Sig-DEG for Distillation: Making Diffusion Models Faster and Lighter
Sig-DEG is a generator designed to distill large diffusion models into faster versions by approximating their reverse diffusion process at a coarse time grid. It uses partial signatures, which summarize Brownian motion over subintervals, combined with a recurrent structure based on stochastic Taylor expansions. This allows for fast inference while maintaining high sample quality.