PixelatedGAN: A Generative Adversarial Network For Pixelated Images

Abstract

This work presents a generative adversarial network which generates images in a pixelated output space. The results of this project have both utility in allowing for more accurate training and generation when based upon input images which are pixelated, and also for creating uniquely intelligently pixelated outputs when trained on non-pixelated input images. Pixelated images are used often in video games and art. Pixelated images are also uniquely useful for image compression since they do not lose any visual information when made smaller. At the minimum, a pixelated image can be compressed to a quarter of its original size without losing any data. Several attempts have been made by researchers in the field of generative AI, prior to this paper, to create a neural network which generates pixel art. However, these attempts focused more on the artistic value of images stylistically similar to pixelated images rather than on actually having the network create images which were properly pixelated.