This Person Does Not Exist
February 14, 2019 2:34 PM   Subscribe

This Person Does Not Exist is a single-serving website that does only one thing: algorithmically generate fake human faces. Refresh for more.

Some of them look pretty real. Most of them have uncanny (but frequently fascinating) flaws. It seems particularly challenged by hairlines.

Some analysis and context here and here.
posted by penduluum (96 comments total) 56 users marked this as a favorite
 
Now the twitter bots can have individualized avatars, instead of stealing ones off stock photo sites.
posted by msbutah at 2:37 PM on February 14, 2019 [29 favorites]


After reloading a few times, my strongest impression is that maybe you shouldn't source your dental training data from UK photos...
posted by turbowombat at 2:39 PM on February 14, 2019 [6 favorites]


I'm not getting uncanny valley at all. This is a first for me with these things.
posted by clawsoon at 2:41 PM on February 14, 2019 [11 favorites]


This is really useful for creating fake personas when doing design mockups. fakenamegenerator.com should add it to their generator.
posted by Foci for Analysis at 2:43 PM on February 14, 2019 [6 favorites]




Oh well holy shit it just served me this. Thanks for the nightmares...
posted by penguin pie at 2:46 PM on February 14, 2019 [11 favorites]


How to recognize fake AI-generated images

Wait, I'm asymmetrical and have messy hair. AM I REAL
posted by grumpybear69 at 2:50 PM on February 14, 2019 [15 favorites]


I'm pretty far along the facial-blindness continuum. (Roughly at the "can't follow an action film featuring three white guys with similar haircuts" level. Not the "can't recognize my colleagues in the hallway" level.) So, I'm not terribly well equipped to judge this project.

But. . . I never would have suspected these aren't real photographs of people if it weren't pointed out. The arXiv paper in the second link sure makes it seem legit. Otherwise, I'd be deeply skeptical that this isn't just a prank. Neat!
posted by eotvos at 2:50 PM on February 14, 2019 [7 favorites]


Part of the amusement (?) of this site is to reload for a new face and see either something really plausible or else ... errr/ewww .
posted by King Sky Prawn at 2:51 PM on February 14, 2019


Ha, I didn't get anything nearly as bad as that one, penguin pie. The only screwups I saw were a few minor patches of droopy and unconvincing skin texture. About half the pictures I looked at were totally passable as real peeps.
posted by MillMan at 2:52 PM on February 14, 2019 [1 favorite]


Genuinely afraid I'm going to hit refresh one more time, and it's a picture of me, and then I no longer exist
posted by theodolite at 2:52 PM on February 14, 2019 [76 favorites]


Most of them have uncanny (but frequently fascinating) flaws.

As we've seen here on MeFi previously, this doesn't seem to the state of the art. In fact, it's not even all that good.

Is "thisalgorithmisnotallthatgood.com" available?
posted by The Bellman at 2:53 PM on February 14, 2019


I should have said "hairlines and pupils." More than half of the pictures I get look like screencaps from those YouTube videos where conspiracy cases try to prove lizard people are real because their pupils change shape on third-generation encodes of SD TV broadcasts.
posted by penduluum at 2:54 PM on February 14, 2019 [3 favorites]


MillMan - you just need to keep on reloading, you'll eventually get one that's ... oh my. Other than that, the most common anomalies I see are accidental androgyny and monster faces and limbs in the background. Like a tin god that only has enough power to concentrate on the foreground object and just fakes the rest.
posted by King Sky Prawn at 2:54 PM on February 14, 2019 [3 favorites]


The link from gwint above is useful and the next half dozen generated faces after this all showed telltales: asymmetrical earring choices, skin tone differing on opposite sides of a smile line, and odd backgrounds (one cheerful woman seemed to have been photographed in front of a bead curtain consisting of human teeth strung together).
posted by ricochet biscuit at 2:57 PM on February 14, 2019 [6 favorites]


BRB. Just gonna generate a billion of these, upload them to Facebook and tag myself in all of them.
posted by suetanvil at 2:58 PM on February 14, 2019 [36 favorites]


Pretty sure I went to college with some of these people and/or would say oh hello...Jennifer? If I saw them on the street.
posted by CMcG at 2:59 PM on February 14, 2019 [1 favorite]


Genuinely afraid I'm going to hit refresh one more time, and it's a picture of me, and then I no longer exist

Not until you refresh again. This scene was cut from Avengers: Infinity War for time.
posted by ricochet biscuit at 2:59 PM on February 14, 2019 [6 favorites]


Going by the results I got, white faces predominate among the nonexistent.
posted by Morpeth at 3:01 PM on February 14, 2019 [3 favorites]


I can't help but make judgements about what kind of people these fake people are. Some of them look to keep it real!
posted by GoblinHoney at 3:01 PM on February 14, 2019 [1 favorite]


Is anybody else noticing that a weirdly high percentage of the faces have pronounced scars?
posted by saladin at 3:03 PM on February 14, 2019


anyone else see the one with a guy in the frame, pretty normal looking, but there is another face, half cut off, to the left? I just wanted to warn the guy!! run! run!! melting weird face person is next to you. I was going to link but accidentally clicked back in browser...
posted by supermedusa at 3:03 PM on February 14, 2019 [9 favorites]


I have seen calls to deal with the problem of deepfakes by regulating the underlying technology. Stuff like this is just a rearrangement of the same building blocks, though; it is very, very closely related.

It would be like trying to deal with 3D printed gun parts by regulating all 3D printers.
posted by vogon_poet at 3:05 PM on February 14, 2019 [1 favorite]


There's a sci-fi story to be written about the person clicking through this thing who is suddenly confronted with their own face.
posted by nubs at 3:06 PM on February 14, 2019 [7 favorites]


algorithmically generate fake human faces

I can't tell you how happy I am that this isn't a post about AI generating fake human feces!!
posted by Greg_Ace at 3:08 PM on February 14, 2019 [18 favorites]


The way this site works btw is it’s got a big batch of generated faces and it shows you one at random, but every 2 seconds a new one is created.

That constraint may be why they arent quite as good as NVIDIA’s best; don’t want to go broke paying for cloud GPUs.
posted by vogon_poet at 3:09 PM on February 14, 2019 [5 favorites]


suetanvil: best idea I've seen all day. Would you pay for it? Culture Jamming As A Service.

Does anyone know if these are being generated in real-time? They can't be, can they?
posted by Leon at 3:10 PM on February 14, 2019 [1 favorite]


a post about AI generating fake human feces

Now I'm disappointed.
posted by biogeo at 3:11 PM on February 14, 2019 [2 favorites]


thisshitisunreal.com

Someone get on this.
posted by biogeo at 3:15 PM on February 14, 2019 [3 favorites]


Some of these photos seem very "alive". Like there's a soul there, behind the eyes. And now I am officially freaked out.
posted by Leon at 3:15 PM on February 14, 2019 [6 favorites]


I've looked at probably as many faces in my lifetime, and I certainly can't create a better one than these, especially in 2 seconds. This is pretty damn cool and quite impressive imho.
posted by OnefortheLast at 3:22 PM on February 14, 2019


I'm curious about how the 'discriminator' works that judges whether a face is real or not. I assume that's a part of the GNA? It's another algorithm?
posted by crazy with stars at 3:26 PM on February 14, 2019


I cannot tell. At all.
posted by agregoli at 3:29 PM on February 14, 2019


Genuinely afraid I'm going to hit refresh one more time, and it's a picture of me, and then I no longer exist

First, you have to go to the Magnus Archives and make a statement.
posted by GenjiandProust at 3:30 PM on February 14, 2019 [2 favorites]


I found a really cute one. It wouldn't have killed them to add a fake phone number attached to a text generator.
posted by zompist at 3:32 PM on February 14, 2019 [12 favorites]


This is what happens when a hobbit gets into a bar fight with John Wick.
posted by the webmistress at 3:33 PM on February 14, 2019 [1 favorite]


I'm curious about how the 'discriminator' works that judges whether a face is real or not. I assume that's a part of the GNA? It's another algorithm?

Big neural network that takes an image, and at the end spits out 0 for fake and 1 for real. You train it by showing it a mix of fake and real until it seems to be getting it mostly right.
posted by vogon_poet at 3:39 PM on February 14, 2019 [1 favorite]


Some of these look really good but even the really good ones have telltale signs of being generated. Like it's already been said this is a couple years behind the state of the art.

I still find it fascinating that nobody seems terribly interested in utilizing this technology for the things people are afraid it's going to be used for. Fake Facebook accounts use easily-sourced pictures of scantily-clad real women. Idiots making fake Linkedin accounts for his dumbass fake intelligence company use real celebrities. Fake twitter accounts set up by ostensibly savvy social media firms use stock photos of real people. To me, "This is fake because it's a fake AI generated face" is a harder argument to make than "This is fake because it's just a cropped headshot from this stock photo site" so even at this point I'm surprised it's not more common.
posted by Mr.Encyclopedia at 3:42 PM on February 14, 2019


I'm pretty far along the facial-blindness continuum.

As am I (to the extent of being a subject in quite a few studies of it.
Although it turns out that I don't have face blindness, but face amnesia.
It manifests in things like not getting Pareidolia. Basically no automatic face circuits. It's all manual.

One of the studies tested my face seeing power, and I'm especially good (like significantly better than normal humans) at recognising upside down faces. This seems similar in that I recognise that these faces are not right in very specific ways. There's no uncanny valley, because I'm not using the automatic face circuits that humans have. But I can specifically spot small flaws that do things that faces don't do.
posted by Just this guy, y'know at 3:46 PM on February 14, 2019 [7 favorites]


The faces this thing creates are really quite convincing OH GOD WHAT IS THAT THING?.
posted by The Tensor at 3:51 PM on February 14, 2019 [13 favorites]


Is there one of these out there that one could tune for specific ages and features? I would really love to generate some specific character archetypes with something like this.
posted by Caduceus at 3:56 PM on February 14, 2019


Yeah, the whole “other people’s faces in the frame” aspect of this reeeeeally needs some work.
posted by obfuscation at 3:59 PM on February 14, 2019 [4 favorites]


Each time I refresh, I’m scared/excited I’ll see a face that looks like me.
posted by theappleonatree at 4:05 PM on February 14, 2019 [2 favorites]


The creepiest thing about this is that the eyes are always in the exact. same. place. (and I think the same distance apart, no matter what angle the head is at).
posted by oulipian at 4:05 PM on February 14, 2019 [4 favorites]


Here is the amount of progress for state-of-the-art over the last 4 years.
posted by vogon_poet at 4:09 PM on February 14, 2019 [2 favorites]




The faces this thing creates are really quite convincing OH GOD WHAT IS THAT THING?.

These clearly aren’t bring generated afresh each time: I got exactly the same image an hour ago.

ETA: the Cthulhubombed one above, too.
posted by ricochet biscuit at 4:21 PM on February 14, 2019 [1 favorite]


I like how in this image from the "How to recognize fake AI-generated faces" article linked above, the GAN at one point was just like, fuck it, Jon Hamm is a hunk, that's what you're getting this run.
posted by invitapriore at 4:22 PM on February 14, 2019 [1 favorite]


I'm not getting uncanny valley at all. This is a first for me with these things.

Keep going until you find faces of kids. There are some glassy-eyed moppets from the pleasant little town of Uncanny Valley in there, I can tell you.
posted by ricochet biscuit at 4:25 PM on February 14, 2019 [3 favorites]


There should be real faces mixed in, and a Turing-type test attached to the feed. When human observers can no longer tell which ones are fake and which are real, the algorithm has won.
posted by beagle at 4:28 PM on February 14, 2019 [6 favorites]


The face on the bottom right of invitapriore's linked image looks like James May and Willem Dafoe had a love child.
posted by Greg_Ace at 4:30 PM on February 14, 2019


Some variation of this could be great for bias testing if you could use them to replace faces in videos. Somebody has a gun... quick, threat or hero?
posted by clawsoon at 4:31 PM on February 14, 2019


Oh man. Cthulhubombed - yes. No more perfect description is possible.
posted by obfuscation at 4:42 PM on February 14, 2019




Hair lines seem to be the big tell for me. It looks like GAN gets really confused on foreground/background boundaries when it's trying to deal with a few strands of hair, or textured/patterned shirts. Just came across a woman who had an eldritch horror hiding in her long hair. Then there's the unfortunate guy whose hair, beard, and skin color not quite similar enough to hide his tentacles, but I'm not quite sure why he has a mouse on his head.

I realized looking at Janelle Shane's GANcats post that GAN draws cats like a toddler, getting confused a bit about proportions because balancing detail and gestalt is hard, and then makes it exceptionally creepy by putting a photorealistic texture on it.
posted by GenderNullPointerException at 4:49 PM on February 14, 2019 [7 favorites]


Fun fact: after the first hundred, it is all just pictures of Stephen Miller.
posted by ricochet biscuit at 4:56 PM on February 14, 2019 [9 favorites]


It seems to have a surprising amount of trouble with ears. Ears not on the same level, earrings not matching, etc.
posted by tavella at 4:59 PM on February 14, 2019 [5 favorites]


Some of these are pretty good but there are some very dodgy titfers.
posted by unliteral at 5:02 PM on February 14, 2019 [1 favorite]


Totally unrelated but amazing (from the linked image set...)
posted by I_Love_Bananas at 5:04 PM on February 14, 2019 [3 favorites]


So far, all the kids I've seen look like the drank from the wrong grail.
posted by robocop is bleeding at 5:12 PM on February 14, 2019 [5 favorites]


I can't think of any good use for this. The people working on this should be mildly ashamed.
posted by great_radio at 5:28 PM on February 14, 2019 [2 favorites]


Yeah, the whole “other people’s faces in the frame” aspect of this reeeeeally needs some work.
I'll say
posted by unliteral at 5:28 PM on February 14, 2019 [1 favorite]


This guy seems completely realistic but what the hell is wrong with the person beside him?
posted by Pater Aletheias at 6:25 PM on February 14, 2019 [1 favorite]


a post about AI generating fake human feces

...and every single one of them smack dab in the middle of the Uncanny Valley.
posted by CynicalKnight at 6:31 PM on February 14, 2019 [1 favorite]


a post about AI generating fake human feces

...and every single one of them smack dab in the middle of the Uncanny  Valley 
Privy.

FTFY.
posted by Greg_Ace at 7:11 PM on February 14, 2019


1. The program has an issue with aligning the teeth properly. In a smile the gap between incisors should fall pretty much directly below the center of the nose. Almost all the images have the incisors gap off center enough to be noticable.
2. The pupillary diameters of both eyes are not equal in many of these. Also the ai really like heterochromatic eyes.
posted by Chrischris at 7:16 PM on February 14, 2019 [5 favorites]


I got elf-Bjork with a cotton flower in her mouth and a slender finger embedded in her bangs.

Good night everybody!
posted by RobotVoodooPower at 8:55 PM on February 14, 2019 [4 favorites]


I can’t wait to use these in UX personas and customer journey maps. must resist the temptation to go with my first five choices though …
posted by iamkimiam at 9:10 PM on February 14, 2019


There should be real faces mixed in, and a Turing-type test attached to the feed. When human observers can no longer tell which ones are fake and which are real, the algorithm has won.

For this test to be valid, ten percent of the humans would have to standing next to ILM- or WETA-produced maquettes of severely misshapen faces, just visible at the edge of the frame. And no one could be wearing matching earrings either.
posted by ricochet biscuit at 9:39 PM on February 14, 2019


Along with the mismatched earrings, I’m especially digging the people in eyeglasses. Inevitably one lens is round and the other is square, or one is trapezoidal and the other is a cat-eye.

Warby Parker, are you listening?
posted by armeowda at 10:05 PM on February 14, 2019 [1 favorite]


They sure only trained this on white people
posted by latkes at 10:10 PM on February 14, 2019 [1 favorite]


The creepiest thing about this is that the eyes are always in the exact. same. place.

Which means you could very easily make a super-creepy montage of each face morphing into the next one.
posted by straight at 10:38 PM on February 14, 2019 [2 favorites]


They sure only trained this on white people

Yeah, I was going to make a joke about using these for the person-on-the-street quotes in The Onion (those poor stock-photo people have had so many horrible quotes put next to their faces), but then realized I would be suggesting they replace all of the people of color with a simulated white person.
posted by straight at 10:45 PM on February 14, 2019


The inevitable AI video is going to be even creepier.
posted by iamck at 10:45 PM on February 14, 2019 [1 favorite]


What will be really interesting is how fast Facebook can train an algorithm to distinguish these pictures from real people.

Right now the easiest way would be to measure where the eyes are, so if you want to use these for suetanvil's project, you might want to crop or stretch them a bit.
posted by straight at 12:36 AM on February 15, 2019


Wow, the logo on this baseball cap is definitely one the computers designed to use when they rise up and make us wear hats like this.
posted by straight at 12:41 AM on February 15, 2019 [3 favorites]


What will be really interesting terrifying is how fast Facebook can train an algorithm to distinguish these pictures from real people.

posted by straight at 7:36 AM

There. Fixed it for you.
posted by Faintdreams at 1:06 AM on February 15, 2019


Holy crap, somebody dug up the corpse of Dinesh D'Souza and brought him back to life with Frankenstein bolts in his neck.
posted by straight at 2:10 AM on February 15, 2019


I’m especially digging the people in eyeglasses. Inevitably one lens is round and the other is square, or one is trapezoidal and the other is a cat-eye.

You say that like it's a bad thing.
posted by radwolf76 at 2:11 AM on February 15, 2019 [2 favorites]


Since looking at a bunch of these, I now find that I'm looking more closely at ears and hairlines and teeth in photos of real people, which I never did before. Help!
posted by Samarium at 3:49 AM on February 15, 2019 [7 favorites]


I see an unusually high number of unknown people each week with something on the order of 12000 passengers disembarking from cruise ships at the pier next to my office. My brain just decided that the first ten faces on the generator were that crowd, and quickly dismissed any weirdness about them.

I think I need a vacation.
posted by halfbuckaroo at 4:03 AM on February 15, 2019 [5 favorites]




That one, and the one I saw with just a corner of rimless glasses coming from only one side of the face, really give a glitch in the matrix vibe where I can imagine seeing this being in the world and realizing my whole fucking world is fake
posted by latkes at 6:54 AM on February 15, 2019 [1 favorite]


That's really astonishingly good. The first several I got wouldn't have made me even consider they were computer generated.

We're edging ever closer to some really interesting and/or disturbing levels of tech. Beginning with the above mentions of realistic avatars for bots that **DON'T** show up on a reverse image search as the typical low effort bot with an avatar does. But also moving into other potentially nifty prospects. Games and movies would benefit from having the ability to populate a crowd with procedural generated unique faces, especially if you can apply some animation to the face which doesn't seem entirely impossible.

I'm wondering about future efforts that generate a skull, muscles, and so on so as to provide an entirely articulated and animatable face. We're already working on superior chatbots for specific commercial purposes (making appointments, ordering fast food, low level tech support, various call center jobs, etc). One with a face might be useful for video chat or what have you once we've got the basic AI part worked out.

Plus, again, the entertainment industry which would doubtless find all manner of ways to make use of such tech.
posted by sotonohito at 7:29 AM on February 15, 2019 [1 favorite]


I did get one yesterday of a black man. he looked pretty normal on the whole except for his head...it just kept going a little to much...up...and...out?? either that or very peculiar afro...
posted by supermedusa at 8:46 AM on February 15, 2019




How does it know?
posted by filtergik at 9:26 AM on February 15, 2019 [2 favorites]


Further observations: male hairlines are seldom the right age for the rest of the face; hats appear to be repurposed tote bags or sports cars; children’s shoulders seem to go on forever, as if they are a rolling terrain out of which young heads grow.
posted by armeowda at 9:30 AM on February 15, 2019 [4 favorites]


The "how to spot.." was amazing but then, I think Donald Trump and his inner circle are hard to tell from bad AI.
posted by bonobothegreat at 6:57 PM on February 15, 2019 [1 favorite]


"This Person Does Not Exist" is probably mostly true, I think.

But I'm not sure if it's as true as it might first appear.

When you look at videos of the same type of network, StyleGAN, trained on kittens or cars, you can see image watermarks and logos appear at certain points when apparently unrelated features like size or orientation are changed.

To me that's sign that sometimes it's picking up on certain individual images, rather than making completely generalised observations.

I can't seem to find how big those cat or car datasets were, but the faces were made with a dataset from Flickr containing 70,000 images.

Remember the morphing scene from Black or White?

That seems very dated now, and it would seem hyperbolic to freeze frame halfway through one of those morphs and say "this person does not exist", because we can clearly see it's based on two people who do exist.

Without trying to find generated people in the dataset, it's difficult to tell how novel the StyleGAN people really are. For example, one image could be real person A with different coloured hair, real person B's nose, glasses, more wrinkles, and the background from image C.

If that was the case it would definitely be more novel than just a morph between A and B, but if you were person A you might not fully endorse the assertion that this person doesn't exist.
posted by PJMcPrettypants at 5:39 AM on February 16, 2019 [1 favorite]


This AirBNB Does Not Exist by Christopher Schmidt, complete with textual listings.

This Cat Does Not Exist, which appears to be a side project of This Person.
posted by Nelson at 8:23 AM on February 20, 2019 [4 favorites]


I wish this cat existed, wearing his fez and showing off his minimalist hooked rug art.
posted by moonmilk at 10:28 AM on February 20, 2019


I was enjoying this until I got a cute toddler with a bullet hole in the center of his forehead. PM me for a screenshot.
posted by JimInLoganSquare at 2:36 PM on February 20, 2019


Judge Fake People. 2,000 harvested faces for you to rank. Don't bother looking at highest and lowest, there are so few votes at the moment that all the faces appear in both.
posted by unliteral at 4:18 PM on February 25, 2019


Yeah, but note that the top 15 are all young, "attractive" (if you can get over the uncanny valley) women.
posted by straight at 12:25 PM on February 26, 2019 [1 favorite]


Dudes are in front now [this will change over time].
posted by unliteral at 2:51 AM on March 2, 2019


« Older Victorian vinegar Valentines: Why send me such...   |   Take Me To Church except he’s practicing it alone... Newer »


This thread has been archived and is closed to new comments