Generative models that gradually add noise to categorical data and learn to reverse the process for sampling.