SWINIR as Generator in GAN : Real world

[jojupiter]

Hello,
thank you very much for sharing your work.
Please, I would like to know why we use a SWINIR model as a generator in a GAN (with a discriminator) in order to create a less performing real_world swinir model? Because the SWINIR model alone gives very good performance. When it is integrated as a generator in a GAN with a Unet or PatchGAN discriminator, the PSNR on the tests decreases. So what is the purpose of the GAN here?