Questions about training degradation decision

[Luciennnnnnn]

Hi, I'm wondering why first train stage1 and stage2 model with codeformer degradation, and then train stage1 model with RealESRGAN degradation. Intuitively, it will cause input distribution shift of stage2 model.