I’ve recently bought into the hype over Generative Adversarial Networks and decided to create a GAN that can learn to generate sprite art. For this project I’m specifically training the network with sprites drawn from a JRPG style so I’m calling the project JRPG GAN (simple right?). Training this network will be particularly interesting for me since I remember the team I worked with for most of my college career seemed to be talented in many things, but art was not one of them. It unfortunately became a recurring problem since we liked video game based projects. I’m not expecting this project to create state of the art (*ahem*) art, but it would be cool to serve as an aid.
Currently, I have collected 4000+ sprites for training the network and that took a while to say the least. I’m testing out DCGAN and WGAN architectures for the training and it took quite a bit of work to get the training to stabilize for both, but so far I like where the training is headed. Once I figure out the architecture and hyperparameters I want to use, I will probably look into moving the training to AWS Sagemaker. My laptop can handle the training, but it’s also brand new and I don’t want to wear it down after just buying it.
Hopefully I’ll will be able to integrate this project into this site soon, but for now I will just give a sample from a training.
Epoch 0
Epoch 35
