I made GPT4 generate thousands of prompts I could use for training. Specifically, these wildcards intentionally are not overloaded with super elaborate prompts.
The intent is to use these prompts to generate images and then use these same prompts as caption input to the training data. I feel like using auto-captioning tools like "Interrogate CLIP" often produce garbage, and I don't want to be bothered with cleanup. Going from text to image seems more consistent than image to text, hence these wildcards.
v2.0-sneak-peek: Unsorted subset of upcoming v2.0
v2.0: Used GPT4 through the OpenAI's API. Threw 100 USD at it.
v1.0: Used ChatGPT3.5 and ChatGPT4 manually through a few sessions.
Description
GPT4
FAQ
Comments (5)
Hello, thank you for the contribution to the community, I'm trying to understand the use of this better. Where precisely do you use this for training?
I use this to generate images with the A1111 wildcards extension. Then I right-click the "generate" button to select "generate forever" to generate thousands of images. Then I manually curate the results. After curation, I extract the prompt saved in the image metadata of each image. Now I have everything I need for training: a prompt and an image.
You don't need to use these for training. They are useful as just base-prompts for generating images without all the "high quality, 8k" style baggage baked in.
@thomaseding Smart idea. thank you for sharing and explaining
Thank you for such a smart idea.
I am comfy user, any workflow to use it better there?