Apple drops a paper on how to speed up image gen without retraining the model from scratch. Does anyone knowledgeable know if this truly a leap compared to stuff we use now like lightning Loras etc