Text generation moved from sentences to long documents quickly. Video generation has stayed at a few seconds per clip, and the constraint is a combination of computation, memory and what the training data contains.
Coherence has to hold across every frame pair
A video is coherent only if each frame is consistent with the others, and the number of relationships to maintain grows faster than the number of frames.
Attention mechanisms that compare every element with every other therefore become expensive very quickly as duration rises.
Architectures that limit comparison to nearby frames make longer output affordable and reintroduce the drift that full comparison was preventing.
Memory is consumed during generation, not just training
Producing video requires holding the working representation of the whole clip in accelerator memory at once.
That representation scales with resolution, frame rate and length together, and doubling any one of them consumes the headroom that the others needed.
The visible consequence is the trade offered in most products: longer clips at lower resolution, or short clips at high resolution, rarely both.
Training material is mostly short
Models learn from clips segmented out of longer footage, and the segments are kept short for the same computational reasons.
A model trained on four-second segments has seen very few examples of what happens across a minute, so it has learned little about how longer sequences develop.
Continuous single takes of any length are also rare in stock footage, which is edited material by construction.
Extension produces accumulating error
Longer output is commonly assembled by generating a clip, taking its final frames as the starting condition, and continuing.
Each continuation inherits whatever inaccuracies the previous one contained, and small errors in colour, lighting and geometry compound across the joins.
The characteristic result is a sequence that slowly changes appearance, which is more distracting than a hard cut would have been.
Editing conventions make the limit less binding
Conventional film cuts every few seconds, so a shot length of several seconds matches how footage is actually used.
Production workflows generate many short shots and assemble them, which is the same process used with real footage and requires no new skill.
The limitation bites hardest on continuous action that cannot be cut — a single unbroken movement, a long dialogue take — and those are exactly the cases where drift is most visible anyway.