>>1086587
Realized I should expand this when not tired:
When you train a LoRA, you're actually teaching the output
patterns and stuff that's consistent between your selection, with some filtering and sorting by tags. If all of a LoRA's training data is of a single style, the result will adopt that style (especially for the character, but the BG too to some degree). For characters this is generally perfectly fine and often preferable if you're doing their native style (indeed, a lot of animated characters are totally unrecognizable outside their native style) and lessen the effect of any style LoRAs applied (essentially it's wearing layers) and especially styles just from tags. This was a bigger issue in SD1.5, but some styles are still pretty extreme and make it much harder/impossible to combine with other stuff. Many of these can be isolated by hitching them to style tags the base model knows ("3D" is a big one. If you make a LoRA using primarily 3D training data and don't tell it that training data is 3D,
everything you make with it will be a weird pseudo-3D, but if you say its 3D it works fine. Same for "realistic, photo (medium)", "chibi", "sketch", "pixel art", and a few others). It will really be an issue if you try teaching clothing, poses, etc. using a single source (properly tagged "realistic" aside) because your result will be as much a (bad) style LoRA as it will be one for clothes (etc.).
(One caveat is that if a character has warped proportions (e.g., very big head) as part of their style, those tend to carry over and look strange in styles incompatible with such. Pic related)