CartoonGAN

One or two years ago, I tried CartoonGAN to convert photos into anime like pictures. “GAN” is an abbreviation for “Generative Adversarial Network”, a deep learning technology that learns features from prepared images and generates new pseudo images. When I saw my recent photos of black dragonflies, I wondered what would happen to them, then I tried it.

The code for CartoonGAN was downloaded from the GitHub below. The explanation is also there.

https://github.com/Yijunmaverick/CartoonGAN-Test-Pytorch-Torch

It is MIT license. The width of the generated image is 450 pixels as default, so I changed it to 800 pixels in test.py file because it seems a little bit small. The pictures above are the results (the original photo is in the previous article). I like Hayao and Shinkai flavors more than others by the way.

This is interesting itself, but during working on it, I further wondered what would happen to photos taken by old lenses with a strong swirly bokeh. So, I took pictures of a road with a tunnel of cherry trees. This is the normal photo.

And this is the shot with Helios 44M-6 with a swirly bokeh.

After processing the Helios 44M photo with CartoonGAN, the result is as follows.

I thought combination of old lenses and the new technology is even more interesting because image distortion created by old lenses could be exaggerated by computing.