Exploring the Latent Space of Facades. DigitalFUTURES 2021 (paper)

StyleGAN2 generates plausible building facades without any conditional input; the latent space is then steered, visualised and explained.

Launch work ↗

Generative adversarial networks (GANs) are widely used to generate architectural images, but output quality, controllability and model interpretability can still be improved. This study trained StyleGAN2 on 9,772 building facade images at 1024 × 1024 and generated plausible facades without any conditional input; GANSpace then found directions in the latent space that control high-level properties, for generated images and for new images outside the training set alike.

Image embedding and PCA projection make the generation and control visible: generated images can be classified without supervision, and the relation between images and their latent vectors is laid open. It is the earliest starting point of Dimensional Recasting: wandering a high-dimensional space and projecting what is found back into ours.

Paper: Shengyu Meng, Exploring in the Latent Space of Design: A Method of Plausible Building Facades Images Generation, Properties Control and Model Explanation Base on StyleGAN2, DigitalFUTURES 2021, Springer. Springer · ResearchGate


Series: Dimensional Recasting

Search