Comparative and experimental analysis of leading text-to-image generative artificial intelligence models for regional residential architectural designs

• Evaluates the performance of leading generative AI models (Midjourney, DALL·E 2, and Stable Diffusion) for regional residential architectural design. • Provides a comparative analysis based on design accuracy, cultural representation, and architectural fidelity. • Uses a mixed-methods approach combining expert visual evaluation and user surveys. • Identifies strengths and limitations of e model in capturing regional architectural elements. • Offers insights into the integration of AI in professional architectural workflows. The integration of generative artificial intelligence into architectural design workflows is transforming the way designers conceptualize and visualize ideas. Despite rapid advancements in text-to-image models such as Midjourney, DALL·E 2, and Stable Diffusion, limited research has investigated their comparative performance in domain-specific applications like regional architecture. This study addresses this gap by evaluating the capabilities of these models in generating modern villa designs inspired by Emirati vernacular traditional architecture. A standardized experimental setup involving a total of 44 design students was employed to generate and assess both interior and exterior design outputs using textual input prompts. The evaluation focused on four key criteria: image quality, architectural accuracy, prompt adherence, and creativity. Midjourney demonstrated the strongest overall performance, achieving the highest scores in image quality, prompt adherence, and creativity, with outputs consistently visually compelling, imaginative, and closely aligned with design prompts, resulting in an impressive overall average score of 4.75. DALL·E 2 also performed well, slightly outscoring Midjourney in architectural accuracy, with a solid overall average of 4.13 despite being somewhat less creative and visually refined. In contrast, Stable Diffusion lagged across all categories, especially in prompt adherence and creativity, with outputs lacking the distinctiveness and detail seen in the other models, leading to the lowest overall average score of 2.75 and demonstrating its relative limitations without further fine-tuning. A major limitation noted across all models was their inability to consistently and accurately represent Emirati architectural elements, underscoring the need for future development of custom models fine-tuned on region-specific architectural design data. These findings emphasize the importance of localized training data and culturally aware model development to advance generative AI applications for architecture.