Generate millions of graphic design images for training (github.com)

We are interested to solve the reverse problem of finding the mask of text given an image with text on it. Any ideas? Machine-learnt or otherwise.