A study on three-stage integration of generative artificial intelligence in architectural design based on human-in-the-loop

The swift advancement of artificial intelligence has posed new demands on the architectural industry’s transformation. Generative AI (GAI) enables the digitization of architectural design experience, offering fresh inspiration for architectural scheme generation. This study develops a human-in-the-loop three-phase framework that distinguishes concept generation, scheme deepening, and result expression while adding explicit decision gates and cross-cutting controls for architectural knowledge, performance verification, data provenance, and professional accountability. This study starts by leveraging VOSviewer software to visualize and analyze keyword co-occurrences in GAI-related literature, summarizing GAI’s development trends and hotspots. Then, it outlines GAI’s application stages in architectural design, namely concept generation, scheme deepening, and result expression. It also lists 20 generative design software useful for architectural scheme design. Subsequently, taking Midjourney, NOAH, and Stable Diffusion as examples, it explores GAI’s specific application scenarios across these three design stages. Finally, it proposes a new architectural design model shaped by GAI. The contribution is not a new sequence of conventional design stages, but an operational phase-control matrix that specifies inputs, outputs, human decisions, verification requirements, and feedback loops. It accelerates inspiration image generation, thereby influencing design thinking and processes, meanwhile it provides a cautious, reproducible framework for selecting and governing GAI tools in architectural design.