I have no face and I'm offering a screaming deal on this armored personnel carrier

This is the result of applying Efros and Freeman to dong style transfer.

• There isn't really an easy way to stop it from looking like an impressionist painting, regardless of what you're transferring from
• Applying this to video gives results that are surprisingly temporally stable

• Doing nearest neighbor over the 16 principal components of patches works well, you don't need to scan over all patches

• Results using large patches (eg 32x32) are garbage (but look cool sometimes)

• Results using small patches (eg 16x16) are less garbage but are too close to the target image (and videos look rotoscoped)

This is a negative result but maybe it's interesting anyway.

Here's the IPython notebook.