LLM-based framework for automated and customized floor plan design

Abstract

Interpreting diverse and ambiguous natural language (NL) user requirements into precise floor plans poses significant challenges for design automation. This paper presents a large language model (LLM)-based framework to automate and customize vectorized floor plan design. This framework utilizes a syntax tree for NL parsing and automated dataset enrichment. A dual LLM approach involves using a recognition model for automated dataset augmentation, while a generation model interprets user inputs to create diverse, geometrically precise vectorized floor plans that align with complex semantic preferences. Experimental results demonstrate the proposed LLM-based models' effectiveness, achieving high accuracy in interpreting user requirements and high quality in generating corresponding vectorized floor plans. Additionally, a large-scale NL-based dataset is generated through the automatic recognition of existing floor plans. The proposed method can advance automated and user-centric floor plan design by enabling direct NL interaction and generating readily usable vectorized outputs.