Quality Control for AI Product Images

A batch of 200 product images comes back. Most of them are good enough to publish today. A handful are not: one hand reads wrong, one shadow falls in the wrong direction, one brand patch came out soft. You cannot ship a catalog where one image in twenty has something visibly off, and you cannot personally open all 200 either.

That gap, between "generated" and "safe to publish", is the last mile. It is not a model problem. It is a workflow problem, and it is the part most AI imagery tools leave to you.

What the last mile actually is

Generation gets you a set of candidates. Publishing needs a decision on every single one: is this good enough for a product page, or does something need fixing first?

Doing that well requires three things that have nothing to do with image quality:

• A consistent standard. Two people looking at the same image should reach the same verdict.
• A record. Which images were checked, by whom, what was wrong, and what changed.
• A path to the fix. Flagging a problem is only useful if someone can act on it and the corrected file lands back in the same place.

On-Model handles this as a built-in stage after generation rather than a separate tool. Every job can be reviewed, and anything flagged can be retouched and re-reviewed, until the whole set is approved.

Three ways to close it

The review stage is the same in all three cases. What changes is who does the work.

	Who reviews	Best for
Self	You and your team	Small batches, or when the brand standard lives in your head
Shared link	A freelancer, agency, or colleague you invite	You already have a retoucher or a brand approver you trust
Managed	Our vetted QA pool	Volume, or when you would rather not staff the last mile at all

Self-review needs nothing set up. Open a finished job, start reviewing, and decide on each image yourself.

Shared-link review sends a specific person a private link to one job. They do not need an On-Model account, and you choose exactly what they can do: view only, review only, retouch only, or both. That is the cheapest way to bring in a freelance retoucher you already work with, or to get a brand manager to sign off without giving them access to everything.

Managed review and retouch hands the job to our QA pool instead. You request it from the job results page, and the work comes back through the same interface. Managed service is available on Pro and Enterprise plans.

Review and retouch consume credits from the same balance as generation. Self-review is free. See pricing for current rates.

How review works

Reviewing is not editing. A reviewer issues one of two verdicts on each image: Approve, or Needs Retouch. Nothing about the image changes.

When an image is flagged, the reviewer says what is wrong using a fixed set of nine categories: artifact, shadow, logo, skin tone, background, garment distortion, pose, lighting, and other. The categories matter more than they look. They are what makes two different reviewers describe the same defect the same way, and they are what a retoucher reads first.

A reviewer can also paint directly on the image to mark the affected region. That mask carries through to whoever fixes it, so nobody has to hunt for the problem.

The house standard for notes is specific over impressionistic. "Hands look weird" gives a retoucher nothing. "Left hand has 6 fingers; thumb is doubled" tells them exactly what to fix and how to know when they are done.

The quality bar itself is deliberately blunt: if a major retailer would refuse the image on their product page, flag it.

How retouch works

A flagged image goes to whoever holds the retouch permission. They see the reviewer's categories, notes, and marked region, plus the brief the image was generated against, so a fix is measured against what the image was supposed to be.

Starting a retouch downloads a package: the flagged output, every input that produced it, and the reviewer's mask. The fix happens in whatever editor the retoucher already uses. The corrected file is then uploaded back as a new version of that image, and the image returns to the queue for re-review.

Retouching has scope limits, and they exist to protect the brief rather than to restrict the retoucher. In scope: artifacts, shadows, logos, skin tone, backgrounds, garment distortion, pose corrections, uneven lighting. Out of scope: changing the identity, altering the garment's design or colour, recomposing the shot, or "improving" it beyond what was asked for.

There is a third option beyond approve and fix. Some images are not retouchable, because a limb is missing or the garment is unrecognizable. A retoucher can send those back for regeneration instead of trying to rescue them.

Not the same as Detail Repair

Both involve fixing an image, so it is worth separating them. Detail Repair is an automated tool: you mask a region and the platform regenerates just that region. Retouch here is a person opening the file and correcting it by hand. Detail Repair is often what a retoucher reaches for first, but the review workflow does not assume any particular method. It only cares that a corrected version comes back.

Rounds, and when a job is done

Review and retouch alternate in strict rounds. A review round has to reach a verdict on every image before any retouching starts. A retouch round has to resolve every flagged image before re-review opens. Then it repeats.

This sounds bureaucratic and is the opposite. It means there is never a moment where half the set is being reviewed while the other half is being edited underneath, and it means "what state is this batch in" always has one answer.

The job closes when a review round ends with everything approved. The owner gets a summary of what was approved, what needed fixing, and what it cost.

Which one should you use

Use self-review when the batch is small or the standard is genuinely subjective, like a seasonal campaign where you are the only person who knows what "right" looks like.

Use a shared link when you already have the person. A freelance retoucher who knows your products will almost always beat a generalist, and this is the cheapest way to plug them in. It also works for approvals: give a brand manager review-only access and they can sign off without touching anything.

Use managed when the volume makes the last mile a staffing problem. If you are publishing thousands of images a season, the bottleneck stops being generation and becomes the checking.

Most teams end up mixing them. Self-review for quick internal batches, a shared link for the freelancer who handles the tricky categories, managed for the bulk seasonal push.

Common questions

What is image quality control in e-commerce? It is the step between producing an image and publishing it: checking every image against a standard, flagging the ones that fall short, fixing them, and confirming the fix. For AI-generated product imagery it matters more than for photography, because the failure modes are less predictable and a bad frame can look plausible at thumbnail size.

Do I have to review every image myself? No. You can review them yourself, invite a specific reviewer or retoucher by link, or request managed review from our QA pool on Pro and Enterprise plans. The interface is the same in all three cases.

Can I invite someone without giving them an account? Yes. A shared review link works without an On-Model account, is scoped to a single job, and expires. You choose whether the recipient can view, review, retouch, or both.

What is the difference between review and retouch? Review issues a verdict on an image and changes nothing. Retouch fixes a flagged image and uploads a corrected version. A job typically alternates between the two until everything is approved.

What happens to an image that cannot be fixed? It gets sent back for regeneration rather than retouched. That decision sits with the retoucher, who is the one best placed to judge whether a defect is repairable.

Does a retouch overwrite the original? No. A corrected file is stored as a new version of that image, so earlier versions stay available.

Getting started

Open a finished job in On-Model and choose Review. Start with self-review on a batch you have already published, to see how your own standard maps onto the categories. When you want to hand it off, create a share link or request managed service.

If you are the one doing the reviewing, the QA guides cover the process in full, including the quality bar and the annotation standard.

New to On-Model? Start with the Flat-to-Model guide, the Create-Packshot guide, or the Model-Swap guide. If several people work on your imagery, Teams covers shared credits and roles.