A Case Study of Generative Adversarial Networks for Procedural Synthesis of Original Textures in Video Games

Abstract

Procedural content generation is helping game developers to create significant quantity of high quality dynamic content for video games at a fraction of cost of the traditional methods. Procedural texture synthesis is a sub category of procedural content generation which helps video games to have significant variations in textures of the environments and the objects across the progress of the game and to avoid repetition. Generative Adversarial Networks are a class of deep learning algorithms which are capable of learning the patterns and creating new patterns. In this paper, Generative Adversarial Networks is used for procedural content generation for original textures synthesis for video game development. This method is used by video game designers for autonomous redesigning of objects and environment textures. This process saves significant time and cost in video game development. The particular attention in this paper is on procedural synthesis of ground surface textures. The generated texture samples are visually acceptable and have mean score of 2.45 with 0.1 standard deviation after 2K iteration. Also the discriminator loss of generated samples reached 0.74 at the final stage of training. The proposed framework can be used as an effective procedural texture synthesis framework for video game design and development.