search

allenai / satlas-super-resolution

Apache License 2.0
190 stars 24 forks source link

Adversarial loss for ESRGAN #29

Closed yunseok624 closed 3 months ago

yunseok624 commented 3 months ago

Hi!

I see that in the paper you mentioned using ESGRAN with clip loss for optimal model for upscaling. In the original ESRGAN paper RaGAN (Relativistic average GAN) loss is used, however your code uses a standard GAN loss. Is there a special reason why you decided to use standard GAN loss?

piperwolters commented 3 months ago

Hi, we originally were basing our code off of the Real-ESRGAN paper, which has slight differences from the original ESRGAN paper, one of which was using the standard GAN loss (in addition to Perceptual and L1). You could certainly try RaGAN if it's already implemented in BasicSR package.