Run a winning ad concept through 50 generated variants and you'll usually find the same bug: the bottle's proportions shift half a size somewhere in the middle of the batch, the label font looks a shade off in another, a necklace that reads gold in cut three reads silver in cut twelve. None of it looks broken on its own. All of it is obvious to anyone who has actually held the product. I dug into why this happens and how to stop it in a longer writeup on DevToolLab, and the short version is that it's not a rare glitch, it's the default outcome of generating each variant independently with no shared reference to return to.
Product consistency is a harder bar than character consistency
A face can drift a little across scenes and a viewer still reads it as "the same person," because human identity perception has real tolerance built in. A product gets none of that slack. Packaging text, exact color, logo placement and material behavior all have to match precisely, not approximately, because most people watching a product ad have already seen the real thing on a shelf or a previous ad. invideo agent handles this by locking a reference at generation time instead of letting every variant reinterpret the product from scratch, but a product reference has to capture something a character sheet doesn't: exactly how the material looks and moves under real light.
Skip the website photo, use the real thing
Where the reference image comes from matters more than people expect. A product photo pulled from a website or marketing page has already been cropped, color-corrected and compressed for the web, and that process strips out the fine texture a model needs to reproduce the item accurately. Real product shots, taken of the physical object from multiple angles under different lighting, hold onto detail that a polished marketing asset already lost.
Establish real scale, and reference every packaging layer
A product reference set needs to do something a character sheet doesn't: establish true scale. A photo of a hand holding the item gives the model a real-world size anchor that a product shot alone can't provide. For anything with layered packaging, a box inside a sleeve, a case around something smaller, each layer needs its own reference photo, or the model tends to flatten the whole thing into one simplified shape. That's exactly the kind of detail people notice in an unboxing-style shot.
Describe the material, not the product name
"A sweater" or "a necklace" tells a model almost nothing about how that object should move or catch light on camera. Swap in physical description instead: "lattice yarn, soft and fuzzy" instead of sweater, "faceted crystal, hard and reflective" instead of necklace. That single change is often the entire difference between fabric that moves naturally and fabric that looks stiff or rubbery in the output.
Lock the first shot, then split the hard cases into two stages
Once you have a solid reference, the first generated shot in a campaign becomes the standard every later variant has to match. This is where a persistent context engine earns its keep: invideo agent holds that locked shot in memory across the whole project, so every later variant inherits the exact look instead of re-deriving the product from nothing. Treating each new variant as a fresh generation, rather than a continuation of that locked reference, is the single most common reason a product drifts across an otherwise solid campaign.
For jewelry and other reflective, highly detailed items, one generation trying to nail both the scene and the exact product look tends to produce something plausible but not quite right. The more reliable route, and the deeper explanation of it is in the original piece, is a two-stage pipeline: build the scene's base aesthetic in one model, then run a dedicated product-locking pass in a second model suited to that job. invideo agent's model routing is built for exactly this split, working across 200+ integrated models including Veo 3.1, Sora 2, Kling 3.0, Seedance, Runway and Nano Banana, with a model like Nano Banana specifically suited to the locking pass.
The mistakes that show up most
- Sourcing the reference from a website image instead of a real photo, which loses the fine detail a model needs.
- Using one flattened shot for a product with multiple packaging layers, so the model collapses the structure.
- Describing materials by product name instead of physical behavior.
- Treating every variant as an independent generation instead of inheriting a locked first shot.
- Asking a single model to both build the scene and lock a reflective, highly detailed product at once.
The takeaway
Treat the product the way a film continuity department treats a hero prop: photographed properly up front, referenced at true scale with every layer accounted for, described in physical terms, and locked at the first shot so later variants inherit it rather than reinvent it. A product that drifts across two variants is a minor annoyance. The same drift repeated across fifty is a campaign that quietly stops looking like it came from one brand.
References
- How to Maintain Product Consistency in AI-Generated Ads and Marketing Videos - the original DevToolLab article, with the full breakdown of reference-set construction and the two-stage pipeline
- invideo agent
- Color Picker - check the exact color value from a reference photo before locking a product's look across variants
- Image Color Extractor - pull the dominant colors straight from a product photo to confirm they match across ad variants
Top comments (0)