[Paper Review] Architext: Language-Driven Generative Architecture Design

Theodoros Galanos, Antonios Liapis+1 more|arXiv (Cornell University)|Mar 13, 2023

BIM and Construction Integration4 citations

TL;DR

Architext introduces a language-driven generative design system that uses fine-tuned large language models (LLMs) to generate valid and diverse residential floorplans exclusively from natural language prompts. The method achieves near-100% layout validity and up to 83% semantic correctness with the largest GPT-J model, demonstrating that LLMs can effectively learn architectural geometry through semantic supervision alone.

ABSTRACT

Architectural design is a highly complex practice that involves a wide diversity of disciplines, technologies, proprietary design software, expertise, and an almost infinite number of constraints, across a vast array of design tasks. Enabling intuitive, accessible, and scalable design processes is an important step towards performance-driven and sustainable design for all. To that end, we introduce Architext, a novel semantic generation assistive tool. Architext enables design generation with only natural language prompts, given to large-scale Language Models, as input. We conduct a thorough quantitative evaluation of Architext's downstream task performance, focusing on semantic accuracy and diversity for a number of pre-trained language models ranging from 120 million to 6 billion parameters. Architext models are able to learn the specific design task, generating valid residential layouts at a near 100% rate. Accuracy shows great improvement when scaling the models, with the largest model (GPT-J) yielding impressive accuracy ranging between 25% to over 80% for different prompt categories. We open source the finetuned Architext models and our synthetic dataset, hoping to inspire experimentation in this exciting area of design research.

Motivation & Objective

• To develop a flexible, open-source system that enables language-driven generative design in architecture using only natural language prompts.
• To evaluate the capacity of fine-tuned large language models to generate valid and semantically accurate residential floorplans without visual input.
• To investigate the impact of model size and architecture on semantic accuracy and design diversity in architectural generation.
• To provide a publicly available synthetic dataset of annotated floorplans and fine-tuned LLMs to advance research in semantic generative design.
• To explore the potential of LLMs as scalable, accessible, and intuitive tools for performance-driven and sustainable architectural design workflows.

Proposed method

• Fine-tune a range of pre-trained language models (from 120M to 6B parameters) on a synthetic dataset of residential floorplans paired with natural language descriptions of spatial and geometric properties.
• Represent architectural layouts using a semantic, text-based geometry representation that encodes spatial relationships and room configurations.
• Train the models to generate floorplan descriptions conditioned on input prompts, using autoregressive language modeling to predict sequential token outputs.
• Evaluate model outputs by converting generated text back into geometric layouts and validating structural validity and semantic alignment with input prompts.
• Use a diverse set of 56 natural language prompts to test generalization across various design constraints and intent categories.
• Leverage established LLMs such as GPT-J and smaller variants to assess scalability effects on performance and diversity.

Experimental results

Research questions

• RQ1 Can fine-tuned large language models generate valid and semantically accurate residential floorplans using only natural language prompts?
• RQ2 How does model size influence the semantic accuracy and diversity of generated architectural layouts?
• RQ3 To what extent can LLMs learn and generalize architectural design rules from purely semantic supervision without visual input?
• RQ4 How does the performance of LLMs vary across different types of design constraints expressed in natural language?
• RQ5 Can a synthetic, semantic-annotated dataset enable effective zero-shot or few-shot generalization in architectural layout generation?

Key findings

• All fine-tuned models achieved a layout validity rate of 99% or higher, indicating robustness in generating structurally sound residential floorplans.
• The largest model, GPT-J (6B parameters), achieved semantic correctness scores ranging from 25% to over 80% across different prompt categories, with the highest reaching 83%.
• Semantic accuracy improved significantly with model scaling, demonstrating a clear correlation between model capacity and performance in architectural design generation.
• The system successfully generated diverse and valid layouts across all 56 tested prompts, confirming the feasibility of language-only generation for architectural design.
• The open-sourcing of the synthetic dataset and fine-tuned models enables reproducibility and further research in semantic generative design.
• The results suggest that LLMs can serve as powerful, scalable, and accessible generators for design exploration, especially when integrated into iterative workflows like Quality Diversity optimization.