Consistency is not a fixed cost. It is a requirement that has to hold between every pair of neighboring frames, so the number of agreements the model must maintain grows with the length of the clip. A two-second clip asks for a few dozen consistent transitions; a twenty-second clip asks for hundreds. Each one is a chance for a small disagreement to creep in, and small disagreements do not stay small — once a frame drifts slightly from the one before it, the next frame is refined against the drifted version, and the error compounds down the sequence.
This is why short clips are easier than long ones, and it is a structural reason rather than a limitation of any particular product. The same effect explains the characteristic artifacts.
Flicker is the mildest form: brightness, color, or fine texture shifts slightly from frame to frame, so the clip appears to shimmer even when nothing in the scene is moving. It usually comes from tiny frame-to-frame disagreements in pixel values that the eye picks up as a pulse.
Morphing is more serious. An object gradually changes shape or identity across the clip — a face slowly becomes a different face, a hand gains or loses fingers, a logo's letters rearrange. This is the temporal version of the alignment failures seen with still images, except that here the model has no single frame to get right; it has to keep getting the object right while the object also moves.
Drift is the slow version of the same problem: the whole scene slides, the subject's color shifts, or the background gradually changes, so the clip no longer looks like one continuous shot. Warping appears when the model tries to reconcile a moving object with a constraint it cannot satisfy, and the object bends or stretches unnaturally in the attempt.
All four artifacts share one cause. The model is not tracking objects the way a person does; it is producing pixel values that look locally plausible, and local plausibility does not guarantee that frame fifty still depicts the same thing as frame one.