As this picture show, subtract the result from the input to produce the "diff" layer picture.
i.imgur.com/jE9M4e1.jpg
Since edges look better at -n 3 and some gradients look better at -n 1.
So it will be a great help if all n 1~3 diff layers can be produced in one go, and let user or other app merge them.
And if it's possible, in -m scale mode, separate the deconvolution (in upconv7?)/bicubic base and CNN-enhanced diff layer as well.
As this picture show, subtract the result from the input to produce the "diff" layer picture. i.imgur.com/jE9M4e1.jpg Since edges look better at -n 3 and some gradients look better at -n 1. So it will be a great help if all n 1~3 diff layers can be produced in one go, and let user or other app merge them.
And if it's possible, in -m scale mode, separate the deconvolution (in upconv7?)/bicubic base and CNN-enhanced diff layer as well.