I am trying to implement this paper.
The loss function described in the paper depends on the output of an image and a generated noisy version of the image. The base model
