This model is based on the MotionModule of v3_sd15_mm.ckpt, but instead of training or merging, I subtracted 0.0001 from all key parameters in the checkpoint using Python. It produces different results from v3_sd15_mm.ckpt, which might be interesting. Enjoy.