I am curious if it could be easily implemented now that it's part of neural-style? I still love CNNMRF btw. Thanks for the great work.