search

xunhuang1995 / AdaIN-style

Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization
https://arxiv.org/abs/1703.06868
MIT License
1.47k stars 192 forks source link

Improvement #16

Open hristorv opened 7 years ago

hristorv commented 7 years ago

Hello, I am currently trying to improve AdaIN in a Tensorflow implementation. What do you think can be done to improve the algorithm? What can be done in order to have quality similar to this: https://github.com/lengstrom/fast-style-transfer I am trying to implement as stated in the paper, histogram losses. I implemented method for histogram matching on 2 images, but i have no idea how to move on. I need some ideas how to improve the algorithm.

diggerdu commented 7 years ago

S+U GAN may make output more naturally.

hristorv commented 7 years ago

Is it even possible for this algorithm to achieve the same quality as the fast style transfer?

ArturoDeza commented 7 years ago

A workaround I did was to add a SuperResolution pix2pix module on top of the output of the Decoder to try to map back to the original image. This was used in a recent paper we submitted to NIPS: https://arxiv.org/abs/1705.10041

CJHFUTURE commented 7 years ago

Interesting, what would be the effect on the output from this?

hristorv commented 7 years ago

@ArturoDeza What were the results ? Do you have any pictures we can compare ?

ArturoDeza commented 7 years ago

@Sugarbank @hristorv I still need to add the result of adding the SuperResolution module in the Supplementary Material for the ArXiv version 3. As of yesterday I added some extra images that the model produces on the updated ArXiv version (Version 2). I guess the Base_Line images (which use the super resolution module) in the paper are the closest approximation to the original.

ZhongliangGuo commented 9 months ago

see the paper EFDM, I think its matching function or norm (in the context of adain) is quite same as HM