Perceptually lossless (talking head) video compression at 22kbit/s

The author discusses the recent LivePortrait deepfake model, highlighting its potential applications in social media and video compression. The model utilizes 2D avatar/portrait animation to animate still images with minimal data transmission requirements. Despite the impressive compression efficiency, there are concerns about the potential for misuse, as the model could generate inaccurate or distorted reconstructions. LivePortrait’s training dataset of 69 million frames and regional GAN losses contribute to its success in avatar controllability. The author suggests possible practical applications for the technology, such as creating formal avatars for remote work or immersive virtual meetings. Despite some limitations, the LivePortrait model shows promise for future developments in video compression and generative technologies.

https://mlumiste.com/technical/liveportrait-compression/

To top