An Attention-Based Constrained Diffusion Model for Accessible Floor Plan Generation

Abstract

Designing floor plans is a critical part of the building planning and design process, involving numerous design constraints. Although significant efforts have been made to automate floor plan creation, existing methods often struggle to meet specific design constraints and frequently overlook critical accessibility requirements, which are essential for creating inclusive and usable spaces. Furthermore, these methods often fail to generate vector-format floor plans suitable for industry-standard tools such as building information modeling (BIM). To address these limitations, this paper introduces a novel deep learning-based approach to automatically generate vector-format floor plans that comply with geometric and topological constraints including building code accessibility requirements. The proposed approach combines a constrained diffusion model that leverages a transformer architecture with newly introduced boundary-to-corner and minimum-distance-to-room attention mechanisms to capture geometric and topological information from design constraints, with specialized postprocessing algorithms to improve both visual quality and compliance. The input embeddings of the design constraints and the diffusion process ensure the generation of vector-format floor plans. Experiments demonstrated that the proposed approach outperforms baseline generative design methods in terms of visual quality, adherence to design constraints, and compliance with accessibility regulations.