On the Challenges of Generating Pixel Art Character Sprites Using GANs

Abstract

We pose the problem of generating pixel art character sprites facing one side (e.g., right), given their images facing another one (e.g., front), as an image-to-image translation task and investigate the use of the Pix2Pix architecture to solve it. Aiming to improve the results over unseen data, we propose and investigate two architecture modifications: (a) representing images using color palettes and (b) adding a histogram loss term to the generator. We compared the results qualitatively and quantitatively using FID and L1 distances between the generated and target images. Results indicate that representing images with color palettes encourages overfitting, and the histogram loss leads to slightly improved results.