Example output images from my CycleGAN reimplementation.

Together with Ross Bernstein, a colleague at NYU's MSDS program, I wrote a reimplementation of CycleGAN for my Computer Vision class project. We used human faces and anime faces as the unpaired datasets. The main objective was to determine whether it was possible to reproduce the paper; secondary objectives were exploring different versions of CycleGAN to determine how parameters affected performance, and how the network performed in certain sub-scenarios such as with expressive faces or for two images of the same person, but in different scenes or poses.

As you can see from the image to the left, our implementation did pretty well. There are a few oddities, though: for instance, all the generated anime eyes look very similar, and in most of the generated human faces, the lips are also identical.

The full writeup can be found on my Github. The full code repository is also there.