Attention is how a model decides which parts of the input matter for the part it is currently producing. In video, that means every frame can be conditioned on every other frame, which is what makes coherence possible at all.
The cost grows roughly with the square of the sequence length. Double the clip and you do not double the compute, you roughly quadruple it. That single fact explains most of the pricing and most of the length limits in generative video.
It also explains why coherence degrades with length: the model is holding an increasingly expensive relationship in mind, and the practical compromises made to keep that affordable are exactly where drift creeps in.
Why do longer AI video clips cost more than proportionally?
Attention cost scales with roughly the square of sequence length, so doubling the clip length roughly quadruples the compute rather than doubling it.
Is attention why faces drift across a long clip?
Partly. Maintaining relationships across a long sequence is expensive, so the compromises that make it affordable are where coherence starts to give way.