Teaching Creativity — Building an AI that can draw Minecraft skins
On many occasions have I tried to train GANs before. Mostly to create game assets. One day, this may be an entirely new level of procedurally generated game content. The challenge in those projects is often the little amount of available training data and the lack of consistency.
API Farming millions of skins
Minecraft is the exception. It is trivial to farm APIs for existing player’s skins. With some Shell and Python magic, you can easily collect millions of them. I’m happy to share my training set in case someone may be interested!
Training the network
Training the network was straightforward. There were some corrupt skins that I was able to filter out using Pillow. The top and bottom 20% (in terms of file size) were removed from the data set. I then was able to use the DCGAN PyTorch example with little modifications.
Generating new skins
The code (including the pre-trained network) to generate new skins is available on GitHub. I also sell them as NFTs, so far with limited success 😉.