Fast CPU inference with Latent Consistency Model.4 steps is usually enough — crank it to 8 for more detail.