WTF?! Machines' ability to generate fake videos of people has become alarmingly impressive. ByteDance, the Chinese tech giant behind TikTok, just showed off a new AI system called OmniHuman-1 that can create deepfake videos almost indistinguishable from reality to the average person. We may be well past the uncanny valley point right now. OmniHuman-1's fake videos look startlingly lifelike, and the model's deepfake outputs are perhaps the most realistic to date. Just take a look at this TED Talk that never actually took place.The system only needs a single photo and an audio clip to generate these videos from scratch. You can also adjust elements such as aspect ratio and body framing. The AI can even modify existing video footage, editing things like body movements and gestures in creepily realistic ways.Of course, the results aren't 100% perfect. Some poses do look a bit off, like this awkward example of holding a wine glass.There's also this AI-rendered lecture from Einstein where his hands twist in odd directions. His face is rendered almost perfectly, though.Still, the quality overall is way ahead of previous deepfake techniques. // Related StoriesUnder the hood, OmniHuman-1 was trained on 18,700 hours of video data using a novel "omni-conditions" approach that lets it learn from multiple input sources like text prompts, audio, and body poses simultaneously. The ByteDance researchers say that this wider training data helps the AI "significantly reduce data wastage" compared to older deepfake models.The implications of this technology are concerning. Deepfakes have already been weaponized for misinformation campaigns, fraud, and all sorts of nefarious purposes over the past few years. There were numerous incidents during the 2024 election cycle of deepfake audios and videos being spread to mislead voters.Financial scams conned people out of billions last year, too. One notable case involved a scammer using AI to pose as Brad Pitt, tricking a woman into sending $850,000 last month.Considering these incidents, hundreds of AI ethics experts pleaded for deepfake regulations last year. Several US states have already passed laws against malicious deepfakes, but there's still no overarching federal legislation.California, for one, was on the verge of enacting a law that would let judges force people to take down deepfakes and potentially face fines for posting them. However, that bill has stalled in the legislative process.ByteDance hasn't released OmniHuman-1 to the general public, but you can read a paper about the model.