Our memories are notoriously unreliable, and subject to outside influence. I want to see how this would hold up in controlled experiments. Simple experiment: have a person briefly appear in a room full of students. Run each student through this process. Quantify how close the produced images were to the person's actual face.

If the faces are more fleshed out but not accurate, then that's just as bad as having a generic likeness.

There's an academic paper on the work (George et al., EFIT-V: Interactive Evolutionary Strategy for the Construction of Photo-Realistic Facial Composites, Genetic And Evolutionary Computation Conference 2009, http://portal.acm.org/citation.cfm?id=1389095.1389384), but it only mentions using the software as a trial with police departments. There's too many variables there to give me confidence in the results.

If I were a clever defence lawyer I'd argue it would invalidate any later line-ups. This sounds like a stellar way to destroy a memory of a face.

Each time you show a person a face, you're adding new memory trace that will conflict with the real one. This will reduce the confidence in a later recognition task. You're also re-activating the real trace, allowing it to be disrupted. This would increase the false-positive rate in a line-up.

All good points and a subject to an experiment.

A possible (somewhat sci-fi) future extension to this idea would be directly monitoring the brain of the subject, in conjunction with eye movements, in an attempt to quickly detect a subconscious signal of 'recognition' while scanning the computer-generated faces. Might be that certain signals in the brainwave (e.g. the P300 or something similar: http://en.wikipedia.org/wiki/P300_%28neuroscience%29 ) could be used for this purpose.

I bet that a web application built from this would be enormously popular.

It would also be useful as a way to create and share an online avatar likeness.