AI Product Photography: What Holds Up and What Breaks

Forty products on the kitchen table, a phone, a window with north light and a bedsheet taped to a chair. Somewhere in a browser tab there is a quote from a studio, and it has more digits in it than you expected. That is the moment this question gets asked, and it is a fair one.

AI product photography is usually pitched as a straight swap for a shoot. It is not one, and the useful version of the question is narrower. What follows is the answer with both sides costed and the rules attached, because there is a compliance deadline three days from now that changes the arithmetic and almost nobody selling you an image generator has mentioned it.

What does a product photo have to survive?

Four different situations, and they impose completely different requirements. Almost every bad decision in this area comes from optimising for the first one and forgetting the fourth.

Read the third row twice. Every other row is a marketing question with a marketing cost. That one is a money question, and it is the reason the answer to this whole topic is not simply yes.

Where does generated imagery visibly fail?

Generated images fail in four specific places, and they are consistent enough across models that you can plan around them rather than test each release.

Fabric and surface texture. Weave, pile, grain and knit render as an impression of texture rather than a texture. A customer buying a jumper is buying the texture, so this is a category-level disqualification rather than a quality complaint.

Text on the product. Packaging copy, labels, care instructions, anything printed on the item. Recent models handle short display text far better than they did a year ago, but a label a buyer will read is a different requirement from a headline that merely has to look like words.

Hands, and anything holding the product. The classic failure, still the classic failure, and it matters commercially because scale reference is one of the main jobs a lifestyle image does.

Reflections and transparency. Glass, gloss, chrome, liquid. The physics of what should be reflected in the surface is the part that goes wrong, and it goes wrong in a way that reads as cheap rather than as fake, which is arguably worse.

What rules already apply to a generated product image?

Two that bite today and one that starts on 2 August 2026. None of them is optional, and the first one is enforced automatically by software rather than by a regulator, which means you find out by having your products disapproved.

Shopping feed policy

Google's image requirements for Merchant Center products are specific in a way that matters here. Images must accurately display the entire product with minimal staging, with the product occupying between 75 and 90 percent of the frame. Placeholder or generic images that do not represent the actual product are prohibited, along with promotional overlays, watermarks, borders and single-colour blocks.

The line most merchants have not read is the one about provenance: images created with generative AI must retain metadata tags indicating AI origin, using IPTC DigitalSourceType properties such as TrainedAlgorithmicMedia, CompositeSynthetic or AlgorithmicMedia. In practice that means the export step matters. Strip your metadata to save a few kilobytes, or run the file through a tool that discards it, and you have removed the declaration the policy asks for.

The EU transparency rules from 2 August 2026

The European Commission's summary of the AI Act transparency rules puts the application date at 2 August 2026. Providers of generative systems must apply machine-readable marks to synthetic content and enable its detection, with an exception where the system performs an assistive function for standard editing or does not substantially alter the input or its semantics. Deployers, which in this context includes you, must clearly label content that falsely appears authentic. Systems already on the market before that date get a grace period until December 2026 for the marking obligations. Enforcement sits with national authorities and the AI Office, with penalties reaching 15 million euros or 3 percent of worldwide annual turnover.

Read the assistive-editing exception carefully, because it is where most of a small shop's actual work lands. Removing a background from a photograph of your own product, straightening it, adjusting exposure: that is standard editing and it does not substantially alter the semantics of the image. Generating a picture of a product that was never photographed is a different act entirely. The exception is the line between the two, and it is a more useful line than any vendor's marketing.

What if the customer does not receive a physical thing?

Then most of this loosens considerably, and it is worth saying so because the strictest advice on this topic is written for people selling objects.

A restaurant photographing a dish is representing a thing the customer will receive, so the same rule applies and generated food imagery is a bad idea for the menu. A service business is in a different position: an image of a clean, well-lit treatment room is staging, and generating it is closer to hiring a stock photo than to misrepresenting a product. A digital download, a course, a ticket, a subscription: there is no physical object to mismatch, and the honest constraint becomes the ordinary advertising one, which is that the image must not imply something the buyer does not get.

The awkward middle is anything where the photograph doubles as a specification, which is also where the advertising rules bite hardest, as our breakdown of which catalogue jobs a generated product image can honestly do works through job by job. A made-to-order table, a print in a frame, a hamper whose contents vary by season. In those cases the image is doing two jobs at once, illustrating and specifying, and the specifying job is the one that ends up in a dispute. The practical fix is to split the jobs: one honest photograph that specifies, plus whatever staging you like around it, with a line of copy saying which is which. That line costs nothing and settles most arguments before they start.

How do you keep the origin metadata on a file?

By treating export as a step with a rule rather than a button, because almost every convenience in the image pipeline strips metadata by default, which is also why a ComfyUI workflow embedded in a PNG survives some transfers and not others.

The places it disappears are predictable. Screenshotting a generated image instead of downloading it produces a new file with none of the original tags. Pasting into a chat app and saving the result usually strips them. Many compression tools remove all metadata as a size optimisation and describe it as cleaning. Some resizing scripts keep the pixel data and discard everything else. Each of these is a normal thing a busy person does, and each of them quietly removes the declaration the shopping feed policy asks for.

So the workable habit is short. Download the original file rather than capturing it. If you compress or resize, use a setting that preserves metadata, and check one file afterwards rather than assuming. Keep the generated originals in one folder so you can re-export if something downstream eats the tags. And record in your product data which images were generated, because your own record is the only version of this that survives a tool change.

None of this is difficult. It is unfamiliar, which is different, and it will stop being unfamiliar quickly now that a feed policy and a regulation both point at the same field. Detection is closing from the other side at the same time, as our report on sharper AI content detection and what it means for merchants sets out, and image detectors are now shipping alongside the text ones.

What does a real photographer cost for forty items?

Build the estimate rather than trusting a quote you cannot check. The US Bureau of Labor Statistics puts median photographer pay at 42,520 dollars a year, or 20.44 dollars an hour as of May 2024, and notes that commercial and industrial photographers documenting merchandise are part of that occupation, with many working self-employed and running their own business.

That hourly median is a wage, not a rate. A self-employed photographer quoting you covers equipment, insurance, editing time, travel and the hours they are not booked, so a commercial rate sits well above the wage figure. What the wage figure gives you is a floor and a sanity check: if you are estimating how long forty simple product shots take on a table with controlled light, that is a day of shooting plus a day of editing for most catalogues, and you can price two days of skilled self-employed time against your own quote and see whether it is reasonable.

Then price the other side honestly, which people rarely do. Doing it yourself is not free. Forty items at ten minutes each of shooting, sorting and cropping is nearly seven hours, plus the learning curve, plus the reshoots of the six that came out badly. Generating images is also not free: prompt iterations, the ones you discard, the metadata step, and the review pass where you check nothing invented a feature your product does not have.

So which should a small shop actually do?

Split the catalogue by what the image has to prove, not by what it costs.

1. Photograph anything the customer receives. The hero shot, the variant shots, the detail shot, the scale reference. Your phone with window light and a clean sheet is genuinely adequate for most product categories in 2026, and it is honest by construction.
2. Generate the staging, not the product. Backgrounds, surfaces, seasonal scenes, colour-matched banners. This is the assistive-editing zone and it is where the time saving is real.
3. Never generate texture, packaging text, or a hand holding the item. Those are the four failure modes and they map exactly onto the shots that decide a purchase.
4. Keep the metadata on export and note which images are generated. One column in your product sheet. It costs nothing now and it is the thing you will wish you had if a feed disapproves or a rule tightens.

Why the returns dispute is the whole argument

Marketing thinking treats a product image as persuasion. Operations thinking treats it as a representation, and a representation is a claim you can be held to.

When a buyer opens a dispute saying the item is not as described, the description under scrutiny includes the picture. A photograph of the actual item is evidence in your favour. A generated image that is close but not identical is evidence against you, and the closer it is the worse it looks, because a small deliberate difference reads as intent in a way a bad photograph never does.

This is why the cheap answer, generate everything and see what happens, is expensive in the one place it matters. The cost does not show up as a policy warning. It shows up as a refund plus return shipping plus a metric on your seller account, six weeks later, on the product that was selling best.

One more asymmetry worth naming, because it decides how much care each shot deserves. A generated image that is too flattering costs you a return. A photograph that is too honest costs you a sale you were probably going to lose at the door anyway. Those are not the same size of mistake, and the whole catalogue should be biased in the direction of the smaller one.

The version that keeps working

Real product images photographed on a table, generated staging around them, metadata intact, and one column in your catalogue recording which is which. That combination is cheap, compliant on both sets of rules, and does not need revisiting when the next model release changes what is possible. Whether your catalogue has somewhere to put that column at all depends on the starter you began from, which is what a commerce starter has to include before it is useful.

It also puts the work in the right place. The reason product photography feels expensive is rarely the photography. It is that the images then have to be resized, named, uploaded, matched to variants, given alt text sorted by the role each image plays and kept in sync when a colourway changes, which is catalogue plumbing rather than creative work. If your images live inside the same system as your product data, that plumbing disappears, which is the argument for an AI built storefront that owns its own catalogue rather than a shop stitched to a separate image library. The credit costs are worth setting against two days of a photographer's time and the hours of file management nobody quotes for.

If you are also generating the words that sit next to these images, the same honesty rule applies there and the failure modes are similar. Our piece on how much editing AI written shop copy needs before it ships covers the specificity problem, and the broader question of which hours of a founder's week AI genuinely gives back puts photography in context against the rest of the list.