search

xunhuang1995 / AdaIN-style

Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization
https://arxiv.org/abs/1703.06868
MIT License
1.46k stars 192 forks source link

Strange Result without protecting the detail. #8

Closed dovanchan closed 7 years ago

dovanchan commented 7 years ago
222

girl_stylized_5 girl_stylized_9 photo_stylized_5

The result is so broken,I think a great art style should protect more detail(such as eyes ,nose ,mouth)...So I suggest you can reference this website: https://www.reddit.com/r/deepdream/comments/4ky9zp/scarlett_johansson_x4/?st=j11ntfv1&sh=fb7f6881

It's talking about how to protect detail and get a great art result;http://imgur.com/a/cH91s

image

I hold you can try to improve the algorithm for protecting detail ,It will be very beautiful, It will product a real art like this: http://imgur.com/a/cH91s Thank you

dovanchan commented 7 years ago

Hi,One more question: -crop: Center crop both content and style images beforehand , What 's the meaning of crop???When should I use it?

gsssrao commented 7 years ago

@dovanchan You need to use -crop when you would like to do a center crop before scaling (If you use -contentSize 0 or -styleSize 0 scaling will not take place for content and style images respectively). Say you have a content image of size WxH (say 640x480). Then, using -crop will crop the image into a size of min(W,H) x min(W,H) (i.e 480x480) about the center of the image. The same happens to style.

gsssrao commented 7 years ago

@dovanchan Regarding your first query, have you tried the newer model i.e using -decoder models/decoder-content-similar.t7. I seem to get better results with it for faces. Also, experiment with -alpha. Keeping smaller alpha like 0.5-0.6 along with the newer decoder model produces better results. You can see the comparison here:

woman_stylized_asheville20 6 woman_stylized_asheville10 6

Besides this for more improvement, I think that the network will have to retrained with more face data.

dovanchan commented 7 years ago

!!!! I can't agree with you anymore!

photo_stylized_te3 photo_stylized_te4 Here is the result with face ;It's strange～

dovanchan commented 7 years ago

@gsssrao Have you ever known about dreamscope?https://dreamscopeapp.com/ These results from ds are amazing~~

image

Do you have some implement can get close to it?or we can share the great parameters in neural style each other ;

gsssrao commented 7 years ago

@dovanchan They look cool. I agree some artifacts are appearing like curvy lines etc. I also got them in video. I think that the network will have to be retrained with more face data and some slight change to the loss function to remove them.

I haven't experimented with neural style a lot. But, I agree that there are a lot control parameters in it and optimizing them to look aesthetically pleasing will take a lot of time.

dovanchan commented 7 years ago

Hi,I found that you can styled the Video in the demo, pretty cool! Can you tell me how you implement that?

gsssrao commented 7 years ago

Try out this repo for now.

dovanchan commented 7 years ago

Cool,Thank you; In my opinion，If AdaIN can improved by training more and better data,and made more flexible parameter for getting great result, AdaIN will be a best algorithm。 With taking so much time and memory in neural style,AdaIN is a great research direction Seems all of big problem can be sloved in neural style.The only problem is the result; What we need to do just prepared more useful data for traing! I am looking forward a subversive revolution in neual style