The Possibilities of Text-to-Image Tools for the Generation of Floor Plans
-
-
[1] Universidad de Zaragoza
Universidad de Zaragoza
Zaragoza, España
-
- Localización: Graphic horizons, Vol. 2, 2024 (Graphics for education and production / Luis Hermida González (ed. lit.), Joao Pedro Xavier (ed. lit.), Antonio Amado (ed. lit.)), ISBN 978-3-031-57575-4, págs. 297-307
- Idioma: inglés
- Enlaces
-
Resumen
This study builds on previous research to assess whether text-to-image technology can correctly generate images of residential floor plans. Three tools are tested: Midjourney, Stable Diffusion and Dall-E. The process involved: (1) using reference images to generate text descriptions, (2) crafting prompts from these descriptions and testing them on the three AI systems, (3) merging text requests with reference images, and (4) using hand-drawn sketches to create technical architectural drawings.In general, the tools showed potential but were deemed not yet suitable for producing architectural designs due to a lack of syntactic and functional logic. Midjourney emerged as the most effective, consistently generating 2D planimetric images and producing quality results when combining textual descriptions with reference images. On the other hand, Dall-E underperformed in responding to text requests and deviated significantly from delivering the desired images, although it excelled at describing images via ChatGPT, a task at which Midjourney faltered. Stable Diffusion was noted for striking a balance, offering quality close to Midjourney and better text descriptions through Artbot. It also showed promise with its unique ability to create images from hand-drawn sketches, a feature not available in the other tools.The improvements shown by those tools within a short time suggest that they will continue to advance and might soon generate accurate architectural drawings from text descriptions and rough sketches, constituting an important help tool for architects.
