EvolGAN: Evolutionary Generative Adversarial Networks

TL;DR

EvolGAN introduces an evolutionary approach combined with a quality estimator to enhance image quality in GANs trained on small or challenging datasets, maintaining diversity and outperforming previous models.

Contribution

The paper presents a novel evolutionary method that improves GAN image quality using a quality estimator, applicable to any GAN and scorer, especially for limited datasets.

Findings

Human raters preferred EvolGAN images significantly

EvolGAN produces higher quality images while preserving diversity

Effective on various datasets including Cats, FashionGen, Horses, Artworks

Abstract

We propose to use a quality estimator and evolutionary methods to search the latent space of generative adversarial networks trained on small, difficult datasets, or both. The new method leads to the generation of significantly higher quality images while preserving the original generator's diversity. Human raters preferred an image from the new version with frequency 83.7pc for Cats, 74pc for FashionGen, 70.4pc for Horses, and 69.2pc for Artworks, and minor improvements for the already excellent GANs for faces. This approach applies to any quality scorer and GAN generator.