Improved Denoising Diffusion Probabilistic Models

ABSTRACT

Denoising diffusion probabilistic models (DDPMs) have recently demonstrated high-quality image generation, but they suffer from notoriously slow sampling times and sub-optimal log-likelihoods. We propose several improvements to DDPMs to address these issues. By learning the variances of the reverse diffusion process and employing a novel cosine noise schedule, we show that our models can achieve competitive log-likelihoods while retaining high sample quality. Crucially, our modifications allow for a drastic reduction in the number of sampling steps, enabling fast, high-fidelity image generation that matches or exceeds the performance of state-of-the-art GANs.