Generated video looks convincing within a single clip and falls apart across several. The reason is that the model has no persistent representation of anything in the scene.
What the model is actually tracking
A video model produces frames conditioned on the prompt and on nearby frames. It maintains coherence over a short window because those frames are directly visible to it.
There is no internal record of a character as an object with fixed properties. The jacket is consistent from frame to frame because consecutive frames resemble each other, not because the model knows there is a jacket.
Once the temporal window ends, that support disappears, and the next clip is generated with no memory of the last.
Why the failures cluster at cuts
Within a clip, drift is gradual and often unnoticeable. Across a cut, the generation restarts and the discrepancy arrives all at once.
Details with weak textual anchors go first. Hair length, jewellery, the number of buttons, the arrangement of objects on a table and the direction light falls from are each unlikely to be specified in a prompt.
Viewers detect these instantly even when they cannot name what changed, because continuity is something audiences have been trained on by a century of edited film.
How reference conditioning helps
Supplying a fixed reference image of a character or set constrains each generation towards the same appearance, which reduces drift substantially.
The constraint is a pull rather than a guarantee. Strong conditioning holds appearance and restricts pose and angle, so the shots begin to look repetitive.
Productions that need genuine consistency generate more takes and select from them, which converts a modelling problem into a search problem with a real cost in time.
Why physics compounds the problem
Continuity is not only appearance. An object set down in one shot should still be there in the next, and a spilled liquid should not restore itself.
The model has no state tracking causes and effects, so consequences do not persist unless the prompt restates them, and restating every consequence is impractical.
This is why generated sequences work best when each shot is nearly independent, and worst when the story depends on something having changed.
Where the constraint pushes production
Teams working seriously with these tools have adapted by writing for the limitation rather than against it.
That means shots without recurring characters, sequences where visual variety is stylistically acceptable, and heavy use of generated material as background or texture within conventionally shot footage.
The projects that go badly are the ones that assumed continuity would arrive during the edit, which is the one stage where it cannot be added.