'This Person Does Not Exist' Website Uses AI To Create Realistic Yet Horrifying Faces (inverse.com)

A website that uses AI -- Generative Adversarial Networks (GANs) -- to generate photos of people who do not exist is circulating on social media and forums this week. A news writeup adds: Every time the site is refreshed, a shockingly realistic -- but totally fake --picture of a person's face appears. Uber software engineer Phillip Wang created the page to demonstrate what GANs are capable of, and then posted it to the public Facebook group "Artificial Intelligence & Deep Learning" on Tuesday. The underlying code that made this possible, titled StyleGAN, was written by Nvidia and featured in a paper that has yet to be peer-reviewed. This exact type of neural network has the potential to revolutionize video game and 3D-modeling technology, but, as with almost any kind of technology, it could also be used for more sinister purposes.

Comments

Should I be concerned?

[Nidi62]

I pulled up the website and the picture looked a lot like me.....

Re:Should I be concerned?

[sinij]

Activate protocol 2! This is not a drill, activate protocol 2! The human consciousness simulation escaped the containment.

Re:Should I be concerned?

[ShanghaiBill]

I just tried it, and got a picture of a normal looking young woman. Not "horrifying" at all.

The site is really slow, and sucks up a lot of local CPU running JavaScript.

Re:Should I be concerned?

[Red_Forman]

Dude, calm the fuck down a bit. We know it's not a drill. It's a screwdriver.

Re:Should I be concerned?

[dgatwood]

No, they're actually kind of horrifying in many cases, in a manner that is subtle and creepy.

It looks like the algorithm is basically combining two people's faces algorithmically, using the upper half of one and the lower half of another. They might be picking one skin tone and mapping it across the other part, or they might just always pick people whose skin tone is close enough be plausible. I can't really tell.

The problem is, their algorithm isn't always combining pictures taken from exactly the same angle. As a result, the upper half of the face is just far enough off from the bottom half to put the resulting face squarely in the uncanny valley.

They are all almost plausible, but only a few of them are close enough to not cause cognitive dissonance. When I look at most of them, my eye shifts from one part to the other trying to figure out the perspective, unable to do so. They quite literally give me a headache.

Re:Should I be concerned?

[cayenne8]

I'm curious....are these images 'copyrighted' as that they are generated?

Just thinking it would be a useful tool if you needed to just put a face on an image or pamphlet, etc......

Re:Should I be concerned?

[ShanghaiBill]

The picture itself isn't "horrifying", it's the fact that the picture (which as you say looks perfectly normal) isn't real.

I don't see anything "horrifying" about it. GANs are interesting and the results are sometimes impressive. But "horror"? No.

Also, the only thing "new" about this website is that the images are supposedly generated on-the-fly. The faces are not much different than published results for other face GANs. GANs have been around since 2014, and high quality face generators have been around since 2017.

Re:Should I be concerned?

[Capsaicin]

are these images 'copyrighted' as that they are generated

That's an interesting legal-theoretical question and it might make a good topic for a law school essay.

As a matter of mere practice, however, given that no two runs of the program should produce identical faces, it seems unlikely that anyone would even think to look for a watermark (assuming there is one) or other identifying feature connecting the image to the the putative copyright holder. TLDR: You're unlikely to be caught.

The Technology Is Already Being Used Negatively

[dryriver]

A friend who lives in a repressive country where over 90% of the media is pro-government propaganda told me that the when he does a reverse image-search of the little "author image" next to the opinion pieces lauding the government's actions, neither Google nor Tineye can find the person online. It seems that the opinion pieces are authored by supposed "journalists" who's face can only be found next to the opinion piece - and nowhere else online. The authors names and their face images are fictional, even though their faces appear to belong to a real person. It seems like this tech was around before Nvidia supposedly "pioneered" it.

Re:The Technology Is Already Being Used Negatively

[93 Escort Wagon]

This actually sounds a bit like the bulk news outlet services which a lot of small newspapers use nowadays - I believe it was covered in an NPR story a few years back. The news items are collected and aggregated overseas, sometimes rewritten slightly to "localize", and then released using a made up "generic white American" name (and sometimes a stock photo) for the byline.

Re:The Technology Is Already Being Used Negatively

[ScentCone]

Or ... someone could have spent 30 seconds in Photoshop with half a dozen normal face shots of different people, and produced exactly what you're talking about while setting up some fake journo profiles. That's been done for years. And is not the same as an AI-ish thing synthesizing faces.

Not 100%

[Nidi62]

It's not 100% realistic yet. The faces look good, but with the hair and body there are a lot of issues. I pulled up one picture and the face is a significantly lighter tone than the neck/chest area which leads to issue number 2, the woman does not appear to have a neck. The ears are just close enough to realistic to notice they aren't right, and the hair coming down over the body isn't correct either. But the facial features are really good.

Do something useful with it

[Comboman]

Someone should write a bot to post each new photo to Facebook (along with a randomly generated name) to salt their facial recognition algorithm.

Horrifying maybe for the wrong reason

[FilmedInNoir]

Some of the pics have noticeable artifacts that give the appearance of a severe scar... or possibly gills in some cases.

Obvious stock image input

[spaceyhackerlady]

It's obvious the neural network was trained on stock images, including lots of celebrities. I saw Angelina Jolie's eyes, Brad Pitt's jaw, Caitlin Jenner's hair.

...laura