Non-issue: Transfer style but not the colors

[Yotooon]

Is it possible to apply the style of an image but keep the colors of the original?
I've recently read http://blog.deepart.io/2016/06/04/color-independent-style-transfer/ and the results are just awesome.

[jcjohnson]

Based on the comments on reddit, it looks like you can achieve this effect with some simple post-processing with HSV or YUV color-spaces; I might try to implement this over the weekend:

https://www.reddit.com/r/MachineLearning/comments/4nbdvz/colorindependent_style_transfer/

[Yotooon]

Thank you! It might not be as hard as I originally thought... 👍

[htoyryla]

Others appear to be working on this too right now https://github.com/pavelgonchar/color-independent-style-transfer

[ProGamerGov]

Along with keeping the colours of the original image, would having a potential third -color_image parameter be possible? So that you could use the colours of a third image along with your style image and content image.

Also, a -color_type parameter with HSV and YUV as the two options would be an excellent addition to controlling the color of the output image.

[jcjohnson]

@ProGamerGov The implementation of color-independent style linked above simply converts the content image and generated image to YUV, then replaces the V channel of the generated image with the V channel from the content image. This will not give good results if you try to transfer color from a third image, since the color will be not be aligned to the image content in any way.

[ProGamerGov]

@jcjohnson I see, so it is limited to the traditional style and content image combo.

[jcjohnson]

I added support for this feature here: 8250d35

It's pretty simple: just use the Y channel from the generated image and the UV channels from the content image.

[ProGamerGov]

@jcjohnson

This will not give good results if you try to transfer color from a third image, since the color will be not be aligned to the image content in any way.

What if the third image was a previously generated output image which used the same content image, but a different style image? Or would that still not work?

[jcjohnson]

That might actually work!

[ProGamerGov]

@jcjohnson How difficult would this be to implement for testing?

[sebrammer]

Thank you for implementing the V channel translation JC! I've really been enjoying neural-style over the past few months.

Getting an error after replacing neural-style.lua with the new commit of 6/12/16.
/usr/local/bin/luajit: neural_style.lua:5: unexpected symbol near '<'
The error occurs using the new commit and is unaffected by the presence or absence of the -original_colors flag, with or without a valid value.

I merely downloaded and replaced my copy of the neural-style.lua file. Should I have run through the complete dependency install / update?

edit: and I see I just posted this under a 'non-issue' thread. Apologies.

[htoyryla]

sebrammer [email protected] kirjoitti 14.6.2016 kello 17.54:

Thank you for implementing the V channel translation JC! I've really been enjoying neural-style over the past few months.

Getting an error after replacing neural-style.lua with the new commit of 6/12/16.
/usr/local/bin/luajit: neural_style.lua:5: unexpected symbol near '<'
The error occurs using the new commit and is unaffected by the presence or absence of the -original_colors flag, with or without a valid value.

I merely downloaded and replaced my copy of the neural-style.lua file. Should I have run through the complete dependency install / update

The symptoms suggest that you have downloaded the html page, not the raw text file. Have a look.

[sebrammer]

I will now slink off and find a rock to call home. Thank you for pointing out the forest hiding behind all those trees.

[htoyryla]

Happened to me too (how else would I have known).

[jcjohnson]

Thanks @htoyryla for figuring this one out! I was confused.