Cloth Simulation with Deep Learning

We wear clothes everyday but we could easily ignore the fact that this seemingly mundane object is actually governed by a rich and complex dynamics. Its thin surface deforming, energies smoothly transferring, collisions permanently happening, all to create wrinkles that makes it undeniably a cloth. Not as stiff as a paper, not as chaotic as fluids. Representing and visualizing this rich expressiveness of cloth materials has been an active research topic for decades. Cloth Simulation with Deep Learning research tackles the issue of how far deep learning techniques could learn this cloth dynamics and our goal is to integrate realistic realtime cloth behavior in the virtual world.

We are inspired by the physics-based simulation that has been traditionally used for visualizing clothes. We use physical elements (like internal forces and energy) that would effectively represent the current state and history of the cloth and integrate it with our deep learning architecture. We also try novel techniques in deep learning such as shared discrete codebook to compress large datasets of cloth information and apply motion matching techniques in the cloth domain.