Do you use AI to make your product photos?

We may restyle or relight a real photograph of the real product, for example replacing a cluttered background or correcting a colour cast. We never generate a product that does not exist, and never alter true shape, fit, fabric or colour.

How can I tell if a product photo was generated?

Check hands, hems and seams for lines that do not continue, look for texture that repeats unnaturally, and look for printed labels that are nearly letters. Then reverse image search the photo and ask the seller for a plain daylight shot.

Why does black clothing look different in different photos?

Because cameras have nothing neutral to balance against on a very dark subject. Black velvet in particular scatters light and commonly photographs blue or grey, so the same honest photograph can read as several colours.

How we work

Our policy on AI images, and what we will never do to a product photograph

2 November 2026 · 5 min read

A product photograph is a claim. It says this object exists, it is this shape, and it is this colour. Every editing decision either keeps that claim true or quietly breaks it, and image generation has made breaking it almost free. So here is the line we work to, in enough detail that you can hold us to it, plus the checks you can run on anybody else's photographs.

The rule, in one paragraph

We may restyle or relight a real photograph of a real product taken by the supplier or by us. We will never generate a product that does not exist. And we will never alter a product's true shape, fit, fabric or colour.

Those three sentences do a lot of work, because they separate the two things people lump together. Changing how an object is presented is editing. Changing what the object is is a lie. The first has been part of photography since photography existed. The second is new only in how cheap it has become.

What restyling actually means in practice

Supplier photographs are usually bad, and they are bad in consistent ways: a cluttered background, a creased backdrop, a prop nobody needs, strong colour cast from whatever lighting the factory had, dust, and a crooked horizon.

So we will replace or clean a background, crop, straighten, remove dust and lint, remove an unrelated prop, and correct exposure and white balance. We will relight an image, meaning adjust how light falls across the object, to make its actual form legible. On a dark object that is often the only way to show any form at all, because a black piece photographed flat is a silhouette.

None of that changes what the thing is. A charcoal object with the colour cast removed is still the same charcoal object, photographed better.

What we will never do

We will not change a silhouette. Not narrowing a waist, not lengthening a hem, not making a shelf look deeper, not straightening a curve the real object does not have.

We will not change a fit on a body, in either direction. We will not add a sheen that a matte fabric does not have, deepen a black to make it look richer, add a highlight to make brass look more polished, or invent a detail: a buckle, a stitch line, a texture, a piece of hardware.

We will not generate an object and present it as a photograph of a product. We will not generate a person wearing something nobody has worn, and we will not generate a room, place a product in it and let it read as a photograph of the product in a real room. If an image of that sort ever exists here it will say what it is, on the image, and we would rather not make them at all.

Colour is the hard part, and it is harder than it sounds

The rule says we will not alter true colour, and honouring it is genuinely difficult, because there is no single file that is the true colour of anything. The supplier photographed it under their lights, your screen is calibrated differently from ours, and a dark grey under warm tungsten is brown.

Black and near-black are the worst case. Black velvet photographs blue or grey almost every time, because the pile scatters light and a camera's white balance has nothing neutral to lock onto. A dress like the ruched velvet can honestly look three different colours in three honest photographs.

Hard shiny black is the opposite problem. A coat like the patent trench blows out into white highlights, and pulling those highlights back to recover the shape of the garment is legitimate, while flattening them away until the coat reads as matte is not.

What we do is correct towards what the object looks like in neutral daylight, in front of us where we have it in front of us, and say so on the page where a colour is hard to photograph. Brass is the other one worth flagging: cast brass shifts between yellow and brown depending entirely on the light, so a pair of cast brass bats will not look identical on your shelf and in our photograph, and no amount of honesty in editing can fix that.

Labels, and why we do not use many

Where an image has been produced or substantially altered with generative tools, we say so next to the image. Any realistic image of a person or of a lifestyle scene gets labelled, on every platform we post it to, without exception.

The reason you will not see many labels here is not that we are hiding them. It is that we have very little use for generated imagery. The only thing a product page needs is an accurate picture of the product, and the tool that produces that is a camera.

How to check any shop's photographs

Start with hands and hems, which is where generated images still fail most reliably. Count fingers. Follow a seam from one end to the other and see whether it stays on the same line. Look at where a hem meets a lining, where a strap meets a buckle, and whether the hardware could actually fasten.

Then look for repetition. Generated texture tends to repeat at a scale that real fabric does not, especially in knit, lace and anything woven. Printed text melts: a label, a care tag or a logo that is nearly letters but not letters is conclusive. Check that every shadow in the frame falls in the same direction, and that the reflections in any shiny surface match the room the object is supposedly in.

Then do two cheap practical things. Run a reverse image search on the main photograph, which usually reveals either the supplier's original or a hundred other shops using the same file. And email and ask for a photograph of the actual item on a plain background in daylight. A shop that has the object can send you one in a day. A shop that has only files cannot. For something like a printed flannel blanket, where the question is how the print sits on the nap, that one photograph tells you more than the whole gallery.

Why this matters more here than elsewhere

This shop has no reviews yet, so there is no crowd of previous buyers to correct a misleading photograph. That puts almost all of the evidential weight on the images and the measurements.

Which is a good reason to be strict rather than a reason to be flexible. If the photograph is the main evidence, then editing the photograph towards what we wish the object looked like is not marketing. It is removing the only thing you had to go on.

Questions people actually ask

Do you use AI to make your product photos?: We may restyle or relight a real photograph of the real product, for example replacing a cluttered background or correcting a colour cast. We never generate a product that does not exist, and never alter true shape, fit, fabric or colour.
How can I tell if a product photo was generated?: Check hands, hems and seams for lines that do not continue, look for texture that repeats unnaturally, and look for printed labels that are nearly letters. Then reverse image search the photo and ask the seller for a plain daylight shot.
Why does black clothing look different in different photos?: Because cameras have nothing neutral to balance against on a very dark subject. Black velvet in particular scatters light and commonly photographs blue or grey, so the same honest photograph can read as several colours.