Abstract
Denoising diffusion probabilistic models (DDPM) are a class of generativemodels which have recently been shown to produce excellent samples. We showthat with a few simple modifications, DDPMs can also achieve competitivelog-likelihoods while maintaining high sample quality. Additionally, we findthat learning variances of the reverse diffusion process allows sampling withan order of magnitude fewer forward passes with a negligible difference insample quality, which is important for the practical deployment of thesemodels. We additionally use precision and recall to compare how well DDPMs andGANs cover the target distribution. Finally, we show that the sample qualityand likelihood of these models scale smoothly with model capacity and trainingcompute, making them easily scalable. We release our code athttps://github.com/openai/improved-diffusion