An AI image can be beautiful and still be wrong. The jacket has one zip puller where yours has two. The stitching turned from tonal to contrast. The logo on the chest is almost right, which on a product page is the same as wrong. At thumbnail size none of this shows, and that is exactly the problem: the customer finds it when the parcel arrives.

The fashion industry has noticed. In August, Zalando described how it approaches AI content accuracy: an automated layer that compares every generated image with the original product photos, flags changes to colour, construction and brand marks, and leaves the creative call to people. Their rule is simple, and it is ours too: the product itself is non-negotiable.

Most brands cannot build that layer themselves. So we built it into On-Model.

Every job on Pro and Enterprise is now checked automatically. When your images are ready, On-Model compares each one with the photos you uploaded and tells you which ones are worth a second look before you publish. It costs no credits, it never changes an image, and when everything is fine it says nothing at all.

What happens when a job finishes

Nothing changes in how you work. The job completes, your images arrive as always, and a few seconds later a Quality line appears on the job results page.

On-Model job results for a Flat-to-Model job with four polo shirt images, the last one outlined in amber, and a Quality row in the run summary reading 1 of 4 worth a look — Four images, one worth a second look. The rest need nothing from you.

Each image gets one of three outcomes:

Nothing. The image matches your inputs. No badge, no noise.
Worth a look. Something may differ, and a person should decide.
Likely issue. A clear difference from your product, such as a missing detail or the wrong colour.

The results start hidden behind a click, so nobody on your team is forced to see them, and anyone can hide the hints for good. Checks also run again automatically on every new version of an image, so a retouched or repaired image is checked like the original.

It checks your product, not "does it look good"

Generic image scoring asks whether a picture looks realistic. That is the wrong question for a catalog. A realistic image of the wrong jacket is still the wrong jacket. So the check does not grade taste. It compares every output with your own inputs, the photos you uploaded, and looks for where they disagree.

It works at three levels.

This image. The whole frame, against what you asked for: the right model, the shot instruction followed, clean hands and faces, the background you requested, the size you ordered.

Product close-ups. This is where most of the value is. The check finds every product in the output, crops it, and puts it next to the same product in your inputs, including your detail and macro shots. Then it goes through the garment part by part:

Part	What it compares
Shape and cut	Silhouette, fit, proportions
Collar, closure, pockets	Count, position and type
Sleeves, cuffs, hem	Length and finish
Hardware and trims	Zip pullers, buttons, snaps, stitching
Print, logo or text	Placement, spelling, clarity
Colour	Hue and shade against your photo

Going part by part matters. Asked "is this the same jacket?", a model tends to spot one difference and stop. Asked to confirm the collar, then the closure, then the pockets, it finds the second and third.

A quality check close-up comparing the polo shirt in the output with the flat-lay, fabric detail and collar detail photos it was made from, above two findings about the buttons on the placket — The polo in the output, next to the three input photos it was checked against. The placket buttons do not match.

Here is what it caught, up close. In the full-body shot the placket is a tiny strip of the frame, and at a glance the shirt looks right: blue polo, collar, buttons. Zoom in and the placket has two buttons sitting side by side, where your polo has a single button under the collar. It is exactly the kind of detail a tired eye slides past on image 180 of 200, and exactly the kind a customer notices when the parcel arrives.

Across the set. Some problems only show when images sit side by side: the model's face changing between shots, a background drifting from grey to beige, a garment that is slightly longer in image 3. The check compares the whole set and names the images that stand apart.

What the check means for each tool

"Faithful" means something different for each tool, so each one is judged against its own promise.

Tool	What it holds the output to
Model Swap	The new model matches the identity you chose, and everything they wear stays exactly as in your photo.
Flat-to-Model	Your flat-lay or packshot garments appear on the model with their real colour, cut, prints and trims.
Create Packshot	The product is the same product, presented in the style you picked.
Garment Recolor	The garment reaches the colour you asked for, and nothing else about it changes.
Detail Repair	The fix does what you asked. If you asked to remove a logo, a clean removal counts as success, not as a missing logo.

From a flag to a fix

A flag is only useful if you can act on it. Every finding says three things: what it saw, why it matters, and what to do, usually "regenerate this image" or a pointer to Detail Repair.

Open the job in Review and each image has a collapsed Quality check card with the evidence, including the close-up pairs, so you can judge for yourself instead of taking our word for it. From there, one click on Add to review turns a finding into a review flag, with the right category selected, the note filled in and the product marked on the image. It is the same review and retouch workflow described in Quality Control for AI Product Images, with the searching already done.

The On-Model review screen with the product outlined on the image, the Quality check card listing two findings, and a Garment Distortion issue pre-filled below after clicking Add to review — One click on Add to review: the category is chosen, the note is written and the product is marked.

And when the check is wrong, you say so. Dismiss a finding as not an issue, acceptable, or in the wrong spot, and it disappears for everyone working on that job.

Built to stay quiet

A checker that cries wolf is worse than no checker. The second time you dismiss a warning that turned out to be nothing, you stop reading warnings. So the check is tuned to speak only when it has something to show, and we measure it against the verdicts of human reviewers.

It is also repeatable. Check the same image twice and you get the same answer, so a result you saw yesterday is the result your colleague sees today.

The quality check only reads your images. It never edits them, never regenerates them, never holds a job back and never decides for you. A person always makes the final call.

It is honest about what it cannot see, too. Some details are smaller than the image can show: a collar tab on a full-body shot at 1K is a handful of pixels. Rather than guess, the check tells you the detail was too small to verify and suggests generating at 2K or 4K, where there is enough detail to check it properly. Very small parts are still at the edge of what any vision model can judge reliably, which is why it flags and a person decides, and never the other way round.

Before, after, and people in between

This is the second half of a pair. The readiness check reads your inputs before a job starts and tells you what will not work for the tool you picked. The quality check reads the outputs after the job ends and tells you what did not come out right. Between them sits the part that should stay human: the creative judgement of whether an image is on-brand, and the approval to publish it.

That is what we mean by infrastructure for fashion imagery: not a single clever generation, but a pipeline you can trust at the scale of a catalog, where every image is produced, checked against the real product and approved, in the same place.

Common questions

Does the quality check cost credits? No. It is included on Pro and Enterprise plans and never touches your credit balance, however many images a job has.

Which plans include it? Pro and Enterprise. On other plans you can see where the check would appear on your results, and upgrade to turn it on.

Which tools are checked? Model Swap, Flat-to-Model, Create Packshot, Garment Recolor and Detail Repair. Each is judged against what that tool promises.

Does it change my images? No. It only reads them. Your images, your job and your credits are untouched whatever the result.

What does it compare my images with? With your own inputs: the product photos, detail shots and references you uploaded, plus the identity and instructions you chose. Not with a generic idea of a good photo.

What if it flags something that is fine? Dismiss the finding. It is removed for everyone working on the job.

Can it catch every defect? No, and it does not claim to. It is very good at colour, prints, logos, missing or added parts and set consistency. Tiny details on small images are harder, and there it tells you it could not verify rather than guessing. It points your reviewers at the images that need them; it does not replace them.

Does it slow my job down? No. Your images are delivered as soon as they are ready, and the check runs afterwards. Results usually appear within a minute.

Try it now

Open a finished job in On-Model and look for the Quality line on the results page. If you are on Pro or Enterprise, your next job is checked automatically. Not yet? See pricing for what each plan includes.

New to On-Model? Start with the Flat-to-Model guide, the Create-Packshot guide, or the Model-Swap guide.