Latent Morphologies: Encoding Architectural Features and Decoding Their Structure through Artificial Intelligence

With the advent of Artificial Intelligence, new methodologies have been introduced to the architectural discipline, expanding the current possibilities of design processes. Specifically, generative models created a paradigm shift wherein, instead of spending numerous times designing the entire system for a specific task, designers allowed the overall principle and system to remain in the black box and instead focused on the desired results. These attempts, however, strongly rely on randomness and could not achieve overall controllability so those problems have hindered getting meaningful results. This paper started with building an encyclopedic architectural dataset that can represent general architecture for a general understanding of architectural styles, maintaining its variation. The dataset includes an image and a text together to stretch its application and versatility to the extent of multimodal. Several statistical methodologies are utilized to understand and unveil characteristics in massive data. It also suggests two methodologies to achieve controllability in StyleGAN, which are multi-class StyleGAN for general controllability of StyleGAN and multimodal StyleGAN+CLIP for its specific controllability. Multi-class StyleGAN helps navigate latent space to find hidden patterns we cannot identify and their regularity in architectural discourse and StyleGAN+CLIP shows numerous possibilities of text-integrated generative models. The concept of latent space shows incredible possibilities, generalizing architectural features and generating their continuous morphologies, presenting theoretically infinite variations.