@faye_nsfw That's fine, assuming you're not dropping repeats with that many images (some trainers will automatically lower the default repeat count if there's enough images). I do note that you absolutely do not need non-SFW content in training data for a model to make non-SFW content with it. The idea is a holdover from SD1.5, which struggled to change a character's outfit if all the data was trained with only a single set of clothes (so to make non-SFW content it needed multiple outfits in data, not necessarily without an outfit). Having multiple outfits in data still helps get a character's physique correct outside their main clothes (especially if the main outfit is relatively bulky or their body is relatively unusual, like a western cartoon), but it absolutely isn't a requirement (though you will, of course, need data of a particular outfit to replicate it and if a woman has an official under/swim clothes it should absolutely be in the dataset with consistent tagging)

using more pictures is abolutely better, it improves versitility by miles, listens better to more dynamic requests. just make sure they are clean readable images and the sequence of images build up in complexity. the point is to also use good detailed prompt wording with it, to expand its understanding. using less only works for cartoon art where you want to match a very limited style and dont want it to do anything else.