Highly realistic talking head video generation
github.com
github.com
Look at the rapping example near the end. The lip sync is nearly flawless. The first black and white lady singing is also almost perfect. It even gives them the subtle jerk to pause for breathing. Unless you know and really are looking for flaws, you won't find anything that stands out making them look real.
The juxtaposition of modern facial expressions of an influencer type singer covering Ed Sheeran at some X factor type television show are what makes it creepy. It is somehow doubly fake and extremely out of character if you are familiar with her.
Both 'move' in ways that are very unnatural.
This singular peace of technology makes me pessimistic towards the future. Until now, video record was considered to be a very good evidence. Let's say you argue with a person about what X person said. You show them a video and they will be like "ok, he did say that, but...". You could at least set some facts straight, and then discuss the interpretations.
But that will be now gone. You can now generate mass amount of real looking fakes and at the same time label anything you don't like as fake. There's really no independent evidence now, you can only put trust into the medium of your choosing (youtube channel, newspapers, tv station) that they report honestly.
This seems to have only minimal benefits for the society, but huge negatives. But there's no stopping here...
But as usual, we shall adapt and overcome.
I don't believe this is a problem we can "overcome". We will need to learn to live with the "alternative facts" being more prominent than now, but I'm not looking forward to it.
Digital signatures can not remedy this problem? When you login to your bank how do you know you are logging into your bank? In the future a recording without signatures will be like a bank login without https is today.
Just because you initiated the connection how do you know the other end is you bank? Do you trust every internet company that carries your packets to the bank? Trust their employees? Trust their security practices? Do you trust firmware on all the devices involved?
> People are routinely defrauded by scammers who claim to be their bank,
I have read about this in the news just like I read about snakes with two heads etc yet I have yet to meet someone that has had this happen to them. What fraction of people that you know have had this happen?
Could it be that these people believe like you do that "I know I’m logging into my bank because I initiated the connection" as opposed to checking the digital signatures on the connection?
For this to work, you need to have an already trusting relationship with the media. Like, ok, I can trust NYT, so I will trust videos signed by them. But another person distrusts NYT and trusts only Truth Social. In the past, we could at least agree on basic facts like January 6th actually happening, but I think this generative AI will make laying out the facts much more difficult or even impossible.
Many high-end smartphones, such as iPhones and some Android devices, incorporate secure enclaves or trusted execution environments (TEEs). These are isolated, tamper-resistant hardware components that can securely store and process sensitive data. When a recording is made on such a device, the secure hardware can associate the recording with additional metadata, including the specific date, time, GPS coordinates, and user account information. This metadata is cryptographically bound to the recording itself.
Furthermore, the device can digitally sign the recording and its associated metadata using a unique key stored within the secure hardware. This digital signature serves as a testament to the recording's originality. Companies like Apple or Google, who manage the secure hardware and signing keys, can then vouch for the authenticity of the recording.
While this approach doesn't completely eliminate the possibility of fake recordings, it significantly raises the bar for creating convincing forgeries. Modifying the recording or its metadata would invalidate the digital signature, making it evident that tampering has occurred.
Of course, as you mentioned, trust in the entity verifying the signatures (e.g., Apple or Google) is still required. However, this trust is based on their reputation and the security measures they employ, rather than on the content of the recording itself.
You believe this is easy when the device has multiple recording sensors and multidimensional information (such as spatial information, changes in focus sensors durning recording, etc) is part of recording that is digitally signed?
The key idea is to leverage the array of sensors built into modern smartphones and tablets - accelerometer, gyroscope, GPS, WiFi/cellular signal data, etc. - to cryptographically sign the sensor readings along with the visual data itself at the time of capture. This extra layer of verifiable sensor data would help establish that a recording originated from a real physical device in a particular place and time, as opposed to a purely digital fabrication.
Historically, technologies like digital signatures and public key cryptography started out in niche military/government applications before becoming ubiquitous in the computer era. In a similar way, sensor-level authentication of audiovisual media could follow an adoption curve driven by the growing need to combat sophisticated AI forgeries.
https://www.polksheriff.org/news-investigations/polk-county-...
The conference video at 1:00 starts off with a generated clip of Elon Musk saying he's going to move to Polk County. The Sheriff highlights your concerns as well as many others.
Conference video (29:56): https://www.youtube.com/watch?v=DHj18pOcXHc
You may say it's important only for public officials. But why is it important? Because you're giving huge amount of power to single individuals and somehow we're taught that's a good thing - or at least that it's inevitable for keeping peace or to keep crime at bay. What a load of bs. I hope distrust in centralisation increases. It should have been there in the first place.
I think it's a dangerous mindset to doubt everything you can't see with your own eyes. There was this fringe group claiming that the war in Ukraine is fake, that there's no war. With this mindset, such claims will become more mainstream, you could even call it reasonable.
The page has a link to github[1] right at the top but the repo is basically empty.
It's truly bizarre to watch these talking heads because their lips are moving, but their eyes and cheeks aren't moving along with them, except for blinking.
Real people speak with their whole face, not just their lips.
Of course, to do that "right", you need to actually understand the emotional content of what is being spoken. And I'm not talking about highly "emotional" content like in TV drama -- even in a technical presentation, the speaker's face contains lots of emotional signals. Whether warmth, or a sense of humor, or being proud, or excitement of what they're about to reveal, or curiosity about whether the audience understands, etc.
Not true?!
Please rewatch the video at the top: especially the eyebrows are quite animated, and the contours of the face also change as well as the shadows that indicate muscle movement in the phase.
They look kind of lobotomized, sure with maybe some random eyebrow raises thrown in. It's nothing like whole-face expression.
PS: With insightface models in your reqs the oss aspect is also pretty much void for any use other than research and your readme should reflect that.
Is this what we're in for? Human movement that is just normal enough to pass, but not really natural enough to be comfortable with?
A huge amount of people who may be users of a platform such as Facebook are very likely able to watch a video generated using a technology such as this and not consider it possibly to be fraudulent.
Even if the content is somewhat strange or farfetched, these people, the majority of people are likely to not only "not notice" there's something off about the video, but believe what is being said.
They are unlikely to act on the information or regurgitate it unprovoked, but likely to just remember it as a small fact they have received in the past.
Then, my little brother, 12 years younger than me, in 2nd grade, sits down and 2 days later he's programming circles around me. He could make that atari dance like Fred Astaire.
And I was the one who wanted to be a computer programmer. He had no inclination in that direction at all, and became a businessman.
I learned a very painful and embarrassing lesson about how learning the wrong programming language can give you brain damage.
If an organization/entity releases media that is (by them) marked as "real" (or genuine, or trustworthy, original, not tampered with, etc), the stream can be checksummed and committed to a ledger. This way, any content coming from sources participating in this "verity scheme" can be validated against a distributed record (which is presumably incredibly difficult to modify surreptiously).
I don't mean to encrypt the media, I mean to make a hash stream of the media stream, and store that in a way that cannot be adulterated.
Ugh... paired with laws that restrict audio recording but not video recording.