CS 523: Multimedia Systems - Generative models for creative intelligence
Spring 2017
http://creativecoding.evl.uic.edu/courses/cs523/
GANmons generates new images and attributes of 'monsters' using Generative Adversarial Networks (GANs) trained on a dataset of existing creatures and attributes. The goal is to train our neural network on a dataset of existing Pokémon images and attributes (name, type, abilities). After being applied to our GAN, we expect our new GANmons images and attributes to have a similar form to the originals in terms of how appearance and attributes are associated. For example a bird or flying-type GANmon would likely have wings or feathers of some kind. Similarly a grass-type would perhaps have a leafy or plant like appearance. In addition we hope to find interesting results in how names and other attributes such as abilities are associated.
In addition to our training model, we would also like to allow the user of our GAN to be able to interactively sketch out what the desired GANmon appearance similar to edges2cats an implementation of pix2pix by Isola et al.. Our model will then display an image of the GANmon and display its attributes.