In cooperation with the University of British Columbia, Apple has created an AI neural network that can generate a real person’s virtual movements.
After being trained on short videos, the NeuMan neural network can create a deepfake of a person not only walking, but performing any other kind of movement, such as dancing or tumbling. None of these movements need to be featured in the original videos that AI uses for training.
For this purpose, AI relies on two kinds of videos – one with a background scene, and another one with a person that AI will copy. After training, it creates a virtual space where the human model can perform as instructed.
For those who are curious about trying it for themselves, Apple has posted the source code on GitHub.