Exploration and Refinement: A Dual-Diffusion Approach for Multi-Occupancy Floor Plan Generation

Designing floor plans for multi-occupancy buildings involves balancing functionality, constraints, and creativity. Generative design offers a promising solution by leveraging algorithms to explore vast design spaces. Extensive research has focused on generative design for floor plan generation. However, current approaches struggle to generate complex multi-occupancy floor plans, fail to balance creative exploration with strict adherence to constraints, and lack support for iterative refinement. This paper introduces a novel vision transformer-based dual diffusion model to address these challenges. The first model generates diverse floor plans from boundary inputs using inpainting, while the second refines them based on visual feedback from designers. The approach is trained and evaluated on the Modified Swiss Dwellings dataset. To validate the method, we employ Fréchet inception distance to assess visual quality and mean intersection over union to measure alignment with input constraints. Results show improvement in both metrics compared to the baseline.