Which face is real?

Originally published at: https://boingboing.net/2019/09/09/which-face-is-real.html

6 Likes

image

image

It’s so hard to decide!

ETA: I wish it was “neither”.

12 Likes

“2morrow blongs 2 me”

6 Likes

For me the teeth are always the giveaway if you are not sure.

6 Likes

What counts as “real” here? Some of the ‘real’ faces seems to have seen a lot of dental- and plastic surgery.

3 Likes

I wonder if you can distinguish an mp3 from a CD, a very good vinyl or a very good tape recoding.

I’ve got a friend who was working at Ilmenau, with the guys who invented mp3. They’ve got a reference lab there, where high-end audio equipment can be tested. They also have a perception psychology branch. One of the takeaways he dropped on me was that people actually prefer the sound of compressed formats nowadays, since it’s the usual way of listening.

Also, even so-called audiophiles can’t distinguish between the original high-end recordings and higher bitrate compressed formats. He basically said all above 160 kBit/s is definitely perceptually indistinguishable for nearly everyone, and even the best professional classical musicians can’t hear any difference if you go above 192 kBit/s. Even on their reference equipment. But that’s just a sidenote, the point is rather if we probably get used to the uncannily borked pics and prefer them above normal photos…

7 Likes

I quickly realized that the reason I beat it is because I’ve become familiar with the telltale indicators of a neural-net creator

It’s also a hell of a lot easier if you know that one out of two is fake instead of just having to spot a fake in the wild without being made aware beforehand.

12 Likes

People prefer the compressed formats, but they also can’t tell the difference? The argument doesn’t really work both ways.

1 Like

None of them. They’re all reproductions online.

12 Likes

Yes, the game quickly ceases to be “Which face is real?”, and becomes “Which background is real?”

3 Likes

I would guess: Pretty much nobody can tell the difference between a high bit-rate MP3 and the original recording, but the artifacts of lower bit-rate recordings are preferred by some listeners.

1 Like

Neither of these is the real voice of anyone and we must never believe that it is. They are both electronic images of the real voice preserved on magnetic tape; just as we never see a dead body on television news and we must never believe that we do.

5 Likes

Came in this joint for the mouth/eye swappage and left well satisfied… background doucheberry is one of my favorites. Sublime. 10/10, would browse again.

1 Like

…crowdsource the training of the algorithms…

I feel bad now that my initial reaction was that the kid on the right was fake b/c his mouth seemed off…but of course the creation on the left’s hair-slick-with-replication-tank-goo and earring-insufficiently-separated-from-garment sealed it upon examination…

2 Likes

I guess my question is how similar the AI generated faces are to actual faces. I could easily make a program that, given sufficient “training data” could “make” photos of faces that you couldn’t possibly distinguish from “real” photos. (e.g. it “learns” by compressing the “training data” then “creates faces” by simply reproducing an image it’s seen while lightening a few pixels 2%)

So is this AI making the photo pixel by pixel? Presumably not. Is it saying “I need eyes” then being trained to draw one of about twenty different eyes it has absorbed from stock photos? Is it basically taking existing images (maybe in a complex, obfuscated way) and making small alterations to them?

I’m really not sure what to make of this technology. It could be anything from a meaningful, powerful method of creating images to artificially created sentience to crass stage magic. And the way we build these things, sometimes I’m not sure the creator would even know which it was.

2 Likes

8 Likes

Oh, these things are so dumb. I can easily spot the real faces! I am much smarter than some stupid algorithm.

I went to site and spotted the real face a whopping 60% of the time. (12/20.)

Hmmf.

Now imagine there is an AI out there trying to identify faces to pass these tests. Our face generating AI wants it’s faces to pass as real, the pretending-to-be-human AI wants to select real faces, and yet another AI is classifying the images as human or non-human in the first place.

Let those all run for a couple of years and imagine the uncanny monstrosities that will be identified as “faces” in the end.

2 Likes

image

7 Likes

I’m pretty sure if you expanded this to full body compositioning and gave it some user adjustable inputs the amount of porn would increase ten fold…