There is a reason Google's initial DNN project was called "deep dream"- It's because it essentially uses "dreaming" to generate new content based on a source database of unrelated images.

It is making up new image details (highly convincing and realistic new image details) and hence this tech is completely inappropriate for handling criminal evidence.

It's definitely not appropriate for suggesting that a certain person is in the video. Instead, what it can do is help (or poison) the imagination of a viewer, by inserting what a person matching the few details that are in the image/video would look like with average other details added.

Maybe your person has a darker pixel on the one side of their face you see: might be a mole, a face tattoo, or a five-o'clock shadow, or maybe they've just got a bit of mud on their face from falling in the dark. DNN will 'helpfully' provide that large facial moles and tattoos are less common than facial hair in their database, and most of their models wipe mud off their face before having their picture taken, so it will give your suspect a bit of realistic salt-and-pepper growth. Or maybe it has a bunch of interesting pictures of a guy with some facial hair and face tattoos that it can fit to the video. Suddenly you find you've been burglarized by Post Malone.

It's only useful if you're trying to get an idea of what an average situation looks like because you can't imagine what you're seeing in the noise.

By your logic, criminal sketches would be deemed inappropriate for criminal investigations as well. They’re both imperfect, but useful at the same time

Thank you for this. I'll take a look at Pixelmator Pro and their "ML Super Resolution." It seems credible to me that some kind of ML approach ought to be useful here.