I have been playing around the algorithm with tensorflow in this paper. I tried to convert a photo to a Chinese ink and wash painring, but I got some strange patterns in the output picture(those in the top left corner). enter image description here I totally followed the algorithm in the paper except that I add another loss term for total variation denoising. I have some confusions about this:

  1. In the paper it used the 16-layer vgg net. I am going to try some other networks, such as googlenet or alexnet. But I was wondering whether it would avoid these patterns if I use some other networks.

  2. In the paper, it used the gram matrix to represent the style loss. But I don't know the reason here. I was also wondering if there's any other way to stand for the style loss.

  3. Since I have used total variation denoising here, and the result is not satisfactory. I was wondering if there're some other ways of denoising to avoid these patterns.

(I'm sorry that I can't upload my content input image because I don't have enough reputation in stack overflow as a new user. If you're interested, you can search the google image with"Shanghai Jiaotong university" and find the original image there.)

  • $\begingroup$ I think that is an impressive result, but I have only seen other outputs of the style transfer work, not studied it. If you link the original photo and style source in a comment here, I will edit them into the question for you (I'm not going to go searching for matches). I think the both source images are important in case they have some noise element that contributes to the patterns that you do not want. $\endgroup$ – Neil Slater Jul 21 '16 at 7:32
  • $\begingroup$ original photo(the second photo in the link): m.v4.cc/News-1500016.html style source:ivsky.com/tupian/shuimo_fengjing_v3890/pic_119419.html Thank you.@NeilSlater $\endgroup$ – Xiangyu Wang Jul 22 '16 at 8:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.