Quality control
AI video quality control: the checks before an AI commercial airs
For the brand manager, agency producer or legal reviewer signing off an AI-made commercial or brand film, who wants a checklist that catches what generated footage gets wrong.
The short answer
Quality control for an AI commercial is a shot-by-shot check against the boards, the character sheet and the real product, then a technical and legal pass on the finished cut. Check faces, hands, the product and its label, any text, continuity between cuts and physics; then sound, loudness and each channel's specs; then the AI labels each platform and market requires and the licenses behind every model, track and voice. The campaign stills get the same checks at full size.
Why it is different
Why AI footage needs its own checks
A live-action camera records what was in front of it. A video model draws every frame, so it fails in ways a camera cannot: a hand gains a finger for three frames, one letter of a label changes mid-shot, a cup that goes behind a shoulder never comes back. The model makers publish some of this themselves. Runway's Gen-4.5 announcement lists objects that disappear or appear across frames, effects that come before their causes, such as a door opening before the handle is pressed, and actions that succeed when they should not, and calls these common to AI video models. Kling's own guide puts setting up one consistent character at around 10 to 20 attempts.
So the check runs frame by frame where it matters, and it runs before the edit, because a bad shot found after the grade costs a regeneration, an edit, a mix and a grade.
Pass one
The shot-by-shot checklist
| Check | What to look for | How to check |
|---|---|---|
| Faces | The same person as the character sheet; eyes, teeth, ears, hairline | Shot and sheet side by side; watch close-ups at half speed |
| Hands | Finger count, grip, a hand fused to an object | Step frame by frame through every touch |
| Product and label | Shape, color, logo, letters, count | Against the reference pack; eyedropper the brand colors |
| Text on screen | Spelling, legal lines, prices | All type set from files in the edit; proofread every super |
| Continuity | Wardrobe, props, light direction, time of day between cuts | Scrub cut to cut; line frames up side by side |
| Objects | Things that vanish or appear | Watch the background, not only the subject |
| Cause and effect | Doors opening before the handle turns; a pour before the tilt | Watch every action at half speed |
| Crowds and extras | Melted faces, duplicated people | Pause on every wide shot |
| Texture | Flicker, smearing on fast moves, surfaces that crawl | Full size, on a good screen |
The product row deserves its own page: product and label accuracy in AI commercials. The faces row starts with the cast: see character consistency in AI video.
Pass two
The technical pass
- Resolution and frame rate. Know what the model delivered before anyone upscales it: Veo 3.1 outputs 720p, 1080p or 4K at 24 frames a second. Check upscaled masters for halos and invented detail.
- Ratios and safe zones. Every version at its own ratio, with text and logos inside each platform's safe zone; for Reels and Stories, Meta asks for about 14% at the top, 35% at the bottom and 6% at each side to be kept clear.
- Sound. Lip-sync on any dialogue, voice and music levels, no clipped words at cuts.
- Loudness. US broadcast delivery is measured to ATSC A/85, which the CALM Act had the FCC make mandatory for stations and pay-TV distributors; ATSC revised it in July 2026 with new guidance for streaming. European broadcasters work to EBU R 128, which targets -23 LUFS.
- Captions. Word for word against the final mix, including legal lines.
- Files. Codec, size and length limits for each channel. See ad campaign deliverables.
Pass three
Labels and rights
| Where | The rule | What to do |
|---|---|---|
| YouTube | Disclose realistic content made or meaningfully altered with AI | Set the disclosure when uploading realistic AI scenes |
| Google Ads | An AI label setting; campaigns targeting the EU, India and New York get a visible overlay on AI-marked assets | Mark AI assets; Google may also label them itself |
| Meta | AI info labels on ads made with its own AI tools; photorealistic AI people get the label next to Sponsored; third-party AI detected through industry signals | Expect labels; keep provenance data in the files |
| TikTok | Significant AI content needs the AIGC label or a clear disclaimer; undisclosed ads are rejected or restricted | Label every AI-made ad |
| New York | Advertisers who know an ad uses a synthetic performer must disclose it conspicuously | Disclose AI-made people in ads that run there |
| European Union | From August 2, 2026, deep fakes must be disclosed by those who use AI professionally | Ask counsel which shots count as deep fakes |
Rights. Before delivery, confirm in writing that the video model's terms allow commercial use of the outputs, that the music license covers every channel and territory in the media plan, that the voice is licensed, and that any real person's likeness has written consent. Leave provenance watermarks such as Google's SynthID in place.
Who signs
Who checks what
| Stage | Who | What they sign |
|---|---|---|
| Every shot, before the edit | The studio | The shot-by-shot checklist |
| The cut | The brand | The look, the cast, the product, the story |
| The cut and every version | Legal, and MLR in pharma and medical work | Claims, what the product is shown doing, labels |
| The files | The media team | Specs, safe zones and label settings in each platform |
Every frame Ruminate X delivers is made with generative AI. There are no film crews, sets or location shoots. In the Ruminate X pipeline, each shot is generated, checked against its board and generated again until it holds; faces, hands and the product get the most passes; the real logo, label and on-screen text are set in the edit from the brand's files; and every frame is reviewed by a human editor before delivery. The full pipeline is on how we make an AI brand film.
The stills
Where Overs comes in
The key visual and the campaign photos need the same scrutiny as the film, and Overs builds much of it into making them. It is the sister product from the same founder: an AI tool that makes product and campaign photos that match a brand, without a photo shoot, built on the process 100creatives used to make hundreds of photos a week for hundreds of brands.
Before anything renders, a quality check step reads each planned photo against your brand rules; it proofreads and never rewrites the art direction. On Pro and Team, a reviewer then checks each finished photo against the references before you see it, and can retry on its own up to a limit you set. Then you approve a photo, ask for changes, or reject it with a reason; in a shared workspace Overs records who decided, comments sit beside the photo, and the decisions become brand memory for later sets. A wrong photo can be remade on its own with a note, in 1 to 4 versions side by side. The export carries a record of the prompt, model and reference pictures behind each photo, a paper trail for legal review.
The free plan makes 40 photos a month, with the AI billed to your own OpenRouter key at a few cents a photo and no markup from Overs. For the checks themselves, see Overs' QA checklist for AI images. Put every campaign still through review before it goes live: start on Overs free at www.overs.studio.
Should Ruminate X make and check your film?
Hire Ruminate X if
- A brand that needs an AI commercial or brand film checked shot by shot before its legal team sees it.
- A pharma, pharmacy, lab or medical marketer whose AI film goes through MLR review.
- An agency producer who needs a partner that delivers every version to spec, with the labels settled.
Hire someone else if
- You need an independent audit of a film another studio made: a post-production house or a broadcast QC service fits better.
- Your ad runs on broadcast TV and needs a certified QC report for a clearance body: add a broadcast QC house to the chain.
- You only need stills checked: the checklist and Overs cover it.
Questions buyers ask
How do you check an AI-generated video before it airs?
Shot by shot, then as a finished cut. Compare every shot with its board, the character sheet and the real product at full size, stepping frame by frame through hands and product contact. Then check sound sync, loudness and the delivery specs of each channel, set the AI labels each platform and market requires, and confirm the licenses behind every model, track, voice and likeness.
What are the most common artifacts in AI video?
Faces that drift between shots, hands with the wrong number of fingers or fused to objects, product labels whose letters change, objects that vanish or appear, and physics that runs out of order. Runway lists three of these for its own Gen-4.5 model: objects that disappear after being hidden, effects that come before their causes, and actions that succeed when they should not.
Do AI-generated ads need a label?
Often. TikTok requires its AIGC label or a clear disclaimer on ads with significant AI-generated content and rejects or restricts undisclosed ones. Google Ads adds visible AI overlays for campaigns targeting the EU, India and New York when assets are marked as AI-made. New York requires disclosure of synthetic performers in ads, and the EU AI Act requires deep fakes to be disclosed from August 2, 2026. This is not legal advice.
Who should sign off an AI commercial?
The studio checks every shot before the edit. The brand checks the look, the product and the cast. Legal, and MLR in pharma and medical work, checks every claim, including what the product is shown doing, and the AI labels. The media team checks the specs and the label settings in each ad platform. Fix the number of review rounds in the contract.
Can I remove the watermark from AI video?
Leave it in. Google's Veo and Gemini Omni Flash put an invisible SynthID watermark in every video they make. Meta and Google both say they use industry signals like these to label AI-made ads, and the EU AI Act requires AI providers to mark their outputs in a machine-readable way. Keep other provenance data in the files as well.
How does Ruminate X check its films?
Each shot is generated, checked against its board and generated again until it holds, with faces, hands and the product getting the most passes. The real logo, label and on-screen text are set in the edit from the brand's files, and every frame is reviewed by a human editor before delivery. Every frame Ruminate X delivers is made with generative AI.
Send the cut, or the brief
Send the brief, the channels and your review process, and we will plan the checks with you.