Conda Environment

To create the style-gpu environment, run:

conda create -f envs/style-gpu.yml

To create the style environment, run:

conda create -f envs/style.yml

Command-Line Usage

After installing the requisite packages (see "Conda Environment"), the style transfer algorithm can be run by navigating to the project base directory and running

./transfer.py <content_img> <style_img>

Where <content_img> and <style_img> are the filenames of the content and style images, respectively. By default, the result is saved to out.png. This can be changed using the -o argument.

A complete list of options can be seen by running

./transfer.py -h

Python API

The transfer Python function can be imported via

from transfer import transfer

This function takes the same arguments as the command-line interface and has the signature

def transfer(
        content_img,
        style_img,
        content_size=None,
        style_size=None,
        content_layer="block4_conv2",
        content_weight=1.0,
        style_weight=10.0,
        variation_weight=0.01,
        noise=0.0,
        n_iter=500):

content_img and style_img should be NumPy arrays of shape (w, h, 3). The function returns a NumPy array whose shape matches that of content_img (after being downscaled according to content_size).

See the playground.ipynb Jupyter notebook for examples of Python API usage.

Tips

The computational cost of the transfer is proportional to the number of pixels in the content image. I've found that, with a good GPU, the transfer takes about 5 minutes on a 512 by 512 image. I wouldn't recommend making the content image much larger than 1024 by 1024. The size of the content image can be changed by setting content_size.

The "brush size" (e.g. zoom level of the texture) can be controlled by setting the value of style_size. The absolute size (in pixels) of a texture matters!

The balance between matching the content image, matching the style image, and reducing noise can be tuned with the content_weight, style_weight, and variation_weight arguments, respectively. You might not notice a difference until you change a weights by a factor of ~100.

Sample Images

The content and style subdirectories contain images for testing. Many of these are quite large, so it's very important you set content_size and style_size to reasonable values!

The style images (paintings) should all be public domain. The content images are all photos I've taken myself. Feel free to use them without restriction.

Attributions

This code is primarily based on A Neural Algorithm of Artistic Style (Gatys et al. 2015). This blog post was also a source of guidance. The total variation loss term comes from Perceptual Losses for Real-Time Style Transfer and Super-Resolution (Johnson et al. 2016).