Deepfake videos are getting shockingly good | TechCrunch
Deepfake videos are getting shockingly good | TechCrunch
techcrunch.com
Deepfake videos are getting shockingly good | TechCrunch
cross-posted from: https://lemm.ee/post/54578863
17 comments
-
-
This is scary indeed. We may someday soon need something like an active tattoo on our face, or a badge on clothing, with a pattern that changes each second based on a private/public key pair, so videos can't be easily faked of our own likeness with a valid visual signature.
That could actually work -- a QR code that updates at regular intervals, encoding an ever-changing signature. It could be validated to certify the video of a person was genuine.
Of course that would also mean any authenticated video can never be truly anonymous :(
-
Of course that would also mean any authenticated video can never be truly anonymous :(
With modern face recognition, having your face in a video already deanonymizes you so there is no further deanonymization caused by authenticating your video cryptographically.
If you want anonymity then don't include your face (or signature) in the video.
If you want attestation that a video is real and not generated then anyone who witnessed the scene depicted can sign the video to attest to it. Then we would need a web of trust to determine whose signatures are trustworthy.
-
With modern face recognition, ... If you want anonymity then don't include your face (or signature) in the video.
Duh, good point :)
Web of trust -- it's always been so hard to make easy enough to use for the non-technical public, sadly... but yeah that might be the only/best way to really give attestation.
-
-
a QR code that updates at regular intervals, encoding an ever-changing signature
That's brilliant. Any sci-fi authors in the crowd? The private key would be used to hash and sign a record of body movements including all one's crazy hand gestures and every word spoken. Location could optionally be encoded as well. Some kind of algorithm would be devised so that some reasonable loss in fidelity of any video recording would still result in a valid signature. It wouldn't be technically all that useful to display the signature as a QR code shown on a badge, but it'd be a fashionable thing to do and anyone who didn't would be seen as slightly suspicious.
Key infrastructure would be tricky, but anonymous and pseudonymous keypairs could certainly be allowed for if we go with the assumption that instant biometric identification of everyone isn't quite feasible for whatever reason. Maybe it's just banned, punishable by exile to the orbital asteroid mining colonies.
All we need is for everyone (except the underclass) to get neural implants that record their every movement.
-
exile to the orbital asteroid mining colonies.
Can we start with Sam Altman please? Hah.
I was amazed at first with ChatGPT, outpainting, and the early stuff; it was fun making 'paintings' and playing with other imagery, but the main uses are taking such a dark turn I really think we're going to regret this technology's existence.
-
-
-
Trained on 19,000 hours of video content from undisclosed sources
I'm gathering their "undisclosed sources" are social media (YouTube, Instagram, TikTok, Facebook, etc.) and copyrighted content (DVDs, BRDs, etc.), all without permission or legal rights. I wonder when pirating becomes just another brick in the wall...
-
Well, now they have the ability to create new derivative content on the fly… next up is China mass producing movies and TV shows for all regions, with the content subtly different for each country or target demographic.
-
Pretty sure china won't be the only one...
-
-
-
Gesticulating is apparently a key element.
-
This is scary as hell. Someone is going to get screwed big time by this shit.
-
Someone...
Is it us?
It's us, isn't it.
-
Ding! Got it in one.
-
-
Refreshingly, that includes multinational firms: https://www.cnn.com/2024/02/04/asia/deepfake-cfo-scam-hong-kong-intl-hnk/index.html
-
-
Albert Einstein saying "science does not answer 'why' questions" isn't realistic at all.
-
"According to the ByteDance researchers, OmniHuman-1 only needs a single reference image and audio, like speech or vocals, to generate a clip of an arbitrary length."
I am SO glad I never did a youtue channel, a vine, an IG, or really anything. This is creepy as fuck.
-
Hands, hair, camera angles, and voice sync, still have some way to go... but they do seem like they're slowly climbing out of the Uncanny Valley.
Interesting times.
17 comments
I really really really don't like where this is heading. Hell, I don't like like where it already is. 😖