2 ms·

The main difference of this method is doing the transfer locally - while most off-the-shelf filters only create global effects.

• nitrogen 12y ago
  As well as, I'm gathering from the summary, applying local filters to each image frequency band independently ("Laplacian pyramid").
  • yichang 12y ago
    Exactly. In my viewpoint, the dense correspondence is another key factor that makes this technique work.
• jacobolus 12y ago
  How does this compare to Soonmin Bae’s work from a few years ago? (I can go read the paper, but maybe you have a quick summary?) (Another of Frédo Durand’s grad students who graduated a while back, http://people.csail.mit.edu/soonmin/photolook/ http://people.csail.mit.edu/soonmin/photolook/) Do you have any plans for future improvements which might reduce the really distracting color ringing artifacts you get in this current work? Are there any photographers whose style doesn’t transfer as well? (The ones you chose all use very high contrast and harsh lighting, with an emphasis on fine structure detail.)
  • yichang 12y ago
    Both papers use multi-scale approach to capture textures in different scales. While Bae's work is global transfer, this paper uses local transfer. This paper first uses a technique called "sift flow" in computer vision to compute the correspondences between example face and input face, and then transfer the image statistics locally, based on the correspondences.
  • yichang 12y ago
    We are looking at gain maps to see if we can reduce the artifacts. Styles with hard shadow are more challenging. eg Olaf Blecker http://i1.dripimg.com/t/460000/6554/96/450324_500_8a28bd.jpg http://i1.dripimg.com/t/460000/6554/96/450324_500_8a28bd.jpg (Note that high contrast doesn't mean hard shadow)