (University of Utah), Yin Yang (University of Utah)

OptimizationComputational EfficiencyMesh

🎯 What it does: This paper proposes a GPU-based fabric simulation framework based on Projective Dynamics, which can real-time simulate high-resolution clothing models at interactive frame rates while ensuring that all triangles remain untangled;

Efficient Image-Space Shape Splatting for Monte Carlo Rendering

Xiaochun Tong, T. Hachisuka

Computational EfficiencyGaussian SplattingImage

🎯 What it does: Propose a general framework to efficiently reuse ray paths within 2D shapes to improve Monte Carlo rendering efficiency.

Efficient Neural Path Guiding with 4D Modeling

Honghao Dong, Sheng Li

Computational EfficiencyRepresentation LearningDiffusion modelNeural Radiance FieldAuto EncoderImageVideo

🎯 What it does: Propose a path-guided method that utilizes compact neural representations and neural feature decomposition to model high-dimensional spatiotemporal distributions, and achieve low-dimensional subspace approximation through a progressive training strategy, exploring its application effects in distributed ray tracing scenarios such as motion blur and spectral rendering.

EgoAvatar: Egocentric View-Driven and Photorealistic Full-body Avatars

Jianchun Chen (MPI for Informatic), C. Theobalt

Image TranslationRestorationGenerationPose EstimationGraph Neural NetworkTransformerSupervised Fine-TuningDiffusion modelContrastive LearningGaussian SplattingOptical FlowImageVideoPoint CloudMesh

🎯 What it does: Real-time inference of human 3D skeletal motion using egocentric video captured by a monocular head-mounted camera, reconstructing animatable full-body geometry and material, and achieving photorealistic rendering from arbitrary viewpoints through high-quality 3D Gaussian splatting.

EgoHDM: A Real-time Egocentric-Inertial Human Motion Capture, Localization, and Dense Mapping System

Handi Yin (Hong Kong University of Science and Technology (Guangzhou)), Pan Hui (Hong Kong University of Science and Technology (Guangzhou))

Pose EstimationDepth EstimationAutonomous DrivingOptimizationConvolutional Neural NetworkTransformerContrastive LearningGaussian SplattingSimultaneous Localization and MappingOptical FlowImageVideoPoint CloudTime Series

🎯 What it does: Developed an online egocentric inertial and visual fusion system called EgoHDM, which uses six IMUs and a head-mounted RGB camera to achieve simultaneous estimation of human motion capture, global localization, and dense 3D scene reconstruction.

elaTCSF: A Temporal Contrast Sensitivity Function for Flicker Detection and Modeling Variable Refresh Rate Flicker

Yancheng Cai, Rafal Mantiuk

OptimizationComputational EfficiencyRepresentation LearningContrastive LearningTabularTime SeriesPhysics Related

🎯 What it does: This paper proposes the elaTCSF model, which predicts flicker visibility under different conditions by incorporating factors such as brightness, eccentricity, and image area into the existing IDMS TCSF model, and combining it with a spatial probability summation model.

ELMO: Enhanced Real-time LiDAR Motion Capture through Upsampling

Deok-Kyeong Jang (MOVIN Inc.), Dong-Ha Shin

Data SynthesisPose EstimationComputational EfficiencyGraph Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningVideoPoint Cloud

🎯 What it does: Propose the ELMO framework, achieving real-time motion capture at 60fps using a single 20fps LiDAR point cloud;

End-to-End Hybrid Refractive-Diffractive Lens Design with Differentiable Ray-Wave Model

Xinge Yang, Wolfgang Heidrich (UNC)

OptimizationDiffusion modelAuto EncoderOptical FlowImagePhysics Related

🎯 What it does: Propose a differentiable ray-wave model for accurately simulating aberrations and phase modulation in hybrid refractive-diffractive lenses, and achieve end-to-end co-optimization of optical systems and image reconstruction networks.

End-to-end Optimization of Fluidic Lenses

Mulun Na, Wolfgang Heidrich

OptimizationDiffusion modelOptical FlowPhysics RelatedOrdinary Differential Equation

🎯 What it does: A new shape space model for liquid-forming optical lenses was developed and integrated into an end-to-end optical design pipeline based on differentiable ray tracing through differentiable implementation; subsequently, the model and design process were extensively evaluated on simulations and preliminary physical prototypes.

Enhancing the Aesthetics of 3D Shapes via Reference-based Editing

Minchan Chen, Manfred Lau

GenerationData SynthesisDiffusion modelScore-based ModelGenerative Adversarial NetworkPoint CloudMesh

🎯 What it does: Propose an automatic beautification framework based on a reference model, which first collects 3D shape aesthetic scores to build a dataset, and then makes the input shape more aesthetically pleasing through reference-guided global deformation and optional local part transplantation.

Evaluating Visual Perception of Object Motion in Dynamic Environments

Budmonde Duinkharjav, Qi Sun

Human-Computer InteractionOptical FlowVideo

🎯 What it does: Conduct a crowdsourced psychophysical experiment to measure and model observers' perceptual accuracy in determining the direction of object motion in dynamic 3D environments, and propose a general model to guide users to better understand the motion of target objects.

EVSplitting: An Efficient and Visually Consistent Splitting Algorithm for 3D Gaussian Splatting

Qi-Yuan Feng, Shi-Min Hu

OptimizationComputational EfficiencyGaussian SplattingPoint CloudMesh

🎯 What it does: Proposes an efficient and visually consistent segmentation algorithm called EVSplitting for 3D Gaussian Splatting.

Exact and Efficient Intersection Resolution for Mesh Arrangements

Jia-Peng Guo, Xiao-Ming Fu

Mesh

🎯 What it does: A new method is proposed to accurately and efficiently solve intersections and self-intersections in triangular meshes.

FabricDiffusion: High-Fidelity Texture Transfer for 3D Garments Generation from In-The-Wild Images

Cheng Zhang, Fernando de la Torre

Image TranslationRestorationGenerationTransformerDiffusion modelScore-based ModelImageMesh

🎯 What it does: A method called FabricDiffusion was developed, which can extract undistorted, tileable fabric textures from a single garment image and map them onto 3D clothing of arbitrary shapes.

Fabrig: A Cloth-Simulated Transferable 3D Face Parameterization

Changan Zhu, Chris Joslin

GenerationData SynthesisDiffusion modelScore-based ModelMesh

🎯 What it does: Proposes a 3D facial parametrization method called Fabrig, based on facial anatomy, which is fast to set up, transferable across multiple characters, editable, and compatible with physical simulations.

FaçAID: A Transformer Model for Neuro-Symbolic Facade Reconstruction

Aleksander Płocharski (Warsaw University of Technology), Przemyslaw Musialski (New Jersey Institute of Technology)

GenerationData SynthesisOptimizationTransformerLarge Language ModelPrompt EngineeringDiffusion modelContrastive LearningImageText

🎯 What it does: Use a Transformer model to convert segmented building facade images into editable procedural definitions, achieving the inverse process from static segmentation to parameterizable facade modeling.

FaceMap: Distortion-Driven Perceptual Facial Saliency Maps

Zhongshi Jiang, Alexandre Chapiro

RecognitionCompressionMesh

🎯 What it does: Conducted a quantitative study on the importance of facial regions, and created a dataset containing 960 unique facial models with geometric and texture distortions, followed by more than 18,000 subjective comparison experiments to evaluate the visibility of distortions.

Fashion-VDM: Video Diffusion Model for Virtual Try-On

Johanna Karras (Google Research), Ira Kemelmacher-Shlizerman (Google Research)

Image TranslationGenerationData SynthesisPose EstimationConvolutional Neural NetworkTransformerPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningOptical FlowImageVideoMultimodality

🎯 What it does: Proposes Fashion-VDM, a video virtual try-on method based on diffusion models, which can synthesize high-quality try-on videos that retain the person's identity and movements under the condition of given clothing images and person videos.

Fast and Globally Consistent Normal Orientation based on the Winding Number Normal Consistency

Siyou Lin (Tsinghua University), Yebin Liu (Tsinghua University)

OptimizationComputational EfficiencyDiffusion modelScore-based ModelAuto EncoderContrastive LearningGaussian SplattingPoint CloudMesh

🎯 What it does: Propose a new property of the Winding Number — Winding Number Normal Consistency (WNNC), and use this property together with the energy function of Parametric Gauss Reconstruction (PGR) in an alternating iterative manner, to quickly obtain globally consistent and high-quality point cloud normals.

Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillation

Axel Sauer (Stability AI), Robin Rombach (Stability AI)

GenerationData SynthesisComputational EfficiencyKnowledge DistillationTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextMultimodality

🎯 What it does: Propose Latent Adversarial Diffusion Distillation (LADD), which compresses large-scale latent diffusion models into single-step or few-step high-speed models, enabling high-resolution multi-aspect-ratio image synthesis; applied to Stable Diffusion 3 to obtain SD3-Turbo;

Filtering-Based Reconstruction for Gradient-Domain Rendering

Difei Yan, Kun Xu

RestorationImage

🎯 What it does: A gradient-guided filtering method is proposed for color image reconstruction in gradient domain rendering, avoiding instability caused by directly using noisy gradients.

Fluid Implicit Particles on Coadjoint Orbits

M. Nabizadeh, Albert Chern

Diffusion modelOptical FlowPoint CloudMeshBenchmarkPhysics RelatedStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: Propose CO-FLIP: a high-order structure-preserving fluid simulation method based on conjugate orbits, utilizing Hamiltonian dynamics, affine interpolation, and geometric time integration to achieve strict conservation of energy, circulation, and Casimir constants at low resolution.