That's not what I did.

I used a genetic algorithm with facial detection/recognition algorithms as fitness functions

Recognition algorithms like eigenfaces, fischerfaces will reproduce one individual.

More general detection algorithms like YEF realtime object detection as a fitness function will result in more general faces which represent some of the learned features from the LFW database. I'm not just combining images, there's lots of machine learning involved.

Recheck my site in two weeks or so I'll be posting my MS thesis about the subject with more technical details.

(Wrote this on my phone excuse the brevity)