This repository implements a research paper that analyzes and clarifies the design space of diffusion-based generative models. It provides training and sampling code for score networks that generate images via stochastic denoising processes. The work introduces improved preconditioning, sampling schedules, and training methodologies that together achieve FID scores of 1.79 on CIFAR-10 (class-conditional) and 1.97 (unconditional), with faster convergence than prior approaches.