Generative Text-to-Image Models in Architectural Design: A Study on Relationship of Language, Architectural Quality and Creativity

Text-guided generation of images with deep learning technology has made significant advancesand has seen an increasing interest since 2021. With these mostly web-based models, users cansynthesise photorealistic and high-quality digital images from natural language descriptionswith no or little understanding of the underlying technology. Although these AI technologiesare in the early phases, there is already an explosion in AI-generated architectural activity.While generative AI technologies propose a new design method for designers and architects, itwill undoubtedly redefine the skills, knowledge and competencies that designers shouldequipped with. This research focuses on understanding the “artificial intelligence – architect”interaction as a design method, specifically the “language as a design driver”, and interrogatesthe role of the designer in AI-driven design. In the context of the research, the textual inputs(“prompts”) and the outputs of the architectural design studies of 36 subjects generated inMidjourney – a text-to-image latent diffusion model – were analysed in terms of the possiblerelationships between the language of the prompts, (1) prompt length, (2) descriptive language,(3) specific architecture-related indicators, and the quality of the outputs in two terms ofarchitectural quality and architectural creativity.