Abstract: We propose to use a quality estimator and evolutionarymethods to search the latent space of generative adversarial networkstrained on small, difficult datasets, or both. The new method leads tothe generation of significantly higher quality images while preserving theoriginal generator’s diversity. Human raters preferred an image from thenew version with frequency 83.7% for Cats, 74% for FashionGen, 70.4%for Horses, and 69.2% for Artworks - minor improvements for the alreadyexcellent GANs for faces. This approach applies to any quality scorer andGAN generator.