Quality assessment of residential layout designs generated by relational Generative Adversarial Networks (GANs)

Abstract

Architectural design, a complex optimization process, employs computational design tools to converge to a subset of design options within a large design space. Recent efforts incorporate deep learning to replace traditional hard-coded rules, yet the nascent and opaque nature of data-driven architectural design requires evaluating its alignment with implicit architectural requirements. This paper assesses whether current data-driven image generation models comply with these non-explicit, domain-specific requirements. Over 80,000 architecturally created floor plans were compared to a similar number of floor plans generated by the authors using the state-of-the-art generative design models. This assessment was based on metrics such as the distribution of space sizes based on their shared/private nature, visibility based on privacy requirements, and room connectivities. The employment of Mann-Whitney U tests and descriptive statistics revealed significant disparities, indicating a gap in the generative design models' understanding of the nuanced, latent rules present in real-world data. These findings highlight the need to develop domain-specific metrics to evaluate the true performance of generative design models, offering insights that will aid in the adoption and advancement of data-driven algorithms in the architecture design domain.