Analysis of state-of-the-art deep generative models for images

Abstract

In this work we will try to break down the fundamentals of deep generative models for image generation. For the sake of simplicity and understanding we will be using the MNIST dataset, which can be easily trained and understood. The algorithms that we will be explaining and comparing are an autoregressive model, a flow-based model, a variational autoencoder, and a generative adversarial network. Within each of these categories, we have chosen simple algorithms for the sake of clarity as well. Keep in mind that this is a wide but introductory work to the area of image generation. However, these algorithms were certainly state-of the-art around 2016.