Why Extended Clips Drift, Flicker, or Crumble — and How to Prevent It

The extension ran. The clip is longer. And something is wrong — not obviously wrong, the kind of wrong you notice on the third viewing and cannot unsee afterwards.

Drift is the defining failure of AI video extension, and it comes in four recognisable forms. Naming them makes them fixable, because each has a different cause and a different fix.

Why extension drifts at all

A video model does not have a memory of your scene. It has a conditioning signal: the last frame you gave it, plus your text description of what happens next. From that it samples forward.

The last frame is a single instant. It tells the model what is on screen, the light, the colour, the composition — and nothing about what has been happening. Which way the camera was moving, how fast, whether the subject is accelerating, how the light has been changing over the last second: none of that is in one frame.

So the model infers. Most of the time it infers well, because most shots have a dominant motion and a stable grade. When it infers wrong, you get drift.

And when you extend an already-extended clip, the second run conditions on a frame that is itself a guess. Error compounds. This is why three consecutive extensions almost never work.

Signature 1: motion discontinuity (the “jump”)

What you see: the clip plays smoothly, then at the join the camera speed or direction changes. A dolly becomes a slow drift. A pan accelerates. The subject's gait changes rhythm.

Cause: acceleration and velocity are properties of the last several frames, not of the last frame. The model sees a camera mid-move and must guess whether it is speeding up, coasting, or arriving. It picks one.

Fixes that work:

• Describe the velocity explicitly. “Camera continues dollying in at the same slow constant speed” gives the model a target. “Camera continues moving” does not.
• Avoid joining at a velocity extreme. Trim the source so the last frame sits mid-move rather than at the start or end of a movement. Mid-move extension has the most information to condition on.
• Prefer cut over continue. If the join is visible, a cut at that point hides it entirely and costs nothing.

Signature 2: identity drift

What you see: a person, animal, or product changes subtly over the extension. The face is a near-miss. The jacket gains a detail. The product's proportions shift slightly.

Cause: the model has one frame of that subject. A face seen from one angle at one instant is not enough to reconstruct the subject; the model generates a plausible continuation of the appearance, which is not the same as the person.

Fixes that work:

• Keep the subject small and far. Identity drift is proportional to how much of the frame the subject occupies.
• End the source with the subject in a stable pose. Mid-turn or mid-gesture frames give the model more freedom to wander.
• Do not extend clips with a speaking face. This is the worst case and the least recoverable.

Fixes that do not work: describing the person at length. The model is not matching against a memory of them; detail in the prompt does not restore identity that was never encoded.

Signature 3: grade and grain shift

What you see: the join is not a motion problem but a colour one. The new seconds are slightly cooler, or warmer, or cleaner, or noisier. Skin tones read differently. Shadows lift or crush.

Cause: a different model, or a different generation pass of the same model, made different assumptions about the look. This is the most common cause of "the extension is technically fine but I cannot use it" — the seam is stylistic, and stylistic seams are very hard to hide.

Fixes that work:

• Extend with the model that made the source. This is the single most effective fix, and it is the one people skip to save credits. A Runway extension (90 credits) on Veo footage will produce a seam you then pay 450 credits to redo.
• Apply a grade after joining. A subtle look-matching pass across both sections — slight contrast and saturation matching — often closes a grade seam that is otherwise obvious.
• Add matching grain to the extension. Grain is the tell. Clean frames next to grainy ones look pasted.

Signature 4: the crumble

What you see: not a subtle drift but a collapse. Colours separate. Shapes lose coherence. Text becomes alien. The scene dissolves into something that is technically video but not a shot.

Cause: under-determined scenes plus compounding error. Busy environments, fast motion, many objects, conflicting lighting. The model had too many plausible continuations and picked one that contradicted the rest of the frame, then built on it.

Fixes that work:

• Treat it as a source problem, not an extension problem. A crumbling extension is telling you the last frame was ambiguous. Change the source's ending, or regenerate the clip, or cut.
• Never extend a crumbled extension. The second run inherits the collapse as its ground truth.
• Avoid extension entirely on dense scenes. Crowds, text-heavy environments, fast action. The failure rate is high enough that the attempt costs more than it saves.

The compounding rule, quantified

Runs on the same clip	Expected outcome
1	Join usually invisible or nearly so
2	Motion energy shifts; grade may drift; usually still usable
3	Identity and lighting drift; joins visible on inspection
4+	Effectively a different clip; discard

The rule that follows is short: one extension per clip, then cut.

If you need more length than one extension provides, generate a second source clip and edit them together. A cut hides every drift signature at once, because a cut is supposed to be a discontinuity. A continuous take gives the viewer no such allowance.

The four inputs that decide the outcome

Before you extend, four things determine whether it will work. All four are checkable in under a minute.

1. The last frame. Look at it. If it is mid-motion-blur, has a compression artefact, or contains a hallucinated logo, the extension starts from a defect. Trim one frame — or one beat — and start from a clean instant.

2. The subject's position in frame. Fully visible and close to the centre? Best case. Partly out of frame or turned away? Expect identity drift.

3. The scene's ambiguity. One moving element in a stable environment extends reliably. Twelve moving elements do not.

4. Whether the source model matches the extension model. Mismatched is the most predictable seam there is.

What “good” looks like

A successful extension has one property: you cannot find the join on a first viewing. Not that it is invisible under frame-by-frame inspection — it will not be, and no extension method currently achieves that. The standard is that a normal viewer watching at normal speed does not register a discontinuity. That is achievable, repeatedly, with clean sources that are simple and short.

It is not achievable on every clip. Knowing which clips those are, before you spend the credits, is most of the skill.

Open the AI Video Extender and run it on your most complicated shot first, not your simplest — if it survives, everything else will.

Related reading: How long can you actually extend a clip? · What an AI video extender actually does · Extend, loop, or regenerate?