Diffusion-Based Generation of Multi-Layered Building Layouts for Accessibility Design and Checking in BIM

Building layout design is a complex and iterative process that must balance various requirements, which makes manual design time-consuming and inefficient. While recent AI-based generative approaches have improved efficiency by learning layout patterns from data, most produce single-layer images where walls, rooms, and openings are entangled. This requires heavy post-processing and limits downstream applications such as Building Information Modeling (BIM) conversion and accessibility analysis. This paper introduces a diffusion-based model that generates multi-layered floor plans in a single pass. It can produce a structural layer for walls, a room layer with color-coded room types, and an opening layer with color-coded windows and doors. Each layer includes a transparent background to enable flexible stacking and easy editing, similar to graphic design software. Evaluated on the Modified Swiss Dwellings (MSD) dataset, our approach achieves coherent, semantically separated, and constraint-compliant layouts, advancing the utility of generative design for architectural workflows and downstream analysis.