Follow

Eye contact // THREAD: How to recognize AI-generated faces? 

A few years ago, a site called "This Person Does Not Exist" (thispersondoesnotexist.com/) has popped up, using a neural network called StyleGAN to serve a new image of an AI-generated picture of a person’s face after every refresh.

People online have started taking pictures from this site with malicious intent in order to look more legit to perform scams, harassment or other type of deception.

However the generated pictures are not perfect, in this thread I will show a few telling signs with illustrations that might indicate a picture has been lifted straight from TPDNE. Keep in mind that people doing this might not be taking the most egregious examples and that these are only pointers, not an exact science!

Eye contact // THREAD: How to recognize AI-generated faces? Part 1: Dentition 

Of course not everybody in real life has perfect dentition, but even then, AI-generated pictures often have uncannily irregular teeth. Look out for uneven sizes and shapes, and especially take a close look at the edges of the mouth: partially obscured teeth often look pretty flawed with unrealistic edges

Eye contact // THREAD: How to recognize AI-generated faces? Part 2: Hair 

Generally, the farther away the hair is from the face, the more unnatural and smudgy it will look. This is especially noticeable on pictures of people with longer and/or messier hair.

Eye contact // THREAD: How to recognize AI-generated faces? Part 3: Ears 

Ears might display irregular shapes, lack usual features or be very asymmetric, especially if they are partially obscured by hair.

Eye contact // THREAD: How to recognize AI-generated faces? Part 4: Irises 

Irises might be of different sizes between the eyes and might be in pretty irregular shapes. This is especially visible on people who have light-colored eyes

Eye contact // THREAD: How to recognize AI-generated faces? Part 5: Glasses 

When it generates pictures of people wearing glasses, it usually makes mistakes like the frame blending with the skin, or inaccurate light distortions. The frames may also be asymmetric.

Eye contact // THREAD: How to recognize AI-generated faces? Part 6: Apparel 

Clothes on shoulders will often have irregular, asymmetric and confusing shapes. Ear-rings or other kind of apparel will also often look amorphous on closer inspection.

Eye contact // THREAD: How to recognize AI-generated faces? Part 7: Backgrounds 

Pictures with more than a uniform background will often have a background that kind of looks like a real place at a glance, but upon observation is amorphous and confusing. It’s impossible to make out anything out of it.

Eye contact // THREAD: How to recognize AI-generated faces? Conclusion 

There might be other details, such as colored spots on the face, or whenever the image is generating a picture with other people on the side of the frame, they often look mangled and pretty uncanny. I’ve chosen not to include example of those because they’re very blatant, and people who are trying to look real would certainly not use pictures with those features. If you go on the website, you might encounter those after a few refreshes.

I strongly encourage you to train yourself to recognize them for a few minutes on whichfaceisreal.com, once you notice the telling signs, it becomes pretty easy to discern an image generated from the StyleGAN neural network.

I think this thread I made a few months ago might be important as image-generation AIs are getting more advanced

re: Eye contact // THREAD: How to recognize AI-generated faces? Conclusion 

@Siph that's it
I'm calling the AI police, you know too much
oops my hand er no look the other way *runs*

@Siph thank you for the thread! It has been really instructive :blob_cat:

@Siph There are also some software-based methods of detecting images created by tools like StyleGAN, if you are interested in such nerdery.

This paper is fairly old but gets into one of the more reliable methods:
https://arxiv.org/pdf/2007.10466.pdf

And here is a newer paper from 2020 that gets closer to the state of the art:
https://personal.utdallas.edu/~shao/papers/joslin_dls20.pdf

There is an ongoing cat-and-mouse game between the generation and detection software.

@Siph Good idea for a webtoy: show the user a selection of real and AI-generated images and ask them to identify which ones are fake

@Mandrake that's a thing and I linked it at the end of my thread

Sign in to participate in the conversation
Rage.love

Generalist Hometown instance with a strong focus on community standards. No TERF, no SWERF, no Nazi, no Centrist.