Abstract

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

 

Quick Read (beta)