HOOD: Hierarchical Graphs for Generalized Modelling of Clothing Dynamics

Figure 1 . We combine graph neural networks, a hierarchical graph representation, and multi-level message passing with an unsupervised training scheme to enable efficient prediction of realistic clothing dynamics for arbitrary types of garments and body shapes. Our method models both tight-fitting and free-flowing clothes draped over arbitrary body shapes. At test time, the method generalizes to new, entirely unseen, garments (left), and allows dynamic and unconstrained poses (right) and changes in material parameters and topology.