This looks to be specifically optimized for the Moon, but they are very close to a generic algorithm that they can feed blurry cell-phone images of common photography subjects on the input, match them to high-resolution known-good stock photos. Those stock photos may be taken by pros using studio lighting and full-frame DLSRs with lenses containing more glass than the mass of the entire cell phone. Regardless of whether you're copy-pasting in a rectangular bitmap from a stored stock photo or whether your convolutional neural net applies stock-photo-ness to input data, the effect is the same: you can generate photos that look really astonishingly good given that the input optics stack has a 5mm deep total track and your sensor has less than 10mm across the diagonal.
Once you have that general framework, though, I'd expect the Samsung to want to 'enhance' Samsung phones and smartwatches (especially), Coke cans, candles, and cars, butterflies and birds, the Moon or a Macbook, rainbows or diamond rings, and anything else that's both hard to take a good picture of and also a common photography subject. Speaking of Coke cans, that gives my pessimistic, dystopian-leaning imagination an idea: I wonder if Coke could or other brands might in future develop an exclusive partnership with Samsung to send stock photos of their product, resulting in photos of your meal where the familiar red can looks especially glossy, saturated, reflective, crisp, and refreshing...while an Instagram of your meal from the food truck that has cans of Pepsi still looks like a cell phone photo.
Upscaling photos of printed pages by matching against a font library seems like a similar opportunity, though you run into the old Xerox copier issue that the transformation removes information about uncertainty. I'm sure they'd love to have pictures of people look better too, but it's more difficult when there are billions of subjects to ID.
I'm not saying that they should or should not do any of this, but when you want to sell phones with cameras that make your shots look really good, this kind of technique lets you cheat the laws of physics.
Sorry for being late in reading your response, but yes I agree that those dystopian scenarios are not outside of the realm of possibility. As I said before, what makes the moon unique is that every photo of the full moon is essentially the same image from the same angle, whereas a photo of a Coke can is going to be from all different angles, sides, lighting conditions, even sizes, and thus much harder to convincingly "enhance." But I have no doubt that Samsung Research is up to the task of figuring out even more ways to monetize our eyeballs, so your point is well taken.
Once you have that general framework, though, I'd expect the Samsung to want to 'enhance' Samsung phones and smartwatches (especially), Coke cans, candles, and cars, butterflies and birds, the Moon or a Macbook, rainbows or diamond rings, and anything else that's both hard to take a good picture of and also a common photography subject. Speaking of Coke cans, that gives my pessimistic, dystopian-leaning imagination an idea: I wonder if Coke could or other brands might in future develop an exclusive partnership with Samsung to send stock photos of their product, resulting in photos of your meal where the familiar red can looks especially glossy, saturated, reflective, crisp, and refreshing...while an Instagram of your meal from the food truck that has cans of Pepsi still looks like a cell phone photo.
Upscaling photos of printed pages by matching against a font library seems like a similar opportunity, though you run into the old Xerox copier issue that the transformation removes information about uncertainty. I'm sure they'd love to have pictures of people look better too, but it's more difficult when there are billions of subjects to ID.
I'm not saying that they should or should not do any of this, but when you want to sell phones with cameras that make your shots look really good, this kind of technique lets you cheat the laws of physics.