@red5cheat Thanks. By the way, you don't actually need "clean" images to do comic book characters: It's fine if there's speech bubbles, so long as you blank them out and tag the images "speech bubble, blank speech bubble". Same with panel divides and the like as long as they're tagged "comic" . I've got 9+ LoRAs (not counting PD to IL retrains) that used this for all their training data, and a few more that used it to bulk out their data. I've yet to see it produce an issue (aside from Illustrious not being as good as Pony at filling in majority monochrome data, but that's not applicable to western comics).

Also your metadata says all your training images had the same exact tags. You should both exclude tags that aren't visible (boots and gloves not being in a portrait) and tag both stuff the character is doing (standing, fighting stance, jumping etc.) and stuff around the character as best as tags the base model recognizes allows (such as adding "table" to an image with a table).

@red5cheat No problem. Ask me (in the comments to the article) if you've got questions on LoRA making my training guide (or the linked articles) doesn't cover something: I'd have loved to have had better advice when starting out (and publicly posting my processes means that if I'm wrong, someone will tell me how).