The AI photo looks great.
It just is not quite your product.
The label reads almost right. The lid is a shade off. The proportions drifted just enough that a buyer who has one in their hand would notice. That is fine for a mood board and a problem on a listing, where a photo that oversells what arrives is a return, a one-star and eventually a policy warning. And it is only half the job anyway: one good image is not a listing. A listing is seven, of the same object, lit the same way, in three aspect ratios because you sell in three places.
The shot list first, the prompts second
Each slot carries a selling point, not an angle
Hero on white for the category rule, one in context, one in a hand for scale, a detail crop on the part buyers ask about, a size comparison if returns say people get it wrong. Written that way, the fourth photo has a job. Written as more angles, the fourth photo is the same product turned slightly and nobody scrolls to it.
Reference editing, so the object stays yours
GPT Image 2, Seedream 5.0 and Nano Banana all take a reference image and edit rather than invent, which is the property that matters when the thing in the frame has to be the thing in the box. You bring one decent photo from a phone; the backgrounds, the surfaces and the lighting are what change.
Then the boring part, done in one go
Square for one marketplace, taller for another, a caption in a second language for the third, plus the A+ modules and the first frame of the short video. This is the part you put off until Sunday night. It is also the part that is purely mechanical once the set exists, so it runs while you do something else and the files come out named in a way you can still read in six weeks.
Where a generated photo gets you in trouble
A listing pulled for a misleading image costs a great deal more than the photoshoot you skipped.
Label text and geometry are not guaranteed
Reference editing keeps the product recognisable. It does not guarantee that the packaging copy, the logo and the exact proportions come through intact on every generation, and the tighter the crop the more that shows. Everything that goes on a listing needs your eyes on it before it goes up.
No reference means it is guessing at your product
If you have no photo of the item, what comes out is a plausible object of that description, and putting a plausible object on a marketplace listing is a compliance problem rather than a design one. One decent phone photo is enough to start. Nothing is not.
Competitor images are not reference material
Unauthorised likenesses, competitor material, music, IP and trademarks are out, and that is a fixed line rather than a setting you can flip. Looking at a competitor's set to decide what yours should show is a different thing entirely and perfectly sensible.
Twenty-one generations cost twenty-one generations
Seven shots across three aspect ratios is not one image, and high-volume or high-cost runs ask for confirmation before they start rather than after the invoice. You can point the work at your own provider account if you would rather pay them directly.
What people ask before generating a set
Why not just use a free prompt list?
Use them. They are good and they are free, and for one image in a browser tab that is the right tool. The gap shows up at the twentieth file, when you need the same product across a set, at three sizes, with consistent lighting and a naming scheme you can upload from. That part is a workflow, and a folder on your own disk holds it better than a tab does.
Does it know the marketplace image rules?
It works from the constraints you give it and what your category page shows, and it writes them into the brief, so the hero comes back on pure white with the product filling the frame the way the rule asks. It is not an official checker and it does not promise your images pass review. Treat it as the first pass, not the approval.
How much cheaper is this than a photographer?
That depends on your category and we are not going to invent a number. What is worth saying is what it does not replace: if the product has fabric drape, glass, food texture or a finish that has to be seen, a real shoot still wins and generated frames go around it rather than instead of it. Where this earns its place is the twentieth SKU and the seasonal re-cut of images you already have.
Can it do the short video too?
It writes the script, the shot breakdown and the prompts, and it can drive Seedance, Kling 3.0, Veo 3.1, Runway Gen-4.5, Hailuo or Vidu to generate. Same caveat as the stills, more so: motion is where product geometry drifts most, so review before anything goes near an ad account.
You already have one decent photo of it
Free, macOS and Windows. Image models run on your key or an Orkas plan.