In work exhibited at the NeurIPS 2018 Machine Learning for Creativity and Design, three-dimensional architectural massings for single-family homes are generated by a generative adversarial network. This GAN is trained on a small dataset of three-dimensional models of homes falling into seventeen architectural styles, and that are represented as multi-view heightfield images. A process is developed for converting from 3d CAD model to 2d tiled heightfield image, and from the 2d heightfield images generated by GAN back to three-dimensions in voxel format.

This work is based on the product of a workshop at the 2018 Smart Geometry Conference, led by Adam Menges, Kat Park, Samantha Walker, and Kyle Steinfeld. Thanks to workshop particpants Ben Coorey, Marantha Dawkins, James Forren, Timothy Logan, Antoine Maes, Jenessa Man, Sebastian Misiurek, Gabriel Payant, Aseel Sadat, and Nonna Shabanova.

The dataset of CAD files was hand-crafted by a group of undergraduate students at UC Berkeley: Ricardo Ayala, Natya Dharmosetio, Jenny Zhu, Loryn Cook, Cheuk Ng, Hunter Paine, Saya Coronado, Merve Heyfegil, Karine Yedikian, Yi-En Chen, April Liu, Tien Nguyen, Jerry Chen, Reem Makkawi, and Janeth Miranda.