>>41212258Depends.
Generally speaking, adding the character tag will not be a bad thing as long as it's an unique tag.
Without tags you are basically dumping all the training into tags like "1girl" or "park" or "checkered background". So yeah you'll be able to get the character without doing anything fancy but if the tags aren't clean that's gonna do into random shit and have unexpected effects.
Adding the tag is extra "glue" for the concept, which can be weighted to use it as a little boost.
So yeah it's slightly better with (specially if your character is rare and already not in the current models) but if your training set is good and varied enough it won't matter much. But not all characters have that luxury so it's good to keep it in mind.
BUT
If the character DOES exist in NAI or your base model in any capacity, even one unable to make a good image, you DO want to tag. It'll benefit from whatever training is there (or guide it in the direction you want).
As for specifics, the whole thing works with pixel nonsense and the tags to guide its attention. It's gonna attempt to find the things specified in tags in the pixel soup.
Mandatory minimum work for quality is removing everything that doesn't belong to the image, specially if it was captioned automatically (deepdanbooru, etc), because that thing makes a lot of goofs, inserts unrelated character names which will make a mess internally, and so and so. You'd be spreading the training into random nonsense and losing on "general purposeness" for it or directly contaminating if it's out of hand.
Then depends on how perfectionist you get. I am training more LORAs for flexibility so I can change costumes and backgrounds easily, so there's a bit of tedium involving teaching about the AI about the different layers of clothing (and without any clothing if possible) and accessories (and without any accessories) and have tags reflect the difference.
As said before, it's best to make test runs and YOLO it and see what specific details need fixing. For example I trained a LORA on Shauna from Pokemon and the thing absolutely insists on putting bows on every piece of upper clothing, so to fix it I'll need to edit or draw a picture of her without the decorative bows on her shirt (and no bow tag. Then make sure every picture with the bows does have the tag) for it to get a clue and stop doing that. Did it for other characters, it fixed the issues 90% of the time (can do it more times to probably eliminate the issue or keep it to 99% good)
It also has a compulsion to keep alternate costumes "in the same color theme" but I actually like that, so not gonna bother to fix it too hard, but you can do it the same way.
Note you can train more than one concept in a single LORA (use separate folders and tag accordingly) if you want to reinforce a specific object, background or concept to go with the character. In my experiments the multiple concepts won't overlap unless prompted for (and it doesn't step into other tags the character is training). Specifically I trained one of Windy from Panel de Pon with a bunch of glitter and sparkles (because that game is girly as fuck, was fitting) and I can turn them on and off at will from the prompt. Helps giving that fairy energy.
It was the "Polka eyes" theory, I finally got to implement it now that it doesn't take weeks to train, lol.
I'm terribly verbose and I'm sorry if I digress too much, but hope this helps. It gave me "is this fucking magic!?!?!" results from absolutely lacking image sets. (That's also why I post about findings in /vtai/. I'm sure someone has an unpopular oshi and this might help)