Game effect sprite generation with minimal data via conditional GAN

Abstract

Image generation using convolutional neural networks has gained popularity in various industries in the field of computer vision over the last decade. However, thus far, generative adversarial networks (GAN) have not been utilized to their maximum potential to generate computer game sprites, especially game effects. This is because typically the amount of available open-source game sprite data is considerably less than that of other problem domains such as image classification or object detection tasks, which use hundreds of thousands of images to train classification models. This study demonstrates that GAN can be utilized to generate game effect sprites adequately with a small set of input data, thus increasing the productivity of large-scale 2D image modification work in game development. In this study, we propose a simple 2D game effect sprite generation technique called a Game Effect Sprite Generative Adversarial Network (GESGAN). The proposed model generates a new style-translated image based on the structure of an effect image and a style label. The model consists of an encoder designed to extract the feature representation of an input image and a label; a decoder to generate a synthetic image; and a discriminator that learns to determine the authenticity of images with an auxiliary classifier for the label. Experimental results show that GESGAN is capable of reliably generating style-translated images for various shapes of object images and drawing styles and can perform 2D image sprite generation and modification tasks in near real-time, thereby reducing game development costs.