Polynomial Speedup in Diffusion Models via Multilevel Euler-Maruyama Method
arXiv·medium signal
Introduces the Multilevel Euler-Maruyama (ML-EM) method for computing SDE/ODE solutions using a hierarchy of approximators with increasing accuracy and cost, requiring only a few evaluations of the most expensive model. This provides a rigorous mathematical framework for achieving polynomial computational speedup in diffusion model sampling — moving beyond heuristic acceleration techniques to provably efficient methods. Applicable to any diffusion-based generative model pipeline.