Tencent offers deepfakes as a service based on 3 minutes of footage
Tencent Cloud is currently running a pilot where customers can order deepfakes based on images submitted by them. For this, the Chinese company uses its own deep learning platform consisting of ten AI algorithms. A deepfake costs about 131 euros.
Customers must provide about three minutes of footage and at least 100 spoken sentences, after which Tencent’s algorithm delivers a deepfake video within 24 hours. confirms The Register. One can choose from displaying the deepfake person from the hips or full screen. There would also be options to change the background of the video and the tone of the video. Customers can choose from multiple global style presets, including 3D images that are realistic or cartoonish, or 2D deepfakes that are realistic or cartoonish in style.
The voices of the deepfakes should sound as realistic as possible with Tencent’s own algorithms. For example, the timbre of the imitated person should be less easy to distinguish from real voices thanks to deep learning models and vocoders based on neural networks. Many speech models currently generate quite mechanical sounding voices. In contrast, there is no example available of what Tencent deems realistic.
It is not clear whether Tencent will eventually roll out the deepfake service more widely. As far as we know, there is no official announcement of the service yet. It is also not clear what restrictions the Chinese company has with regard to the content of deepfakes and whether anyone can have images of another person generated without supervision and possibly unsolicited.
Image via Tencent / Jimian