How do I make sure AI product photos show my actual product?

This question came up on a sales call this year. The owner had paid an agency before me, and its AI images showed the product at the wrong dimensions. When the owner pointed it out, they were told nobody would look that close.

Customers look that close. So do the people who make the product. Everything below is the check I run before a render goes anywhere, and each step exists because something got past me or past a check I trusted.

Start with references you’ve confirmed

A render can only be as accurate as the photos you feed the model. I assumed mine were clean. They weren’t.

On one catalog I render for, I audited the reference photos the pipeline was using. On 54% of the products, at least one image labelled as an angle of that product was a photo of a different product. The store’s product pages show related items, and the image scrape picked those photos up with the same labels as the real ones. Each one quietly told the model to draw something else.

It had been happening for a while. One product carried a hand-written note telling the model not to render a feature that product doesn’t have. That feature was exactly what its contaminating photo showed.

The fix was to keep only images whose file name matches the product’s own front photo. About one in five products dropped to a single trusted photo after that. One correct photo beats three where one of them is the wrong product.

What to do: before any render, open every reference image for that product and confirm it’s the product. If a vendor can’t tell you where their references came from, they haven’t checked.

Automate what you can measure, and make it fail closed

Some defects are measurable. Overall proportions, which side a feature sits on, whether something is mirrored, how many of a repeated element appear. I write those as automated checks against a spec file that holds the product’s real geometry.

Two lessons cost me real work here.

A check has to fail when it can’t answer. On the same catalog, an automated vision check asked whether a render had a feature most of the products don’t have. It returned “no” whenever the QA model errored out. Since “no” was the passing answer for nearly every product, an overloaded model didn’t stall the gate. It passed it. A render with the defect plainly visible passed five runs out of five, and every one of those answers turned out to be a server error. With retries added, the same question caught that render three times out of three. An error now counts as a failed check.

Test a new check on a known-bad image first. The first version of one check passed the exact render it was written to catch. It was averaging in pixels from a neighbouring feature. I now keep a stress test that takes a known-good render and breaks it on purpose: mirrored, flipped, squashed, stretched, tapered, a hole filled in, a hole made too big, an empty frame, a second part in frame. Fourteen deliberate defects, and each one has to trip the right check. I re-ran it on 15 October 2026:

Defect introduced on purpose	What the gate has to do	Result
None (the clean render)	Pass it	Passed
Image mirrored	Catch the feature on the wrong side	Caught
Mounting slot flipped	Catch the slot direction	Caught
Part squashed	Catch the length-to-width ratio	Caught
Part stretched	Catch the length-to-width ratio	Caught
Part tapered	Catch the width changing along its length	Caught
Hole filled in	Catch hole size and shape	Caught
Hole made too big	Catch hole size and shape	Caught
Empty frame	Refuse to grade nothing	Caught
Low-contrast background	Still find the part	Found it
Second part in frame	Still find the part	Found it
Background the same color as the part	Fail rather than pass blind	Failed it
A real earlier render with the slot on the wrong side	Catch it	Caught
Wide scene shot from an angle	Hand angle-dependent checks to a person	Handed off

All 14 behaved correctly. The last row matters as much as the others. In a scene shot from an angle, perspective makes outline measurements unreliable, and on its first run the gate confidently reported a slot on the wrong edge purely because of the camera. A check that’s confidently wrong is worse than one that admits it can’t tell.

On the catalog with the contaminated references, the three vision checks that run on every render each take three votes, fail closed, and were validated 15 out of 15 against a labelled test set that includes a known-bad image.

The side-by-side is the part that can’t be skipped

Automated checks measure. They can’t judge shape.

On a small hardware part, a render passed all three of the automated checks I had at the time and went live. It was wrong in three places none of them looked at: a curved lip rendered flat, a mounting hole drawn too large and too low, and a fastener standing proud of the surface instead of seated in it. Later that day a mounting slot came out flipped. That one wasn’t cosmetic. With the slot upside down, the part can’t hang.

Every one of those was obvious the moment the render sat next to a real photo at the same scale.

I tried to automate the curved-lip check anyway, and the numbers are why I stopped. Measured from the outline, the good render scored 0.05, the rejected render scored -0.01, and a photo of the real product scored 0.01. The real thing looked like the bad render, because the lip points toward the camera and barely shows in an outline. That check would have failed the actual product, so I didn’t ship it. The gate prints its own blind spots on every run instead, and the side-by-side stays mandatory.

How I run the side-by-side:

• Render and real photo on one sheet, at matched scale, with gridlines.
• Zoom in. On another job a render added a third wall light that the source scene doesn’t have. At thumbnail size it was invisible. A cropped, upscaled band with gridlines at 25%, 50% and 75% showed it immediately.
• Count discrete things out loud: holes, handles, drawers, lights. Judging the overall look is how extra parts get through.

Checks I built and threw out

Most of what I’ve built to check renders didn’t survive testing. I’m publishing the failures because they’re the reason the side-by-side stays manual.

Check	What it was meant to catch	Why I pulled it
Lip profile from the outline	A curved lip rendered flat	The real product scored 0.01, the rejected render -0.01, the good render 0.05. It would have failed the real product.
Camera angle	Renders shot from the wrong angle	Two of five test images graded wrong and a third flipped between runs. On one product the question compared the front with a side panel the product doesn’t have, so no answer could be right.
Color cast from the wall	Renders that came out too warm	One render read as a strong warm cast. The wall in that scene is specified as a warm taupe, so the check was measuring the room’s own color. On a surface that’s actually white, that render was the most neutral of the five.
“Has this feature?” vision check, original version	A feature most products don’t have	Returned a pass on every server error, so an overloaded model passed everything. Rebuilt to fail closed.

The pattern is the same each time. A number that seems to measure the defect is really measuring something else: the outline instead of the shape, a side panel that isn’t there, the paint instead of the light. Compute what’s genuinely measurable, and put a person on the rest.

Know how the product is used before you put it in a scene

Getting the part right isn’t enough once it’s in a scene. I once rendered a product in use at the wrong scale, because I had a spec for the part and had never looked at how people use it. The client’s own site had real photos of every step of the job the whole time.

Now I pull real in-use photos first and write scale anchors into the spec, like how big the product is next to a hand or a person. In the prompt I name the geometry with numbers (“exactly eight times longer than it is wide”) instead of describing it (“slender”).

Where a rule sits in the prompt matters too. On one product, four renders in a row failed the gate. Those rules lived about 20,000 characters down, in the guardrails. I moved them to the top, next to the geometry, and the next render passed clean. That’s the one prompt change I’ve measured, so it’s the one I’ll vouch for.

There’s a limit worth knowing before you pay anyone for this. A generative model won’t reliably hold a feature a quarter inch wide on a ten-inch part in a wide scene. Ten rounds of generate-and-correct kept breaking it, and each fix to one detail broke another. Renders are dependable for the studio shot of the product on its own, where the geometry can be checked. For detailed in-use images of small precise parts, the owner’s own photography beats any render, and I ask for it first.

The rules when a render fails

• A failed render gets regenerated, never retouched.
• Never mirror an image to fix a feature on the wrong side. Mirroring moves every other asymmetric detail too.
• A render that fails doesn’t go live “for now.” If one is already live, it comes down.
• Passing every automated check isn’t a pass. The side-by-side decides.

What to ask whoever makes your images

• Can I see the side-by-side sheet for every image, render next to the real product?
• Where did the reference photos come from, and who confirmed each one shows this product?
• What does your automated check do when the QA model errors or can’t tell?
• What do you do when a render fails: regenerate, retouch, or ship it?

If the answers are vague, the images will be too. For when AI imagery is worth using at all, see whether AI product photography is good enough for a Shopify store.

When to bring me in

Bring me in when most of your catalog has no lifestyle imagery, when shoots can’t keep up with new products, or when ads run on flat packshots.

I'm Conner Crowe, and I write every audit myself. Send the free catalog audit request with your store URL. I read the public store, so there’s nothing to grant, and within 24 hours of the request you'll have a written read on which pages need imagery first. No call needed, and nothing in your account or store changes.

Request the free catalog audit →

Not a home or furniture store? Get a free ads audit instead