I wonder if the "deep learning" is getting any hints from the bayer patterns that may be detectable even in the desaturated image. It would be interesting to see what it did from a true b/w sensor, a Fovean sensor, or a scan of black and white films.

On a related "deep learning" topic:

We do a lot of work with 3D modeling tools and scanning large objects, including people, at our facility.

One thing we realized that should be possible with "deep learning" is taking a standard human computer model, and configuring it to match the position and shape of a scanned human model, or a photo of a human being -- in order to add back missing bits. Imagine a website where you can upload a swimsuit image and get back a computer generated nude image with the obscured body parts replaced. This should be very doable today, and would make a very popular website!