Product photography is the one commercial use where generated images are already good enough to ship — and the one where getting it wrong has consequences beyond a bad picture.
- Never generate the product itself. Photograph it once, then generate everything around it. This is the whole method.
- A generated product is a misrepresented product — the label will be wrong, and that is a returns and advertising-standards problem, not an aesthetic one.
- Backgrounds, scenes and lifestyle shots are where the saving is, and they carry almost no risk.
- Match the light to the plate. If the scene light and the product light disagree, the composite reads as fake instantly.
- Marketplaces have rules about what the main image may show. Check before you build a workflow on it.
- Consistency across a catalogue matters more than any single image being beautiful.
Treat the product as a fact and the world as a set.
The customer is buying the object. Get the object right by photographing it, once, properly. Everything else in the frame — the marble counter, the morning light, the linen — is set dressing, and set dressing is what generation is genuinely good at.
The rule that makes this safe
The product is photographed; the world is generated
A generated version of your product will be nearly right. The logo will be close. The label text will be plausible. The number of buttons may differ. The stitching will be in the wrong place.
"Nearly right" is the worst possible outcome for a product image, because nobody notices until a customer receives something that does not match, and then it is a return, a complaint, and potentially a claim that your advertising was misleading.
So: one clean photograph of the real item on a plain background. Phone-quality is usually fine if the light is even. That plate becomes the fixed element in everything you produce afterwards.
What you generate instead
Generate the room, not the thing being sold. Every euro of risk in this workflow sits in the object — and the object is the one part you can photograph in five minutes.
Making the composite hold up
Light is what gives a fake away
The most common failure is not a bad background. It is a product lit from the left sitting in a scene lit from the right. The eye catches it immediately even when the viewer cannot say what is wrong.
Two ways to solve it, and the first is much easier:
Shoot the plate with the scene in mind. Decide the light direction before you photograph — "window light from the left, soft" — then prompt every background to match that description.
Or generate the scene first and match your lighting to it. Slower, but useful when the scene is fixed by a campaign.
Also match the shadow. A product with no contact shadow floats. A generated surface with a shadow in the wrong direction is worse than no shadow at all.
A prompt shape that works for backgrounds
Backgrounds are the workhorse, and they want the opposite of a dramatic prompt: specific, plain, and deliberately unexciting, with room left for the product.
The "leave space" line is the one people forget, and without it you get a beautifully composed scene with nowhere to put the thing you are selling.
The rules that apply to you, not to the picture
Marketplaces set their own image standards, and they are stricter about the main listing image than about secondary ones — commonly requiring the product on a plain background, filling most of the frame, without added text or graphics.
Check the specific marketplace's current image policy before you build a pipeline on it, because a workflow that produces non-compliant main images is worse than no workflow.Marketplace image requirements differ by platform and change; this page deliberately quotes none of them. The pattern — strict main image, freer secondary images — is common, but read the policy for the platform you sell on. Checked 31 Aug 2026
And advertising rules apply to generated imagery exactly as they do to photography. An image that materially misleads about what arrives in the box is a problem regardless of how it was made. Generation does not create a new category of allowed exaggeration.
Hands holding your product, a model wearing it. Two things follow. A fully synthetic person is not a real person, so there is no likeness to license — but if it is modelled on a real one, see cloning a voice or a face.
And a synthetic model is a claim if it implies a real customer. A stylised lifestyle shot is normal advertising; a generated "customer photo" in a review section is fabrication.
Before a generated product image goes live
No marketplace rules, no prices and no tool ranking are quoted here — all three change often enough that a specific claim would mislead within weeks. What is durable is the method: photograph the object, generate the world, match the light, and check the label.
The compositing and lighting advice is craft reasoning rather than measurement. The legal points are pointers to the pages that carry the sources and dates. Checked 31 August 2026.
The through-line: the customer is buying the object, not the picture. Photograph the object and generate everything else, and the risk in this workflow drops to almost nothing.