Some implementations of CNN based single-image style transfer.
The best model was trained using ideas from these two papers, as well as adding two GAN discriminators on top to improve visual quality.
The AdaIn paper + GAN was implemented in style2/adain.py.
It is actually amazing how well transferring just the first two moments is able to copy style, but
it definately does not get the job fully done.
The second paper, which I call AdaInRA (rigid alignment), prescribes an additional step to AdaIN's moment matching that
allows the model to retain global style artifacts. It (plus two more GAN discriminators) are implemented in style2/adain2.py
It requires doing an SVD with columns proportional to square image size, which unforunately is very slow.
It definately gives better results, but is not suitable for running on videos. I haven't found a way to
run the svd at a lower resolution then upscale. Perhaps you can train an MLP to approximate the SVD instead.
AdaInRA results below (rows are content, style, transferred style-to-content)
