Generating images with recurrent adversarial networks

TL;DR

This paper introduces a recurrent generative model inspired by gradient-based image optimization, trained with adversarial methods to produce high-quality images, and proposes a new quantitative comparison method for adversarial networks.

Contribution

It presents a novel recurrent generative model for image synthesis and a new adversarial network evaluation framework.

Findings

The recurrent model generates high-quality images.

Adversarial training effectively trains the recurrent generator.

A competitive evaluation method for adversarial networks is proposed.

Abstract

Gatys et al. (2015) showed that optimizing pixels to match features in a convolutional network with respect reference image features is a way to render images of high visual quality. We show that unrolling this gradient-based optimization yields a recurrent computation that creates images by incrementally adding onto a visual "canvas". We propose a recurrent generative model inspired by this view, and show that it can be trained using adversarial training to generate very good image samples. We also propose a way to quantitatively compare adversarial networks by having the generators and discriminators of these networks compete against each other.