Implementing layer-level normalization of gradients similar to neural_artistic_style?

[ProGamerGov]

When people are discussing style transfer projects, I have seen people mention that the "weird muddy areas" they see on artwork produced by Neural-Style, are an issue which can be solved. Is this something that is possible for Neural-Style, seeing as Neural_Artistic_Style already solved this problem? How serious is this issue with Neural-Style?

The gradient normalization on neural_artistic_style being discussed in more detail: andersbll/neural_artistic_style#39 (comment)

Regarding the visual style, I perform layer-level normalization of gradients. In the VGG-net, the features in each layer may exist on different scales. By L1-normalizing the gradient signals, I get a more even contribution across the different layers. I suspect this is the secret sauce. :)

[jcjohnson]

L1 normalization of gradients coming from content or style losses has been in neural-style for a long time; you can just add the flag -normalize_gradients.

[htoyryla]

Neural_artistic_style also seems to use scaling of gram matrices to
compensate for different sizes of content and style images. If I
understood the code correctly, it does not scale the size of the matrix,
but only the values in the matrix. I might give this a try as I have
already experimented with variations of the gram matrix. But this is not
likely to have a major impact on quality; to me it seems more like an
automatic adjustment of style weights depending on the image sizes.

[MrZoidberg]

@jcjohnson is there any samples of using -normalize_gradients with the commands? I'm getting just weird results with the following command:

th neural_style.lua -style_image $HOME/sync/styles/style7.jpg -content_image $HOME/sync/iteration1/selfie2.jpg -output_image $HOME/sync/style7_sw150_sc15_norm.jpg -image_size 600 -backend cudnn -normalize_gradients