It's also comparing two overly compresses images with encoders which produce very different kind of artificats (bluriness vs. blocks of wrong colors). Which one is better is always going to be subjective (or at least application-specific). Really, as long as the model rates both as shit it is not necessarily a bad model.