A convincing still cannot establish what happens when a hand lifts a cup. The clip must preserve the cup’s identity, keep its contact with the hand coherent, and show the result of the lift. A failure in any one of those relationships needs a more specific description than “the video looks wrong.”
The inspection below follows a single imagined shot through those checks. It is a method for judging an output against its intended use, not a ranking of products or a claim that every generated clip fails.
Separate appearance, continuity and physical behavior
First inspect a representative frame for image quality. Then watch the clip for continuity. Finally ask whether the action follows the physical relationship it is meant to depict. Passing one layer does not establish the next.
The VBench paper separates frame quality from temporal dimensions such as subject consistency, background consistency, flickering and motion smoothness. It also measures how much motion occurs, because a largely static clip can score well on consistency checks while avoiding difficult movement. Its 2023 experiments concern the systems tested then, not today’s complete market.
A smooth movement can still be wrong. A cup might glide through a tabletop without flickering. Its texture would remain consistent while its interaction with the table would fail. Conversely, an otherwise sensible action might have distracting texture changes. Each needs its own description.
Do not reduce both cases to “the AI looks bad.” Name what changed and when. That gives an editor a concrete decision and makes comparisons between versions easier.
Write the promises in the shot
Before reviewing, state what the shot must show. Consider this original hypothetical example: a person lifts a blue cup, moves it behind a book and places it on a clear part of the table.
The requirements are modest but connected. There should be one cup throughout. The hand should maintain a plausible grip while moving it. The book should conceal part of the cup, rather than transform it. The cup should reappear with the same recognizable features and end on the tabletop.
Write only the requirements that matter for the use. A surreal music video may deliberately transform objects. An instructional clip demonstrating a grip cannot treat the same transformation as harmless decoration.
A reviewer should also distinguish a visible event from an inferred one. If the grip is hidden for most of the shot, the clip may not show enough to judge it. Marking that requirement “unverifiable” is more accurate than awarding it a pass because the beginning and ending look good.
Use a few checkpoints without losing the whole action
Watch once at normal speed to understand the intended action. Watch again with attention to a particular relationship. Then inspect frames around a suspected break, rather than choosing only the most attractive still.
For the cup example, a small worksheet could look like this:
| Checkpoint | Question | Useful failure description |
|---|---|---|
| Before contact | Where are the cup and hand? | The hand begins inside the handle |
| During the lift | Does contact stay consistent? | The cup moves before the grip closes |
| Behind the book | Is concealment coherent? | The book’s edge shifts with the cup |
| After reappearance | Is the object still recognizable? | The handle has changed sides without a turn |
| Final placement | Does the cup meet the surface? | Its base sinks through the table |
These are proposed inspection questions, not measured findings about a particular model. A failure description should include enough detail that another reviewer can locate it. If it appears only in a single compressed frame, check whether you are seeing a transient artifact rather than a lasting change.
Physical plausibility needs an explicit test
T2VPhysBench uses prompts tied to physical laws and human judgments of the generated results. Its 2025 study tested ten systems on short clips and found substantial difficulties. More detailed hints did not consistently improve the tested behavior. Those experiments illustrate why visual quality and physical compliance need separate evaluation.
The paper is a preprint with a defined sample and protocol. Its scores are not universal failure probabilities, and the tested product versions should not be silently treated as current versions. You also cannot infer the exact internal cause of one faulty clip merely by recognizing an artifact.
For your own review, turn “realistic motion” into an observable condition. If a demonstration needs a ball to fall and strike the floor, check that the ball’s path, contact and subsequent motion are shown. If the impact is missing, you cannot establish it from a plausible final position. For people, the human-motion walkthrough separates joint geometry, contact, and physical support.
Prompt changes can be worth trying, but a better instruction is a proposed fix, not verification. Review the new output against the same requirements. Otherwise a rewritten prompt can produce a prettier shot while dropping the difficult part of the action.
Match the acceptance decision to the use
A landscape illustration may tolerate a drifting background detail. A product demonstration may require the item’s geometry and operation to remain faithful. A science lesson needs independent verification of the mechanism it teaches; an attractive simulation-like image is not experimental evidence.
Keep a short record of the requirement, the observed event and your decision. For example: “Cup identity fails after concealment; unsuitable for continuity-sensitive product footage.” That statement is more actionable than an overall score without an explanation.
If you shorten or crop the clip to avoid a defect, check the resulting sequence again. The edit may remove needed evidence of the action. Label generated material appropriately for its context, and do not present an invented physical event as recorded evidence.
The decisive review follows relationships over time. Find the object, follow its contact, inspect what happens while it is hidden and verify the result. A strong still frame begins the assessment; a coherent sequence is what lets the shot do its intended job.
