SAGS-GNN: Graph Neural Network for self-collision and anisotropy in dynamic garment simulation

Garment is an essential component of digital humans, and the accurate representation of dynamic simulation details and wrinkle characteristics is crucial for enhancing the realism of virtual scenes. However, this task remains significantly challenging in complex simulation scenarios. Therefore, we propose a novel garment simulation method based on Graph Neural Networks (GNNs), referred to as SAGS-GNN, which effectively simulates self-collision and cloth anisotropy. To tackle the self-collision problem, we present the repulsive loss term and the maximum depth loss term. These terms effectively simulate the interactions between the vertices of the cloth mesh by jointly constraining their positions, thereby facilitating the self-collision handling of garments. Furthermore, our approach utilizes the Neo-Hookean StVK method to achieve anisotropy in cloth, further reflecting the different wrinkle details of multiple materials during motion. In summary, our SAGS method effectively mitigates the issue of interpenetration among garments, facilitates the realization of anisotropic properties in a variety of fabric materials, and significantly enhances the visual realism of virtual apparel. We evaluate our method on various garment types and materials, demonstrating competitive qualitative and quantitative results. • A new garment simulation method based on Graph Neural Networks , SAGS-GNN. • A novel self-collision loss term is proposed to manage vertices that are prone to or already in collision, effectively addressing self-collision issues in garments. • An anisotropic loss term based on the Neo-Hookean StVK model generates wrinkles for various fabrics, enhancing the physical realism of garment dynamics.