They are a way of training generative models similar to GANs or autoencoders. My understanding is that if you train an autoencoder with a gaussian likelihood t

They are a way of training generative models similar to GANs or autoencoders. My understanding is that if you train an autoencoder with a gaussian likelihood then you will tend to get fuzzy samples, but using an iterative process where each step is a gaussian conditioned on the previous step can give you nicer samples.