LICENSE

Abstract

The Generative Adversarial Network (GAN) is a machine learning model that has introduced the possibility of artificial intelligence-created art. However, direct generation methods fail to create convincing artworks that are realistic and structurally well-defined. Here, we present a GAN variant, CompositionGAN (CGAN), which originates edge-defined, artistically-structured paintings without a dependence on supervised style transfer. CGAN is composed of two stages, edge generation and edge-to-painting translation, and is trained on a new dataset of traditional Chinese landscape paintings never before used for generative research. A 242-person human Visual Turing Test study reveals that CGAN paintings are mistaken as human artwork over 55% of the time, significantly outperforming paintings from a baseline GAN model. Our work highlights the importance of artistic composition in art generation and takes an exciting step toward computational originality.