How to animate loose clothes without physics simulations

Pose alone can't explain how loose clothing drapes and moves; the same body position produces different garment states depending on momentum and contact. This work adds a learned dynamics model to neural avatars that evolves a latent code over time, decomposing motion into driving, restoring, and damping forces—mimicking physics without explicit simulation. Across nine sequences of everyday motion, the method generates temporally coherent, history-dependent animations with sharper details than prior data-driven approaches, verified through metrics and user studies.