# AI system generates accurate building designs from text prompts by retrieving real architectural examples

> Japanese researchers built an AI that turns text descriptions into architectural renderings with 70.5% accuracy on floor counts. In tests with 56 graduate students, image quality and detail scored above 4 out of 5.

- Type: AI news
- Page: https://completeaitraining.com/news/ai-system-generates-accurate-building-designs-from-text/
- Published: 2026-03-31
- Updated: 2026-04-04
- Author: Jeroen Erne
- Jobs: Science and research
- Reading time: 1 min

### Researchers Build AI System That Generates Accurate Building Designs From Text

A team at Japan Advanced Institute of Science and Technology has developed an AI system that converts text descriptions into realistic architectural renderings with correct floor counts and facade details. The work addresses a persistent problem: existing text-to-image models often fail at basic requirements, such as generating a five-story building when asked.

The system uses a retrieval-augmented generation approach, combining text prompts with reference data from real architectural examples. This method helps the AI understand spatial constraints that standard training datasets lack, such as floor placement and window positioning.

#### How the System Works

The framework mirrors how architects actually design buildings. It begins with a simple structural sketch showing overall form and floor count. The sketch is then refined by adding specific elements-windows, doors, facade components-drawn from a database of real building components. Finally, the refined sketch and original text description combine to produce a finished rendering.

Researchers tested the system on campus building designs, where accuracy in floor count and entrance placement matters. They built three datasets: one with 2,200 building images, another with 4,000 images showing window and entrance variations, and a third pairing 1,600 sketches with text prompts and final renderings.

#### Results and Real-World Application

The system achieved 70.5% accuracy in generating the correct vertical configuration. In a study with 56 architecture and design graduate students, participants rated the system above 4 out of 5 for image quality, prompt alignment, and architectural detail accuracy.

The work could shorten design cycles significantly. Architects could revise schemes during client meetings without manual rendering. Planners and developers could visualize dozens of design alternatives before committing to detailed modeling.

The research was [published in Frontiers of Architectural Research](https://completeaitraining.com/tag/generative-ai-and-llm/) on March 26, 2026, by a team led by Associate Professor Haoran Xie from JAIST and Associate Professor Ye Zhang from Tianjin University.

"High-quality architectural visualization currently requires significant expertise and expensive software," Xie said. "With this work, individual designers and smaller teams can participate meaningfully in their own built environments without needing a large professional team."

---

This is the Markdown form of https://completeaitraining.com/news/ai-system-generates-accurate-building-designs-from-text/. The whole site is mapped at https://completeaitraining.com/llms.txt.