Generated video handles a subject moving in front of a fixed camera far better than it handles a camera moving through a scene. The difference reveals what these models have and have not learned about space.
The model works in frames, not geometry
Training data is flat video, and what the system learns is how pixels tend to change over time.
It has no explicit representation of a scene as objects at positions, so it has nothing to consult when the viewpoint changes.
Camera motion is therefore reproduced as a learned pattern of pixel movement rather than as a consequence of moving through a place.
Parallax is the specific failure
When a camera moves laterally, near objects shift across the frame faster than far ones, and the exact rates encode depth.
Getting this wrong makes a scene feel like painted layers sliding past each other, which is the most common artefact in generated tracking shots.
Errors are also revealing: a background that moves too slowly reads as much further away than the shot otherwise implies.
Occlusion demands invented detail
Moving the camera reveals surfaces that were previously hidden behind something else.
The model must generate that material with no information about it, and whatever it produces must remain stable for the rest of the shot.
Because generation is probabilistic, revealed regions frequently change as more of them becomes visible, which viewers see as the background rewriting itself.
Returning to a previous view exposes the absence of a scene
A camera that pans away and pans back should show the same thing, and this is where the lack of an underlying model is clearest.
What returns is something of the same description, differing in the specifics, because nothing preserved the original arrangement.
Short shot lengths conceal this, and it is a direct reason why generated footage favours motion that never revisits.
Explicit camera control is the current workaround
Newer systems accept a camera path as a separate input, so the trajectory is specified rather than inferred from words like sweeping or handheld.
This produces motion that is repeatable and matchable across shots, which is what a production actually needs.
It does not supply the missing geometry, so the parallax and occlusion problems persist — the camera now moves where it was told, through a scene that is still being improvised frame by frame.