The smoothness loss is also different: I kept getting nan with the logarithmic operation so I deleted that part, the left part could be seen as total-variation loss (tv-loss)

[lilin-git]

You normalize the data in (-1,1) and the parameter of tf. log () is greater than 0, so loss is nan.

[SystemErrorWang]

sorry for late reply, i also found this problem in my training. I suggest you to try deep guided filter as it only use tensorflow default ops and no additional self-defined layers. I got good result with it