my colleagues have now published not only the paper, but also, full source code (!) for their new ML Video Codec model that, to the best of our knowledge, is the first ML Video Codec that actually works reliably across different hardware implementations (so, is actually usable in a real world setting where two people in a call might have different models of computer). At lower, but still *relevant* (for videoconferencing) resolutions like 540p, this also runs fast enough for actual low latency, real time operation!
If you're interested in making stuff like this work reliably in real world settings, this might be interesting for you to read!
Github: https://github.com/microsoft/MLVC
Paper: https://arxiv.org/html/2606.28027v1
They'll probably also write a blog post, at some point
