4 ms·

No, you have to register yourself with a video where you're required to say a unique code. So unless you've posted a video of yourself online saying every numb

by pants2 1y ago

No, you have to register yourself with a video where you're required to say a unique code. So unless you've posted a video of yourself online saying every number from 1 to 99 they won't be able to copy your likeness

noduerme 1y ago
So only the heads of companies who lead shareholder meetings are vulnerable to this exploit? Cool.
solid_fuel 1y ago
That's more than I expected from them, genuinely. But it still doesn't seem like a very solid solution. I wonder how much variation in look and voice it accepts? My partner likes to cosplay, and some of the costumes are quite extensive. If they want to generate a video in a specific outfit will they need to record a new source video? The problem exists in the other direction, too. If someone looks a lot like Harrison Ford, will they be able to create videos with their own likeness? I wonder how this extends to videos with multiple people, as well. E.g. if both my friend and I want to be in a video together.
WmWsjA6B29B4nfk 1y ago
It’s not like making a video of someone saying a number, given a single photo and any voice sample is a very difficult problem today. We can just fast-forward a few weeks into a world where this „registration“ is already broken.
kouteiheika 1y ago
This seems... pretty easy to get around? There are already open weight models which can take any photo and audio and make a video out of it with the character speaking/singing/whatever, and it runs on normal consumer hardware.
- pants2 1y ago
  So you wouldn't know what the three numbers are ahead of time, you'd have to be using a real time face replacement model (or I guess live-switching between pre-rendered clips) and somehow convince the app that you're the iPhone selfie cam. But at that point you might as well just use WAN 2.2 Animate and forget about Sora.

4 ms·

No, you have to register yourself with a video where you're required to say a unique code. So unless you've posted a video of yourself online saying every numb

by pants2 1y ago

noduerme 1y ago
So only the heads of companies who lead shareholder meetings are vulnerable to this exploit? Cool.
solid_fuel 1y ago
That's more than I expected from them, genuinely. But it still doesn't seem like a very solid solution. I wonder how much variation in look and voice it accepts? My partner likes to cosplay, and some of the costumes are quite extensive. If they want to generate a video in a specific outfit will they need to record a new source video? The problem exists in the other direction, too. If someone looks a lot like Harrison Ford, will they be able to create videos with their own likeness? I wonder how this extends to videos with multiple people, as well. E.g. if both my friend and I want to be in a video together.
WmWsjA6B29B4nfk 1y ago
It’s not like making a video of someone saying a number, given a single photo and any voice sample is a very difficult problem today. We can just fast-forward a few weeks into a world where this „registration“ is already broken.
kouteiheika 1y ago
This seems... pretty easy to get around? There are already open weight models which can take any photo and audio and make a video out of it with the character speaking/singing/whatever, and it runs on normal consumer hardware.
- pants2 1y ago
  So you wouldn't know what the three numbers are ahead of time, you'd have to be using a real time face replacement model (or I guess live-switching between pre-rendered clips) and somehow convince the app that you're the iPhone selfie cam. But at that point you might as well just use WAN 2.2 Animate and forget about Sora.