Latent morphologies: Encoding architectural features and decoding their structure through artificial intelligence

Abstract

This article explores the impact of Artificial Intelligence (AI) on the architectural discipline, focusing on generative models and their controllability. While generative models have revolutionized the design process by freeing designers from specific tasks and allowing them to focus on desired results, the reliance on randomness frequently hinders controllability and meaningful experimentation. To address this challenge, the article proposes the construction of an encyclopedic architectural dataset, encompassing various architectural projects and combining images with text for multimodal applications and two methodologies, multi-class StyleGAN and multimodal StyleGAN+CLIP to enhance controllability. Utilizing specific conditions, multi-class StyleGAN enables designers to navigate latent space and identify hidden patterns, while StyleGAN+CLIP integrates text to achieve specific controllability and generate diverse architectural features. Through experimentation, the research showcases the potential of generative models to create structured designs that incorporate existing architectural styles.