While I try not to make LORAs of celebrities that have plenty of LORAs already, I've wanted to try this out for about a week now.

I took the GPT-4 captioned dataset found here: https://civarchive.com/models/281464?modelVersionId=321185, made by steffangund, and then duplicated the entire zip file. I then re-tagged every photo in the set using WD14 tagging conventions, and then trained the LORA using both sets - thus training with short, one & two word descriptors along with more verbose phrases to describe the same images.

The result, hopefully, is a more responsive model that will listen better to your img-to-txt input, no matter how you input it (so long as it's in English).