ByteDance, the company best known as the owner of TikTok, has introduced a new generative AI system called OmniHuman-1, capable of creating incredibly realistic deepfake videos. Unlike many other systems, which often leave obvious traces of AI interference, OmniHuman-1 impresses with its ability to avoid the uncanny valley. In demo videos released by the ByteDance team, you can see fictional performances by famous personalities such as Taylor Swift or Albert Einstein.
OmniHuman-1 requires only a single image and audio to generate a video of any length. You can adjust the video’s aspect ratio as well as the proportions of the subject’s body. The model was trained on 19,000 hours of video content from unknown sources. It can even edit existing videos by changing people’s limb movements. This is an impressive leap forward in the world of deepfakes.
However, the system is not without its drawbacks. ByteDance notes that “low-quality” images do not yield the best results, and some poses may be problematic for the model.
The rapid development of deepfakes is causing concern. Last year, political deepfakes spread around the world, including incidents in Taiwan, Moldova, and South Africa. Deepfakes are also used for financial crimes, deceiving consumers and companies. According to Deloitte, in 2023, AI-generated fraud led to losses exceeding $12 billion, and this figure could rise to $40 billion in the US by 2027.
The growing number of deepfakes makes them harder to detect, and most people are worried about being deceived. While some social networks and search engines are trying to limit their spread, the volume of AI-generated content continues to increase.