Auto Captions for SDXL

This script can help you generate captions for a person or subject by using GPT-Vision.

Usage

First you must make sure you have your Open AI API key set. The program assumes it is set to OPENAI_API_KEY

You can temporary set it by running this:

export OPENAI_API_KEY=<YOUR_KEY>

You can expect to pay around $0.01 per image captioned. This is not a guarantee and you should check the Open AI pricing page.

After that is set up, clone the repo and run the caption command.

1. git clone https://github.com/GeorgeNance/gpt-caption-sdxl
2. cd GPT-CAPTION-SDXL
3. python3 caption.py <token> <image_path>

Replace token with what you want to use for the subject you are training. If you are not sure what to use, Google is your friend. There is a lot of discussions on the best tokens to use and that is far beyond the scope of this project.

Replace image_path with the path to the image folder you want to caption.

Example

python3 caption.py "sks" "C:\Users\user\training_data"

To change the extension of the output files, you can pass the --ext flag to the command.

python3 caption.py "sks" "C:\Users\user\training_data" --ext=md

Optional parameters

Parameter	Description	Default
--ext	The extension of the output files	txt
--overwrite	Overwrite existing caption files	False

Output

The outputted captions will be written to a .txt file that has the same name as the image it is captioning.

Example:

C:\Users\user\training_data\image1.jpg
C:\Users\user\training_data\image1.txt

This is perfect for using with kohya_ss

Notes

This script is not perfect and will not always generate the best captions. Looking over the captions and making sure they are correct is important.

Roadmap

• Add more options for the captions
• Work on the prompt to make it give better results
• Add style support

License

I don't mind you taking this code and using it for your own projects. Its optional to give me credit, but it would be nice.

MIT License