How to Keep the Product Accurate in an AI Commercial

Why products change in AI video

A video model keeps no copy of your product. It generates every frame from what the first frame, the references and the prompt suggest, so anything the model is unsure about can change: the curve of a bottle, the color of a cap, the number of gummies in a hand. Small text suffers first. Letters on a label become shapes that look like letters, and a logo can soften or morph as the camera moves.

Three things make it worse. A turn that shows a side of the pack the model never saw forces it to invent that side. Reflective and transparent surfaces, glass, foil and liquids, change with every move of the light. And hands that hold the product press into it, cover the label and drag it out of shape.

Why it matters to juries and clients

In a contest, the product is part of the score. Atlanta AI Ad Fest weights entries 40% idea, 30% craft, 20% brand and 10% impact, asks of the brand score "Does it make the brand matter?", and warns that "Technical novelty alone does not win." The Formula E Creators Challenge requires elements of its GEN4 Asset Kit on screen for at least 4 seconds, so the car in your film has to match the car in the kit. Golden Dunes' 2025 listing judged its Best AI-Commercial category on "creativity, visual execution, brand alignment, and innovative use of AI technology."

For paid work, accuracy is a legal matter. In FTC v. Colgate-Palmolive (1965), the US Supreme Court held that "It is a material deceptive practice to convey to television viewers the false impression that they are seeing an actual test, experiment or demonstration" when they are not. In 1970 the FTC acted against Campbell Soup and its agency over ads with "clear glass marbles which prevent the solid ingredients (garnish) from sinking to the bottom" of the bowl. In March 2026 the UK's Advertising Standards Authority upheld 22 complaints about ads for a robot dog toy whose footage of "real and/or AI-generated puppies" gave "the impression that the product was highly realistic." And in the EU, the AI Act's transparency duty for deepfakes, whose definition covers content that resembles existing "objects" as well as people, has applied since August 2, 2026.

An AI label does not rescue an inaccurate product. The UK's Committee of Advertising Practice wrote in May 2025 that "disclosure alone is very unlikely to mitigate the harm caused by a fundamentally misleading message." Stylize the world around the product as much as you like, and show the product as it is.

Build the reference set before any prompt

Photograph the real product yourself, even for a spec spot. Shoot the front, the back, both sides, a three-quarter view, the top, a straight-on close-up of the label and one shot in a hand for scale, in soft, even light on a plain background, with the phone at 2x zoom so wide-angle distortion does not stretch the pack.

Then write the spec as words the model can follow and you can check against: the exact label text, letter for letter; the colors as names plus hex codes; the materials and finishes; the proportions ("the can is about twice as tall as it is wide"); and a short list of what must never change. Paste the product sentence into every prompt unchanged, the same way you would a character description.

The video tools can carry that reference set into the shot. Runway's product ad recipe takes "One to ten reference images of the product" and treats the first as the primary reference. Google's Veo 3.1 takes "up to three images of a single person, character, or product." Kling 3.0 Omni takes up to seven images or elements when no video is attached, and ByteDance's Seedance 2.0 and 2.5 take up to 9 and 30 reference images. Higgsfield says its product video generator holds "the exact shape, label, and color frame to frame" from one to three references. Treat all of these as vendor claims and check the output yourself.

Approve the product as a still, then move it gently

Make each product shot's first frame as a still and check it at full size against the real pack before it becomes video. A wrong label caught on a still costs a re-render measured in cents. The same mistake caught in a finished clip costs the clip, and on a deadline it can end up in the film.

Choose moves that keep the label facing camera. Push-ins, slow slides and light sweeps across a static product hold up better than spins, because a spin shows sides the model has to invent. Keep product clips short, and let hands touch the product at the edges, not across the label. The agency 100 Creatives saw the spread in its own 2026 numbers. Gummy supplement ads took 3.2 renders per usable image, while surf apparel worn by one model across studio, cliff and ocean scenes took 13.6.

One product shot, start to finish

Here is one shot worked through for a hypothetical supplement jar in a 30-second spot, a four-second push in on the jar standing on a bathroom shelf.

• References: The front of the jar, a straight-on crop of the label and one three-quarter view, sent in that order.
• Product sentence: The exact label text, the lid color as a name and hex code, the jar's proportions, and "the label never changes."
• First frame: A 16:9 still with the label facing camera and large enough to read, approved at full size against the real jar.
• Motion prompt: A slow push in, the jar stays still, the label stays facing camera and unchanged, steam from a mug drifts behind it.
• Tries: Three video renders, the best one trimmed to four seconds.
• Fix: If the label softens in the last second, cut earlier or track the real label artwork over it.

That is six decisions for four seconds of film, and each one costs less to get right at the still stage than in the edit.

Put the real label back in the edit

When the model cannot hold the label, stop asking it to. Track the label or logo from the brand's real artwork onto the product in the edit, or cut to a real photograph for the moment the label must read. Match the colors to the hex values with an eyedropper during the grade, since a model reads a hex code only as a rough hint.

Check the contest rules before you composite. Some contests require AI in every shot, and some require all video to come from one tool, so a tracked label may need a line in your submission notes.

The final check, frame by frame

• Read every visible word on the pack, letter by letter, at 100%.
• Compare logo shape and position with the brand's own artwork.
• Check the cap, lid, seal and any small parts against the photos.
• Check colors against the hex values with an eyedropper.
• Count what should be counted: pieces, cans in a pack, gummies in a hand.
• Check proportions in wide shots, where products tend to stretch.
• Check the hands: fingers, grip, and whether they cover or deform the label.

Where Overs fits: product-exact stills before any motion

Overs is built for the half of this problem that happens before motion. Its ten steps include a product analysis before any shot is planned, and each shot gets reference slots. The plan names which reference pictures a photo needs, in order, and sends only those. If the plan calls for a view you never supplied, such as the back of the pack, Overs skips that reference instead of guessing with the wrong picture.

Review is where accuracy gets enforced. On Pro and Team plans, a reviewer checks each photo against the references before you see it and can retry on its own up to a limit you set. You approve, ask for changes or reject with a reason, and those decisions become brand memory for the next set. When exactness matters most, pick GPT Image 2.5 Sunburst, the slowest model Overs offers and the most exact at keeping a product identical while the scene changes.

Overs was built on the process 100 Creatives, the agency Abhi Chawla founded, used to make hundreds of photos a week for hundreds of brands. Across four of its 2026 campaigns, 361 renders gave 45 keepers, about 8 tries per usable photo, and the gummy supplement ads needed only 3.2. Abhi's own businesses make their images with Overs today. It makes stills and writes a motion prompt per photo for Seedance or Veo, so approved product frames go straight into your video tool. Overs' guide to why AI changes your product lists seven fixes in order.

Overs is a sister product to AI Film Contests, from the same founder. You see the shot plan and the estimated cost before anything renders, and the free plan covers 40 photos a month on your own OpenRouter key, with no markup. Try Overs free at www.overs.studio with the product from your next spec spot.