TikTok’s parent company ByteDance has released what may be the most realistic AI deepfake maker yet.
A YouTube video posted on 4 February demonstrates the realistic nature of the videos produced by this deepfake AI tool called OmniHuman-1.
The video shows a man giving a Ted Talk on how to design an app that gives users the best experience and improves business metrics.
How it works
The new deepfake maker from ByteDance works very simply by combining a picture and an audio clip to create a video.
All the creator needs is a suitable image and the sound he wants to associate with the picture and the tool can generate near real videos from scratch that viewers can hardly distinguish from the real thing.
The tool can also adjust elements of the video such as aspect ratio and body framing, as well as modify existing video footage, editing things like body movements and gestures.
Although the videos aren’t without faults obviously, they are clearly much better than videos created by any other tool in the past.
This isn’t surprising because the software went through 18,700 hours of video data training using a novel “omni-conditions” approach.
This lets OmniHuman learn from multiple input sources like text prompts, audio, and body poses simultaneously, which significantly reduces data wastage compared to other deepfake models, ByteDance researchers said.
Good but also bad
Deepfake tools like this can significantly make it easier to make videos for advertising and other noble purposes, but they certainly also worsen the concerns around such tools.
For example, Pope Francis has repeatedly spoken on the need to tighten regulation around such technologies as they can be used for bad ends.
The pope spoke even more after deepfake pictures of him were released on social media in a way he did not approve of. With such realistic videos, there’s no telling what can be done next.