DreamBooth: Fine-Tune Stable Diffusion in Google Colab Free (Multiple Subjects at The Same Time)

[ ⏰ time ]
The whole thing took about 40 minutes because there were two subjects with an Nvidia T4 on Google Colab free. It takes about 15 minutes if you've only got one subject.

Some sections are sped up for easier viewing.

[ 📔 important note ]
➞ Training multiple subjects of the same gender on the same model is very likely to lead to blending between them. You may notice Sandman having one eye a bit different, which he "inherits" from Aemond's eyepatch.

➞ To mitigate blending of multiple subjects, the author of the notebook (TheLastBen) recommended using UNet_Learning_Rate: 2e-6 instead of the default 5e-6. He recommends training a subject on a separate model to get the best results.

[ ⚙️settings used ]
Images Used: 10 images Aemond, 10 images Sandman
Stable Diffusion Version: 1.5
UNet_Training_Steps: 2000 (100 steps per image)
UNet_Learning_Rate: 2e-6
Text_Encoder_Training_Steps: 350

Everything else left to default values.

[ ⌛ timestamps ]
00:00 - Check we're using GPU
00:04 - Run first 3 cells
00:11 - Name session and run 4th cell
00:26 - Connect to Google Drive
00:35 - Run "Instance Images" cell
01:42 - Bulk rename images
02:18 - Upload images
02:35 - Adjust training settings
02:56 - Run training & wait (~32 minutes)
04:04 - Run Test Trained Model & wait (~4 minutes)
04:25 - Access Web Interface & Generate Images

[ 🎵 audio ]