AI Images for Amazon Listings: A Slot-by-Slot Permission Map

Two pieces of advice dominate every thread about AI images and Amazon, and both are wrong. The first says AI imagery is banned — that Amazon detects it and suppresses listings for it. The second says nothing has changed — generate whatever looks good and upload it. Neither survives ten minutes with Amazon’s actual product image requirements, because those requirements are almost entirely about what an image shows, not about what drew the pixels.

The short answer, and the one this whole guide unpacks: AI participation is decided slot by slot, not listing by listing. The main image is a photograph of the real product and effectively closes the door on generation. Every other slot — secondary product shots, lifestyle, infographics, A+ modules — opens by a different amount, and each one carries its own hard specs that generation does not exempt you from.

What follows is the process, not a prompt dump: a Slot Permission Pass you run once per slot before opening a generator — claim, canvas, constraint — plus the prompt structure that belongs to each slot.

Sourcing note. The load-bearing rules below come from Amazon’s public product image requirements in Seller Central help (reference G1881), re-checked on 2026-08-27. That page renders only inside a logged-in browser session, so every rule here is a paraphrase, not a quotation of Amazon’s wording. Specification numbers are cross-checked against our own Amazon product image size requirements breakdown.

The Rule That Actually Governs AI Images: Accuracy, Not Authorship

Read Amazon’s image requirements looking for the word “AI” and you will not find a framework. What you find instead is a set of statements about outcomes: images must accurately represent the product being sold; the main image must show only that product against a pure white background; the product must fill at least 85% of the frame; text, logos, borders, watermarks and decorative graphics do not belong on the main image (data checked 2026-08-27).

Notice what those rules are indifferent to: which camera, which retoucher, which compositing tool or which model produced the file. Enforcement lands on the result. A listing gets suppressed because its main image has a grey background, not because of the software that made it grey.

There is exactly one structural exception, and it happens to sit on the most valuable slot on the page. The main image is specified as a photograph of the actual product, with drawings, renderings and illustrations excluded. That is a rule about how the image came to be, not only about what it depicts. Which is why the honest formulation is not “Amazon never cares how images are made” but:

Amazon judges every slot on what the image shows; the main image additionally constrains how it was made.

The practical consequence: the risk of AI imagery on Amazon is rarely a policy risk. It is an accuracy risk — a generator quietly changing a fabric’s weave, adding a port to a device, or seating a 350ml mug next to a kettle at 600ml scale. None of that trips a filter. All of it returns eight weeks later as refunds and customer photos that contradict your gallery.

The Slot Permission Pass

Before any prompt gets written, each slot gets three questions. They take about ninety seconds per slot and they are the whole method.

1. Claim — what factual statement does this image make about the physical product? The main image claims “this is the item, this shape, this colour.” A scale shot claims “it is this big relative to a hand.” A bundle shot claims “you receive these five pieces.” Write the claim down in one sentence.

2. Canvas — which pixels are the product, and which are surround? Surround is background, surface, environment, lighting, props that are not part of the purchase, layout blocks and typography. In a lifestyle image of a kettle on a counter, the kettle is claim and the counter, the window light and the out-of-focus plant are canvas.

3. Constraint — which hard specs does this slot carry? Background colour, frame fill, aspect ratio, minimum resolution, text permissions, mobile legibility. These are the same numbers whether the image is photographed or generated; the full specification table is worth keeping open in a second tab rather than reproduced here.

The rule that falls out: AI may rework the canvas. AI may not author the claim. And the output must land inside the constraint. Every row of the table below is that rule applied to one slot.

The Slot-by-Slot Permission Table

Two deliberate omissions: the specification numbers stay in the size requirements guide rather than being duplicated here, and advertising creative is out of scope — Sponsored Brands, display and off-Amazon channels run on separate policies this page does not evidence.

Slot 1: The Main Image — Where AI Participation Stops

The main image is the only slot where the question “was this generated?” is answerable in principle rather than by inspection, because the requirement is that it be a photograph of the product you are actually shipping. A synthesised product image is not a photograph of it, however photorealistic the render.

That does not reduce AI’s role to zero — it relocates it. Almost every professional main image in 2026 has had AI touch it, on the surround:

• Background removal. Cutting the product from a shot and placing it on RGB 255,255,255. This is the single most common AI operation on Amazon and it changes nothing the rules care about, provided the mask does not eat product edges.
• Shadow and reflection cleanup. Removing a stray studio reflection is maintenance. Generating a soft contact shadow that the object never cast is a decision — keep it subtle, keep it physically plausible, and never let it change the apparent footprint of the product.
• Upscaling. A 900px archive photo upscaled past the zoom threshold is legitimate, right up to the point where the upscaler starts inventing texture. Check fabric weave, printed packaging text and any mesh at 100% before accepting the output.

One failure signature is worth naming: edge erosion. Automated cutouts trim a pixel or two off translucent, furry or wiry edges, and after two rounds of cleanup the silhouette is measurably slimmer than the product. Compare it against the original photo at 100% before export.

If the main image needs to change materially, that is a photography job, not a prompt job.

Slots PT01–PT08: Secondary Product Shots

Secondary slots are where most sellers under-use AI, because they assume the studio rules from the main image carry over. They do not. Backgrounds here can be any colour, any texture, any environment. Text overlays are allowed. The constraint that does carry over is accuracy — and its sharpest edge in this slot is the angle problem.

If you photographed a device from the front, the left and the top, you have three angles. Ask for a three-quarter rear view and the generator invents whatever is back there: a port that does not exist, a vent in the wrong place, a label of plausible nonsense. That is a false statement about the product, and the most common way a competent-looking AI gallery becomes a returns engine.

Two rules keep this slot safe:

1. Every angle in the gallery must correspond to an angle you actually photographed. Generation restyles the surround around a real capture. It does not add viewpoints.
2. Bundle and contents shots are claim-only images. If the slot’s job is “here is what is in the box,” every object in the frame is claim, not canvas — the box, the cable, the manual, the spare filter. Nothing in that frame may be generated, and nothing may be quietly omitted for composition.

Within those bounds, generated surfaces are a genuine upgrade: one consistent slate or brushed-metal surface across six slots reads as a brand system and costs one prompt instead of a studio day. Settle the slot logic first — which image does which job, checkpoint 5 of our listing image audit checklist — because generating the wrong slot beautifully is still a wasted slot.

Lifestyle Images: The Widest Lane on the Listing

Lifestyle is where AI image tools were built to work and where the economics are obvious: a generated kitchen, nursery or workshop costs minutes instead of a location fee. The whole scene is canvas. The product, composited in from a real photograph, is claim.

Three accuracy traps specific to this slot:

Scale. Generators have no idea how big your product is; they size it to suit the composition, so a 20cm cutting board becomes a 40cm one next to a generated chef’s knife. Shoppers calibrate size from lifestyle images more than from any dimension line, which makes this the error behind most “smaller than expected” reviews. Anchor instead: composite against a scale reference you measured — a mug, a hand, a door frame.

Multiplication. Diffusion models duplicate. A second, subtly wrong copy of your product appearing on a shelf in the background implies a multi-pack. Sweep every generated frame at full resolution for extra units before it goes near an upload.

Implied use. A bathroom scene for a product not rated for humidity, or an outdoor one for a product that is not weatherproof, is a durability claim you did not intend. Choose environments that match the specification sheet, not the mood board.

On people in lifestyle scenes, one deliberately limited note: third-party reports describe a metadata disclosure keyword expected on photorealistic AI-generated people in listing media, and no official Amazon page confirming that requirement was reachable when this page was checked on 2026-08-27. It is not treated as settled here and it does not carry any argument in this guide. Checkpoint 13 of the listing image audit checklist covers how we currently handle it; until an official page states the rule, the conservative route is a real model or a scene without a photorealistic face.

Infographics: AI Layout, Human Numbers

Infographics are the slot where the division of labour is cleanest, because a feature callout is made of two completely different materials: composition and claims.

Composition — grid, colour system, iconography, background, the arrangement of callout blocks around a product cutout — is canvas, and generators handle it well. Claims — the 90-day filter life, the 40°C rating, the 12-hour battery, the dishwasher-safe icon — are facts about the product that must come from your datasheet, typed by a human, and set in real type.

The practical rule that saves the most rework: never let a generator render your text. Image models produce text that is convincingly shaped and frequently wrong — a “10” that becomes “1O”, a unit drifting from mm to cm, a certification badge that looks exactly like a real one. Generate the layout with placeholder blocks, then set the copy in a design tool where you control the glyphs. It is also the only way to guarantee this slot’s other hard constraint: legibility at 160px in search results and on a phone screen.

Fabricated trust marks deserve their own line. A generated “certified” seal, a made-up test-lab logo or an invented award ribbon is not a stylistic flourish — it is a factual claim about third-party validation. Icons that imply certification belong in the same category as the numbers: sourced, not generated.

A+ Content Modules: Different Canvas, Same Test

A+ modules are the widest creative surface on a listing — full-width banners, brand-story imagery, textured backdrops, atmospheric scenes that would never pass as gallery images. Generation fits this brief better than any other slot on the page.

The tests that still apply:

• Accuracy carries over. The product in your A+ art must be the same product as in the gallery. A restyled hero shot with a slightly different finish reads as two products to anyone comparing carefully.
• Comparison modules make claims. A comparison chart against your own range is a set of factual statements about your own products, and the imagery in it must match each variant. Naming or depicting competitors stays off the table regardless of how the image was made.
• Mobile stacking is a spec, not a preference. Premium A+ layouts collapse to one column on phones, and banners composed for a wide desktop crop lose their subject entirely. Compose for the mobile crop and check the exported module at phone width.
• Consistency across modules matters more than individual quality. Six modules generated across six sessions, each excellent, read as a stitched-together page. Lock one scene recipe — same lighting direction, same colour temperature, same surface family — and reuse it.

Deciding which modules to build at all is a research question rather than a generation question; our guide to A+ content competitor research covers how to read a category’s module grammar before you commission art for it.

Prompt Structures by Slot

Prompt lists age badly — they are tuned to one model’s quirks and stop working at the next version. Structures do not, because they encode what the slot requires rather than what a particular model responds to. Every structure below uses the same five components:

[SUBJECT LOCK] — what must not change · [SCENE] — the canvas · [LIGHT/CAMERA] — how it is rendered · [FRAME] — the slot’s hard specs · [NEGATIVE] — what must not appear

Main image (background work only)

SUBJECT LOCK: preserve the uploaded product photo pixel-for-pixel;
              no reshaping, recolouring, relighting of the product
SCENE:        pure white background, RGB 255,255,255, seamless
LIGHT:        neutral studio light; optional soft contact shadow
              directly beneath the object, no cast direction
FRAME:        square 1:1, product occupies ≥85% of the frame,
              output ≥1600px longest side, sRGB, JPEG
NEGATIVE:     no props, no packaging unless the product is packaging,
              no text, no logo, no watermark, no border, no gradient,
              no invented reflections, no edge trimming

Secondary product shot (PT01–PT08)

SUBJECT LOCK: use uploaded photo of <angle you actually captured>;
              product geometry, finish, port layout and branding unchanged
SCENE:        <surface / gradient / studio set>, matching the
              surface family used across the gallery
LIGHT:        <single soft key from upper left>, consistent across all slots
FRAME:        square 1:1, ≥1600px, product centred with even margin,
              headroom reserved for a caption strip
NEGATIVE:     no additional units, no accessories not included in the box,
              no alternate viewpoints, no generated text

Lifestyle / in-use

SUBJECT LOCK: composite the real product photo unchanged into the scene;
              scale it against <named reference: standard mug / adult hand>
SCENE:        <room, surface, time of day, season, two or three props>
              matching the environment the product is rated for
LIGHT/CAMERA: <window light from the right>, 50mm equivalent,
              shallow depth of field, eye-level
FRAME:        square 1:1, product occupies 40–60% of frame,
              clear separation from background
NEGATIVE:     no duplicate units, no readable brand text on props,
              no environment the product is not rated for,
              no photorealistic faces

Infographic

SUBJECT LOCK: product cutout unchanged; callout anchors point to
              real features at their real positions
SCENE:        <two-column layout / radial callouts>, <brand colour>
              background, three to five callout blocks with
              PLACEHOLDER TEXT ONLY
LIGHT:        flat, even, no scene lighting on the graphic layer
FRAME:        square 1:1, ≥1600px, callout blocks sized so text
              remains legible at 160px thumbnail and at phone width
NEGATIVE:     no rendered text, no numerals, no units, no certification
              seals, no award badges, no icons implying third-party testing

A+ content module

SUBJECT LOCK: same product photo and finish as the gallery hero
SCENE:        <brand environment>, reusing the gallery's surface
              family, colour temperature and light direction
LIGHT/CAMERA: <same key direction as gallery>, wide framing with
              empty band on <left/right> for module copy
FRAME:        <module aspect ratio>, subject positioned inside the
              central safe zone so the mobile crop keeps it
NEGATIVE:     no competitor products or marks, no testimonial-style
              staging, no awards, no text, no elements that break
              when the layout stacks to one column

Two habits make these worth more than the sum of their prompts. First, the negative block is where compliance lives — every entry in it maps to a specific way a slot fails. Second, write the FRAME block from the slot’s spec sheet before the SCENE block. Compose first and crop later and you get a beautiful 3:2 lifestyle image that loses a third of its subject when squared.

The Source-Asset Step Everyone Skips

Every structure above starts with SUBJECT LOCK, and a subject lock is only as good as the file you feed it. This is the step that separates galleries that look designed from galleries that look generated: before opening any tool, you need two sets of real assets.

Your own current gallery, at original resolution. Not the derivatives the product page serves — the originals behind them, the files that survive compositing, upscaling and 100% inspection. You also need to see what is live right now, slot by slot: a variant carrying last year’s packaging, a slot empty for a year, a hero swapped during a promotion and never swapped back.

Five to ten category leaders’ galleries, foldered by slot. Not for reuse — study, never reuse; competitor imagery is their asset, and the point of pulling it is to read the category’s conventions. Which slot carries the scale shot in your category? Does everyone put the infographic at PT02? Are lifestyle scenes indoor or outdoor here? Those answers set your slot plan, and the slot plan is what your prompts implement. Our competitor image analysis workflow covers the scoring side of that read.

Collecting both sets by hand — right-clicking, renaming, tracking which file came from which variant — is the tedium that stops most sellers doing it at all. ASINCrate exports a listing’s full media set as one ZIP: originals at highest available resolution, named by slot (MAIN, PT01), foldered by variant and ASIN, A+ module images labelled separately, plus a CSV of URLs and dimensions. The extension is free to install; bulk ZIP and CSV export sit in the paid tier, single-image download and variant browsing do not.

For a comparison of the generators themselves — what Pebblely, Photoroom and the rest cost and where each one’s limits show up — see our Pebblely alternatives breakdown. That page is the tool-and-pricing view; this one is the permission view, and its short section on where AI images are allowed is the compressed version of the table above.

What Crossing the Line Actually Looks Like

Almost no one gets caught by a policy filter. The failures are quieter, and they all look like this:

Accuracy drift. The gallery product differs slightly from the one in the box — a deeper colour, a cleaner seam, a weave the real item does not have. Each image is individually defensible; the aggregate over-promises, and it surfaces as returns coded “not as described.” Checkpoint 12 of the audit checklist is the review-photo contradiction test that catches this from the outside.

Scale collapse. The most expensive single error in AI lifestyle imagery, because it is invisible to the seller (who knows how big the product is) and decisive for the shopper (who does not).

Style incoherence. Six slots generated in six sessions, with six lighting directions and six colour temperatures. Each looks professional; together they look like a scraped listing. This is a trust signal shoppers read without articulating.

Text artefacts. Generated glyphs that survived into an uploaded infographic. Rare enough to feel unlikely, common enough that it is worth zooming to 100% on every numeral before export.

Suppression. The one genuine policy outcome, and nearly always the main image: an off-white background, a prop, a border, a watermark, a product filling half the frame. Nothing to do with AI; everything to do with the main image being the one slot whose rules fail automatically.

The pattern across all five: the compliance question is answered by the specification, and the quality question is answered by the source photo. Generation sits in between, and its job is the canvas.

Frequently Asked Questions

Does Amazon allow AI-generated images on listings?

Amazon’s product image requirements are written about what an image shows — the product, the background, the frame fill, any text — rather than about which software produced the pixels. The structural exception is the main image, which is specified as a photograph of the actual product, so a fully synthesised main image fails on the photograph requirement itself. Secondary, lifestyle, infographic and A+ imagery are governed by accuracy, not authorship (data checked 2026-08-27).

Can I use an AI-generated main image on Amazon?

No. The main image has to be a photograph of the real product on a pure white background with the product filling at least 85% of the frame, so a generated or illustrated product is out. AI can still touch that slot in a narrow way: background cleanup and white-canvas compositing around unmodified product pixels, which changes the surround rather than the product.

Which listing images can be AI-generated?

Lifestyle scenes, backgrounds, surfaces, infographic layouts and A+ module backdrops are the widest lanes, because the product itself can remain a real photo composited into a generated environment. Secondary product-only shots allow generated surround but not generated angles the product never had. The main image is the one hard stop.

Do I have to disclose AI-generated images on Amazon?

Third-party reports describe a metadata keyword expected on photorealistic AI-generated people, and no official Amazon page confirming that requirement was reachable when this article was checked on 2026-08-27. Treat it as an unconfirmed obligation, keep it out of your compliance narrative until an official page states it, and follow checkpoint 13 of our listing image audit checklist for the current handling.

What specs do AI-generated Amazon images still have to meet?

Exactly the same ones as photographed images: 1,000px on the longest side to trigger zoom, 1,600-2,000px as the working standard, sRGB, JPEG preferred, under 10MB, pure white and 85% frame fill on the main image, and no text or logos on the main slot. Generation changes nothing about the specification.

What is the most common way an AI listing image goes wrong?

Accuracy drift rather than a policy breach: the generator subtly restyles the product — a different weave, an extra port, a cleaner logo, a mug that is now the wrong size next to a kettle. The image looks excellent and describes a product that does not exist, which shows up later as returns and review contradictions.

Running the Pass

The Slot Permission Pass is deliberately boring: name the claim, mark the canvas, attach the constraint, then generate. It converts an anxious question — is AI allowed here? — into a mechanical one with a different answer for each slot on your page.

Start from real assets, keep the product photographic wherever it is making a claim, let generation own the surround, and audit the exported set rather than the previews. That sequence produces galleries that are cheaper than a studio day, consistent across slots, and accurate enough that the images and the parcel describe the same product.