>be me>decide to do a model merge>decide to use bayesian merger>want to do a character model/LoRA merge into another model>manually implement character based scoring based on a contrastive text-2-img model that was designed specifically to compare and contrast reference images to a generated character, and then report a score>let it gen overnight>come back to high character comparison scores>genned images look different to the reference images but look like the character(s) they are supposed to be genning>kept all of the style of my original modeland now I'm happy again.
Also HLL anon should look into adding CCIP into his training pipeline if he hasn't already.
(pic-related is an example, not my work)
https://deepghs.github.io/imgutils/main/api_doc/metrics/ccip.html 