Text-to-Image Generative Αrtificial Intelligence As Conceptual Mechanism For Architectural Design

This study investigates the potential of text-to-image generative AI applications to be used as exploration mechanisms for architectural design rather than mere visualisation tools. We examine whether text-to-image generative AI applications can "understand" architecture's spatial and geometric configuration in meaningful and creative ways. Our experimental method used a series of architectural plans and artistic patterns as image prompts, combined with varied text prompts and parameters to evaluate Midjourney's capacity for geometric and spatial interpretation and creativity. The experiments focused on the application's capacity to (1) generate diagrams from architectural plans, (2) create variations of two-dimensional drawings and patterns, and (3) transform two-dimensional layouts into three-dimensional spaces. Results indicate that Midjourney can generate meaningful abstractions and spatial configurations but struggles with strict spatial fidelity, particularly for asymmetrical or fragmented layouts. The study highlights the significance of short prompts and increased image weight parameters for improving fidelity and structural coherence. While Midjourney exhibits potential as a conceptual tool for highly creative architectural exploration, it lacks the precision required for deterministic spatial manipulation. Dall-E performs significantly better in fidelity and alignment to the original drawing, but further research is needed to determine its potential use as an assistive tool for architectural design. The findings contribute to the discourse on AI-assisted architectural design for creative ideation and suggest further research into the design capacities of other similar models.