I dont think it's supposed to be consistent. It can't easily be with StableDiffusion. LCM sampling is very fast and has brought StableDiffusion closer to real time. It will be helpful iterating fast, I am more interested in seeing the effect of prompt while I am still typing.