This is my WWDC 2018 scholarship submission, which has been accepted by Apple.
The playground has a short introduction to what a Generative Adversarial Network, or GAN for short, is along with a demo of one. I've trained this deep convolutional GAN on images of all 802 Pokémon, and the generator has then learned to create images of something that resembles a new Pokémon.
One caveat, as you'll see, is that the generator currently spits out 32x32 images. I attempted to up the resolution but couldn't get the network to train properly. For an 8-bit game, however, they might just be convincing enough. 👾
The model is created using Keras on top of TensorFlow and then exported as a CoreML model that runs in the playground. The Jupyter notebook used for training along with various python scripts can be made available if someone wants it. 😊
