DreamBooth: Fine-Tune Stable Diffusion 1.5 in Google Colab Free (Single Subject - Human Face)

Описание: A demo training Stable Diffusion 1.5 with Dreambooth on a single subject. Full written tutorial here https://bytexd.com/how-to-use-dreambo...

[ ⏰ time ]
The whole thing took about 20 minutes with an Nvidia T4 on Google Colab free. The training took just 16 minutes.

Some sections are sped up for easier viewing.

[ 📔 important note ]
➞ When training similar subjects, like same gender faces, it's recommended to train them on separate models for best results.

➞ When training multiple subjects on the same model, then reduce the UNet learning rate to 2e-6 or lower. For this purpose you can see this video tutorial where I'm training SD 1.5 on Aemond Targaryen and Sandman (from Netflix)    • DreamBooth: Fine-Tune Stable Diffusion in ...  

[ ⚙️settings used ]
Images Used: 10 images Aemond
Stable Diffusion Version: 1.5

All training settings left on default values.

[ ⌛ timestamps ]
00:00 - Run first cell & wait to connect to Google Drive
00:18 - Run 2nd cell to install dependencies
00:20 - Run 3rd cell to download SD 1.5 model
00:23 - Name your session & run cell
00:29 - Run Instance Images cell & wait for upload button
00:48 - Upload subject images
01:02 - Start DreamBooth training & wait (~16 min)
01:27 - Run Test The Trained Model & wait (~2 min)
01:36 - Access the web UI and start generating images