Amazon Product Image Requirements: Why AI Photos Get Rejected

Here is a photo a supplier sent for a canvas tote bag: five bags of different colors hanging off a clothing rack, with somebody's denim jacket in the frame.

You cannot sell from that. So you do what everyone does now — you ask an image model for a proper studio shot, and about a minute later you get one. Clean, evenly lit, no rack, no jacket. It looks like it cost a photographer three hundred dollars.

It also would not have passed Amazon's check, and the second one we generated was not even the right product.

That is the part nobody tells you about AI product photography. Making a good-looking image is the easy half and it is solved. The hard half is making an image of your product that a machine will accept — and those are two different problems.

AI photo generators draw a product like yours, not yours

We ran this on a real dropshipping catalog. One of the products is an electric cat toy: the supplier's photo shows an orange plush crayfish, ribbed body, black stitched eyes, fabric tag on the side.

Asked for a studio shot of it, the model produced a pink plush shrimp. Smiling. Standing on two little plastic legs with feet, which the real toy does not have.

Left, the toy that ships. Right, what came back when we asked for a studio shot of it.

It is a lovely image, and an advert for a product that will never arrive. The buyer opens the parcel, finds something orange with no legs, and files a not-as-described claim — on Amazon a policy matter, not an edit.

The same happened to a phone stand. The real one folds flat and turns on a round base; that rotating disc is the selling point and it is in the product's name. The generated version was a single solid curve with no hinge and no disc.

This is not a bad model or a bad prompt. Text-to-image does what it is asked: it draws a thing matching the words. Your product's identity lives in the supplier's photo, and a fresh generation never looks at it.

Amazon's product image requirements are numbers, not opinions

Now the second failure, which is sneakier, because these images do show a plausible product.

Amazon's main image rule is not a matter of taste. The background must be pure white — RGB 255, 255, 255 — the product must fill 85–100% of the frame, and only what the buyer actually receives may appear.

We measured the studio shots the model returned. Not one contained a single pure-white pixel: the backgrounds came back as light greys, which look white on a monitor and are not white to a validator. The product filled roughly half the frame in two of the three shots, against the 85% Amazon asks for. And the phone-stand image had a phone in it — a prop that does not come in the box, on the one image where props are banned.

The near-white failure is worth understanding, because it is the one that will keep happening. An image model is drawing a photograph, and photographs of white backdrops are never pure white — they have falloff, a warm side, a shadow gradient. The model is being faithful to what studio lighting looks like. The validator is not looking at a photograph; it is sampling pixels for one exact value. Both are behaving correctly, and the image is rejected.

eBay's rules are looser and still mechanical: at least 500 pixels on the long side, no borders, no added text or artwork. A "designed" AI shot with a caption baked in trips that without anyone noticing until the listing is blocked. So the same file can be fine on one channel and blocked on the other — which is the same trap that catches supplier titles on three different channels: the field looks fine to a person and is checked by a machine.

So a photo can be beautiful, honest, correctly lit — and still fail, on arithmetic. Which is good news, because arithmetic is checkable before you upload.

Where each tool actually belongs

The main image stays tied to the real product. Take the supplier's photograph, cut the background out, put the product on true white, crop until it fills the frame. That is a mechanical operation on a real photo, and it is the only route that guarantees the buyer sees what ships.

Lifestyle shots are where generation is worth the credits — with one check. We asked for the tote bag on a café table and got a lovely image: morning light, wooden table, coffee, open book. The kind of second image a small seller never has.

Look at the top of the bag in each: the drawstring the supplier sells is missing from the generated one.

Look at the bags, though. The supplier's are drawstring bags — that gathered, cinched top is the product. The generated one is a plain open tote. Beautiful photo, wrong bag, again.

So the check for a lifestyle shot is the same, just less strict: does the product in it still have the features you are selling? If not, either the scene hides that detail, or you are back to the real photograph.

A camera is still the answer when the supplier's photos are too poor to cut out, and whenever the product carries printed text. Lettering is the thing that improved most and is still the thing to check: we asked for a retail box with a brand name and some small print, and the brand name came out perfectly.

The headline is real type. Enlarge the line under it and the letters stop being letters.

The small print did not. Magnified, it is a convincing texture of letterforms that spell nothing — and the small print is exactly what a cautious buyer zooms in on. Real branding with nonsense under it reads as counterfeit, which is worse than no packaging shot at all.

All of the above came out of one conversation. SellerClaw is a team of AI agents connected to the stores you already sell on, plus your suppliers, that you talk to in a chat; asked in one sentence for two photos each for three products, it wrote its own prompts from the catalog, because it can see the products — the supplier's photograph, the title, the attributes that say what the thing actually is. A generator on its own starts from whatever you can type into a box, and that is exactly where the product's identity goes missing.

What to take away

Before any generated image goes on a listing, two questions in this order. Is this my product — not a product like mine? And does it pass the numbers: true white, filling the frame, nothing in shot that is not in the box?

The first is why the main image should come from a real photograph. The second is why the pretty one you just generated might still be rejected.

To see the split for yourself, take one product with weak supplier photos, ask SellerClaw for a lifestyle shot, and put it beside the original. That is a couple of minutes, the free credits cover it, and you will know after one product which half of the job you have actually automated.