What are you using as your training data out of interest? Very impressive generation.
I'm always wondering that. Where do people get 660 million labelled images from?