Test-time training 3D reconstruction

I think the project page is a better URL, it contains more info and demos for what this is and it has a link to the code anyway:

https://rover-xingyu.github.io/TTT3R/

This looks really interesting, especially when combined with gaussian splats.

Can it be used to make a stereoscopic video, e.g. via making a 2D video input have a reconstructed 3D depth map?