ML BS
I am actually giving it one more chance by trying it as the merge block for upsamples and residuals in a U-net, but when it was the final head-to-tail residual merge, the early training of the network was almost flat, so my hopes are not super high.