GraphNeuralCloth: A Graph‐Neural‐Network‐Based Framework for Non‐Skinning Cloth Simulation

Cloth simulation has a wide range of applications, spanning from virtual animation (e.g., video games and movies) to robotic manipulation (e.g., automatic cloth folding). Conventional physics‐based methods involve computationally exhaustive iterations, making real‐time simulation challenging. Deep learning provides an alternative for fast and realistic cloth simulation but most of these learning models are still trained using virtual data only. In this article, we present a cloth motion capture system and a point‐cloud‐to‐mesh processing method to support the prediction of real‐world fabric deformation. We also propose GraphNeuralCloth, a graph‐neural‐network (GNN)‐based framework, capable of estimating the cloth morphology change in real time. Our framework has been evaluated on three datasets, showcasing its performance not only in avatar try‐on (≈4.2 mm in nodal distance error using Santesteban's dataset 1 ), but also in real fabric free falling (≈2.3 mm using fabric draping dataset). In addition, our network model can accurately predict changes of real‐world cloth morphology (≈8.55 mm using cloth motion dataset). The ablation study demonstrates that the key components of our model, namely spiral graph convolution and identity connection, contribute to faster training convergence and reduced inference time (≈1000 fps).