In other news. I think the right way to train a Wan 2.2 LoRA is to use the low noise model. You can plug them back in to the high noise model without much damage to the motion, the reverse feels not so true with with the high noise model training.