Generative Isovist Transformer: Machine learning for spatial sequence synthesis

While isovists have been used widely to quantify and analyze architectural space, its utilization for generative design still needs to be explored. On the other hand, advanced deep learning has shown opportunities for data-driven generative design. This research revisits the isovist capacity to represent architecture as a series of spatial sequences and extends the role of isovists beyond merely a perception model to projective agents. This paper presents the development of GIsT: Generative Isovists Transformer in sampling, learning, and generating architectural spatial sequences. By coupling isovists with discrete representation and generative deep learning models, we untapped the generative potential of isovist representation for spatial sequence synthesis. We demonstrated its capacity to learn the architectural spatial sequence and extendability via few-shots learning. The results show a promising direction toward integrating data-driven experiential spatial synthesis in future computational design tools.