Diffusion Models

Overview

Diffusion models. The encoder (forward, or diffusion process) maps the input x through a series of latent variables z1 . . . zT . This process is prespecified and gradually mixes the data with noise until only noise remains. The decoder (reverse process) is learned and passes the data back through the latent variables, removing noise at each stage. After training, new examples are generated by sampling noise vectors zT and passing them through the decoder.