Motion exposes everything a still image can hide.

The uncanny valley in generated video is not mainly about faces. It is about violated expectations of how physical things behave, and the expectations are extremely precise.

You know how a coat should swing, how a foot should plant, how a head should move when the body turns. These are learned from a lifetime of watching people, and they are not conscious knowledge.

A still image asks one question: does this look right? A clip asks a hundred: does it move right, does it stay the same object, does it obey the light, does the sound belong to it?

Each question is a chance to fail, and the failures compound. That is why short clips of simple scenes work and long clips of people talking rarely do.

Weight is the first thing to check. Feet slide, contact is soft, and objects are picked up without the hand appearing to grip them.

Momentum is the second. Cloth, hair and loose objects do not lag and settle; they move in sync with the body or drift independently of it.

Consistency is the third. Objects change shape across frames, props come and go, and background elements rearrange when the camera moves.

These read as wrong even to viewers who cannot articulate why, which is precisely the uncanny response: the scene is legible and not believable.

Eyes fail in small ways that matter enormously: blinks too regular or absent, gaze that does not converge, pupils that change size between frames.

Mouths fail during speech. Teeth merge, the count changes, the jaw moves without the lip shapes matching the sounds, and the tongue appears in places it should not.

Micro-expressions are usually absent. A real face holds tension around the eyes and mouth between expressions, and a generated one returns to a neutral mask that no person has.

If the clip needs a speaking face, treat it as the hardest case in the whole field and plan for extensive review or a different shot.

Watch for flicker: subtle changes in brightness, colour or texture between frames that a real camera would not produce.

Watch for drift: the same object slowly changing shape, a pattern on a shirt gradually rearranging itself, a logo mutating across a pan.

Watch the seams. Longer clips are often assembled from shorter generations, and the join shows as a jump in framing, lighting or body position.

Half-speed playback is the fastest way to see all three, and it is the first thing a professional reviewer does.

Mismatched audio destroys a convincing clip faster than any visual defect, because the ear is unforgiving about timing and space.

Check lip sync within a frame or two. Check that the ambience matches the room: a large hall and a small office do not sound alike, and generated audio frequently ignores this.

Check continuity. Traffic, birds and hum should behave consistently across cuts rather than restarting or vanishing.

Watch muted first, then with sound. Reviewing with audio first is how bad visuals get approved.

Wide shots, slow camera movement, simple subjects and short durations are the reliable territory. A landscape with a single figure at a distance can be excellent.

Close-ups of hands doing work, conversations and anything with complex occlusion are the unreliable territory.

Cutaways, inserts and abstract motion are the cheapest wins: they carry a video without asking the viewer to scrutinise a face.

Design the edit around the strengths. Three good six-second shots cut together beat one fifteen-second shot that has to hold attention through its weakest frames.

Cut earlier. The failure is often in the last second, and trimming to the last good frame is free.

Grade it. A consistent colour treatment unifies mismatched shots and hides minor flicker, without pretending the footage is something it is not.

Replace the audio with real sound: room tone, foley, music. Sound design fixes more perceived realism than any visual repair.

If the failure is structural, reshoot the concept rather than repairing the clip. Regeneration with a simpler brief is usually faster than frame-level rescue.

Frame-level repair has a hard limit. Retiming or warping a face to fix a blink introduces a smear that reads worse than the blink, because the viewer's eye is drawn to exactly the region you changed. Fix the shot, not the frames.

The creator stops seeing errors after the fourth viewing. A fresh viewer notices the sliding foot in one pass, because they are not watching for what they intended.

Give them one instruction: say what looks wrong, in order, without being polite. Polite review produces approvals.

Show it on a large screen and on a phone. The two expose different defects, and the placement decides which matters.

Record the notes against the clip. Three viewings and a list is a process; a feeling that it is fine is not.

Where a video could reasonably be mistaken for real footage of real people, say what it is. The requirement varies by platform and country, and the editorial obligation is broader than the legal one.

Do not present a generated clip as documentary evidence of an event, a product in use or a person saying something. That is not a quality problem, it is a misrepresentation.

Where the clip is clearly illustrative, an on-screen label or a line in the description is enough, and it costs nothing.

Keep records of how the footage was produced. The question arrives eventually, and reconstructing it later is much harder than noting it at the time.

Generated clips are usually short, and the instinct is to stitch several together to reach a usable duration. Stitching is where the illusions break, because each clip was resolved independently.

Light changes between shots. A subject's clothing detail drifts. The camera height moves by ten centimetres. A viewer will not name any of these, and will feel all of them as amateur editing.

The professional answer is the same as in any edit: cut on action, use a cutaway between shots, keep shot sizes different so the eye does not compare them directly, and hold each shot long enough to be intentional.

Budget for the audio you did not generate. Room tone, foley and music cost a fraction of another generation pass and they do more for perceived realism than any visual repair.

If the piece needs a person speaking for more than a few seconds, plan a real shoot for that segment. One authentic close-up anchors a sequence of generated cutaways, and the reverse does not work.

Watch muted. Watch at half speed. Follow the eyes, the feet and the contact points. Then watch with sound and check sync and ambience.

If it survives, check the background for life and the shadows for consistency, which are the last two failures most people find.

Then ask the only question that matters for publishing: at this length and this size, does the viewer have time to see the flaw?

The review list on this page runs the checks in that order, so a batch gets the same scrutiny as the first clip.