>>105919838
always train on the base model
all (or rather most) derivative models will work if you do that
if you train on a derivative, you're using whatetever changes are in that derivative from the original model, and learning all its bad habits, which like wont work with either the base model or other derivatives
that's why "they" always said to train on base sdxl to use on downstream models.
although if you're only going to use one derivative model by all means train on that