Pixel-wise L2 loss

[vmirly]

So, regarding the loss function on the U network, the following is mentioned in the paper (page 4):

"The pixel-wise l2 loss is employed to enforce image intensity similarity between the upsampled HR faces and their ground-truth images."

However, I am wondering if using Pixel-wise L2 loss is valid or not since the network uses STN modules and therefore the output of the network may undergo some spatial transformation, while the original HR image has not been transformed.