Automatic generation of graphical game assets using GAN

Abstract

This paper presents an application of the Generative Adversarial Networks (GAN) approach to automatically generate realistic-looking fantasy and science fiction game icons. In this study, we explored ways to avoid the manual drawing of graphics, instead producing synthetic images indistinguishable from artist-made ones. We developed the method based on multiple experiments and commonly known GAN improvement techniques. To enhance the training process an altered version of the Adaptive Discriminator Augmentation was used. Achieved results were measured by Fretchet Inception Distance (FID) metric. As realism is a subjective metric, a visual evaluation was performed, where a group of 50 observers assessed a mix of generated and original (drawn) images in order to recognize how many of the generated images are indistinguishable from the human-created ones. As a result 69% of generated icons were perceived as ‘real' ones, in comparison to 70% for drawn images. These outcomes clearly indicate that generated illustrations can be of high quality and be indistinguishable from drawn graphics. The developed solution enables users to significantly decrease costs of asset creation in commercial projects, and could even potentially empower designers to produce more distinguished and unique experiences for each player thanks to a vast amount of generated graphics.