Procedural content creation in the age of generative AI

Abstract

Within the domain of Computational Creativity, Procedural Content Generation (PCG) in multimodal domains presents challenges in modal alignment, diversity, and coherence. Although recent generative AI models, including Large Language Models (LLMs), have demonstrated improved cross-modal capabilities, they frequently struggle with balancing autonomy and control to produce creative, high-quality artefacts. This thesis explores how generative AI can be structured to enhance diversity, coherence, and evaluation in multimodal PCG, with a focus on text and image generation. This work addresses the fundamental question: how can AI be guided to generate novel, high-quality, and semantically aligned artefacts across modalities? To this end, the research introduces four key approaches: Firstly, evolutionary quality diversity algorithms are combined with generative models to improve output diversity in image generation while maintaining semantic relevance. Secondly, the MAP-Elites algorithm is augmented with Transverse Assessment to explore multimodal search spaces, discover diverse solutions, and improve coherence across text and image generation. The third approach demonstrates CrawLLM, a zero-shot generative pipeline that uses LLMs to orchestrate the creation of game levels, narrative elements, and visual assets for a video game, ensuring thematic consistency and content structure. The last approach builds on CrawLLM by incorporating the LLM-driven pipeline into an adaptive evaluation loop to dynamically assess and regenerate underperforming artefacts, improving overall quality and thematic fit. These contributions advance methodologies for multimodal PCG by integrating generative AI with structured evaluation and optimisation techniques. The findings provide novel insights into search-based creativity, LLM-driven generative orchestration, and AI-assisted evaluation frameworks, with applications in computational creativity, game design, and digital media generation.