Negative Tagger Concept?
Hi!
something about the original clip interrogator that is really interesting is the negative tagger: https://huggingface.co/spaces/pharmapsychotic/CLIP-Interrogator
https://github.com/pharmapsychotic/clip-interrogator
I've found that when using a loRA, running the original image style through the negative tagger, then using that string of text as the negative prompt can significantly stabilize the LoRA when prompting with it
If done on a dataset-wide batch of images, the most common negative tags can then be used as a custom catch-all negative tag list for the model, again offering it a lot more stability.
where I'm going is I think the same concept could be applied to booru tags and I certainly don't know how to get it to generate the inverse of any tags, but if anyone knows where to get started it would probably be you, so just putting it out there if it sounds interesting.