> The model needs to "understand" geometry and physics to output a video.

No it doesn't. It merely needs to mimic.