Yep. Almost everything I make is hand-tagged (I'll get advice and background details from the auto-labeler, since its found a lot of good tags, but the bulk of the work is done by me). Only things that weren't were Yuri Shall Conquer the Earth, and Confused looking anime girls with interrogation marks on their heads (where the format is consistent due to the nature of the subject, so apply all+using auto-label for things like hair color/style works), and a chunk of the no humans pics for the Gundam X style (where I couldn't really add anything useful beyond what the auto label had. I was exhausted tagging everything else in it). I've got a good enough knowledge of booru tags that it only takes ~15 minutes or less for a character/outfit/etc. that isn't full of details only visible from certain angles (e.g., a print on the shirt that isn't visible except from straight front). I'm so into characters obscure enough to have limited data sets I could never let auto-label everything.

I also do cleaning of my data set (by far the biggest cost most of the time), cropping out extra characters and obliterating text.

@schmede For whatever it's worth, I view data cleaning beyond the most basic jobs (clear text off a white background/speech bubble, crop to a single character) as a way to increase the number of images that are usable, rather than a requirement (37 is actually a lot of images by my standards, especially when I only have one variant included). Also note that I train 15 epoches, but publish something before that (generally 10, 11, 12 or 13) so reported stats may be slightly off.

As for LoRA compatibility, I think having the training data be a bunch of different styles in the first place (which this very much was) helps improve compatibility with styles.