PGN-Cloth: Physics-based graph network model for 3D cloth animation

Graph neural networks have been applied to learning-based simulations of cloth and have received significant attention recently. However, some existing learning-based graph networks lack essential information on the cloth's structure, leading to generated results that are implausible and prone to penetration issues. To address these challenges, we propose the PGN-Cloth model, which leverages meshes to represent the cloth system state and computes dynamics through graph neural networks. Our contributions include the following: (1) PGN-Cloth combines the ideas of physics-based deep learning and mass-spring models, incorporating EdgeLoss, CosLoss, and BendLoss to produce more realistic results. (2) An additional penetration loss is added to address the penetration problem that is prevalent in current state-of-the-art methods. (3) Our approach significantly improves the training speed of the network while only requiring a small increase in computational cost. Furthermore, it exhibits better stability than traditional learning methods. The experimental results demonstrate that our method outperforms the state-of-the-art in several indicators and significantly reduces the incidence of penetration in the generated results.

Automated summary

In this study, a PGN-Cloth model based on graph neural networks is proposed for cloth simulation. The model represents the cloth system state using meshes and computes dynamics with graph neural networks. An additional penetration loss is introduced to address the penetration problem. Experimental results show that the proposed method outperforms the state-of-the-art in several indicators and significantly reduces the incidence of penetration in the generated results.