Generated video handles motion and lighting convincingly and still fails at keeping the same person looking like the same person. The problem is identity persistence, and it is structural rather than a matter of resolution.
Each clip is generated from scratch
A video model produces a sequence conditioned on a prompt and, usually, a starting image. It does not retain anything from a clip it generated earlier.
Asked for the same character in a new shot, it reconstructs a face matching the description, and descriptions do not specify a face precisely.
The result is a plausible person of the right general type, differing in the details that identity actually rests on.
Faces are recognised on fine geometry
Human recognition depends on relationships that are hard to state: the exact distance between the eyes, the shape of the philtrum, the asymmetry of a smile.
Those measurements are far more precise than any text prompt can express, and a model that gets the hair, age and expression right can still miss all of them.
Viewers register the mismatch without being able to say what changed, which is why generated characters read as unsettling rather than as wrong.
Drift also happens inside a single shot
Within one clip the model conditions each frame on the previous ones, which maintains continuity across short spans.
Small errors accumulate, so a face at the end of a long clip has usually migrated from the face at the start, particularly through a turn or an occlusion.
This is a direct reason for the short default clip lengths, since consistency degrades faster than image quality does.
Reference conditioning is the working fix
The practical approach supplies reference images of the intended face alongside the prompt, so the model has geometry to match rather than a description to interpret.
Adapters trained on a specific person work better still, encoding the identity into the model rather than passing it as input.
Both require the face to exist beforehand, which suits productions with a defined character and does not suit generating a new one on demand.
Consistency competes with control
Constraining identity tightly reduces the model's freedom to change pose, expression and lighting, and outputs become stiff.
Loosening the constraint restores the range of performance and lets the face drift again, and no setting satisfies both.
Production workflows resolve this by generating loosely and applying a face replacement pass afterwards, which is an admission that the two requirements are handled better by two separate systems.