I roughly followed the deep convolutional GAN framework in this paper.
On the left, icons from the game NetHack. On the right, DCGAN generator output after being trained on the NetHack dataset.
The output of a similar deep learning algorithm (has more layers in its generator and discriminator networks) trained on landscape paintings can be found here.
Looking for a good image dataset? Try CelebA, CIFAR10, or MNIST.