The first frame is a persuasive sales pitch. It offers composition, light and a recognisable subject all at once. A production shot has a harder job: it must remain coherent while something happens, join the surrounding sequence and survive the final delivery format. Approving a clip as a picture can postpone its real problems until the timeline is already built.
We propose three viewing contexts for acceptance: the clip on its own, the clip in sequence and the exported programme. Each context answers a different question. None is replaced by a general model ranking.
First write the shot contract
Before generating or commissioning a shot, describe the information it must communicate. Name the subject, required action, camera behaviour, duration and the details that must remain stable. Distinguish essential content from a visual preference. “The package stays recognisable while the camera moves” is an acceptance condition. “It should feel premium” is a direction that still needs interpretation.
Keep the contract short enough for a reviewer to hold in mind. If the shot must show an exact interface, label or technical procedure, decide how that detail will be supplied and checked. A vague prompt followed by an exacting review is an expensive way to discover that the brief was never agreed.
Context one: continuity inside the clip
The VBench research programme separates video-generation quality into dimensions including subject consistency, motion smoothness, temporal flickering and spatial relationships. Its original CVPR 2024 paper describes sixteen dimensions rather than treating quality as a single visual impression. Read the VBench paper.
Borrow the discipline of separating dimensions, without pretending that a studio review reproduces the benchmark. Watch the clip at normal speed first. Then inspect moments where objects cross, leave the frame, touch or become partly hidden. Ask whether the action completes and whether the important object remains the same object.
Do not let a near-static result pass a brief that required meaningful motion merely because it looks stable. Conversely, do not reject a deliberately still shot for lacking action it was never meant to contain. The contract decides what matters.
Context two: the neighbouring shots
Place the candidate in a rough sequence as early as possible. A slight direction mismatch may become obvious only when the next shot appears. Colour, subject size and screen position can each create an unintended jump. The candidate may be attractive in isolation while forcing extra work on every neighbouring shot.
Review the transition into and out of the clip. Is there enough usable material to make the planned cut? Does an important action begin before the edit can show it? Can the shot be shortened without removing the moment that explains it? If the only usable second sits in the middle of a longer generation, count that honestly.
Keep version notes tied to timeline placement. “Candidate 4 approved” is ambiguous if its approved use was a two-second cutaway but someone later stretches it into the opening shot. The accepted interval and purpose belong in the review record.
Context three: delivery
Inspect the exported file, including the intended crop, overlays and caption placement. A flaw hidden by a large editing preview may become conspicuous in a small vertical crop. A meaningful object can disappear under on-screen text. Review on the destination layout rather than treating export completion as acceptance.
Include accessibility in the plan. W3C’s media-planning guidance treats captions, description and transcripts as production decisions to consider early. Our practical takeaway is to reserve time and screen space for the information viewers need, instead of treating those elements as decorations added after picture lock. W3C WAI media-planning guidance.
Keep three outcomes, not two
A candidate can be accepted, rejected or accepted for a narrower use. The third outcome is valuable when recorded honestly. A clip with an unreliable final movement may work as a short establishing shot. That does not make the full generation successful; it makes a defined segment usable.
Record the reason a candidate failed. Separate brief mismatch, temporal defect, continuity problem and delivery problem. A short record reveals whether to revise the prompt, change the shot design, use another production method or stop trying. The objective is an approved sequence, not an ever-growing folder of impressive clips.