Yes, AI product photography works for most of a catalog. It is reliable on opaque, matte, rigid products and struggles on glass, chrome, fine chain and anything with legally required text on the label. The property that predicts the result is the material, not the category. The second thing that decides it is the slot: main listing images carry rules the lifestyle slots do not.
A split such as “use AI for 80% of the catalog and shoot the other 20%” does not help at the moment you need it. It does not say which 80%.
Quality is less often the constraint now. What stops a generated image is often Amazon suppressing the listing, Etsy removing it, or Google Merchant Center disapproving the feed item. This guide sorts your catalog by the property that predicts failure, names the rule for each slot, and prices the job honestly.
Key Takeaways
Material predicts the result, category does not. A ceramic mug and a glass perfume bottle sit in the same homeware catalog and behave nothing alike under a generative model.
Reflective and transparent products are the hard cases. Glass, chrome, polished stone and jewelry require the model to invent an environment reflection that stays physically consistent with a scene it also invented.
The main image is a separate job. Amazon wants a realistic, professional-quality image with the product filling 85% of a pure white frame. Etsy requires at least one unedited photo of the actual item, for most products. Google recommends 75% to 90% frame coverage and does not accept placeholders.
Anything with required text stays photographic. Ingredient panels, dosage, care labels and warnings must come from a real photo of the real unit. Models rewrite small text.
Check left and right on every shoe. Amazon asks for a single shoe, facing left, at a 45-degree angle. An image model can return the wrong shoe or turn it the wrong way, and nothing in the result tells you.
Cost the job, not the image. There is no single price for an image, because each model costs a different amount, so DesignerBox shows the cost of a run before you press Run.
The EU disclosure rule has applied since 2 August 2026. Article 50 of the AI Act asks deployers to label deep fakes. Ordinary retouching, such as color correction or lighting changes, usually does not make an image a deep fake.
Does AI product photography work?
AI product photography works reliably on opaque, matte, rigid products with simple geometry: packaged goods, books, ceramics, candles, luggage, most homeware and most furniture. It works less reliably on reflective, transparent and finely repeated surfaces, where the model has to invent physics it cannot verify. It works poorly on anything carrying small printed text that a buyer or a regulator relies on. Sort the catalog on those three properties before choosing a tool.
That is the whole answer. What follows is how to apply it to a real SKU list. For the basics first, see the beginner’s guide to AI for product photography.
Sort your catalog by material, not by category
Product categories group things by what a shopper wants. Generative models group things by how light behaves on a surface. Those two groupings do not line up, which is why a category-level rule keeps failing on individual SKUs.
Sort every SKU into one of three bands. The band decides how much of the generated shoot workflow you can trust without a camera.
| Band | Materials | What generation does well | What to watch |
|---|---|---|---|
| Straightforward | Cardboard, paper, matte plastic, unglazed ceramic, wood, leather, most textiles | Backgrounds, angles, lighting, scenes, seasonal variants | Color drift on saturated dyes |
| Conditional | Glazed ceramic, brushed metal, matte-coated glass, printed packaging | Scenes and lighting hold, surface detail needs the source photo held tightly | Label text, logo shape, edge highlights |
| Photographic first | Clear glass, chrome, polished stone, gemstones, fine chain, watch dials, filigree, eyewear lenses | Little. Use generation to change the environment around a real photograph | Reflections that contradict the scene, invented facets, smoothed link patterns |
The reason the third band is hard has nothing to do with model quality. A reflective surface carries a picture of the room it sits in. When the model invents the room and the reflection separately, the two disagree, and a shopper reads the disagreement as wrong long before they can name why. That is also the failure mode that survives a casual review and gets caught in a return.
Run the bands against your own inventory before you commit, as an AI proof of concept on your own catalog. Run our five accuracy checks on the hardest product you sell first.
The main image is a different job from every other slot
A product page is four separate image problems, and only one of them is tightly regulated.
The main image identifies the product. Marketplaces treat it as the truth claim, so they write the most rules for it. The alternate slots show angles, detail and scale, and the rules loosen. Lifestyle and scene slots are the loosest, and they are where generation earns its keep. Ad creative sits outside the listing entirely, with platform rules of its own.
Generation strength runs in the opposite direction to regulation. The slots with the most rules are the slots where a model is least useful, and the slots with the fewest rules are where it replaces a shoot outright. Deciding this per slot instead of per catalog is the difference between a rollout that ships and one that gets a listing pulled.
The practical split: derive the main image from a photograph of the real unit and use generation for background replacement and cleanup. Generate the scene slots freely. For the alternate slots, generate only what your accuracy band allows. The full shot list is in seven types of product photography, and the listing images by platform in what each marketplace shows.
Four rules decide whether a generated image can be the main image
These are the rules that bite. Each one is published by the channel, and each one is checkable.
Amazon. The main image must “accurately represent the product as a realistic, professional-quality image”, on a pure white background at RGB 255, 255, 255, with the product filling 85% of the frame and shown once. Images need at least 500 pixels on the longest side, and 1,000 pixels or more turns on zoom. No product image may carry Amazon logos or Amazon badges (sellercentral.amazon.com, September 2026). For clothing, Amazon says “Only photos are allowed”, and it does not accept logos, watermarks or text on the main image (sellercentral.amazon.com, September 2026).
Etsy. For most products, Etsy requires at least one unedited photo of the actual item against a real background. You may not use photo editing software or AI tools on that photo. AI tools may edit the other listing images, as long as they still show the real product accurately (etsy.com, October 2026). Etsy allows changes to light, exposure and color balance that match real life, cropping, and a cleaned-up or new background. It does not allow changes to the product’s color, texture or condition, hidden defects, or AI changes to the core design. A fully generated picture cannot be the required unedited photo.
Google Merchant Center. The feed image must show the product. From 31 January 2027, Google will require at least 500 x 500 pixels for every product image. Until then the minimum is 100 x 100 pixels, or 250 x 250 for clothing (support.google.com, September 2026). Google recommends that the product takes 75% to 90% of the image. It does not accept placeholders, images that do not show the product, or images with promotional elements or overlays (support.google.com, September 2026).
Google also says every image made with generative AI must keep metadata that says AI made it, such as the IPTC DigitalSourceType value TrainedAlgorithmicMedia. Do not remove it (support.google.com, October 2026). These rules reach every Shopify store running Google Shopping, whatever Shopify itself permits.
The EU AI Act. Article 50 has applied since 2 August 2026. Deployers who publish a deep fake, meaning realistic AI images, audio or video that could pass as real, must disclose it (European Commission, September 2026). The Commission’s guidelines, published 20 July 2026 and not binding, say color correction, lighting changes and background replacement for clearly aesthetic purposes usually do not make an image a deep fake (European Commission, September 2026). A real product on an AI background usually is not a deep fake, as long as the ad does not mislead about the product. A photorealistic invented person wearing your garment is the case that can count as a deep fake. This is general information, not legal advice.
There is a fifth rule that only applies to footwear, and a generated image breaks it easily. Amazon asks for a single shoe, facing left, at a 45-degree angle, for all footwear (sellercentral.amazon.com, September 2026). An image model can return a right shoe or turn the shoe the wrong way, and nothing in the result flags it.
For the disclosure picture across Amazon, Walmart, eBay and TikTok Shop, we cover what each marketplace requires for AI product photos in full.
What breaks, and the check that catches it
Five failures cause many of the problems buyers notice. Each has a quick check.
- Label and packaging text. Models rewrite small type. Zoom to 100% and read every word on the pack. Ingredient panels, dosage, warnings and care labels come from a photograph, always.
- Logo geometry. Letterforms drift, kerning slips, a registered mark moves. Overlay the output on your brand asset at 50% opacity and look at the edges.
- Color against the real dye lot. Screens lie and models push saturation. Put the generated image beside a photo of the physical unit shot on the same screen, not beside your brand hex value.
- Repeated fine detail. Chain links, knit stitch, engraving and watch indices get smoothed or invented. Count them. If the real product has twelve, the image needs twelve.
- Handedness and asymmetry. Shoes, gloves, asymmetric garments and anything with a zip on one side. Check which side, every time.
Color can turn into a return, because the shopper only sees the real color after the box is open. Treat it as the blocking check. For the checks around these five, the product photography checklist lists 30 checks in six stages, from the source photo to the live listing.
What it costs per SKU
Multiply a headline studio rate by a large image count, compare it to a marginal generation cost, and you get a large saving figure. That figure skips the two things that move it.
The first is volume. Studios often price by shot and by order size, so a large order is rarely priced at the single-image rate. The second is the full cost of a shoot: retouching, studio hire, sample shipping and coordination sit on top of the quote.
The studio side is priced in what a product photoshoot costs.
The generation side prices differently, because there is no single price for an image. Each model costs a different amount, and the model is your choice per shot, so the cost of a run is shown before you press Run. You plan against a monthly allocation instead of a per-image rate. Every image that fails a check costs another run, and the AI image retry rate shows what that does to the cost of each usable image.
Uploading your own photos and the commercial license start on the Pro plan. Plans are on the pricing page.
There is a free plan, and it runs on sample products. Uploading your own photos starts on the Pro plan, so run the material bands against your own hardest SKU there.
Running this across a whole catalog has its own ceilings, and they are platform ceilings rather than generation ones. Shopify allows up to 250 media items per product against up to 2,048 variants (help.shopify.com, September 2026), which is the kind of number that decides a rollout. We cover those in what actually automates in a catalog run.
The real saving is the SKUs that never got shot at all. A catalog budget already decides which products deserve a full gallery, and everything below that line ships with one supplier image. Those are the listings where generation adds an angle set, a scale shot and a scene where there were none.
A production order that clears every channel
Run it in this sequence. The order matters because each step constrains the next.
- Photograph one real unit per SKU. A phone on a clean surface in daylight is enough. This is the source of truth for color, text and proportion, and the input every later step derives from.
- Sort the SKU into a band. Straightforward, conditional or photographic first. The band decides how much the model is allowed to invent.
- Build the main image from the photograph. Clean it, cut the background, place it on white, size the product to the frame coverage the channel wants. Those are four edits in a row on the same file, and they have to hold resolution and surface detail through all four. Do not generate a new product surface in this slot.
- Generate the alternate slots inside the band. Angles and detail shots for band one, tighter supervision for band two, real photography for band three.
- Generate the scene and lifestyle slots freely. This is the volume win and the lowest-risk work on the page.
- Run the five checks. Text, logo, color, repeated detail, handedness.
- Check the channel rule for the slot, not the catalog. Amazon frame coverage, Etsy unedited photo, Google Merchant Center coverage and format, EU disclosure if a synthetic person appears.
- Save the sequence as a workflow, then run it on the next SKU. Steps three to five are the same decisions every time. Build them once, and SKU 40 needs a run instead of a rebuild. The checks in steps six and seven stay with you on every product.
DesignerBox runs steps three through five from the same upload. The Pictures for a listing app makes six marketplace pictures from one product photo. AI product photography covers listing shots, new angles, flat lays and styled scenes from that one source photo. The saved workflow runs again on the next SKU, with no new prompt. Reviewing the results is the part that scales badly. Critic steps score the results of a run, and best-of-N keeps the best one. Batch runs one workflow over a whole sheet of products, so you keep or discard each result.
For model choice, the same brief run through every image model on one product is more useful than a leaderboard. A leaderboard ranks how good a picture looks. It does not show how faithfully the model held your source. The model list shows the image and video models DesignerBox runs. If you sell across marketplaces, DesignerBox also has a page for ecommerce brands. Our own breakdown of the image spec for each channel covers where two platform requirements contradict each other.
Anyone can make an AI picture. Making hundreds that still look like your brand is the hard part. The source photos, the listing shots, the scenes, the brand rules and the Assets library sit in one account, so a catalog stays together. Start from a finished result, add your brand and your products, and run it. See the templates.
FAQ
Can I use AI product photos on Amazon?
Yes, for most slots. Amazon’s product image guide judges what the image shows, and its AI rule covers photorealistic people made fully by AI. The main image must be a realistic, professional-quality image of the product on pure white at RGB 255, 255, 255, filling 85% of the frame, with no placeholders. Derive it from a photograph of the real unit and use generation for the background and cleanup. For clothing, Amazon says only photos are allowed. Amazon separately asks sellers to tag images that show a photorealistic person made fully by AI, with the keyword contains-synthetic-performer in the XMP dc:subject field.
Does AI product photography work for every product?
No. It works reliably on opaque, matte, rigid products and struggles on clear glass, chrome, polished stone, gemstones, fine chain and watch dials. Those surfaces carry reflections of their environment, and a model inventing both the reflection and the environment produces two versions of a room that disagree. For those SKUs, photograph the product and generate the scene around it.
Do I have to disclose AI-generated product images?
Usually not for an ordinary product photo. The EU AI Act’s Article 50 duty, which has applied since 2 August 2026, covers deep fakes. The Commission’s guidelines say color correction, lighting changes and background replacement for clearly aesthetic purposes usually do not make an image a deep fake. A photorealistic invented person wearing your product is the case that can trigger disclosure, and Amazon asks for its own metadata tag for exactly that.
How much does AI product photography cost per SKU?
It depends on the model you pick, so there is no flat per-image rate. In DesignerBox, the cost of a run is shown before you press Run. When you compare that with a studio quote, count the volume pricing on large orders and the retouching, studio hire and shipping that sit on top of the quote.
Will AI change my product’s real colors?
It can, and the buyer only discovers it after delivery. Models push saturation and shift hue on strongly dyed materials. Check the generated image against a photograph of the physical unit displayed on the same screen. Comparing it to your brand hex value tests the wrong thing.
Can AI generate the main listing image, or only the extras?
Treat the main image as photographic with generated support. Amazon and Google Merchant Center both constrain that slot. Etsy requires at least one unedited photo of the actual item, for most products, so keep one photo with no edits at all. Cleanup, background replacement and reframing usually fit the main slot. Inventing product surface is not. The alternate, scale and lifestyle slots are where generation replaces a shoot.
What does AI product photography not replace?
Brand campaign imagery where the concept is the product, tactile luxury goods sold on unboxing feel, and the source photograph itself. Every workflow here starts from one real photo of one real unit, which is the thing that keeps the output honest about color, text and proportion.
Sources
All read in September 2026 unless stated.
- Amazon product image guide: the main image requirement for a realistic, professional-quality image, pure white RGB 255, 255, 255, 85% frame coverage, the 500 to 10,000 pixel range with zoom at 1,000 pixels, the single shoe facing left at 45 degrees, and the contains-synthetic-performer metadata tag: sellercentral.amazon.com
- Amazon image guidelines for clothing, “Only photos are allowed”, and no logos, watermarks or text on the main image: sellercentral.amazon.com
- Etsy Listing Image Requirements policy on the required unedited photo of the actual item, and the photo editing and AI rules: etsy.com, October 2026
- Google Merchant Center image requirements on placeholders, promotional elements and overlays, the 75% to 90% recommendation, and the metadata rule for images made with generative AI: support.google.com. The interim minimums and the 500 x 500 minimum from 31 January 2027: support.google.com
- EU AI Act Article 50 transparency obligations, the 2 August 2026 application date and the deployer duty for deep fakes: European Commission FAQ. The Commission’s Article 50 guidelines, published 20 July 2026 and non-binding: European Commission
- Shopify limits of 250 media items and 2,048 variants per product: help.shopify.com
- DesignerBox plan gates for the commercial license, video, try-on, team features and seats: DesignerBox pricing page (designerbox.ai/pricing), September 2026
Channel rules verified from Amazon Seller Central, Google Merchant Center, Etsy and Shopify documentation as of September 2026. EU obligations verified from the European Commission’s Article 50 FAQ and guidelines as of September 2026. The Google Merchant Center image rules were re-checked on 2 October 2026, and Etsy’s listing image rules on 22 October 2026. Platform policies in this area change quickly and vary by product category. This is general information, not legal advice. Individual results vary.
