Sorry, what do GANs have to do with this? It is not the same kind of "evaluation".

And anyway, there is no need to have two networks to iteratively refine output: one suffices (like we naturally are meant to do).