Generative models that create images by learning to denoise random noise in a compressed latent space rather than pixel space.