TLDR: This paper introduces an AI workflow that uses Large Language Models (LLMs) to generate architectural floor plans from natural language descriptions. The system converts text prompts into structured layout data, which is then used to create 3D models in Autodesk Revit, complete with parametric attributes. It includes a unique furniture placement algorithm to ensure spatial coherence and is designed for open-source reproducibility, addressing limitations of existing commercial tools.
The world of architectural design is witnessing a significant transformation with the advent of artificial intelligence. Traditionally, drafting architectural floor plans has been a labor-intensive and time-consuming process, often prone to inefficiencies. A new AI-powered workflow, detailed in a recent research paper, aims to streamline this by using Large Language Models (LLMs) to automatically generate schematic architectural floor plans from simple natural language prompts.
This innovative system, named “Text-to-Layout,” interprets textual input to produce comprehensive layout options. These options include the placement of walls, doors, windows, and even furniture arrangements. The workflow integrates several key components: sophisticated prompt engineering to guide the LLM, a specialized algorithm for refining furniture placement, and Python scripting to ensure compatibility with professional design tools like Autodesk Revit.
One of the primary goals of this research is to bridge the gap between language-based design concepts and their practical realization in architectural modeling environments. Instead of generating final, polished models, the system focuses on providing structured draft layouts. This approach empowers architects by offering a robust starting point for the design process, facilitating early-stage ideation and functional spatial planning. Crucially, the generated models retain all Revit-native parametric attributes, allowing for direct integration into professional Building Information Modeling (BIM) workflows without the need for extensive manual adjustments.
Existing AI tools for architectural design often come with limitations. Many are commercial and closed-source, restricting access and customization. Furthermore, even when these tools integrate with CAD and BIM environments, their outputs frequently lack the full range of parametric information, such as complete element types, material specifications, and constraint definitions. This often necessitates substantial manual enrichment to meet construction-ready BIM standards.
The “Text-to-Layout” workflow addresses these challenges by offering an openly documented process. This transparency ensures reproducibility and adaptability for other researchers, fostering community-driven development and lowering integration barriers. The system’s ability to produce Revit-native models with complete parametric information is a significant contribution, enabling immediate use in professional BIM environments.
The methodology begins with a single, well-structured natural language prompt from the user. This prompt is then processed by an LLM, such as GPT-4o, which consistently demonstrated the most reliable outputs in experiments. The LLM translates the textual requirements into a structured JSON file. This file acts as a digital blueprint, encoding spatial coordinates for various architectural elements. For instance, walls, doors, and windows are defined by start and end coordinate pairs, while furniture items are grouped by room with their center-point coordinates.
To convert this JSON-based layout into an executable Revit model, a second-stage prompt instructs the LLM to generate Python code compatible with the Revit API. This script then programmatically constructs the 3D model in Autodesk Revit, significantly reducing manual drafting efforts while maintaining spatial logic and functional clarity.
A notable feature of the workflow is its greedy algorithm for furniture placement adjustment. While LLMs can generate initial layouts, they might not always place furniture in ideal positions, leading to overlaps or illogical arrangements. This algorithm incrementally adjusts AI-generated furniture positions, evaluating clearances and moving items towards intended placements, such as against a wall or centered in a room, until all spatial constraints are met. This ensures that each furniture item is legally within its designated room, has sufficient circulation space, and aligns neatly where appropriate.
A case study involving a 30×40 ft single-story residential building demonstrated the system’s effectiveness. Early experiments revealed issues like ambiguous room boundaries, furniture overlaps, and illogical door placements. Through iterative refinement of prompt syntax and the implementation of post-processing enhancements within the Revit Python script, the system significantly improved the quality and usability of the generated floor plans. These enhancements included evaluating rooms based on expected furniture content, ensuring proper door and window placement, and using a grid-based pathfinding algorithm to assess connectivity between functional spaces.
The automation pipeline progresses through several steps: first, wall structures are generated from JSON inputs; then, doors and windows are automatically inserted; next, furniture items are initially placed; and finally, furniture positions are refined for optimal spacing, clearance, and usability. This systematic approach, as detailed in the research paper available at arXiv:2509.00543, ensures a coherent and functional design.
Also Read:
- Boosting 3D CAD Part Retrieval with Reflective AI: A Training-Free Framework
- Agentic Reinforcement Learning: Empowering LLMs as Autonomous Decision-Makers
In conclusion, this workflow represents a significant step forward in automated architectural design. It provides a transparent and reproducible method for generating BIM-compatible 3D floor plans from natural language, offering a powerful tool for architects to enhance efficiency and creativity in the early stages of design. Future research aims to expand its applicability to more complex geometries, integrate building code compliance, and explore adaptive prompt generation methods.
