AI is very capable when it comes to manipulating videos, and it only shows signs of getting better at it, with the most prevalent examples being face-swapping deepfakes. But the abilities of AI don’t stop there and we are finding more and more uses for it in every part of our lives. The latest attempt? Dancing, of course!
The AI in question was developed by a team of researchers from UC Berkeley and it is capable of reading someone else’s dance moves and then copying them onto a target body.
The system is made out of a few small steps. First, a sub-program turns the movements of a recorded target into a stick figure and then a source video is found and a stick figure is made out of their movements too. With all that data on hand, the swap happens – the video is synthesized and the target individual can move like the target source.
A subroutine smooths the movements of the stick figures so the dancers won’t move awkwardly around, while a completely separate neural network is dedicated to retracing the target’s face in order to ensure realism. There’s a lot of engineering at work here, guys.
The program does come with some limitations, of course. It cannot model loose fabrics, so the target has to wear clothes that are tight-fitting. In the example video above, you can clearly see the moments where the AI was unable to replicate more complex movements but it’s still impressive work nonetheless.
It will probably be turned into an app soon enough.
The project looks like a lot of fun but what are the implications when it comes to manipulating video though? Endless. It makes the recent AI advancements both exciting but also a little bit frightening.