The official description I typed up: I added an updated V2 version. This version is a bit less overtrained, has additional images added to the dataset, and some images were removed (mostly images that were close-ups on ass, boobs, etc.).

More specifically, I added a handful of fan-stitched images (i.e., images where, for example, a panning shot of Olga was stitched together to create a full-body shot). I also removed some pictures of just the girl's vagina, boobs, random close-ups that don't really do much to define the characters, etc. The dataset went from around 900ish to 750, but the images/dataset itself should be a bit more focused now. The original dataset consisted of the first six episodes, where I used ShareX to capture a screenshot every second and removed irrelevant images. As a result, the V2 dataset remains comprehensive, despite the reduced image count.

This does seem to be an issue with V2 for whatever reason. (no crown, no headwear, no tiara:1.5) in prompt and (crown, headwear, tiara:1.5) in negative worked for me.

Hmmm. that'd be interesting to try. Honestly, though, since they have so little screen time and, as you said, are just background characters, I'm not sure how well they'd translate to a LoRA.

Maybe try https://civitai.com/models/1230000/knight-kuroinu paired with my model? I'll probably not be making a V3 model (I have too much other stuff to do lol).