Why does human quality control still matter in AI image production for ecommerce?

Why can't AI product images go straight to your store?

An image model doesn't copy your product. It generates a new picture that matches the patterns it has learned, guided by your prompt and reference images. Most of the time the result looks right at a glance. The problem is the details that only look right at a glance.

In ecommerce the image is part of the product description. A customer who orders the shade in the picture expects that shade in the box. If the label reads differently, the cap is a different shape or the bottle looks bigger than it is, that is a return, a complaint or a bad review. For a beauty brand it can also be a compliance problem.

So the question isn't whether AI images are good enough. Many are excellent. It is whether each specific image is accurate, and that needs someone to look.

What does AI typically get wrong on beauty products?

The same handful of errors comes up again and again on cosmetics, skincare and fragrance. Knowing them makes checking faster.

Our guides to whether AI will change your product and realistic skin and texture in AI images go deeper on the two most common problems.

What do marketplaces and ad platforms expect?

Platforms already ask for accuracy, and some now ask about AI directly. Google Merchant Center's image requirements, at the time of writing, say your main image must clearly show the exact item being sold, and that each variant needs its own image matching its colour, finish and material. They also say images made with generative AI must carry metadata that marks them as AI generated, such as the IPTC digital source type, and that you shouldn't remove that metadata.

Advertising rules add a second test. In the UK, the ASA has ruled against beauty ads where filters exaggerated what a product does, and it has said a "filter used" style disclaimer is unlikely to fix that. The same reasoning applies to AI. If an image shows a result, the result must be real and typical. Read do you have to label AI ads for the disclosure side. This is general information, not legal advice.

Each marketplace sets its own image rules and they change, so check the current requirements for every channel you sell on before a batch goes live.

What does a proper human quality check look like?

A real check is a fixed routine, not a quick look. The studio checks every frame against the real pack before delivery, and a useful in-house routine follows the same order:

• Open the real packshot next to the AI image at the same size and zoom to 100 percent.
• Read every word on the pack. Don't skim. Invented letters are easy to miss.
• Check the silhouette: cap, shoulder, base and any pump or dropper.
• Check colour on a calibrated screen against the physical product or an approved reference.
• Look for anything that shouldn't be there, such as extra products, odd reflections or a second label.
• If a person appears, check skin, hands and hair for anything that looks artificial or improved.
• Ask whether the image implies a result. If it does, confirm the result is real and typical.
• Check file size, crop and format against the channel's spec, and keep any AI metadata in place.

Record the result. A simple sheet with the image name, checker, date and any fixes made is enough. It shows who signed off what if a question comes up later.

Can automated checks replace a person?

Not yet, at least not for final sign off. Software can help. It can flag missing metadata, wrong dimensions, background colour problems and some text errors. It can also compare colour values against a reference, which is useful for large shade ranges.

What it struggles with is judgement. Is this shade close enough for the customer to recognise? Does the skin look like skin? Would a buyer read this scene as a claim about the product? Those calls depend on knowing the product, the brand and the rules, and they are the ones that cause returns or complaints when they go wrong.

The practical setup is automated checks first, to catch the obvious, then a person on every image before it is published.

Who should do the checking?

Ideally two people at two stages. The studio or team making the images checks for accuracy as part of production. Then someone on the brand side who knows the product signs off before anything goes live. For shade heavy ranges, that should be whoever owns colour approval for packaging and swatches.

Claims need their own pair of eyes. If an image could be read as showing a result, the person who approves claims for your ads should see it too. It is much cheaper to change an image before launch than to pull an ad after a complaint.

When you brief a studio, ask exactly how they check images and what they compare them against. Our guide on how to choose an AI product photography agency lists the questions to ask.

How do you build quality control into an AI image workflow?

Start with good source material. Sharp, accurate packshots from every angle make every later image easier to check and easier to get right. Our guide on what photos to send explains what helps.

Then build the checks into the schedule rather than adding them at the end. A few habits help:

• Keep a master reference folder per product, with approved packshots, swatches and pack copy.
• Agree what counts as a fail before production starts, for example any text error or visible shade shift.
• Allow time for fixes. Some images will need a second pass, and that is normal.
• Check the first images of a new product slowly. Once the setup is proven, later batches go faster.
• Recheck when a pack changes. An old reference will pass an image that is now wrong.

If you'd rather not run this yourself, our AI product photography service works from your real packshots and checks every frame against the pack. The brand owns the output.

Questions people also ask