ArtGAN: Artwork Synthesis with Conditional Categorical GANs

This paper proposes an extension to the GenerativeAdversarial Networks (GANs), namely as ARTGAN to synthetically generatemore challenging and complex images such as artwork that have abstract characteristics. This is in contrast to most of the current solutions that focused on generatingnatural images such as room interiors, birds, flowers and faces. The key innovationof our work is to allow back-propagation of the loss functionw.r.t. the labels (randomly assigned to each generatedimages) to the generatorfrom the discriminator. With the feedback from the label information, the generatoris able to learn faster and achieve better generatedimage quality. Empirically, we show that the proposed ARTGAN is capable to create realistic artwork, as well as generatecompelling real world images that globally look natural with clear shape on CIFAR-10.