Learning a Generative Adversarial Network for High Resolution Artwork Synthesis

Artwork is a mode of creative expression and this paper is particularly interested in investigating if machine can learn and synthetically create artwork that are usually non- figurative and structured abstract. To this end, we propose an extension to the Generative Adversarial Network (GAN), namely as the ArtGAN to synthetically generate high quality artwork. This is in contrast to most of the current solutions that focused on generating structural images such as birds, flowers and faces. The key innovation of our work is to allow back-propagation of the loss function w.r.t. the labels (randomly assigned to each generated images) to the generator from the categorical autoencoder-based discriminator that incorporates an autoencoder into the categorical discriminator for additional complementary information. In order to synthesize a high reso- lution artwork, we include a novel magnified learning strategy to improve the correlations between neighbouring pixels. Based on visual inspection and Inception scores, we demonstrate that ArtGAN is able to draw high resolution and realistic artwork, as well as generate images of much higher quality in four other datasets (i.e. CIFAR-10, STL-10, Oxford-102 and CUB-200).