Short answer: AI product photography means generating the image around a product from material you already own, instead of booking a studio day for every variant and every season. It is good at backgrounds, scenes, seasonal sets and packaging mockups. It must never invent the thing in the box. A product image is not a mood board. It is a factual claim about an object you are going to put in a parcel, and the buyer checks the claim on the kitchen table.
This is the narrow version of a subject we have written about more broadly, from what ad creative costs to AI ad creative against a human designer. The difference matters. Ad creative sells an idea and gets judged on whether it works. A product image gets judged against the physical item, by someone who paid for it.
A product image does three different jobs
Stores talk about product photos as one pile of files. They are three jobs with three different tolerances.
The first is the feed image, the one that goes to Google Shopping, to a marketplace, to a price comparison site. That image is the product record. Google publishes image requirements for Merchant Center and it is worth reading them before anything generated goes near a feed, because this is the version that gets reviewed by a platform rather than admired by a customer.
The second is the store image, the gallery on your own product page. Detail shots, the item in a room, the scale reference, the material close up. This is where a buyer decides whether to trust you.
The third is the ad image, the one competing with everything else in a feed on a phone. It has the widest creative latitude and the shortest life, which is exactly why shooting each one is so expensive per useful day.
Generation earns its place from the third job backwards. It belongs in ads. It belongs in parts of the store gallery. It belongs in the feed image only as retouching of a real photograph, if at all.
What generation is genuinely good at
Four things, and they are not small.
Backgrounds and surfaces. One truthful packshot becomes the same jar on oak, on linen, on a bathroom shelf, in flat winter light. The object does not change. Its surroundings do.
Consistency across a catalogue. Four hundred products photographed across two years by three different people look like four hundred products from three different shops. Regenerating the surround from a single reference set is often the cheapest way a mixed catalogue starts looking like one brand.
Seasonal and campaign variants. A seasonal flavour, a Christmas set, a summer scene. These rarely justify their own shoot day, so in practice they never happen, and the campaign runs with the same image as March.
Packaging and concept visuals before print. Showing a label idea on a real bottle shape, in context, before committing to a print run. Nobody ships this image to a customer, so the accuracy bar is a different one entirely.
Where it breaks, including on this page
The failure modes are consistent and they are all about facts.
Colour drifts. A generated version of a specific dye lot or a specific glaze is a close relative of the real colour, not the colour. For anything sold on shade, that difference comes back as a return.
Material gets invented. Weave, grain, stitching and finish are the details a model fills in plausibly rather than accurately. A buyer who ordered brushed steel and got the render of brushed steel notices.
Text turns to mush. Labels, ingredient lists and wordmarks come back garbled or subtly wrong more often than not. The hero image at the top of this post is a worked example. It took three generations: the first came back with black letterbox bars baked into the frame, the second printed a legible hex code across a colour chart that was never meant to carry text, and only the third was clean enough to publish. That image only has to look like a studio table. A product image has to be right.
Counts and assemblies go wrong. How many pieces are in the set, which accessories are included, what the mount looks like. These are contract terms, not composition.
The rule that settles most arguments: the picture must match the parcel
The legal question is not whether a machine made the image. It is whether the image tells the truth about the product. The EU Unfair Commercial Practices Directive treats a presentation that deceives a consumer about the main characteristics of a product as an unfair practice, and a photograph is a presentation. The tool that produced it does not change the test. In Estonia the authority that handles this is the Consumer Protection and Technical Regulatory Authority.
So the working rule is simple to state and easy to check: every claim a customer could measure against the delivered item has to come from the real item. Colour, material, size, contents, condition. Everything else in the frame is staging, and staging has always been allowed to be arranged.
What you have to mark, and what you do not
The EU AI Act adds a separate duty. Article 50 requires providers of AI systems that generate synthetic image, audio, video or text content to ensure the output is marked in a machine readable format and detectable as artificially generated or manipulated. The same article carves out an exemption where the system performs an assistive function for standard editing and does not substantially alter the input data.
Read plainly, that separates two things a shop does every week. Cleaning a background or removing dust on a real photograph sits close to standard editing. Generating a new scene that never existed does not. The marking duty in that article falls on the provider of the generative system rather than on the shop using it, which is why the practical tooling lives at that layer: the C2PA content credentials standard for provenance metadata, and vendor watermarking such as Google DeepMind's SynthID.
None of that removes the shop's own job, which is record keeping. Know which images in your catalogue are generated, per SKU. When a customer or a marketplace asks, the answer should take a minute, not an afternoon.
A workflow that survives a real catalogue
The pattern that holds up in production has six steps and one non negotiable human.
- Shoot the reference once, honestly. Every product gets one truthful set: the object, correct colour, correct scale, correct contents. This is the source of fact and it is the part you do not automate.
- Generate from the reference, not from a description. Image to image from your own packshot keeps the object and changes the surround. A text prompt invents a product that resembles yours, which is the origin of most of the failures above.
- Fix the recipe per product family. One prompt and lighting recipe per family, not per SKU, so a catalogue looks like one shop instead of four hundred experiments.
- Route it through a workflow, not a browser tab. The image API gets called from an automation, the outputs land in your media library named against the SKU, and nothing gets uploaded by hand. This is the same ecommerce automation plumbing that syncs stock and orders.
- Keep one human approval per SKU. Somebody who has held the product compares the generated image to the actual item. This step is the whole quality system. Remove it and the workflow becomes a machine for generating returns.
- Write the generated flag into the product record. A field in the PIM, set at upload time, so provenance is a data question later rather than an archaeology project.
Our own stack for this is unremarkable on purpose: an image generation API called from an n8n or Make workflow, outputs into the asset library, one approval step, then publication. The hero on this page came from exactly that pipeline, which is also how we know what its failure rate looks like on a bad prompt. That is the machinery behind our ad factory work, and the ecommerce version of it lives at ad factory for DTC ecommerce. Food and drink brands, where the product is styled rather than assembled, have their own set of constraints at ad factory for food and beverage.
What it costs, honestly
The per image API charge is the small number and every vendor publishes it. Two costs are larger and neither shows on an invoice. The first is the approval step, which is real human minutes per SKU and does not shrink with volume. The second is rework: a generated image that misrepresents a product costs you the return, the refund, the review and the trust, which is a far worse trade than a shoot day.
That is why the sensible sequencing is catalogue first, volume second. Work out how many images actually block a launch this quarter, how many of those need to be truthful records and how many are staging, and build for the staging half. Our own model starts with a free audit of about 30 minutes that maps the catalogue and where images are the bottleneck. If building makes sense, ongoing build-and-run starts at 600 euros a month plus VAT, quoted after the audit. The pricing logic behind that sits in how many ad creatives you actually need.
When to shoot instead
Three cases, and they are common enough to name.
A small catalogue with a high average order value and one hero product. Shoot it. The whole business rests on ten images and a studio day is cheap against that.
A regulated category. Food claims, supplements, cosmetics and medical devices carry labelling rules on top of consumer law, and a generated label is an unforced error.
A store whose returns already carry the phrase "not as pictured". The image pipeline is not your constraint. The accuracy of your existing photographs is, and generating more of them faster makes the same problem larger.
Frequently asked questions
What is AI product photography?
It is the practice of generating product imagery from material you already hold, such as a packshot, a sample photo or an existing studio set, instead of booking a new shoot for every variant. In practice it produces backgrounds, scenes, seasonal sets and packaging mockups around a product that was photographed once for real.
Can generated images replace a product photo shoot entirely?
No, and the attempt is where stores get into trouble. You still need one truthful reference set per product, shot with the real object, because that is what colour, finish, texture and contents are checked against. Generation extends that reference into more scenes. It cannot be the origin of the facts about the product.
Are AI generated product images legal to use in a webshop?
Using them is not the issue; misrepresenting the product is. EU consumer law treats a misleading presentation of a product's main characteristics as an unfair commercial practice, whatever tool made the image. The EU AI Act separately requires providers of generative systems to mark synthetic output in a machine readable form, with an exemption for assistive standard editing that does not substantially alter the input.
Which product images should never be generated?
The ones a buyer measures the parcel against: colour and finish, material and texture, what is included in the box, scale against a hand or a room, and anything in a regulated category such as food claims, supplements or medical devices. Marketplace feed images should also stay photographic, since they are the version the platform treats as the product record.
What does an AI product image workflow cost to run?
The image API charge per picture is the small part; the real cost is the human approval step and the reshoot when a generated image is wrong. Our model starts with a free audit of about 30 minutes that maps your catalogue and where images actually block launches. If building makes sense, ongoing build-and-run starts at 600 euros a month plus VAT, quoted after the audit.
If you want to know which half of your catalogue is staging and which half is a factual record, that is what the free audit is for. Tell us how many products you sell and how images reach the site today and we will map where the shoot days are actually going.