The honest version of "AI can do your product photography" is "AI can do most of your product photography, and you need to know which part it cannot." A tool that produces a clean white-background image from a phone photograph in a minute is easy to reach for when the job is a lookbook that has to show how a coat moves, or a chrome kettle in a sunlit kitchen, or a claim about a product's texture that a customer will compare to the delivery. Those images come out of the tool looking finished, which is what makes them dangerous.
This page is a list of the cases in which AI product photography is not the right choice, sorted by the reason. It publishes no share of images that fall into them. What it gives is a way to recognise each case before you start — a sign in the brief, not in the output — and what to use instead.
When is AI product photography the wrong choice?
When the image has to show something a photograph of the product does not contain, or has to carry a claim the image itself is evidence for. The scene appearing in a reflective or transparent surface; the fit, drape and movement of a garment on a body; the product in a hand, in use, at its real scale against a person; a texture or a finish the customer will compare to the delivery; a product that does not exist yet or exists only as a prototype; a regulated or evidentiary image; a hero image whose entire value is that it happened. In those cases photograph it. AI product photography is the right choice for everything else: plain grounds, studio sets, scenes generated around a product whose pixels are kept, catalogue volume, variants of a photographed product.
What "not the right choice" means here
It does not mean the tool will refuse or fail visibly. It means the image will look finished and be wrong in a way the customer, the channel or a regulator can find: a scene that should be in the surface and is not, a fit the garment does not have, a detail the product does not have, a claim the image cannot support. The distinction that runs through every case is between what a photograph of the product contains — its pixels, which a good workflow keeps — and what the image is being asked to add.
Case 1 — the scene has to appear in the product
Chrome, mirror-polished metal, gloss lacquer, patent leather, clear glass, liquids, gems. Their surfaces return or pass the surroundings, so the surroundings are part of the product's pixels. A kept product carries its old room; a redrawn product is a different product.
- The sign: a reflective or transparent product plus a scene that is not plain and softly lit.
- Instead: photograph the product in that scene, or on that ground, and generate only what the surface cannot see. Plain scenes from a neutral master remain fine.
Case 2 — the image has to show fit, drape or movement
How a coat hangs open, how a dress moves, how a knit stretches over a shoulder, how trousers break on a shoe. A generated model image shows a garment at rest, on a body it did not meet. It is the right tool for a catalogue view; it is the wrong tool for the question the customer is asking when they look for movement.
- The sign: a brief that uses the words fit, drape, movement, walk, stretch or "how it wears".
- Instead: a fitting on a person or a mannequin, photographed; a short film for movement. Keep the generated images for the flat and the catalogue view, where the garment's own pixels are what matters.
Case 3 — the image is evidence
A claim about the product that the image supports: the finish of a surface, the exact colour of a dye lot, the size of a stone against a scale, the contents of a set, the state of a second-hand item, a before and after. A regulated category where the image is part of the record. A customer dispute where the listing image is compared to the delivery.
- The sign: the word "shows" in the brief — the image shows that, proves that, matches that.
- Instead: a photograph, kept with its original file, and no generated version of the same claim. Generated images of the same product can sit in the gallery as long as the claim rests on the photograph.
Case 4 — the product does not exist yet
A prototype, a render, a sample in the wrong colour, a product that will be made if enough people order it. AI product photography starts from a photograph of the product; without one it starts from a description and draws a typical product, which is the wrong one.
- The sign: no photograph of the final product to upload.
- Instead: wait for the sample, or publish the render as a render and say so. A photographed sample in one colour can be used for that colour only.
Case 5 — the scale or the hand is the point
A ring on a finger, a bag on a shoulder, a lamp beside a chair, a tool in a hand: images where the customer is asking "how big is it" or "what is it like to hold". A generated hand or body is a scene element, and a scene element does not know the product's size.
- The sign: a size question the listing has to answer visually.
- Instead: photograph the product with the person or the reference, or put a dimensioned drawing beside a generated image. A generated product-in-hand image can suggest use; it cannot be the size reference.
Case 6 — the value of the image is that it happened
A hero shot for a launch, a founder holding the first unit, the product in the place it was made, a campaign whose story is a real day. A generated image of these is an illustration of a story that did not happen, and the audience knows the difference when they are told, which they will be.
- The sign: the brief is a story, not a product.
- Instead: a photographer, for that image only. The catalogue around it stays generated.
Ask, sort, shoot
The loop the table below serves, named so the other decision pages can point at it.
- Ask — what must this image show that a photograph of the product does not already contain?
- Sort — nothing: generate it. Something in the six cases: photograph it.
- Shoot — the few images that need a camera, and generate everything else from the same session's masters.
The decision table
| The image has to | Right tool | Why |
|---|---|---|
| Show the product on white, grey or a plain studio ground | AI, from a photographed master | The product's pixels are kept; the ground is generated |
| Show the product in a soft, plain lifestyle scene | AI, from a photographed master | The scene is generated around the kept product |
| Show a scene inside a reflective or transparent surface | Photograph | The surroundings are part of the product's pixels |
| Show fit, drape or movement on a body | Photograph or film | A generated model shows the garment at rest, on no one |
| Prove a finish, a colour lot, a size, a condition | Photograph, kept as the record | The image is evidence; a generated image is not |
| Show a product that has no final sample | Neither; wait, or label the render | There is no photograph to keep |
| Answer "how big is it" or "what is it like to hold" | Photograph with a reference | A generated hand is a scene element, not a scale |
| Tell a story that happened | Photograph | The value is that it happened |
| Fill a catalogue with variants, angles and grounds of a photographed product | AI | The product's pixels are kept across every image |
The first two rows and the last one are most of a catalogue; the six between them are the images a photographer is for.
One brand, one week
A small outerwear brand plans a season: eight coats, each in three colours, each needing a white main image, two studio angles, a lifestyle scene and a lookbook image that shows the coat worn open and walking. It photographs every coat on a mannequin in a white tent with a grey card, and generates the white, the studio angles and a soft daylight scene from those masters — ninety-six images, all with the coats' own pixels. The lookbook is eight images, and those are case 2: a fitting on a person, photographed in a morning. One coat has a patent finish; its lifestyle scene is case 1, so it is photographed on the walnut table instead of generated there. The week's camera time is a morning and an hour; the rest is generated, and no image claims what a photograph did not contain.
Frequently asked questions
Is AI product photography good enough to replace a photographer?
For most of a catalogue — plain grounds, studio sets, soft scenes generated around a photographed product, variants and volume — yes, because the product's own pixels are kept. For the six cases on this page — a scene that must appear in the surface, fit and movement, evidence, a product with no sample, scale in a hand, a story that happened — no, and the sign is in the brief before the first generation.
How do I know before generating whether an image will be wrong?
Ask what the image has to show that a photograph of the product does not already contain. If the answer is nothing beyond a ground or a scene, generate. If the answer is a reflection, a fit, a proof, a size, a product that is not there, or a real event, photograph.
Can I use AI for the catalogue and a photographer for the hero images?
That is the arrangement this page recommends: the same session's masters feed the generated catalogue, and the few images that need a camera — the lookbook, the reflective product in a real scene, the launch shot — are photographed. The camera time shrinks to those images.
What about products that are hard but not impossible — jewellery, glass, labels?
They are AI cases with conditions: a neutral master for reflective surfaces and plain scenes; a master shot on the publishing ground for glass; labels kept as photographed pixels and read character by character. They become photograph cases only when the scene has to be in the surface or the image has to be evidence.
Is a generated lifestyle image dishonest?
Not when the product's pixels are the photograph's and the scene is presented as a setting rather than a claim. It becomes dishonest when it shows something the product does not have — a fit, a reflection of a place, a feature — or stands in for a photograph the customer is entitled to take as evidence.
What this page does not claim
That any share of product images falls into the six cases, or that a photographer costs more or less than a workflow; this page measures neither. What it claims is that the line runs between what a photograph of the product contains and what an image is asked to add, and that the sign is readable in the brief before anything is generated.
To generate the part of the catalogue the tool is for, see AI product photography; for the case of reflective products in depth, AI product photography for products with reflective surfaces; for what goes wrong when the wrong tool is used anyway, a failure taxonomy for AI product photography.