Hacker News Reader

The short answer is that they're only showing meshes that are deforming. Teeth are rigid and would be put back in based off the jaw and skull position. If memory serves, eyeball position tracking isn't great with this approach, it's not all that hard to add back in, and they often want to cheat eyelines anyway.

I think pfranz is on the right track - in one of the other publications they released today they show off how their tech does jaw positioning based on certain constraints (it seems to know where the jaw hinges and how the jaw bone interfaces with the skull).

My guess is that the teeth isn't visible throughout all the frames, so it's hard to reconstruct accurately in 3D. In fact the parts that look like gum and the tongue don't seem to be correct, far from the reconstruction quality of the face itself.

They're not? You can clearly see the teeth in the frames where they are visible in the raw. Teeth are just not commonly seen outside unnatural facial movements.

At 1 minute in the video in the mesh reconstruction, the gums / tongue are modeled, and there's clearly no teeth