I think it gets flexible if you just tag everything accordingly, but I did ensure there were images with different outfits, genders and styles so it could disentangle the pyramid head from the body and whatever else, you can test yourself if you want, I think it was this that I used: https://easyupload.io/95r8v7 (it has no tags, just the trigger word, I don't remember if I trained like this or if I fucked up the captions afterwards while testing out stuff, its probably easier to disentangle stuff if you tag)
I used a higher number of repeats in the hq images, so it would something like a pyramidhead folder, and inside it that folder a 10_hq folder with the hq images and a 4_lq folder with the lq images, it may not even be necessary to do this, I was just experimenting to see what would happen (if you throw a bunch of low quality images just to get be the bulk and use the high quality ones to improve it, kinda like SD was trained)