The two families behave differently in ways that matter on a set. A model that denoises the whole clip at once tends to hold global coherence better but is capped in length. One that generates forward can in principle run indefinitely, and accumulates error as it goes.
That accumulation is the mechanism behind the drift you see towards the end of a long generation: later output is conditioned mostly on earlier output rather than on the original reference, so small deviations compound.
Practically it means the two families want different production discipline. Whole-clip models want you to work within their length. Forward models want you to re-anchor, with a reference or a fresh conditioning frame, before the drift becomes visible.
Why does a long generated clip get worse towards the end?
In a forward-generating model, later frames are conditioned mostly on earlier generated frames rather than on the original reference, so small errors compound rather than being corrected.
How do you work around it?
Re-anchor before the drift shows: use first-and-last-frame conditioning where the model supports it, or cut and start a new generation from a fresh reference.