>>530598993
My personal take is some combination of:

A) Face tracking with a single black and white camera is too hard and the problem is not solvable no matter how smart you are.

B) Only a large company like Apple or Meta is willing to pay for 50-100 people with vastly different faces to take hundreds of clean images for use as training data.

It is honestly impressive that babble works as well as it does given the right circumstances but the problem is extremely hard. Babble is not just trying to detect like open/closed mouth and smile/frown, there are 50 different expressions parameters that it needs to differentiate and convert to stable signals and if any of them wobble it makes your avatar look clowney (unless you delete less reliable sections from the blendtree which is an option).

https://docs.vrcft.io/docs/tutorial-avatars/tutorial-avatars-extras/unified-blendshapes

I've just attached a random screenshot from their discord, but this is the level of image quality that gets fed to the AI.