Text2Form Diffusion: Framework for learning curated architectural vocabulary

Abstract

Stepping towards the machine learning age in architecture, many architects and researchers have developed creative ways of utilizing machine learning with domain-specific architectural datasets in recent years. With the rising popularity of large language-based text-to-image models, architects have relied on different strategies for developing the prompt to create satisfactory images representing architecture, which lessens the agency of the architects in the process. We explore alternative ways of working with such models and put forward the role of designers through the fine-tuning process. This research proses a fine-tuning framework of a pre-trained language model, namely Stable Diffusion, with a dataset of formal architectural vocabularies towards developing a new way of understanding architectural form through human-machine collaboration. This paper explores the creation of an annotation system for machines to learn and understand architectural forms. The results showcased a promising method combining different formal characteristics for architectural form generation and ultimately contributing to the discourse of form and language in architecture in the age of large deep learning models.