Hacker News Reader

Comment by signaru | Hacker News Reader

On a streak so far.

I'm playing "can you tell which picture has a non blurry background and has no artifacts?"

edit: My first mistake is when I thought a piece of fabric on a human was unusually warped.

This. The generator is really bad at compositing people into the image. So while from the actual face it's sometimes hard to tell, backdrop and foreground items (like a mic or toy) are a giveaway. So is face paint or unusual props (fake mustarch or carnival custome). Especially since a lot of images from the real humans dataset seem to contain these.

So next time you're on a video call with someone and you're unsure if they're human or not, ask them to draw a letter on their face or have them dress like a pirate ;-)

>So next time you're on a video call with someone and you're unsure if they're human or not, ask them to draw a letter on their face or have them dress like a pirate ;-)

is that a thing now? my cursory search for "deepfake video call" gave me https://www.youtube.com/watch?v=wYSmp-nrJ7M but other than that, there is just youtubers goofing around with the tech. Do you know of a "good quality" deepfake video call that can fool us like the whilefaceisreal does sometimes?

Some German politicians thought they had video calls with Klitschko. But it was a deep fake.

Maybe there are more instances, but this made the local news.

I only get localized results on the phone, so here is a German link. Use deepl or Google translate: https://www.tagesschau.de/investigativ/rbb/deep-fake-klitsch...

ah.... so it "has" already caused some problems. in that case it does look like an interesting idea....

btw, firefox has a first party translator that is on-device so that works nicely, btw

Yeah. In that case I think the intention was to have them look bad. And since the faked person is a celebrity, enough data was available to produce a fake of suitable quality.

Maybe we will see this in the future for CEO scams. Though in that case maybe a good UI that clearly indicates that the victim is called by an external user "Mr. Big CEO <hackerperson@totally-not-s.us>" might already be helpful.

My game is looking at the eyes, seems like the model has a tendency to make faces that have pretty much mirrored eye shapes and sockets, pupils being clear of imperfections or just plain circles or even both pupils being identical to each other, tilt is also a huge issue, most of the AI images have their output "looking directly into the lens" and nearly perpendicular to the aperture, real humans are off center in more ways than one in all of these aspects and more, as well as having numerous orientations

To me, looking at the background is kind of cheating to sus out facial features, after all we are trying to figure out if the face is real not the background

True, in most cases, backgrounds with a lot of elements (ex: a house in the background) were part of real pictures

Typically in AI generated images another giveaway is that when the head covers the entire height of the picture, the background may randomly change between the left and right side in implausible ways.

>My game is looking at the eyes

I look at the hairs. In real images you can see their fine threaded structure, in fake ones it's rather blurry and inconsistent.

Looking at the ears has worked for me 100% of time.

Outside of obvious facial artifacts the background was the next give away. I got it correct every time.

Same here. Just looked at the background and got all (10/10) right.

Reflections (highlights) in the eyes being different, artifacts and so on. But that said, if I weren't looking for fakes I'd probably accept them as real enough.

As soon as I started paying attention to the faces I lost my streak; they are basically perfectly realistic. But the backgrounds are a dead giveaway.

so next: generate green backgrounds and use the same background in both pictures

Reply on news.ycombinator.com