Distilling SD-1.5 into LCMs for very fast inference. We can get a viable image in 2-5 steps compared to 50-60 steps required from the orginal model.