Recently I have been experimenting with techniques to simplify images, ideally to the bare essentials. Also been more selective with the datasets, often using first a larger set of landscapes and cityscapes, then continuing with a smaller, focused set, such a collection of photos from the Middle east. Sometimes mixing geometrical shapes and real photos. Also tampering with the neural network architecture, such as applying blurring when the image is still on a conceptual level, has not yet developed into the pixel level.
Aside of my usual drive towards simplicity and reduction, I also sometimes let the network expand the image beyond its natural size, so that the inherent features of the image are multiplied, varied and mixed in a seamless way.