AI product photos: what actually changed

my business sells clothing on a marketplace. we have made ai product photos for ecommerce three different ways in about two years, and today a hundred finished cards can appear overnight for roughly a dollar each, sorted into “publish this” and “look at this one”.

one photo of the product goes in. what comes back is a listing: frames from several angles, the infographics, a short video, the text. nobody is at the desk while it happens.

getting there was not about finding a better image generator. the bottleneck never disappeared. it moved twice, and both times it took us a while to notice.

era one: the studio day

get the samples. book the studio. book the photographer, the model, the stylist. shoot. then about two weeks of retouching before anything is ready to publish.

for seventy to a hundred items that ran me about ten thousand euros, and that was with a large discount. but the real cost was the calendar. a new item could not go live until a shoot happened, and a shoot only happened when enough items had piled up to justify one. so the whole catalogue moved at the speed of the slowest photoshoot.

era two: ai, one photo at a time

then the tools got good, and we did what everybody does: generated cards one by one. a person uploads a photo, picks a style, waits, looks, approves or asks again.

per image, this was a transformation. two years earlier that same frame needed a studio and a fortnight. now it took minutes.

fine for one product. useless for a catalogue — you sit there typing prompts all evening for a handful of pictures.

and then we ran a real batch, and the team rejected it.

what does ai get wrong in a product photo?

this is the part that usually gets left out of these articles, so here it is with numbers. thirty-one cards reviewed by the two people who actually publish them. the complaints weren’t random taste. the same defects repeated:

• no close-up frame. about sixteen cards. the brief asked for one every time; the engine treated it as optional.
• the model’s face copied the reference. five cards, and it was the verdict that killed the batch. left alone, the generator borrows the identity from your reference image, so a batch comes back looking like one person in different outfits.
• wrong location. about eight. the brief said street, the output said studio.
• wrong garment. four. a skirt instead of trousers, the wrong length, a wrap version instead of the solid one.
• plus duplicate poses, washed-out colour, the model standing too far away, and one frame that came back as a collage.

eleven distinct defect classes out of one batch.

here is the useful reading of that list. every one of those is an engine problem or an input problem, not a taste problem. which means they’re fixable, but only if somebody collects them in that form instead of saying “the photos are bad”.

and the deeper lesson: we had moved the bottleneck out of the studio and into the reviewer. generating fifty cards was now easy. looking at fifty cards was not. the person checking became the queue.

era three: the factory

what we built next is not a better image model. we did not train anything. the models are the same ones anybody can rent.

three unglamorous things made the difference.

a judge before the money. the prompt is checked against the actual product before a single paid generation runs. no styled brief, a studio scene where the brief said street, no lock on the actual garment, and it gets sent back to be fixed. rejecting a bad prompt costs nothing. rejecting a bad photo costs a generation.

a second judge on the output, which marks each card done, done-but-unverified, or needs-a-human. that third status is the honest one, and it is the reason the system is trusted: it never claims a card is checked when the check did not run.

the complaints go back into the machine. the team writes what is wrong in a column of the same spreadsheet they already work in. the factory reads it, replies with what it understood, and regenerates. no meeting, no ticket, no translation layer between the person who saw the problem and the thing that caused it.

around that: batches run straight from the spreadsheet, up to fifty at a time, roughly six to seven minutes a card, and the results land in dated folders ready to publish. if you want to see how the thing is built, here is the machine itself.

before — the input photo

after — factory card

after — factory card

after — factory card

what do ai product photos actually save you?

this is the part i undersold to myself for months, so let me put it plainly.

a hundred cards overnight. not a hundred images. a hundred finished cards, each with its frames and its text, sitting in a dated folder in the morning. nobody sat at a screen while it happened. compare that with a studio day, which produces one shoot’s worth of items and then two weeks of retouching before anything is publishable.

the judge is honest about which ones are good. it doesn’t hand you a hundred cards and a shrug. it marks each one: accepted, done-but-unverified, or needs a human. so in the morning you aren’t reviewing a hundred cards, you’re reviewing the ones it flagged. that is the difference between a tool that produces work and a tool that produces sorted work.

and it costs about a dollar a card. against a shoot, that isn’t a discount, it’s a different category of expense. against ai content creation one photo at a time, the saving isn’t the money at all. it’s that a person no longer sits inside every single card, uploading, waiting, choosing, re-asking. that person’s day is the resource you are actually buying back.

put those three together and the change is not “we generate images faster”. it’s that catalogue photography stopped being an event you schedule and became something that runs while you sleep, with a queue of exceptions waiting for you in the morning.

that is the honest sales pitch for this whole approach, and every part of it depends on the judging. volume without a judge gives you no idea what’s publishable, which is worse than slow, because now the bottleneck is your own eyes and there is no list telling them where to look.

the number that nearly cost us a batch

before a batch starts, the system checks whether there is enough credit to finish it. that check was originally set at fifty cents a card.

the measured cost, across real batches, was about one dollar and five cents per card of four frames, including both judging passes and the occasional regeneration. our estimate was off by half, in the direction that lets a batch begin and die halfway through with money spent and nothing publishable.

we raised the gate to a dollar twenty and moved on. small thing, but it is the difference between an automation you trust and one that leaves you with a half-finished batch on a friday.

if you take one operational habit from this: measure the cost of your ai on real runs, not on the pricing page.

which products does ai photography quietly get wrong?

the failure modes here are documented and boring, not bad luck, and they are worth knowing before you promise your catalogue to a machine.

reflective and transparent goods come back with reflections that are physically impossible — bottles, glass, glossy electronics. jewellery loses the fine detail that is the reason someone buys it. small logos and text come out garbled. striped and plaid fabric will not stay aligned as it wraps a body, so the pattern drifts across a seam.

for those items, keep a photographer or budget for heavy correction. and understand what a wrong photo actually costs: if the image shows a colour or a texture the real product doesn’t have, the customer orders, receives something else, and sends it back. inaccurate photos push returns up, not down. that is why the culling step is the thing you are paying for, not an add-on to it.

what can ai product photography still not do?

it does not remove the person. someone accepts the batch, and someone still writes the brand style: what the model looks like, what the locations are, what the brand refuses to look like. that judgement is not in any tool.

it needs a clean reference of the actual product. a bad input photo produces a confidently wrong output, and no amount of prompting fixes a reference that shows the wrong garment.

it needs identity forced fresh on every card, and checked. that one is worth repeating because it is the single defect that killed our first batch.

and it does not invent search keywords. our text agent writes the title and the description, and where the keywords go it writes a note saying an editor adds them from real search data. inventing plausible keywords is easy and worthless.

do you have to label an ai product photo?

if you sell into the eu, one date matters: the second of august 2026, when the transparency rules of the eu ai act start to apply. in plain terms: an ai image has to be marked in a machine-readable way and detectable as artificial, and when a picture could pass for a real photograph the shopper needs a disclosure they can see. standard light editing that doesn’t really change the input is exempt. (source)

two honest cautions. the common answer is “just use content credentials” — they help, but a manifest gets stripped by a screenshot or a re-upload, so the eu’s own guidance points at a layered approach: metadata plus watermark plus logging. (source) and regulators themselves admit no watermarking method today is fully robust against removal, so this is a moving target rather than a solved problem. (source)

marketplaces draw their own line, and it is stricter than “ai is fine”. the main product image is expected to show the actual product. background swaps, colour correction and lighting are usually allowed; a fully synthetic representation that misrepresents the physical item can be flagged or rejected regardless of any disclosure. (source) the consumer-protection angle is the same idea from the legal side: an image that materially misrepresents colour, size, features or what is in the box is deceptive whether or not ai made it, and “the ai did it” is not a defence. (source)

none of that argues against ai photos. it argues for accurate ones, disclosed properly — which is the same reason the human stays in the loop.

who is ai product photography for?

not everyone. if you have five products, shoot them properly once and move on.

this earns its keep when you have a wide catalogue, dozens to hundreds of items, new ones every month, and no appetite to pay studio money each time. the machine takes the routine and the night shift. the person stays on the one decision that matters: is this good enough to put my name on.

how can you copy the approach without our tools?

none of this is specific to our tools or our marketplace. it works the same for a shopify store, an amazon listing or any catalogue where photos are the product.

1. write the standard down first. how many frames, what each frame must show, what the model and the locations look like. if it isn’t written you can’t judge output against it, you can only have opinions.
2. put a check before the spend, not only after. it is the cheapest quality gain available.
3. give rejections a structured home. a column in the sheet people already use beats a chat thread. you want defects as a list you can count, because that is what turns complaints into engine fixes.
4. allow a “not checked” state. systems that only say pass or fail will lie to you eventually.
5. measure cost per finished unit on live runs and gate on that number with a margin.
6. keep a person on acceptance. the goal is to remove the studio day, not the judgement.

that is the whole method. the model you use will change within the year; those six things will not.

related: what happens when the robot refuses, what an ai agent build actually costs and when your stock numbers lie to you.

faq

are ai product photos good enough for ecommerce? good enough to sell, not good enough to ship unchecked. in our first serious batch the team rejected most of it, and the reasons were consistent rather than random: a missing close-up, the model’s face copied from the reference, the wrong location, the wrong garment. all of those are fixable in the engine. what is not fixable is shipping without anyone looking.

how much does an ai product photoshoot cost? ours measured about one dollar per card of four frames, including the judging passes and the occasional regeneration. for comparison, the studio route for seventy to a hundred items cost me about ten thousand euros and weeks of calendar, and that was with a large discount. measure your own number on real batches before you budget, because our first estimate was off by half.

what is the real bottleneck with ai product photos? reviewing. generating an image stopped being the hard part. once you can produce fifty, somebody has to look at fifty, and that person becomes the queue. the fix is not a better image model, it is a check that runs before you spend money and a second one that runs on the output.

which products do ai photos get wrong? reflective and transparent goods come back with impossible reflections, so bottles, glass and glossy electronics are the worst cases. jewellery loses fine detail. small logos and text come out garbled. striped and plaid fabric will not stay aligned as it wraps a body. for those items keep a photographer, or budget for heavy correction.

do i have to tell shoppers a photo was made by ai? in the eu, yes, from the second of august 2026. an ai image has to be machine-readable as artificial, and a picture that could pass for a real photograph needs a disclosure the shopper can see. light editing that does not really change the input is exempt. separately, marketplaces expect the main image to show the actual product, so a synthetic shot that misrepresents it can be rejected even when it is disclosed.

how do i stop ai photos from all looking the same? force a new identity on every card and check it. left alone, the model copies the face from your reference image, and a batch comes back looking like one person in different clothes. that single defect was the reason our first batch was rejected.

if you’re producing catalogue photos now, with a studio or with ai one at a time, send me how it works today and what your reviewers complain about. i will tell you where the bottleneck actually sits, which share of your catalogue mix is realistically automatable and which part stays with a photographer, and what a check-before-spend would look like for you. the first conversation is an hour and it is free. book a time.

Created with AI assistance.