To think I have stumbled across the creepiest website ever?

[CitizenFame]

It’s not anything disgusting or dodgy or gross.

www.thispersondoesnotexist.com

Every time you refresh the page you’re presented with a picture of a person. But none of the people are real. The thought of the manipulation and repercussions are mind-boggling to me. It makes me really uncomfortable.

@BigBadVoodooHat

The computer has to generate everything, including the background. Sometimes it gets the background plausible, usually when it's something vague or blurry, and sometimes it gets a background very, very wrong, most often when it is something that we can almost recognise.

It doesn't "know" that it's drawing another person in the background, so it can't apply the same rules as it does to the main face - it just has to guess what pixels should go next to each other and the result is a really distorted version of a face

I’m not very ‘techy’, but if the program is shit at generating additional faces, why is it set to generate additional faces?

Surely a program this ‘clever’ could be set simply to generate generic backgrounds without secondary faces in them, given that they’re absolute failures and serve only to highlight the fakery of the images.

The program isn't "set" to do anything. They will have fed it a load of source images to "study" if you like originally which is why most of the people look American. From there it tries to figure out what makes a convincing face, and attempts to generate images that look right to humans. They will then have a load of real humans judging the output of the program and feeding back to it whether or not it looks realistic. From this it learns which combinations of pixels are successful and which are not, usually in a way we can't really understand or explain. This is called machine learning.

It's likely that first of all the human judges are instructed to judge only based on the actual face rather than the background, but also, the computer is less good at drawing backgrounds, because backgrounds vary much more than human looking faces do.

Most likely in the earlier stages of the program the main faces were coming out like the weird background faces. You can see this with the cat and horse ones which might be in earlier stages of development or simply cats are in many more varied positions in the source images and so the computer is good at drawing a convincing fur texture or paw or whatever but not very good at knowing which parts to put next to each other.

And also because it's a machine and not a human, it doesn't understand which parts of the cat are important and which aren't, so it doesn't know for example that cats need exactly one head, no head looks wrong and multiple heads look wrong. It also doesn't know that it's drawing a head at all and in fact different parts of cat head are probably internally coded as different items and it might not yet know which of them definitely go together and which don't.

Is there a point to this site?

It’s to show how good our current technology is at replicating a human face. It’s disturbing from a ‘deep fake’ point of view, especially when in the hands of certain governments.

Anyone remeber the "toaster" sticker some weird looking thing youd never guess was a toaster but fooled nearly all image recognition software a few years back

www.bbc.com/news/technology-42554735

For some reason I find it hilariously funny. I am nearly crying with laughter!!

And when I look at them I know what their accent is!

This is what I got,wtf is next to him??

This one looks like there is a cat giving birth on the sofa behind him.

This is my new favourite thing! I can't stop laughing at the horse one 🤣

@jellybaby1🤣🤣 it's like a bizarre dream

I've started giving them names and occupations

This is Jenny. She used to be a primary school teacher, but because of "trouble with her joints" she now tutors from home, specialising in 11+ preparation.

OMG, OMG! I just saw my boss, but clicked past!

These two look weird. The chin on the first one!

Oh these are priceless. I've just laughed so hard I've had a wheezing fit with tears of hilarity still in my eyes.

I have what looks like a ten year old with bum fluff and a pierced fleshy tumour alongside.

I got Louis Theroux

They will then have a load of real humans judging the output of the program and feeding back to it whether or not it looks realistic.

There are no humans in the loop - this feedback on whether the image is convincing or not comes from another program, called the discriminator.

The two programs train in parallel - the generator gets gradually better at producing realistic faces and the discriminator gets gradually better at spotting tell-tale signs of fakes (which the generator will then learn to stop making).

This is why the backgrounds look weird - the discriminator doesn't see enough examples of book cases or rooftop pools or whatever to be able to tell if the background is convincing or not, and so the generator doesn't get penalised for implausible backgrounds.

Mr bean after his plastic surgery didn't quite work

@SerenDippitty

These two look weird. The chin on the first one!

OMG. Apart from the weird chin and the three different shades of hair(!) that first picture looks the spitting image of my friends child when he was younger

@hughjackman88

I got Louis Theroux

Genius 😱😂

A lot of the horse ones are really weird.

And more...

@petherbridge

They will then have a load of real humans judging the output of the program and feeding back to it whether or not it looks realistic.

There are no humans in the loop - this feedback on whether the image is convincing or not comes from another program, called the discriminator.

The two programs train in parallel - the generator gets gradually better at producing realistic faces and the discriminator gets gradually better at spotting tell-tale signs of fakes (which the generator will then learn to stop making).

This is why the backgrounds look weird - the discriminator doesn't see enough examples of book cases or rooftop pools or whatever to be able to tell if the background is convincing or not, and so the generator doesn't get penalised for implausible backgrounds.

Oh, this is really interesting. Thanks for the explanation.

How does the discriminator know? It is just comparing against real people?

Something I find really fascinating about this kind of thing is the way they show up biases - I guess that's why most of the people are white and have perfect teeth?

I Just got someone who looked like David Walliams but not quite.
It was really weird!

Oh god! The cats! I'm crying, whilst sniggering like Mutley!