Yet Another INSANE AI From China: First DeepSeek Now ByteDance (Beyond Scary)

China just dropped another AI bombshell. After deep seek, this might be the biggest game changer yet. Omni Human 1 Built by ByteDance, the company behind TikTok, this AI can take a single photo and turn it into a full body video, complete with natural movements, gestures, and even singing. It's insanely realistic, pushing deep fake tech to a whole new level. And that comes with some serious implications. All right, so Omni Human is basically a new approach to AI powered video generation that can create eerily lifelike footage of people from just a single photo. And I'm not talking about just a face replacement or a bit of lip syncing. No, we're talking full body animations, gestures, movements, even singing and playing instruments, all synced up with audio. And all of this from just one image. If you're thinking, wait, this has to be some kind of black magic level AI wizardry, you're not too far far off. Previous deepfake apps and tools typically need a bunch of reference images or video, or they just give you that weird, uncanny look that screams fake. Omni Human, on the other hand, can apparently generate a really convincing deepfake using minimal input. And that, my friends, is a game changer. Now, depending on which write up you read, bytedance says they trained Omni Human on a huge data set of somewhere between 18,780 and 19,000 hours of video. That's a ton of content. They're pretty hush hush on the exact sources, but we can safely assume that with TikTok's massive content library, they aren't running short on examples of human movement or speech patterns. The paper on Omni Human, Posted to the OpenSxarx archive site describes how the researchers used multiple conditioning signals, meaning text, audio, body poses, to teach the AI how humans move and talk. They call this approach Omni conditions, which basically allows them to feed the AI different inputs so it can produce videos for a wide variety of scenarios. Speaking, singing, gesturing, or even playing instruments. In some of the sample videos ByteDance released, you see a fictional Taylor Swift performance that looks unbelievably convincing, like you might do a double take to check if it's real. There's also a clip of Albert Einstein giving a lecture in crisp black and white, looking as if it was shot with modern HD cameras. At one point, Einstein says, what would art be like without emotions? It would be empty, gesturing with his hands and everything. It's borderline spooky because it truly looks like the real Einstein is right there talking about art. Then there's a video that apparently shows a TED talk that never took place. And in another demonstration, they show some random scenario where a person is holding a wine glass, but you can see some weird movements as if the AI's having trouble with that particular pose. So it's not perfect, but the leaps it's making beyond older deep fake methods are huge. They also mention that you can tweak the aspect ratio, the body proportions and the length of the clip. So let's say you have a single image of a friend or a celebrity. You just feed that plus some speech or vocals into the model and Omni human can generate an entire video with that person dancing around doing big gestures or delivering a speech. It's like those apps where you can animate old family photos, except taken to the next level, full blown near Hollywood level video illusions. But of course, while this is extremely cool technology, there is a dark side. We're living in an era where deep fakes have have already begun spreading across social media. Articles reference multiple incidents around the globe. For example, on election day in Taiwan, there was a group with ties to the Chinese Communist Party that posted an AI generated audio clip. It made it sound like a local politician was endorsing a pro Chinese candidate even though he wasn't. In Moldova, there were deepfake videos depicting President Maya Sandu resigning when she actually hadn't. And in South Africa, a deepfake of rapper Eminem endorsing an opposition party circulated around election time. So yeah, the risk for misinformation is serious. And that's just politics. Deepfakes are also getting used in financial scams where fraudsters create fake videos of celebrities pitching bogus investments. Deloitte reported that AI generated content contributed to more than $12 billion in fraud losses in 2023, and it could reach $40 billion in the US by 2027. That's a staggering amount. So it's no surprise there have been calls for strict regulation, including an open letter signed by a bunch of AI researchers last year. Meanwhile, over 10 states in the US have enacted or proposed laws against AI aided impersonation. California, for instance, had a law in the works that would let judges order the removal of deep fake content or even fine the people who posted it, though that legislation has apparently stalled. A major problem is that deepfakes, especially high quality ones, can be really tough to detect. Even though social networks and search engines claim to be implementing AI based detection tools or labeling systems, the volume of AI generated stuff is just exploding. One ID verification firm, Jumio, did a survey in May 2024 and found that 60% of people said they encountered a deepfake in the past year. And get this, 72% of them said they were worried about being fooled by deepfakes on a daily basis. A majority also supported new laws to rein them in. Now you might be thinking so where does bytedance stand on releasing Omni Human? That's a bit up in the air. The folks at ByteDance haven't released the model publicly yet, at least not at the time these articles were written. But as we've seen in the AI community, once a new model is demoed, it's only a matter of time before someone tries to replicate or reverse engineer it's so even if ByteDance holds back, there's a good chance we'll see similar systems sprout up from other labs or from open source efforts, especially given how popular generative AI has become. On the plus side, experts mention some potentially positive applications. For instance, there's talk about using Omni Human in education, maybe resurrecting historical figures like Marilyn Monroe or Humphrey Bogart to deliver lectures, or having virtual teaching assistants that keep students engaged. ByteDance also owns TikTok, so we could see creators using these lifelike avatars or AI driven stand ins so they can take a break while still pumping out content. One professor from the University of Southern California, Freddy Tran Nagger, pointed out that if someone wanted to cast deceased actors in new films, this technology might make that possible. Of course, how that would look on a giant theater screen is still unknown, but on a phone or laptop it could be convincingly real. Meanwhile, Samantha G. Wolf, an adjunct professor at NYU's Steinhardt School, warns of the negative consequences. She highlights the idea of a fake CEO or political leader video that could cause big corporate or governmental disruptions if people actually believe it. So basically the stakes can get pretty high. She also said as AI starts to look more and more realistic, the likelihood of people believing a fake is going to skyrocket. And that's exactly what's causing so much concern. Let's talk a bit more about ByteDance's approach. They mentioned that traditional human animation techniques struggled to scale up as general video generation models, but OmniHuman overcame that by using Omni Conditions training. This means they wasted less data by combining audio, text and body movements all in one training pipeline. Over 18,700 hours of human video content were used to do this, so the neural network can effectively create full body animations that move in sync with speech. That's a big leap from older methods that mainly stuck to facial or upper body animations. One of the articles even pointed out that Omni Human can handle different aspect ratios, so you could have a portrait for a standard vertical TikTok or a wide shot that shows the person's entire body. And they also mention you can adjust how much of the body is shown, controlling whether the final video is a close up or full body. Another cool tidbit is that Omni Human can edit existing videos, so if you have a clip of someone, you could presumably go in and change their gestures, limb movements or maybe the expressions on their face. Like a big do overs of any recorded content. Pretty wild. Of course, there's plenty of concern over misuse. People worry about fake political ads, phony celebrity endorsements, and how quickly misinformation might spread. Some US states already have laws against AI impersonation, but there's no federal rule yet. Meanwhile, experts say we need stronger detection tools, especially before future elections. ByteDance plans to show Omni Human at a computer vision conference soon, and they're not the only ones working on this kind of tech. Google, Meta and Microsoft are also in the race. Still, ByteDance has a huge advantage, thanks to all the data from TikTok. One joke is that if you've ever made a TikTok, you might have helped train Omni human. Ultimately, this AI proves how far deepfakes have come. One photo plus a clip of audio is enough to create a super realistic video of just about anyone doing anything, from speeches to guitar solos. It's exciting for creative uses, but also raises a lot of red flags. Let me know your thoughts in the comments. And if you enjoyed the video. Don't forget to like and subscribe. Thanks for watching and see you in the next one.