id |
caadria2024_293 |
authors |
Xu, Weishun, Li, Mingming and Yang, Xuyou |
year |
2024 |
title |
Can Generative AI Models Count? Finetuning Stable Diffusion for Architecture Image Generation with Designated Floor Numbers Using a Small Dataset |
source |
Nicole Gardner, Christiane M. Herr, Likai Wang, Hirano Toshiki, Sumbul Ahmad Khan (eds.), ACCELERATED DESIGN - Proceedings of the 29th CAADRIA Conference, Singapore, 20-26 April 2024, Volume 1, pp. 89–98 |
doi |
https://doi.org/10.52842/conf.caadria.2024.1.089
|
summary |
Despite the increasing popularity of off-the-shelf text-to-image generative artificial intelligence models in early-stage architectural design practices, general-purpose models are challenged in domain-specific tasks such as generating buildings with the correct number of floors. We hypothesise that this problem is mainly caused by the lack of floor number information in standard training sets. To overcome the often-dodged problem in creating a text-image pair dataset large enough for finetuning the original model in design research, we propose to use BLIP method for both understanding and generation based automated labelling and captioning with online images. A small dataset of 25,172 text-image pairs created with this method is used to finetune an off-the-shelf Stable Diffusion model for 10 epochs with affordable computing power. Compared to the base model with a less than 20% chance to generate the correct number of floors, the finetuned model has an over 50% overall chance for correct floor number and 87.3% change to control the floor count discrepancy within 1 storey. |
keywords |
text-to-image generation, model finetuning, stable diffusion, automated labelling |
series |
CAADRIA |
email |
|
full text |
file.pdf (3,617,387 bytes) |
references |
Content-type: text/html
Access Temporarily Restricted
Access Temporarily Restricted
Too many requests detected. Please wait 60 seconds or verify that you are a human.
If you are a human user and need immediate access, you can click the button below to continue:
If you continue to experience issues, please open a ticket at
papers.cumincad.org/helpdesk
|
last changed |
2024/11/17 22:05 |
|