Using a corpus of Pokemon front/back images and a variation of the DCGAN-Tensorflow implementation found here: https://github.com/carpedm20/DCGAN-tensorflow , we trained the network to generate new Pokemon from all existing Pokemon. We standardized and cropped all our images from sprite sheets found online, making sure the images all followed the same general sizes. After each epoch, a sample image was generated, slowly appearing more and more like Pokemon. The network is far from optimal -- but it shows promise in generating pixel art game assets. The question becomes -- are these works of art considered original? See the youtube video in the videos section (watch at a slower framerate for a better idea), and often you will see Pokemon generated that are very clearly like existing ones.
We also tried a side-by-side with the same network configuration but a corpus of inventory item pixel art (potions, weapons, armor, materials, monster drops, etc). In the end, most of what was generated were nearly identical to a single type of potion in the corpus.