A Taylor Swift video that was never filmed, a TED talk that was never held, an Einstein lecture that never existed. These are not movie special effects, but the incredible creations of OmniHuman-1, the new artificial intelligence system for deepfake videos developed by ByteDance (the “owner” of TikTok) that is redefining the boundaries between reality and digital fiction.
An unprecedented technology
Deepfake videos they are nothing new in the digital landscape. There are already numerous applications capable of inserting a person into a photo or making them say things that have never been said. But most of this manipulated content shows obvious signs of artificiality. OmniHuman-1 seems to have overcome this limitation, producing surprisingly realistic results.
The system requires only two elements to generate a deepfake video: a reference image and an audio file, which can be a speech or a voice track. With these minimal inputs, OmniHuman-1 can produce clips of any length, with adjustable aspect ratios and complete control over the subject’s representation. The system was trained on 19.000 hours of video content (from undisclosed sources).
Advanced manipulation capabilities
OmniHuman-1's capabilities go beyond simple video generation. The system can also modify existing footage, even altering the movements of people's limbs. The results, as shown by the team at ByteDance, are incredibly convincing, although not without limitations: if it is “fed” with low-quality reference images the results are inferior, and some poses are still problematic. In any case, the degree of refinement achieved by OmniHuman-1 raises serious concerns, in an already rather incandescent picture.
In 2024, political deepfakes have already caused significant problems: in Taiwan, a group affiliated with the Chinese Communist Party released doctored audio of a politician; in Moldova, deepfake videos showed the fake resignation of the president Maya Sandu; in South Africa, a fake Eminem he supported an opposition party.
Secondo Deloitte, in the United States alone, AI-generated content has contributed to Fraud losses to exceed $12 billion in 2023, with projections that could reach $40 billion by 2027. Consumers are being duped by celebrity deepfakes promoting fraudulent investments, while businesses are being scammed out of millions by digital impostors. The scientific community is calling for tough regulation, and more than one body is considering legislation that would allow judges to order the removal of deepfake videos, potentially subjecting violators to monetary penalties.
Deepfake Videos, The Uphill Battle of Detection
Detecting deepfake videos remains a significant challenge. Despite efforts by social networks and search engines to limit their spread, the volume of manipulated content online continues to grow at an alarming rate. A survey by Jumio of 2024 reveals that 60% of people have encountered a deepfake in the past year, and 72% fear being fooled by one on a daily basis.
ByteDance has not yet publicly released OmniHuman-1, but experience shows that it doesn’t take long for the AI community to replicate similar systems. This raises crucial questions about the future of digital truth and the need to develop effective tools to protect society from this increasingly powerful, and increasingly dangerous, technology.