Generating Architectural Floor Plans Through Conditional Large Diffusion Model

This paper introduces a novel method for generating architectural floor plans using Conditional Large Diffusion Models to migrate the limitations of existing generative methods, such as restrictions on rectilinear configurations, limited scalabilities, and the simplicity of details. Central to this study is the development of a large-scale dataset comprising high-quality floor plan images with corresponding condition maps and textual captions. The essential step is to setup the conditions in relative to the architectural floor plan. The data collection and processing align with these condition requirements. The development of the dataset includes manual preparation of 1007 floor plan images as an initial set for training the floor plan recognition models that facilitate the automated annotation, algorithmic generation of the 12000 images with Grasshopper as a supplementary pseudo set for testing the generation effectiveness, and several image captioning and visual question-answering models for producing textural descriptions. The generation models trained on this dataset demonstrates the potential to produce diverse planning options using basic constraints as the conditions, which was evaluated by several user studies. The findings underscore the transformative potential of integrating advanced generative AI technologies in architectural design.