Most defects in generated video survive review because reviewers watch once, at a small size, with sound, while thinking about the idea. The defects live in motion, at the edges, and in silence. Twelve checks, ten minutes, before anything ships.
Watch it three times, differently
Pass one: full screen, sound on, do not take notes. You are checking whether it works at all.
Pass two: muted. Most of the audience will see it this way. If it stops making sense, the piece has a structural problem no polish fixes. See how to caption video for silent viewing.
Pass three: on a phone, at arm's length. Where it will actually be watched.
The twelve checks
1. Edge flicker. Play at quarter speed and watch subject edges against the background. Frame-to-frame instability is the most common generated defect and is invisible in a still.
2. Identity drift. If a person or product appears in more than one shot, put the frames side by side. Drift is invisible sequentially and obvious in a grid.
3. Hands and small articulated objects. Pause on any frame where they are prominent.
4. Physics. Anything pouring, falling, clicking, or making contact. If it looks almost right, it is wrong; viewers register it without naming it.
5. Text in frame. Any generated text is suspect. Read every word.
6. Eye contact. Sustained unblinking contact with the lens is the strongest uncanny tell. See why AI video looks fake.
7. Lighting continuity across shots. Generated shots drift in colour temperature and direction. A sequence lit three ways reads as three different days.
8. Audio sync, if generated. Check footfalls, impacts, and any lip movement specifically.
9. Caption safe area. Platform UI covers the bottom of vertical video. Captions sitting at the true bottom get hidden behind buttons.
10. First frame. It is the thumbnail. Check it as a still.
11. Last 0.5 seconds. Generated clips often degrade at the end. Trim rather than fade.
12. Aspect and resolution against the destination. The boring one that causes the most re-delivery.
The defects worth regenerating versus fixing in post
Regenerate: identity drift, physics errors, hands, text. Post cannot fix a model's misunderstanding.
Fix in post: colour continuity, edge softness, grain mismatch, audio, trims. Cheaper and faster than another generation.
Knowing which is which saves more time than any other habit here. A regeneration is 24 to 448 credits and a few minutes; a grade is seconds.
The one check for anything with a person in it
Did we generate a person who is presented as real? In a testimonial, a case study, a founder story, or a recruitment video, the answer must be no. That is not a quality issue, it is a claim about reality, and the checklist is the last place to catch it. See how to protect creative work when a tool shuts down for the adjacent rights questions.
FAQ
How long should QA take? Ten minutes per asset, less once the checks are habit. Considerably less than a re-delivery.
What is the most commonly missed defect? Edge flicker, because it does not exist in a still and reviewers pause on stills.
Should the client do this pass? No. Ship work that has already passed it. A client finding an artefact costs more than the ten minutes.
Three passes, twelve checks, and know which defects post can fix. The 8frame canvas is free and unlimited, and generation is paid from $19/month.