A product shot has one hard requirement that a normal AI image does not: the object has to be your object. A model that invents a plausible bottle has failed, however good the picture is. So the technique is reference-locked restaging — attach the real thing, describe the set around it, and never describe the product itself. This page covers which models do that here, what each of them costs, and the two limits that catch people out on commercial work.
Text cannot specify a label, a bottle's shoulder curve or the exact green of a box. Ask for "a matte black skincare bottle on marble" and you get a beautiful photograph of a bottle nobody sells. The only thing that carries the real object into the render is a photograph of it, which means every product workflow here starts with a reference image and a prompt that talks exclusively about everything else.
The source photo does not have to be good, but it has to be honest: the whole product in frame, in focus, unobstructed, shot straight on or at the angle you want the result to keep. A phone photo on a kitchen table is a fine input. A photo where the logo is half cut off is not, because nothing downstream can invent the missing half correctly.
| Model | Runs on | Price | References | Plan |
|---|---|---|---|---|
| Krea 2 Identity | Our own GPU | from ⚡1 | 1 required, 2 max | Free and up |
| GPT Image 2 | A subscription we own | 0 credits | up to 6 | Free and up, 3 a day rising to 200 |
| Seedream 4.5 | Wavespeed | ✦1.5 | up to 6 | Starter and up |
| Seedream 5.0 Pro | Wavespeed | ✦2.2, ✦3.7 at 2K | up to 6 | Starter and up |
| Nano Banana Pro | Wavespeed | ✦4.7, ✦8.1 at 4K | up to 6 | Starter and up |
| Qwen Image Edit (Local) | Our own GPU | from ⚡1 | up to 3 | Free and up |
What it costs: Own-GPU prices start at ⚡1 an image. They are set at runtime, can carry a demand multiplier when the render queue is deep, and fall back to a gold-billed API lane if the local hardware is unavailable. GPT Image 2 rides a subscription we own, which is why it costs nothing and also why it is limited to a few runs a day and pauses for everyone when that subscription's allowance is spent.
Krea 2 Identity is the free lane and it is genuinely good at this, because restaging an attached subject is the exact job it was built for — it happens to have been built with people in mind, but a bottle is a subject like any other. Its "new location" recipe relights the object for the new setting, which is the hard part. Its "into a scene" recipe takes two images, and the order is fixed: the scene first, the product second.
The hosted models take a different approach — up to six references reasoned across together — which is worth the gold when you have several angles of the same product and want the model to understand its shape rather than copy one silhouette.
Load the studio backdrop prompt — Fills the composer with the studio recipe. Attach your product photo and press Generate yourself.
There is a ready-made version of this in the Presets tab — "Product on seamless" — which loads that prompt, sets a square frame and two copies, and ticks Seedream 4.5, Nano Banana Pro and GPT Image 2 together. On a free account the two hosted rows are locked and GPT Image 2 still runs, so the preset degrades into a working free recipe rather than a dead end.
Which plan: Free and Starter output carries an OpenModels watermark, applied when the result is saved. The switch to turn it off is a Creator entitlement, and it applies to new results only — it does not go back and re-save the pictures you already made. For product photography that ships to a store page, this is usually the deciding line rather than the credit price.
The second limit is storage. A free account holds 50 MB of assets, Starter 250 MB, Creator 2 GB. A product shoot of a hundred 4K renders is not a 50 MB job, and the per-file upload cap is derived from the same number — so the plan that fits a real catalogue is usually chosen by storage before it is chosen by credits.
Yes. Krea 2 Identity and Qwen Image Edit (Local) run on our own GPUs from ⚡1 against a wallet that refills to ⚡20 a day, and GPT Image 2 costs nothing for three runs a day on Free. The catch is the watermark, which stays on Free and Starter output.
Mostly, and not reliably at small sizes. Text is the first thing to soften. Give the model a reference where the label is large and sharp, prefer a model that accepts several references, and check the result at full size before shipping it.
There is no stable answer, which is why the composer runs several at once. Seedream 4.5 is the cheapest hosted option at ✦1.5, Nano Banana Pro is the most expensive at ✦4.7, and Krea 2 Identity is the free one. Run all three on your own product and keep the winner.
Yes — open the finished image and use the lightbox Relight tool, which orbits a key light around the subject and re-renders. It is billed in ✦ gold, so it needs a paid plan or a credit purchase.
Yes, subject to the Terms. What differs by plan is the watermark: it is always on for Free and Starter results and switchable from Creator up.