Self-supervised Garment Dynamics with Persistent Wrinkles

This paper tackles a long-standing gap in neural garment simulation: most self-supervised methods look good in motion, but they wash out the persistent wrinkles that make fabric feel physically grounded. The authors explicitly model plasticity, so cloth can retain creases instead of snapping back to a perfectly elastic state.

The main technical change is a physics-inspired loss that turns training into a moving energy minimization problem. That’s clever, but also tricky: because the loss itself changes during optimization, training can become unstable. To deal with that, they introduce a curriculum where the target material gradually shifts from pure elasticity to elasto-plasticity, letting both the loss and the model parameters converge together.

For game teams, the practical appeal is obvious. You get a self-supervised approach, so there’s no dependency on captured training data, while still improving visual realism on garments, body shapes, and motion types. The paper claims it outperforms prior methods across a range of metrics and, importantly, can generate natural persistent wrinkles for the first time in this setting.

This is still research, not a drop-in engine feature, but it points toward cheaper cloth workflows for cinematics, character customization, and maybe eventually runtime simulation where believable fabric detail matters more than exact mechanical fidelity.