In this article, I will show what “image style transfer” is. Basically, it is the transfer of some internal parameters of CNN model from one image to another image. Usually one image used as a “base” image for content, another image used as a “style” image, result image is the “base” image with applied “style” to it, resulting in new “target” image that has content from “base” image and style from “style” image. It is important to understand that not only one “style” image might be applied to generate the “target” image, but several “style” images with different “weight ratio” might be applied as well, resulting in a truly unique final image that can be really beautiful and unique both in content and in style.
The technical implementation with code example will be published in the next article, while in this article I just show the power of “image style transfer”.
Here the ART.
(generated image in the right column)
|base “content” image||the “style” image||resulting image with “style” applied to “content”|
As I said previously, not only one image might be used as “style”, but several images as well. Here below is just one example where two “style” images applied to one content image, resulting to “target” image that has style from both “style” images. You can define the weight(ratio) of each “style” in resulting image thus creating a truly unique art.
|base “content” image||style one, 0.3 factor||style two, 0.7 factor||resulting image with “style” applied to “content”|
The second part of this article with more image examples can be found at this url: http://datafor.art/image-style-transfer-part2/