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: <a target="_blank" rel="ugc" href="https://civarchive.com/models/281464?modelVersionId=321185">https://civarchive.com/models/281464?modelVersionId=321185</a>, made by <a rel="ugc" href="https://civarchive.com/user/steffangund">steffangund</a>, 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 &amp; 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).

Files with SHA256: 7e5288831043021cfd7fd83455e4c8910b2aceaba6478e85e029a03f642c0423