The more interesting problem is learning doc -> image directly. With Dropbox's scale of data, seems feasible for some data types.