Pytorch-Neural-Style-Transfer

A PyTorch implementation of neural style transfer with color control described in the papers:

• A Neural Algorithm of Artistic Style
• Controlling Perceptual Factors in Neural Style Transfer

Examples

The Neckarfront in Tübingen, Germany

The results were obtained with the default settings except scale_img=0.5. Tested on Ubuntu 18.04 and python 3.6.9

When you reduce the image size, the style becomes more rough. Images from left to right: original image, scale_img=1.0, scale_img=0.4

Also you can control color preservation. Set preserve_colors=True to keep original content image colors. Example: original image, preserve_colors=False, preserve_colors=True

Requirements

• PyTorch >=1.7.0, torchvision >=0.7.0
• numpy, matplotlib, Pillow

You can install this with the command:

pip install -r requirements.txt

Usage

python run.py --content <content_file> --style <style_file>

Example : python run.py --content images/tubingen.jpg --style images/starry-night.jpg

Arguments

Required :

• --content: Path to content image.
• --style: Path to style image.

Optional :

• --output: Filename of the output image. Default : result.jpg
• --iters: The number of iterations to run. Default : 500
• --preserve_colors: Save colors of the content image. Default: False
• --img_size: Size of the output image (set None to match to the size of content image). Default : None. Example : 512 or 768 1024
• --scale_img: Scaling factor of the image size. When img_size is not None, the new size will be scaled. Default : 1.0
• --print_every: The number of iterations to show info (set 0 to not show info). Default : 0
• --content_layers: VGG19 layers used for content loss. Default: r42
• --style_layers: VGG19 layers used for style loss. Default: r11 r21 r31 r41 r51
• --content_weights: Content loss weights for each content layer. Default: 1.0
• --style_weights: Style loss weights for each style layer. Default: 0.2 0.2 0.2 0.2 0.2

Using in code

Single image prediction. predict() method receives a path to image, a pillow image or a numpy array and returns pillow image. You can use all the parameters mentioned above.

from style_transfer import StyleTransfer
from PIL import Image
content = Image.open('images/tubingen.jpg')
style = Image.open('images/starry-night.jpg')
st = StyleTransfer()
img = st.predict(content, style, scale_img=0.5, print_every=100)

You can create a generator and set yield_every parameter as any number to get images during style transfer process.

gen = st.predict_iterator(content, style, yield_every=50)
images = []
for image in gen:
	images.append(image)

Also there is a predict_hr() method, that applies style transfer two times to the same image, but in different scaling, which allows you to get an image with better quality.

Parameters:

• hr_scale - scaling factor of the image relative to original content image after first style transfer. Default : 1.0
• hr_iters - the number of iteration to run second style transfer. Default : 400

img_hr = st.predict_hr(content,style,iters=300,scale_img=0.5,hr_scale=0.9)

See more in example.ipynb

References

The implementation is based on the projects:

• https://github.com/leongatys/PytorchNeuralStyleTransfer
• https://github.com/leongatys/NeuralImageSynthesis/
• https://github.com/hwalsuklee/tensorflow-style-transfer