An implementation of a DCGAN in Keras. Not too many fancy tricks, uses Spectral Normalization and Leaky Relu.
To train, run main.py. I have not provided any pretrained models at this time.
Here's a gif of the network overfit to a single image of a cloud.
Some other images of clouds:
I also ran this GAN on a dataset of tattoos from (this)[http://tattoodles.com/] website. CW: lots of nudity This dataset was much more varied, and produced worse results. See below. Perhaps with some better hyperparameter tuning I could get more than color blobs.
