Is it possible to sample at scale and at high-resolution every single building ever built in a country? If so, can such a dataset be used to train an AI model in generating new yet locally compliant buildings without any explicit regulatory control inputs? The project explores the design agency of deep generative neural networks in learning architectural notions of three-dimensional exteriority and interiority with a redesigned 3D generative adversarial network (3D-GAN) architecture. Trained with a large dataset of 3D digital models of high-rise buildings found in Singapore, it generates not only formally plausible and semantically coherent configurations but begins to also imagine novel and uncanny architectural forms, interpolating and extrapolating among standard high-rise housing typologies such as the slab and point blocks. The work was on display at Singapore’s Arts House which featured the outputs as 3D-printed architectural pieces, 3D latent walk video animations, and full-height digital prints on paper. It was previously exhibited at the 17th Venice Architecture Biennale’s CITYX Venice Italian Virtual Pavilion.
