Let me throw in another name, I think this is essentially just constraint propagation with back tracking and making mostly random choices if there are no more constraints to propagate.

But the idea of extracting constraints from a sample images and using those to generate variants is neat.