I wonder what will happen when somebody combines a GAN with a feature recognizing network like the Tesla cars use, so it can use its own extrapolated map of the surroundings to stabilize its output as the camera moves around.