The importance of context in image generation : a case study for video game sprites

Abstract

In recent years text-to-image generative models have found a wide range of applications, including the generation of video game assets such as sprites. However, sprites are often reused throughout a game, leading to repetitive visuals. Creating custom sprites tailored to different in-game environments can be a resource-intensive task when done manually. In this paper, we address the challenge of generating visually appealing sprites that not only align with the overall theme of the game but also adapt to the varying environments within it. We investigate how different contextual details of these environments, provided as prompts to a text-to-image model, influence the resulting sprites. Our approach is demonstrated through sprite generation for the video game design tool LLMaker. Experiments using objective performance metrics confirm the effectiveness of our method, while a user study shows that the generated sprites resonate with players.