User-centred artificial intelligence for game design and development with GAGeTx: Graphical Asset Generation and Transformation

Abstract

In an increasingly digitalised world, visual media is utilised in a wide array of forms. This visual content ismade up ofmany individual elements, referred to as graphical assets. A wide variety of well established and nascentmethods, referred to as graphical asset generators (GAGs), can be used to automate the production of graphical assets. Video games are a popular and growing application of graphical assets, requiring copious amounts of 3D and 2D visual content. The aim of this thesis is to examine how generativemethods can be applied to the creation of graphical assets for games, and to discover how game designers and developers choose to utilise them. This is achieved through the pursuit of 5 research objectives: first, collating and examining the state-of-the-art of GAGs in the literature; second, developing a framework for using, implementing and evaluating GAGs (GAGeTx); third, obtaining user needs and preferences through a user experiment; fourth, developing a proof-of-concept prototype tool, serving to validate GAGeTx; fifth, refining the framework through further user experimentation using the prototype tool. Contributions of this thesis include: the GAGeTx framework; a systematic literature review state-of-the-art GAG methods; empirical findings on user needs and requirements; a novel, game engine-integrated framework and prototype tool for sword generation; and a method for dataset creation, tailored to unsupervised deep learning for GAG tasks. The GAGeTx framework offers a comprehensive categorisation and conceptualisation of GAG methods, which allows researchers and practitioners to identify or create themost appropriate GAGmethods given their needs and requirements, through a step-wise process built on empirical findings. This is supported by the integration of GAG evaluationmetrics and the consideration of user pipeline applications. The systematic literature review consolidates fragmented research fromvarious domains into a unified taxonomy, identifying key aspects of GAGs. This facilitates cross-over between domains and provides a valuable entry point for both new researchers and practitioners in the field ofGAGresearch. Empirical findings fromuser studies provide guidelines for integrating GAG tools into game design and production pipelines withminimal friction and facilitating the adaptation ofGAGresearch into practical tools. In addition, they identify the appropriatemetrics for evaluating the strength and utility ofGAGtools based on their technique, further aiding in the benchmarking and improvement of GAGmethods. The prototype, named Swordgen, allows users to generate varied sword assets for games via several generative techniques at different levels of user initiative, providing and validating a configurable and extendable framework for game-engine integrated GAG tools. The dataset creation method enables the creation of bespoke content and style specific datasets for training unsupervised deep-learning-based GAGs. Through selecting specific data sources, including concept-art fromthe user’s current project, users can control the design constraints of a GAG model, without compromising on dataset size.