Improved Denoising Diffusion Probabilistic Models

The paper builds on denoising diffusion probabilistic models (DDPMs), a class of generative models already shown to produce excellent samples, and introduces a few simple modifications to improve them. With these changes the models achieve competitive log-likelihoods while maintaining high sample quality, and by learning the variances of the reverse diffusion process rather than fixing them, the authors enable sampling with an order of magnitude fewer forward passes at negligible cost to sample quality.

These improvements matter for practical deployment, since far fewer forward passes are needed to generate samples. The authors also use precision and recall to compare how well DDPMs and GANs cover the target distribution, and show that both sample quality and likelihood scale smoothly with model capacity and training compute, indicating the models are easily scalable; they released their code.