CODEWITHSUNDEEP
imagedeep-learningpytorchneural-networkstyle-transfer
AswinBravissimo
AswinBravissimo

Reputation: 31

Neural Style Transfer Unsatisfactory Results

I am trying to use neural style transfer(NST) to visualize the results of remodelling of a home's interiors and how it will look as per the user's style image. While the code I use gives me good results most of the time, it occasionally gives me noisy and irrelevant results. What can be the cause of this phenomenon?

This is an image where the NST has produced decent results and is able to transfer the stye quite well

The above is an image where the NST has produced decent results and is able to transfer the stye quite well

This is another situation where the NST process has failed terribly and is only outputting noisy results

This is another situation where the NST process has failed terribly and is only outputting noisy results

Is there any technical name for this occurrence? Technical papers and articles on why this occurs and how to solve this will be greatly appreciated.

Upvotes: 0

Views: 109

Answers (1)

shoab ahamed
shoab ahamed

Reputation: 1

I have recently started studying different neural style transfer models. I do not know much. But if you are using A Neural Algorithm of Artistic Style then I have some suggestion for you.

  1. In the paper conv4_2 output(relu output) was used for representing content image and they also generated the stylized image from white noise. So it becomes very hard for the model to generate the stylized image while retaining the original content image's fine details. There are two solution for this first use content image and modify it to generate stylized image or use a shallow layer like conv2_2 alongside the conv4_2 output for representing the content image. Though you did not seem suffer the same problem I suffered as the some things in your image is pretty clear. In my case the stylized image was blurry.
  2. There is also a chance that there is something wrong with your code. Did you clamp the values between 0 and 1? or normalized the input image? There are many things that can go wrong. Hope the suggestion helps although it's little late

Upvotes: 0

Related Questions