Explore Generative AI Models for Vector-Based 3D Architecture Data

In recent years, we have witnessed a notable surge in generative AI from Large language models to image diffusion models. It appears we are on the brink of an era in which, given sufficient data, generative AI models can learn the data distribution of any dataset and synthesize new, high-fidelity examples. This thesis explores the application of this idea in architecture, focusing on the generation of vector-based 3D architectural data. Emerging state-of-the-art generative AI models—such as MeshGPT—have demonstrated the ability to generate 3D vector data, including meshes, wireframes, and boundary representations (B reps). To explore whether such models can learn human spatial habits and formal conventions, I trained MeshGPT on the 3DBAG dataset to generate buildings with diverse roof types and morphologies from text prompts. The thesis evaluates the feasibility of this approach in relation to current model capabilities, architectural datasets, and design needs. Evaluation results show that while the model generates accurate outputs for common prompts, it struggles with rare typologies and more complex descriptions.