5 ms·

How about first making deepfake faces actually believable? Seems like every AI project does something halfheartedly, ponders what the world will be like once i

by runeks 4y ago

How about first making deepfake faces actually believable? Seems like every AI project does something halfheartedly, ponders what the world will be like once it’s perfected, and then starts the next project long before the first project is actually useful for anything but meme videos.

black_puppydog 4y ago
That's what goes to media. "Engineers scrap last little artifacts off deep-fake still images" just doesn't make for "good" headlines. Somewhere, someone is working hard to perfect these. In this particular case probably under NDA... le sigh
BudaDude 4y ago
I didn't know me and AI had so much in common.
walls 4y ago
They're believable enough for video calls: https://www.dw.com/en/vitali-klitschko-fake-tricks-berlin-mayor-in-video-call/a-62257289 https://www.dw.com/en/vitali-klitschko-fake-tricks-berlin-ma...
- kleiba 4y ago
  As far as I remember those calls were actually not made with deep fake tech, but by reusing video material from a previous call, skillfully edited to be believable enough.
jcims 4y ago
I think the problem we have with deepfake believability today it just takes one weak link to spoil it. It turns out that we're somehow still pretty bad at believable audio and not even close with deepfaked 'presence' in the form of persona of motion. But if you pair a believable impersonation with something even remotely state of the art in the visual, you end up with something pretty compelling: https://www.tiktok.com/@deeptomcruise https://www.tiktok.com/@deeptomcruise https://www.youtube.com/watch?v=kjI-JaRWG7s https://www.youtube.com/watch?v=kjI-JaRWG7s https://www.youtube.com/watch?v=VWrhRBb-1Ig https://www.youtube.com/watch?v=VWrhRBb-1Ig https://www.youtube.com/watch?v=bPhUhypV27w https://www.youtube.com/watch?v=bPhUhypV27w (Not the greatest visually but funny nonetheless, esp the end)
armchairhacker 4y ago
Even AIs which have existed for years and been "perfected" are very noticeably not-human. Though they do look believable from far away, up close they are still in the uncanny valley. For instance Siri and Google Voice: they are clearly understandable but they sound noticeably different than real people. Or Stable Diffusion which will supposedly put real artists out of business. It is definitely viable for stock photos, but I can usually tell when an image was made by Stable Diffusion (artifacts, incomplete objects, excessive patterns). thispersondoesnotexist.com faces can also be spotted, though only if I look closely. If they are a profile pic I would probably gloss over them. In fact, I bet you can make an ML model which very accurately detects whether something was made by another ML model. Actually that's a good area of research, because then you can make a deepfake model which tries to evade this model and it may get even more realistic outputs... Ultimately I think we will see a lot more AI before we start seeing truly indistinguishable AI. It's still close enough that the ethical concerns are real, as people who don't really know AI can be fooled. But I predict it will take at least a while before a consensus of trained "AI experts" can't agree on authenticity.
kabes 4y ago
The company behind this post recently got into america got talent finals with a deepfake act. It looked pretty convincing to me. Especially compared to state if the art of just 2-3 years ago.
- thrown_22 4y ago
  >It looked pretty convincing to me. Especially compared to state if the art of just 2-3 years ago. This has been the case for decades now. Much more realistic than x isn't a good enough metric. It needs to be indistinguishable from the real thing. I'm old enough to remember this being called photo realistic: https://static1.thegamerimages.com/wordpress/wp-content/uploads/2021/02/tomb-raider-underworld.jpg https://static1.thegamerimages.com/wordpress/wp-content/uplo... And it was, compared to everything that had come before. Now ... not so much.
  - kabes 4y ago
    https://youtu.be/TVezHTlPMw8 https://youtu.be/TVezHTlPMw8 This is the act I meant. Judge for yourself, but I believe we're close to bridging the uncanny valley

by runeks 4y ago

black_puppydog 4y ago

That's what goes to media. "Engineers scrap last little artifacts off deep-fake still images" just doesn't make for "good" headlines. Somewhere, someone is working hard to perfect these. In this particular case probably under NDA... le sigh

BudaDude 4y ago

I didn't know me and AI had so much in common.

walls 4y ago

They're believable enough for video calls: https://www.dw.com/en/vitali-klitschko-fake-tricks-berlin-mayor-in-video-call/a-62257289 https://www.dw.com/en/vitali-klitschko-fake-tricks-berlin-ma...

kleiba 4y ago
As far as I remember those calls were actually not made with deep fake tech, but by reusing video material from a previous call, skillfully edited to be believable enough.

jcims 4y ago

I think the problem we have with deepfake believability today it just takes one weak link to spoil it. It turns out that we're somehow still pretty bad at believable audio and not even close with deepfaked 'presence' in the form of persona of motion. But if you pair a believable impersonation with something even remotely state of the art in the visual, you end up with something pretty compelling: https://www.tiktok.com/@deeptomcruise https://www.tiktok.com/@deeptomcruise https://www.youtube.com/watch?v=kjI-JaRWG7s https://www.youtube.com/watch?v=kjI-JaRWG7s https://www.youtube.com/watch?v=VWrhRBb-1Ig https://www.youtube.com/watch?v=VWrhRBb-1Ig https://www.youtube.com/watch?v=bPhUhypV27w https://www.youtube.com/watch?v=bPhUhypV27w (Not the greatest visually but funny nonetheless, esp the end)

armchairhacker 4y ago

Even AIs which have existed for years and been "perfected" are very noticeably not-human. Though they do look believable from far away, up close they are still in the uncanny valley. For instance Siri and Google Voice: they are clearly understandable but they sound noticeably different than real people. Or Stable Diffusion which will supposedly put real artists out of business. It is definitely viable for stock photos, but I can usually tell when an image was made by Stable Diffusion (artifacts, incomplete objects, excessive patterns). thispersondoesnotexist.com faces can also be spotted, though only if I look closely. If they are a profile pic I would probably gloss over them. In fact, I bet you can make an ML model which very accurately detects whether something was made by another ML model. Actually that's a good area of research, because then you can make a deepfake model which tries to evade this model and it may get even more realistic outputs... Ultimately I think we will see a lot more AI before we start seeing truly indistinguishable AI. It's still close enough that the ethical concerns are real, as people who don't really know AI can be fooled. But I predict it will take at least a while before a consensus of trained "AI experts" can't agree on authenticity.

kabes 4y ago

The company behind this post recently got into america got talent finals with a deepfake act. It looked pretty convincing to me. Especially compared to state if the art of just 2-3 years ago.

thrown_22 4y ago
>It looked pretty convincing to me. Especially compared to state if the art of just 2-3 years ago. This has been the case for decades now. Much more realistic than x isn't a good enough metric. It needs to be indistinguishable from the real thing. I'm old enough to remember this being called photo realistic: https://static1.thegamerimages.com/wordpress/wp-content/uploads/2021/02/tomb-raider-underworld.jpg https://static1.thegamerimages.com/wordpress/wp-content/uplo... And it was, compared to everything that had come before. Now ... not so much.
- kabes 4y ago
  https://youtu.be/TVezHTlPMw8 https://youtu.be/TVezHTlPMw8 This is the act I meant. Judge for yourself, but I believe we're close to bridging the uncanny valley