simple script for easy flux finetuning from a folder of images. uses replicate for the cloud gpu goodness.

populate .env file: add your openai and replicate keys. openai used for captioning.
pip install -r requirements.txt
put all the images inside data/source_images
adjust constants at the top of finetune.py (your replicate details)
run python finetune.py
wait for the script to finish and it will return the the training url
optionally create embeddings for all the image descriptions and store them in a .csv - so that later you can do semantic searches over your image library

Behind the scenes the script will:

ask the user for details e.g. what to call the model
create a new folder data/training_pack and copy + convert all images and rename to uuid.jpg format
if need be - downscale images to 1024x1024 (max)
run each image through gpt4-o-mini to generate a description
save all descriptions as uuid.txt and put it into the same folder, optionally creates embeddings and adds to the csv
.zip the folder
create a new model on replicate
create a new training job on replicate - and give the user the url to check on the training