These 27 SIGGRAPH 2023 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every SIGGRAPH 2023 paper, free trial on arXivSub.
A Contact Proxy Splitting Method for Lagrangian Solid-Fluid Coupling
Tianyi Xie (University of California Los Angeles), Chenfanfu Jiang
🎯 What it does: This paper proposes a unified Lagrangian method to achieve two-way strong coupling between weakly compressible SPH fluids and nonlinear elastic FEM solids; through optimization-based time integration, barrier-type contact, approximate viscous potential energy, and proxy-based time splitting, the coupling efficiency and stability are significantly improved.
A Convex Optimization Framework for Regularized Geodesic Distances
M. Edelstein, M. Ben-Chen
CodeOptimizationPoint CloudMesh
🎯 What it does: Propose a unified convex optimization framework for computing geodesic distances with regularization, providing theoretical guarantees, optional regularization terms, symmetric full-pair distance implementations, and an efficient ADMM solver.
🎯 What it does: This paper proposes an unsupervised stroke-level disentanglement technique based on StyleGAN, achieving an AI-assisted drawing system called AniFaceDrawing that generates high-quality anime portraits incrementally during the hand-drawing process;
🎯 What it does: Proposes ArrangementNet, a graph neural network, for estimating indoor scene layouts from incomplete point clouds and generating BIM models.
🎯 What it does: Proposes AvatarMAV, a fast 3D head avatar reconstruction method that utilizes motion-aware neural voxels, capable of training a photorealistic 3D head model from monocular video in just 5 minutes.
Bidirectional GaitNet: A Bidirectional Prediction Model of Human Gait and Anatomical Conditions
Jungnam Park (Seoul National University), Jungdam Won (Seoul National University)
CodeGenerationData SynthesisPose EstimationConvolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningMixture of ExpertsDiffusion modelAuto EncoderMeshBiomedical DataMagnetic Resonance ImagingComputed TomographyElectronic Health Records
🎯 What it does: Constructed Bidirectional GaitNet, a bidirectional generative model capable of predicting gait based on human anatomical parameters, and inversely estimating the corresponding anatomical and muscular conditions given a gait.
CodeGenerationData SynthesisAnomaly DetectionTransformerMixture of ExpertsDiffusion modelGenerative Adversarial NetworkContrastive LearningPoint CloudMeshSequential
🎯 What it does: Propose COFS, a controllable furniture layout synthesis model based on Transformer, which supports fine-grained conditional control over any subset of object attributes and achieves non-autoregressive sampling.
🎯 What it does: Propose a new differentiable isosurface extraction framework called FlexiCubes, which can generate high-quality, differentiable 3D meshes during gradient optimization.
🎯 What it does: Proposes a framework for automatically designing thin plate metal (Sheet Metal) structures that can minimize manufacturing costs and generate high-performance manufacturable load-bearing parts while satisfying structural, spatial, and manufacturing constraints.
Listen, Denoise, Action! Audio-Driven Motion Synthesis with Diffusion Models
Simon Alexanderson (KTH Royal Institute of Technology), G. Henter
CodeGenerationData SynthesisPose EstimationTransformerPrompt EngineeringMixture of ExpertsDiffusion modelScore-based ModelVideoSequentialAudio
🎯 What it does: This paper proposes an audio-driven human motion synthesis method based on diffusion models, which can generate dance and gestures synchronized with music or speech, and supports style control and style interpolation.
🎯 What it does: A new SC-L1 regularization method is proposed for achieving local shape editing based on elastic energy, along with a three-block ADMM optimization framework that enables real-time interactive performance.
Masonry Shell Structures with Discrete Equivalence Classes
Rulin Chen, Ying He
CodeOptimization
🎯 What it does: A method is proposed to model masonry shell structures as discrete equivalence classes, and through hierarchical clustering and optimization, reusable shell templates are generated to achieve approximable and manufacturable seamless structures for free-form surfaces.
🎯 What it does: This paper proposes a three-dimensional shape reconstruction method based on template mesh deformation, which uses grid density adaptive energy to densify vertices in complex structures, thereby improving reconstruction accuracy in inverse rendering and non-rigid registration.
🎯 What it does: Propose the NeRO method, which uses multi-view RGB images to simultaneously reconstruct the geometry and BRDF of reflective objects under unknown lighting and without object masks through neural rendering.
🎯 What it does: A low-dimensional perceptual embedding space was constructed, mapping the density and correlation of point patterns into a three-channel image, enabling intuitive editing through image editing software;
🎯 What it does: A low-cost surface display technology was developed, which attaches a flexible PCB to a 3D printed support structure, covering the surface and arranging individually addressable RGB LEDs.
🎯 What it does: This paper proposes a method for real-time stylized intermediate motion synthesis (RSMT), which can generate natural, high-quality, and controllable transition motions given the start and end frames, time duration, and target style.
🎯 What it does: Proposed a reconstruction-oriented automatic scanning method called ScanBot, which utilizes hierarchical deep reinforcement learning for global ROI planning and local NBV planning to improve scanning efficiency and reconstruction quality.
🎯 What it does: Propose StyleAvatar, a real-time high-fidelity full-body portrait video avatar reconstruction and reenactment method based on StyleGAN, which can achieve high-quality portrait reconstruction within a single video segment.
🎯 What it does: This paper proposes an intrinsic triangulation simplification method for solving equations on surfaces, which maintains the approximate intrinsic geometry through local vertex removal, reconstruction, and intrinsic Delaunay flips, generating a multi-resolution hierarchy directly usable for numerical solutions.
CodeGenerationRetrievalRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper creates a dataset called text2fabric, which includes 45,000 rendered images of 3,000 fabric materials and more than 15,000 natural language descriptions, and performs lexical, attribute, and structural analysis on these descriptions; subsequently, the dataset is used to fine-tune CLIP and BLIP for fine-grained text retrieval, image retrieval, and fabric description generation.
Toward Optimized VR/AR Ergonomics: Modeling and Predicting User Neck Muscle Contraction
Yunxiang Zhang (New York University), Qi Sun (New York University)
CodeHuman-Computer InteractionPose EstimationOptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackConvolutional Neural NetworkTabularTime SeriesSequentialBiomedical Data
🎯 What it does: This study collects synchronized data of neck muscle contraction levels (MCL) and head motion (posture and acceleration) by using surface electromyography (EMG) sensors on VR headset wearers. Based on this data, a biophysical model (MCLNet) is constructed to estimate the MCL of completed head movements. Subsequently, TrajectoryNet is used to regress typical motion trajectories between initial and end postures, further predicting the MCL before the head movement begins, thereby assessing the user's potential neck discomfort in advance.
🎯 What it does: Propose an unsupervised spectral shape matching framework that directly learns point-to-point correspondences through deep feature mapping.