EXPLAINABILITY OF GENERATIVE AI FOR ARCHITECTURE: INVESTIGATING THE ALIGNMENT WITH DESIGN INTENTIONS

Date

2023-9-8

Author

KARAOĞLU, MUSTAFA SİNA

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

431
views

176
downloads

Cite This

Generative AI systems have the potential to elevate computational design tools from mere design artifact representation aids to influential collaborators in the design process. However, due to the black-box nature of generative AI, it is challenging for users to grasp its mechanisms and ascertain if the system aligns with their design intentions, thereby limiting its full potential in the architectural design process. Addressing this issue, this thesis presents a specialized text-to-image generative model. Unlike off-the-shelf models, this model benefits from curated and structured training data, facilitating controlled image generation and systematic evaluation. The model is developed by fine-tuning a pre-trained Stable Diffusion model using 85,000 architectural images and structured textual labels, representing diverse architectural design concepts related to the images, extracted from architectural project descriptions via neural topic modeling. Through investigating how well these labels are represented in the generated outputs across various architectural design use cases, the study demonstrated the model's alignment with design intentions.

Subject Keywords

Deep Generative Learning for Architecture, Architectural Design Process, Explainable Generative AI, Diffusion Models, Large Language Models

URI

https://hdl.handle.net/11511/105546

Collections

Graduate School of Natural and Applied Sciences, Thesis