Residential floor plans: Multi-conditional automatic generation using diffusion models

Abstract

Automatically generated residential floor plans using Artificial Intelligence that lower the skill barriers in and facilitate non-professional residential design, have become a significant topic. However, in previous studies, the limitations of RFP generative models have exhibited low controllability in outputs and limited flexibility in input conditions. In this study, a multi-conditional, two-stage generative model, FloorplanDiffusion, was developed to address these shortcomings. Based on Denoising Diffusion Probabilistic Models, a new model structure was established, allowing human designers to intervene for enhanced controllability. Furthermore, we implemented a multi-condition model input with structured information using images, thus significantly enhancing the model's input flexibility. Finally, through experiments we demonstrated that our model flexibly generates high-quality, diverse, and controllable results. A Turing test indicated that our model has the capacity of human experts.