In fact we're only scratching the surface when it comes to the generative capabilities of deep models. See e.g. our recent work on using Deep Boltzmann Machines to learn how to draw object silhouettes: http://arkitus.com/ShapeBM/