• For a general description and pictures: style-transfer-theory.pdf.
• For an explanation of the use of L2-Wasserstein distance instead of the loss function from the Gatys et. al framework see why_wasserstein.ipynb.
• Demo video

tl;dr - Passes a subject and style image through the convolution layers of the vgg19 network. Extracts probabalistic descriptions (first two moments) of the convolution filter activations. Calculates L2-Wasserstein distance between these probability distributions and then modifies the subject image optimally to minimize this distance.

Requires:

1. Python 3 (w/ NumPy and PIL)
2. Tensorflow (tested w/ version 1.3)
3. 'vgg_conv.npy' binary (115.5 MB) which contains calibrated convolution filters from vgg19 network. Can be downloaded here [md5sum: bf8a930fec201a0a2ade13d3f7274d0e]

Basic Usage from Command Line:

python basic_styletrans.py --subject media/wave_small.jpg --style media/kngwa_small.jpg --output media/wave_kngwa.jpg --vggnet vgg_conv.npy

Output:

loss: 118,500,040.00
ALL DONE 
post clip | time: 49.0 final loss: 2,771,486.50
synthesized image saved: media/wave_kngwa.jpg