Consistent Diffusion Models: Mitigating Sampling Drift by Learning to be Consistent