I made a prototype a week ago with a face recognition engine on the frontend, connecting into the Jitsi JavaScript API, running my own Jitsi server on DigitalOcean.

The result: the video conferencing only showed faces! Imagine the possibilities of moving those faces around a stage, or using virtual cutouts where the faces would appear, etc.

It was very very fun, I have screenshots of 10 people connecting at once -- hit me up if you want to pursue something :-)