How-To Guide
How do you check that an AI-generated image of your product still shows the real product accurately before you post it?
A pre-post accuracy checklist for AI product images: reference photos, attribute comparison, scene-claim test, correction moves and platform rules to read.
By the Superdirector editorial team. Google Merchant Center help, two FTC pages, two Adobe Firefly pages and two practitioner guides (BattleBridge; Visualsclipping's Amazon guide dated July 1, 2026) read November 14, 2026.
At a glance
How do you check that an AI-generated image of your product still shows the real product accurately before you post it?
Treat the AI image as a suspect copy of your product until a person has compared it with the approved reference. Set the reference pack first: approved photos of the exact SKU and variant from several angles, plus packaging. Generate from that reference image instead of a written description alone, because Adobe's own guidance says to anchor the product's look with image-to-image before you generate. Then put the result beside the reference at full resolution and check size and proportions, colour and finish, logo and lettering, parts and construction, texture and material, and packaging, the attributes Google's listing rules summarise when it asks photos to accurately reflect the exact version, color, and finish being sold. Finally test what the scene claims, correct what fails, and read each destination's own image rules before posting.
- The attribute check that finds most errors, in order: size and proportions, colour and finish, logo and lettering, parts and construction, texture and material, packaging.
- Two rules worth memorising from the sources: Google requires generative-AI listing images to carry AI-creation metadata, and the FTC judges the ad's overall impression, so a scene that shows performance the product cannot deliver is a problem no label cures.
Six attributes marked one row at a time against the approved reference, the generated column carrying the marks and one accent on a worked lettering failure. Original diagram from the article's own six-attribute list and Google's exact-version, colour and finish wording.
Make it work for your brand
Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.
Set the reference photos for the exact product
Start with the real thing. Collect approved photos of the exact SKU and the exact variant you plan to post: several angles at high resolution, the packaging, and the spec sheet if the product has measured details worth protecting. Name the variant in your notes before generating, because colourways and finishes are where accuracy quietly dies. Google's variant rule states the goal precisely: the shot must ensure the photo accurately reflects the exact color, pattern, finish, and material the shopper is selecting.
The reference pack is doing two jobs later. It anchors generation, and it is the yardstick for the comparison in the next section. Adobe's brand-build guide recommends uploading product photos from multiple angles and contexts so its assistant can make better preset recommendations, and that same multi-angle advice is what you will want when checking a generated output for drifted proportions or invented details.
Generate from the reference image, not a description alone
A text prompt alone asks the model to remember your product from training data it may not hold. The safer generation pattern comes from Adobe's product-shot workflow: upload a reference image and anchor the product's aesthetic with image-to-image before you generate, and use the prompt for what the reference cannot carry, the angle, lighting, background and mood. Adobe's own accuracy line stops short of a guarantee, which is worth internalising: the more specific your description, the more accurate your product shot will be. Specificity improves the odds; it never verifies the output.
Keep the generation provenance: which reference was attached, which model and settings ran, and the prompt. Adobe is candid that you are responsible for determining whether content generated with partner models fits your project requirements, and that accountability is the reason the provenance record matters when someone asks why the published image looked like that.
Compare attribute by attribute at full resolution
Put the generated image and the reference photos side by side at similar scale, zoomed to full resolution, and work a fixed list. Six attributes catch the errors that matter:
- Size and proportions: overall shape, thickness, and scale against any hands, props or people in frame. Perspective and cropping should not imply a bigger or smaller product than you sell.
- Colour and finish: body colour, accents, gloss or matte surface. Lighting changes everything, so compare against the reference rather than your memory of it.
- Logo and lettering: spelling, orientation, relative size and placement. Regenerated lettering is a classic AI artefact and the wrongest thing to publish; swap in approved artwork rather than trusting the render.
- Parts and construction: buttons, ports, seams, lids and handles. Check that nothing is added, removed, duplicated or fused.
- Texture and material: fabric grain, stitching, surface pattern, transparency. Look for the AI-smoothed surface look that marketplace reviewers summarise as unnatural proportions and excess glow.
- Packaging: if the pack appears, it must show multidimensional packaging and not just the front face, per Google's packaging note.
Google's listing standard gives this comparison its target wording: shoppers expect the image they see in ads and free listings to accurately reflect the exact version, color, and finish they are selecting. A practitioner checklist for Amazon images compresses the same standard into one line worth adopting: images depict true color and true proportions. For listings rather than feeds, Google also frames the composition window: the product should take up no less than 75%, but not more than 90%, of the full image.
Mark every attribute pass or fail with the reference file name as evidence. Failures move to the correction step; nothing ships until each row is either fixed or the image is withdrawn.
Test the scene for claims the product cannot support
A pixel-accurate product in an untrue scene is still a misrepresentation. The FTC's deception test asks two questions of any advertisement: is a statement or omission likely to mislead consumers acting reasonably under the circumstances, and is it material, meaning important to a consumer's decision to buy or use the product. The FTC's own illustration of omission is a book collection advertised without disclosing that buyers actually receive abridged versions. Translated to product imagery, an image can overstate what a buyer will hold.
BattleBridge, which advises on AI-made advertising, names the image version of this risk: a generated product demonstration that depicts performance the product cannot deliver, listed among the cases where the depiction itself is the material fact a consumer needs. Its decision table puts the duty beside every generated image, independent of any label: whether the image misrepresents the product or setting.
Run the test as a question per scene: what will a viewer now believe this product does, looks like, or comes with? Any belief the product does not support, durability in a storm, a waterproofing the spec sheet does not claim, an accessory not in the box, is either restaged truthfully or dropped.
Correct the miss: regenerate, restage or label
One of three corrections applies to each failed attribute:
- Regenerate with the reference attached. For graphic misses such as lettering, wrong artwork or a stray prop, regenerate or edit with the reference photo in hand, then rerun the whole attribute list, because edits can quietly shift colour and proportions.
- Restage or use a real photo. When the scene claims more than the product delivers, correct the story rather than the pixels: restage inside what the spec sheet supports or substitute a real photograph of the product in use.
- Keep only the labels the destination requires, positioned where the reader will see them. Google Merchant Center requires AI-created listing images to carry metadata marking the image as AI-generated, and naming the IPTC DigitalSourceTypeTrainedAlgorithmicMedia tag. Where a disclosure is owed, BattleBridge's standard is the right shape: it should appear close to the relevant claim, in understandable language, and in a format people will actually notice. A label does not make a false depiction true; it only names what is there.
Do not ship a placeholder while a real image is pending, either. Google explicitly bans any placeholder, generic image, or graphic that does not show the product as a listing image, which is the same principle: what the shopper sees must be the thing being sold.
One failed item forks into the three corrections, with the restage branch marked as the fix that needs no new generation. Original diagram from the article's correction section and the Google Merchant Center metadata requirement.
Check the rules of where you publish
Feeds are forgiving, marketplaces are not. Google Merchant Center's image requirements demand an unobstructed, clean view of the item for sale, ban overlaid calls to action, watermarks and post-production logos, require AI-created listing images to carry their metadata tag, and from January 31, 2027 enforces a 500 x 500 pixel minimum for all products. Note also that it may test subtle cropping on your image, which is why a centered composition survives platform resizing.
Amazon's rules, as summarised by the practitioner guide Visualsclipping kept current to July 1, 2026, keep the main image text-free and on a pure white background at 85 percent of the frame, push text and graphics to secondary images, and can suppress a listing when a render looks too fake.
Social feeds write fewer product-image rules but still judge deception, which brings you back to the FTC standard that governs them: advertising must be truthful and non-deceptive, and advertisers must have evidence to back up their claims. Read each channel's current page rather than assuming yesterday's rules; image policies move.
Record the person who signs off
The check is not finished when the image looks right. Store the reference pack, the generations, the attribute table with its passes and failures, the corrections applied, and the name of the person who signed off. BattleBridge states the underlying duty in one line: AI changes the speed and volume of advertising, and it does not change who is accountable for the result. A model cannot weigh the evidence or accept the consequence; a named person does, and the record is what makes the approval auditable later.
This is also where the FTC's review rule touches you if imagery meets testimonials. When AI imagery or video shows a person presenting a customer experience, the Consumer Reviews and Testimonials Rule, in effect since October 21, 2024, addresses deceptive and unfair conduct involving consumer reviews and testimonials, and a fake experience stays fake however polished the pixels are.
If you want the publishing half of the workflow handled rather than just the checking, Superdirector (ours) makes branded carousels and X posts and publishes them, at $199 a month for an AI-run account. Its declared rule matches this article's: real product screens stay real, because the image model never sees one.