Skip to content

Deeper Understanding and Modifications proposed to the High-Resolution Network for Photorealistic Style Transfer

License

Notifications You must be signed in to change notification settings

divyanshu092/Photorealistic_Style_Transfer

Repository files navigation

High-Resolution Network for Photorealistic Style Transfer

Code taken from paper "High-Resolution Network for Photorealistic Style Transfer" and Author Implementation

I have used Google Colab Environment for the implementation changes.

Implementation Changes

The implementation changes proposed in this work are:

  1. Different layers of the network store different information regarding the content of the image and the style of the image. So, I analysed the importance of each layer and tried a weighted combination. (Task 1)
  2. I varied the ratio of content weight to style weight in order to generate images in a vast diversity. (Task 2)
  3. I tried different combinations of loss functions in order to improve the content preservation while transferring style. (Task 3)
  4. Along with content loss and style loss, I also included Total Variation Loss.(Task 4)
  5. In order to fasten the training process, I tried to use adaptive content to style weight ratio. (Task 5)

Folder details

  1. Artistic NST Images contains the results of the implementation of paper "A Neural Algorithm of Artistic Style"
  2. Content-Style Images contains all the images I have used for generating results of various combination of content and style.
  3. Content-Style Weight Ratio contains of results of Task 2.
  4. Loss Variations contains the results of Task 3.
  5. More Results folder contains the results after the best combination of all Tasks on different content and style images.
  6. Single Layer Extraction contains the results of Task 1.
  7. Swapping Results folder contains swapping between various Artworks, Day-time Translation and FLower images.

Examples

Here are some results(from left to right are content, style and output):





Acknowledgement by Author

Our work is inspired by Deep High-Resolution Representation Learning for Human Pose Estimation.

The transfer code is based on Udacity.

Contact

Author's contact (Ming Li limingcv@gmail.com)
My contact (Divyanshu Mandowara divyanshu.mandowara92@gmail.com)

About

Deeper Understanding and Modifications proposed to the High-Resolution Network for Photorealistic Style Transfer

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published