This looks very interesting. I wonder how much data would be required in order to imitate a particular handwriting style? i.e. if I know someone with a lovely penmanship and I can ask that person to create such a reference data set, how much should that person write in order for the network to be able to reproduce the style reliably?