Has anyone got experience with this relatively recent image restoration net by NVIDIA?
After reading the paper I figured the only thing they do differently is that, instead of using Xi (noisy images) and Yi (noise-free ground-truth) they use two versions of Xi with different random noise. One is used as noisy input and the other shown to the model at the output. And they use MSE as loss (somewhat contrary to what we learned in the course, but if I follow correctly, necessary for this approach to work (see eqns 4-6 in the paper).
I just wondered what your take is approach on this or if anyone tried something similar yet.