VERY interesting - as soon as I read the title, I thought "finally, this will really tell me whether deep earning is bullshit or not!". Why? Because my understanding is that colorization is basically magic, like, how can you possibly get the color of a telephone handset out of a black and white image? In fact, you can't! Here's the set of the Addams Family:

http://www.fastcodesign.com/3021327/asides/the-addams-family...

Totally all over the map. So I thought this would be such a fantastic, fantastic way to compare my knowledge of the world with a punk algorithm's.

I didn't read the key/legend/explanation - as soon as I saw the first set of 3 images, I knew the middle lamp was the true color, and the right lamp was generated; because lampshades overwhelmingly look like the middle picture (in my mind) - https://www.google.com/search?q=lampshade not that weird blue color.

"Ha, stupid algorithm", I thought. "Who has a blue lampshade". this algrithm doesn't even come close.

Then I kept scrolling with that assumption, and it got worse and worse - wow that photographer's color is like blood orange, the algorithm doesn't even know it's a person! This is terrible. Where does that truck get that green trim, nobody would choose that, straight out of left field.

Until I got to the field with wolves. Why are the flowers' colors missing from the middle picture? This doesn't look right at all.

Then I read the caption. The middle images are the generated ones; the right-hand images are reality.

For five 5 of 6 images, I thought that the generated image was "obviously" an actual photo, and much more plausible in colors than the right-hand real photos. Continuing to scroll, for the park bench also I think the middle image is much closer to how I imagine it.

So we are at the stage where an algorithm generates a much more plausible view of reality, with rare exceptions, than actual reality. This is pretty impressive.