3 ms·

I think that this model used consistency loss during training so that it can yield better results with less steps.