Toward Scalable and Generalisable Garment Simulation with Graph Neural Networks
Access status:
Open Access
Type
ThesisThesis type
Masters by ResearchAuthor/s
Liu, AoranAbstract
Realistic garment simulation is essential for applications such as film production, video games and fashion design. While physics-based methods remain dominant for their accuracy and generality, they rely on costly numerical optimisation at each timestep, limiting their use in ...
See moreRealistic garment simulation is essential for applications such as film production, video games and fashion design. While physics-based methods remain dominant for their accuracy and generality, they rely on costly numerical optimisation at each timestep, limiting their use in real-time settings. Recent advances in data-driven methods offer faster alternatives. Early pose-driven approaches map body poses to garment motion through neural networks but generalise poorly, as they depend on specific garment meshes. In contrast, Graph Neural Networks (GNNs) predict per-vertex dynamics by modelling local vertex interactions, improving both simulation quality and generalisation across garment topologies. However, GNN-based simulators face two key challenges. First, accurate per-vertex prediction requires large receptive fields, often necessitating multiple message passing rounds that reduce computational efficiency. Second, these models struggle to generalise across unseen mesh resolutions, limiting their robustness in diverse scenarios. This thesis addresses both challenges through two studies. The first introduces Laplacian-Smoothed Dual Message Passing (LSDMP) and Geodesic Self-Attention (GSA) to efficiently enlarge the receptive field. LSDMP enhances short-range propagation via Laplacian smoothing, while GSA captures global context through self-attention. Together, these modules improve accuracy and scalability without deep message passing. The second study addresses resolution sensitivity through adaptive message passing that adjusts propagation depth to mesh density, combined with resolution-aware scaling to preserve consistent physical behaviour across resolutions. Extensive experiments show that both methods outperform state-of-the-art GNN-based simulators in efficiency, quality, and generalisation. Collectively, these studies advance the practicality and robustness of GNN-based garment simulation, paving the way for real-time and cross-resolution applications.
See less
See moreRealistic garment simulation is essential for applications such as film production, video games and fashion design. While physics-based methods remain dominant for their accuracy and generality, they rely on costly numerical optimisation at each timestep, limiting their use in real-time settings. Recent advances in data-driven methods offer faster alternatives. Early pose-driven approaches map body poses to garment motion through neural networks but generalise poorly, as they depend on specific garment meshes. In contrast, Graph Neural Networks (GNNs) predict per-vertex dynamics by modelling local vertex interactions, improving both simulation quality and generalisation across garment topologies. However, GNN-based simulators face two key challenges. First, accurate per-vertex prediction requires large receptive fields, often necessitating multiple message passing rounds that reduce computational efficiency. Second, these models struggle to generalise across unseen mesh resolutions, limiting their robustness in diverse scenarios. This thesis addresses both challenges through two studies. The first introduces Laplacian-Smoothed Dual Message Passing (LSDMP) and Geodesic Self-Attention (GSA) to efficiently enlarge the receptive field. LSDMP enhances short-range propagation via Laplacian smoothing, while GSA captures global context through self-attention. Together, these modules improve accuracy and scalability without deep message passing. The second study addresses resolution sensitivity through adaptive message passing that adjusts propagation depth to mesh density, combined with resolution-aware scaling to preserve consistent physical behaviour across resolutions. Extensive experiments show that both methods outperform state-of-the-art GNN-based simulators in efficiency, quality, and generalisation. Collectively, these studies advance the practicality and robustness of GNN-based garment simulation, paving the way for real-time and cross-resolution applications.
See less
Date
2025Rights statement
The author retains copyright of this thesis. It may only be used for the purposes of research and study. It must not be used for any other purposes and may not be transmitted or shared with others without prior permission.Faculty/School
Faculty of Engineering, School of Computer ScienceAwarding institution
The University of SydneyShare