Human and machine collaboration for painting game assets with deep learning

Abstract

Modern games are among the most intricate pieces of software ever devised. However, artists have little to no means of porting their creative work from title to title as the aesthetics change and older models grow to look obsolete. Zooming into the topic of assisted game art generation, the literature is notably scarce, although advances towards automated asset generation are of paramount interest to the field. In this work, we investigate the use of deep learning algorithms to create pixel art sprites from line art sketches to produce artwork of sufficient quality to be used within a game product with little to no manual editing by human artists. Such a problem contrasts with well-known tasks studied in the literature, which are based on natural pictures, boast massive datasets, and are much more tolerant to noise. In addition, we conducted a case study of applying current technology to the drawing pipeline of an upcoming game title, attaining useful and positive results that may fast-track the game development, supporting the argument that current image generation state-of-the-art is ready to be used in some real-world tasks.