Improved generator objectives for GANs

TL;DR

This paper introduces a new framework for understanding GAN training as density ratio estimation and divergence minimization, leading to novel generator objectives that improve sample quality and diversity.

Contribution

It provides a theoretical framework for GAN training, explaining existing issues, and proposes new generator objectives targeting various $f$-divergences for better results.

Findings

New generator objectives improve sample diversity

Framework explains mismatch in GAN objectives

Models achieve better sample quality or diversity

Abstract

We present a framework to understand GAN training as alternating density ratio estimation and approximate divergence minimization. This provides an interpretation for the mismatched GAN generator and discriminator objectives often used in practice, and explains the problem of poor sample diversity. We also derive a family of generator objectives that target arbitrary $f$-divergences without minimizing a lower bound, and use them to train generative image models that target either improved sample quality or greater sample diversity.