One named person per image, usually the art director or creative director, with the brand and legal checks done by the people who own them. Write down who approved what, and why.

Can an AI model do the QA?

A second model can compare an image with its reference and flag drift, and some tools run that check before you see the image. Treat it as a first filter. The claims, likeness and disclosure checks need a person who knows the product and the market.

Do I check every render, or only the ones I keep?

Only the ones you mean to use, but check them as a set. An image can pass on its own and fail next to the others, because the light moved or the product grew.

A QA checklist for AI images before they go live

Checked September 24, 20267 min read

Short answer

Check every AI image against the real product and the brief before it ships. Work through seven groups in order: product accuracy (label, logo, color, shape, count, parts), people (hands, teeth, eyes, jewelry, skin), physics (shadows, reflections, contact, liquids), text, brand, legal (claims, other brands’ marks, likeness, AI labels) and the file itself. Look at 100% zoom, with the reference photo open beside the image.

How to run the check

Open the image at 100% zoom next to the reference photos it was made from. Judge the product against the photo of the product, never against your memory of it.
Check the product first. A wrong label fails the image however good the light is, and there is no point grading the light on an image that has to be remade.
Work down the groups in order and stop at the first fail. Write the fail as a fix someone can act on: "the cap is white in the reference and grey here".
Before sign-off, look at the whole set together at thumbnail size. Light, color and product size can drift from one image to the next while each image passes alone.
Record the verdict, the reason, who decided, and the prompt, model and references behind the image. See how to review AI images as a team.

Product accuracy

Label text matches the printed label letter by letter, including small print, numbers and symbols.
The logo has the right shape, spacing, color and position.
Product colors match the reference. Spot-check with an eyedropper against the brand’s hex codes.
Proportions match: height to width, cap or closure size, the shoulder and neck of a bottle, the toe shape of a shoe.
Counts are right: buttons, eyelets, stripes, pieces in a pack, and anything the pack states as a number.
Every part is present and in place: zips, pockets, straps, seams, hinges, pumps, caps.
Nothing was added: no extra pocket, no second logo, no pattern the product does not have.
Material and finish match: matte stays matte, frosted glass stays frosted, brushed metal stays brushed.
The size looks right next to hands, people and everyday objects.
Back, side and detail views show the real back, side and detail, and each was made from its own reference photo.

When the product comes out wrong, check the reference photos first. The shot needs the right view attached, and only that view. Why AI changes your product covers the other fixes.

People

Each hand has five fingers, the knuckles bend the right way, and the grip could hold the product.
Teeth are the right size and number, with no fused or doubled teeth.
Both eyes look in the same direction, the pupils match, and the catchlights sit where the key light is.
Earrings, cufflinks, shoes and other pairs match left and right.
Skin shows texture, such as pores and fine lines, with no waxy or plastic sheen.
Hair edges are clean against the background, with no smears or halos.
The face matches the approved character sheet in every image of the set.
Wardrobe matches the brief and stays the same between shots unless the plan says it changes.
Limbs connect, bodies have believable proportions, and nobody has an extra arm behind a prop.

Faces drift between shots unless every shot starts from the same approved person. See how to keep the same model in every AI photo and how to get hands right.

Physics

Every shadow falls away from the key light, and there is one shadow per light.
The product has a contact shadow where it meets the surface, so it sits instead of floating.
Reflections show what the surface would reflect, at the right angle, and bend correctly on curved surfaces.
Glass shows what is behind it, bent the way glass bends it, and the top of a liquid at rest is level.
Pours and splashes fall downward and connect to where they came from.
Condensation sits on cold surfaces only, and drops run down.
The product and the table share one horizon and one perspective.
Stacks could stand, and anything resting is resting on something.

Text

Every word is spelled right at 100% zoom, including the brand and product names.
Prices, dates, percentages and legal lines match the approved copy exactly.
No stray letters or fake writing appear on props, signs, background packaging or fabric.
Type sits inside the placement’s safe zone and can be read at the size people will see it.
The key message also appears in the ad copy or the alt text.

The safest route is to render the photo with no words and set the type in a second pass. See how to get text right in AI images.

Brand

The palette stays inside the brand colors, and accent colors are no bigger than the take allows.
The light matches the rest of the set: same direction, same hardness.
Grain, contrast and color grade match the rest of the set.
Nothing on the brand’s never-do list appears.
Casting fits the brief and the brand’s real customers.
No other brand’s product, logo or packaging appears, props included.

Legal and disclosure

The image shows what the customer gets: size, color, contents, and what is in the box. The FTC judges an ad as a whole, "words, phrases, and pictures", and ads must be truthful and not misleading.
Anything the image implies can be backed up. An ingredient on the plate, a result on the skin or a use in the scene is a claim, and the FTC expects evidence for claims.
No real person’s likeness is used without permission, and no face resembles a celebrity. See AI images of real people.
No AI-generated person is presented as a real customer giving a review or testimonial. The FTC’s rule on fake reviews, in force since October 21, 2024, covers reviews that pretend to come from someone who does not exist, though the FTC says it does not ban AI avatars in marketing outright.
No other company’s trademark appears without clearance. See logos and trademarks in AI images.
Platform AI labels are applied where required. TikTok’s ad policy, updated April 2026, requires its AI-generated content label or your own clear disclaimer on ads with AI-generated imagery, and rejects or restricts ads that lack one; background removal and color fixes count as minor edits. Meta requires disclosure on political and social issue ads with photorealistic AI images. See do you need to label AI images in ads.
AI metadata stays in the file where a platform needs it. Google Merchant Center says images made with generative AI must carry metadata saying so, such as the IPTC digital source type "trainedAlgorithmicMedia", and must not have it stripped.
For images shown in the EU, the AI Act’s transparency rules (Article 50) have applied since August 2, 2026. Check with counsel whether an image counts as a "deep fake" under the Act, which covers realistic images that resemble existing persons, objects or places.

The file

The pixel size meets the placement. Meta recommends 1440 x 1800 for a 4:5 feed ad, and Amazon prefers images over 1,000 pixels on each side so shoppers can zoom.
The aspect ratio is exact for the placement. Platforms crop or reject images outside their range.
The file type and weight fit the platform. Google Ads caps uploaded image ads at 150 KB; Shopify takes images under 20 MB.
Product listing images carry no watermark, border or promotional overlay. Google Merchant Center does not allow them.
No artifacts at 100%: banding in smooth gradients, halos around cut-out edges, repeated patches of texture, smeared fine detail, seams where the frame was extended.
Upscaled files are checked at 100% on labels, logos and faces, where upscalers invent detail.
The file name matches the shot list: shot number, deliverable, size and version.

Questions people also ask

Who should sign off?: One named person per image, usually the art director or creative director, with the brand and legal checks done by the people who own them. Write down who approved what, and why.
Can an AI model do the QA?: A second model can compare an image with its reference and flag drift, and some tools run that check before you see the image. Treat it as a first filter. The claims, likeness and disclosure checks need a person who knows the product and the market.
Do I check every render, or only the ones I keep?: Only the ones you mean to use, but check them as a set. An image can pass on its own and fail next to the others, because the light moved or the product grew.

Where Overs fits

On the Pro and Team plans, an Overs reviewer checks each photo against its reference pictures before you see it, and can retry on its own up to a limit you set. A person still approves, asks for changes or rejects with a reason, and those decisions feed later photos.

Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.

Start free How it works

Sources

Checked on September 24, 2026. Prices, specs and rules change; follow the links for the current versions.