Funny, they showed the comparison on the BBC News channel between standard video calls and this approach, and this approach had noticeable glitches in comparison to the real stream. Then I noted they're using artificially throttled h264 for the comparison. But maybe those glitches are the payback for much reduced bandwidth. I'm also not sure why the needed a GAN for this, when picture deformation from point to point is an age old problem that's been solved a long time ago. But I guess that wouldn't handle hair movement well, but it seems to me they could save a lot of CPU churn by only sending that bit to the neural networks and doing the skin movement by more traditional methods.Look! N-vidia invented video chat the same way MMORPG avatars share their conversation!