Variational Autoencoder Model for Image Processing Methods in Game Design

This paper investigates the use of Variational Autoencoders (VAEs) as a deep learning-based generative framework for AI-assisted image processing in game design, focusing on the procedural generation of stylized visual assets. Its application in AI-assisted image generation in game design for the generation of diverse stylized visual assets is explored in this paper. In order to learn stylistic consistent content and generate new art for the Ethereal Monsters game, we propose a deep learning-based generative approach using VAEs to learn latent representations of existing game art. Pre-processing of a curated dataset of 10,000 game sprites spans parsing colour palette and sprite patterns, creating an adapted palette for less sparse variants of sprites, and creating training and testing sets through pooling sprites into images and grouping images for a generation. A convolutional VAE architecture is trained, its (re)construction loss and visual fidelity are evaluated, a prospective error correction test is performed, and the results are analysed. We show that the VAE model can effectively capture the main features of 2D game sprites and if iterated numerous times not only does it produce an endless number of variations, but it also keeps the game-specific aesthetic properties. It is compared with existing generative methods and improved visual coherence is found, whilst diversity is saturated. It adds to the exploration of AI-driven creativity in game design, in particular an increasing number of ways to generate assets and prototypes in a scalable way.