ByteDance unveils OmniHuman-1 deepfake video generation model
ByteDance unveiled OmniHuman-1, an AI video generation model that creates realistic videos of humans from a single still image and can generate matching audio. Experts expressed concern that the technology, if deployed publicly, could facilitate deepfakes and be used to impersonate individuals or deceased figures.
What happened
ByteDance unveiled OmniHuman-1, an AI model capable of creating realistic videos of humans talking and moving naturally from a single still image. The model was trained on over 18,700 hours of human videos and, according to ByteDance researchers, generates videos that lack telltale signs of artificial generation such as issues depicting hand movements or lip syncing, and can potentially evade AI-detection tools.
Sample videos released in the research paper showed OmniHuman transforming a still image of Albert Einstein into a video of the physicist delivering a lecture. The audio used was from a speech by neuroscientist Jaak Pankseep at a TedXRairier event in Seattle.
Reported impact
Experts warned that if made available for public use, the technology could lead to new abuses. According to a world-leading expert on generative AI, the ability to create videos from a single image makes it easier to target individuals, whereas previously hundreds or thousands of images might have been needed.
The technology could potentially be used to replace human influencers or to create deceptive content such as having deceased figures deliver lectures or teach subjects.
Index analysis
The model supports image inputs of any aspect ratio including portraits, half-body or full-body images. Users can generate custom voice audio to match the video output.
According to experts, the video outputs are notably high-fidelity and realistic, particularly on small screens such as phones, though on larger screens potential limitations may be more apparent.
Governance controls
A ByteDance representative told Forbes that if the tool were publicly deployed, it would include safeguards against harmful and misleading content. ByteDance announced last year that the TikTok platform would automatically label AI-generated content and work to improve AI literacy.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.