Gatys' Style Transfer and Universal Feature Transform

This is a personal mini-project. I tried to reimplement the Style Transfer methods described in the Gatys' paper and Adobe's paper. The code is written with Pytorch

Image Preprocessing phase for VGG Nets

Typically for the pretrained vgg19 (without batch_normalization) model of Pytorch, before passing to the network, the image should be:

• rescaled (to size minimum 224x224)
• transformed from BGR to RGB
• normalized with mean=[0.40760392, 0.45795686, 0.48501961] and std=[1, 1, 1]
• the matrix should then be upscaled by a factor 255.0

The postprocess should be a complete inverse of the preprocess above.

Original Gatys' method

You can try the following command:

python extract.py -c <content-img-fn> -s <style-img-fn> --size <height> <width> --lambd 0.5

All images should be located in ./images/ and with format .jpg

For example, you can try the following command

python extract.py -c girl -s stl --size 600 800 --lambd 0.5

, expectedly it will produce

Here are some other results:

compared to original pictures: