This project speeds up the process of manually preparing images for training Stable Diffusion embeddings.
It provides a streamlined UI for selecting a square portion of a picture and then scales that section to 512x512 for training, and then outputs it and loads the next image, all in a single click!
The code has been thrown together quickly. It's disorganized and unprincipled. I made the minimum viable project to speed up my own workflow I was interested in. Pull requests are welcome.
- Install python 3.x from python.org.
- Download this repository as a zip file and extract
- Enter the extracted folder with your file browser
- Double click setup (setup.bat) - This only needs done once
- Double click launch (launch.bat)
Ya'll probably know what to do already...
- Make sure python is installed however is standard for your distribution.
- Git clone the repo or download the zip and extract
- Enter the project folder in a bash terminal
- run
make
- activate the venv with
source ./venv/bin/activate
or alternative script if not using bash... - run the script
python src/training_image_processor.py
- First open a directory with pictures
- Use the buttons at the top to rotate/flip immages as needed
- Resize selection square with mouse wheel
- Click the image to process it and load the next image. The 512x512 image is placed in a new 'outputs' folder and the original goes in a new 'originals' folder, inside the open directory
The contents under the assets folder are from the adwaita-icons project and are released under the license LGPL-v3.0. The rest of the project is GPL-v2 only.