- Build a DCGAN to generator Pokemon images
- Try BEGAN and disentangled VAE for it (recently working on)
- Build a pix2pix app to get a style transfer network
- Do some frontend work
Unlike face, number, landscape of typical kind, pokemons have too many different shape and color, which make DCGAN suffers.
If we choose the yellow pokemon only(you can find in folder pokemon_yellow) to train the network, or use a very small batch(say batch_size < 8), the images you generate will be a little better. But as training with only a few images, no matter how the noise Z changes, DCGAN only generates a few similar images.
If you happen to train your DCGAN with a reletively large batch (say 256), your network will confuse on what you might want it to learn. Adding up 256 different pokemon with different color and shapes will get you mess.
I will try BEGAN and disentangled VAE