Amazon Infographics Without an AI-Generated Product

Sellers ask whether Amazon will ban their AI images. Amazon answers a different question: does the image accurately represent the product being sold? Its image policy is written around accuracy, not production method. A generated background behind your real product photo is unremarkable; an AI-imagined version of your product is a problem, not because a model made it but because the buyer will not receive it. This page maps that line and shows where Graflio sits on it.

What Amazon's rules actually say

The main image carries almost all of the enforceable constraints, and none of them mention AI. It must show the actual product on a pure white background (RGB 255,255,255), filling at least 85% of the frame. No added text, graphics, logos, watermarks or props, and no illustration standing in for a photograph. Files need 500px on the longest side to be accepted and 1000px or more to trigger zoom, with 1600px or more recommended and 10,000px as the ceiling.

Gallery images are governed far more loosely. Text is allowed, and diagrams, comparison tables, dimension callouts and lifestyle scenes are all normal — which is why the infographic set lives in slots two to nine, never slot one. Amazon recommends at least six images plus a video; most listings accept nine, of which seven show by default on desktop.

The substantive rule behind all of it is representational: the image must not show a colour, size, quantity, component or capability the shipped item does not have. That rule pre-dates generative models by two decades, and it is the one they make easiest to break by accident.

The one disclosure rule that now exists

On 22 July 2026 Amazon added a specific AI obligation, narrower than most of the coverage suggested. Sellers must tag listing images, videos and A+ Content containing photorealistic AI-generated people by writing the keyword contains-synthetic-performer into the dc:subject XMP field of the file, using an IPTC-compatible metadata editor, before upload. The trigger was a New York law effective in 2026 requiring advertisers to disclose synthetic performers; Amazon applied it marketplace-wide.

What it excludes matters as much: real people edited with AI, non-photorealistic characters, and images with no people. Background generation, relighting, colour correction and scene composition around a real product need no disclosure. If your set has no human faces or hands in it, this rule never touches you.

Separately, Article 50 of the EU AI Act applies from 2 August 2026 and obliges providers of image-generating systems to mark outputs in a machine-readable format, with systems already on the market given until 2 December 2026. That is a duty on model providers, not a labelling requirement on your listing. Conflating the two is the most common error in current advice here.

Accuracy is a commercial problem before it is a policy one

Suppression is the visible failure; the expensive one is quieter. US retail returns reached roughly $849.9 billion in 2025 on NRF figures, about 15.8% of sales, with online returns near 19.3% — one order in five coming back before anyone mentions imagery. "Did not match the description" sits consistently among the top three stated return reasons, and an Akeneo survey of 1,800 consumers put the share who say they have returned something over incorrect product information at 40% — self-reported, so read it as direction rather than precision.

A returned unit costs outbound shipping, return shipping, inspection and often the unit itself, then leaves a one-star review reading "looks nothing like the photos" at the top of a page you paid to drive traffic to. An image that overstates the product is a loan against your review score.

Where Graflio draws the line, in code

Three paths produce images here, and they treat the product differently.

The set generation path. Research reads your listing and the category best sellers, then writes six or seven image scripts. Each goes to an image model with your listing's own main photograph attached and an explicit instruction: reproduce the product exactly — identical shape, colours, texture and branding. The model composes the whole frame, so the product is faithfully re-rendered rather than pasted. That distinction is real: re-rendering can drift on fine print, small logos and complex textures, and a generated set deserves the scrutiny you would give a designer's first draft.

The layered path. This one composites. A concept is painted, a second pass erases the overlay graphics and leaves only background and product, and that plate is enlarged 2× with Real-ESRGAN — an upscaler, not a generator. A further pass measures the concept against a coordinate grid and returns 20 to 45 positioned objects: text, rules, arrows, badges, magnifier circles, and one photo element that places your actual photo file back on the canvas, optionally with its white background knocked out. The instructions carry a rule in capitals: never redraw the product with shapes. Here the product pixels are your pixels.

The translate path. One feature does not composite, and it is worth saying plainly: translating a finished infographic flattens the canvas and sends the whole image to a generative image model, which redraws it with the text in the new language. The model is instructed to keep the product, layout, colours and photography identical, but the product is re-rendered rather than composited, and what comes back is a flat raster with no layers. Compare the translated version against the original before it goes on a listing — small type, logos and fine texture are where a redraw drifts. The layered path is the one where your real photograph is preserved pixel-for-pixel, so build a localised set there if that matters.

Two further constraints. The research brief runs under a stated prohibition on invented numbers — every statistic, percentage, review count, price and guarantee must appear verbatim in the scraped listing data — so a callout cannot claim a certification your listing never mentioned. And the colour-variant tool repaints the product body only, using the variant ASIN's own photograph from the listing's colour twister as reference, holding every other pixel identical.

What may be generated and what may not

How to decide, per image slot

Main image: photograph only. No generation, no composite, no exception. Pure white, 85% fill, 1000px or more on the longest side (1600px recommended). If your photograph cannot meet that, the fix is a better photograph.

Gallery slots two to nine: composite freely. Real product, generated environment, typeset text. This is where the selling argument lives and where the rules are permissive.

Any image with a photorealistic person: add the metadata keyword before upload, or use a real model and skip the obligation entirely.

No clean photograph at all: a generation tool is the wrong purchase. Shoot it, or pay someone to. Every honest version of this workflow starts from a real image of the real thing.

Common questions

Does Amazon allow AI-generated product images?

Yes. The policy is written around accuracy, not production method, so it does not ask how an image was made — only that it accurately represent the product sold, and that the main image show the actual product on pure white without illustrations or text. A generated background behind a real product photo is fine; an AI-invented product is not.

Do I have to disclose that I used AI on my images?

Only in one case. Since 22 July 2026 Amazon requires listing images, videos and A+ Content containing photorealistic AI-generated people to carry the keyword contains-synthetic-performer in the file's dc:subject XMP field. Background generation and colour correction need no disclosure.

Can I use an AI-generated lifestyle photo as my main image?

No. The main image must show the actual product on pure white with no illustrations, so a lifestyle scene fails before the AI question arises. Lifestyle belongs in the gallery, where it is entirely normal.

My supplier sent renders instead of photos. Can I use those?

Not in the main image, which requires a photograph rather than an illustration or render. In the gallery a technical render is accepted, provided it matches the shipped item. The risk there is not policy; it is that a render flatters a product in ways the arriving box does not.