Laying the foundations for using generative AI images in architectural research: do images convey the intended spaces and ambiances?

Abstract The advent of Generative Artificial Intelligence (GenAI) models, such as Stable Diffusion, Open AI's DALL-E2, and MidJourney, has opened the door to a huge range of new possibilities in architectural research and practice. Before architects and researchers can fully leverage these models, it is crucial to assess their proficiency in generating images that accurately depict real spaces and ambiances. Here, we assessed the proficiency of DALL-E2 in depicting intended spaces and intended ambiances, using real photographs from Google Images as a comparative benchmark. Images of eighteen distinct home spaces, each intended to evoke three different ambiances, were generated and presented to human observers. The images were evaluated in terms of their perceived realism, their accuracy in depicting the space and ambiance intended. Findings highlight GenAI's capability in depicting spaces and point to its potential in convincingly conveying ambiances. Beyond the substantive findings, this research highlights the importance of evaluating the rapidly emerging tools enabled by recent developments in AI.