Researchers at the Japan Advanced Institute of Science and Technology have developed an artificial intelligence system (AI) that turns simple text descriptions into accurate architectural images. The new tool solves a major problem with existing AI image generators, which often produce visually appealing pictures that do not follow real building rules, such as having the correct number of floors or properly placed windows.
Diffusion models are AI systems trained to create images, art, or video from text prompts. However, they frequently ignore precise structural details because their training data lacks clear information about architecture. The new system uses a method called retrieval-augmented generation. This approach combines the text prompt with real examples pulled from special architectural databases, helping the AI reference actual buildings while it creates the image.
The process works in three steps that mirror how architects design. First, the system creates a simple structural sketch that captures the basic shape and correct number of floors. Next, it adds detailed elements such as windows, doors, and facades using a database of real building parts. Finally, it combines everything to produce a realistic, high-quality rendering that matches the original description.
Structural accuracy breakthrough
The researchers built three custom datasets: one with 2,200 basic building shapes, one with 4,000 architectural components like windows and entrances, and one with 1,600 pairs linking sketches, text, and final images. In tests focused on campus-style buildings, the system achieved 70.5 percent accuracy in correctly arranging floors and vertical elements. It also scored higher than standard diffusion models on measures of structural correctness, visual quality, and how well the image matched the text prompt. A group of 56 architecture students rated the results highly, giving average scores above 4 out of 5 for quality and accuracy.
This approach makes early-stage architectural design faster and more reliable. Designers can quickly generate and revise ideas during client meetings, while planners can compare many options under the same rules. The work could help smaller teams and individual designers create realistic visualizations without expensive software or large professional support. This study is published in Frontiers of Architectural Research.