The main limit: the product cannot be generated
Let us start with what will save you money and reputation.
AI should not draw your product. It will draw a similar one: a mug, but with a different handle; a sneaker, but with a different seam; a dress, but with a different neckline. The buyer orders what they saw, receives something else and files a return. Plus a bad review, plus complaints from the platform.
So the rule is simple and not up for debate: a photo of the real product is mandatory. One decent phone shot in daylight, and AI does everything else.
And there is a lot of everything else.
What AI does brilliantly
Replacing the background
The most common task. You have a photo of the product on a kitchen table and need one on a white background or a nice podium.
Upload the shot to an editing model and ask it to replace the surroundings. The model keeps the item and redraws everything around it. Editing models do this best: Nano Banana Pro is specifically strong at changing the context without touching the object.
What to ask for: «the same item, white seamless background, studio light from above, soft shadow under the object». The shadow matters: without it the item looks pasted on.
The product in an interior
The buyer wants to picture the item in their own home. This used to require a location shoot; now a prompt is enough.
An armchair in a Scandinavian living room, a lamp on a bedside table, dishes on a laid table. One product photo gives rise to a dozen usage scenarios, and that directly affects the card’s conversion.
Be careful with scale: models often get sizes wrong. Check that your mug has not grown to the size of a bucket.
Clothing on a model
The most expensive scenario in real life and the most profitable one to replace. Instead of a shoot with a live model, make-up artist and location, you generate a person and dress them in the real item.
This needs special control. Check every frame: has the cut changed, has the colour drifted, has the print disappeared, is it the same material. The model will happily «improve» your item, and that is misleading the buyer.
If the item on the model looks different from real life, better not to publish it.
A video review from one photo
Few people use this, and wrongly so: video in a card noticeably raises engagement.
Take a product photo, give it to a video model in image-to-video mode and ask for a slow camera orbit around the item. You get six seconds in which the product can be examined.
How to do it step by step is covered in the article on how to animate a photo with AI. Ask for minimal motion. A slow push-in or a gentle orbit. Do not ask to rotate the item: the model will start inventing the back it has never seen, and it will invent it wrong.
A special pain: text on infographics
Infographics with captions are half of a marketplace card. For a long time they were a problem: instead of Cyrillic letters, models drew a mess of lookalike symbols.
Now it is solvable. Nano Banana Pro writes Cyrillic cleanly if you meet three conditions:
- Short. A headline of 1 to 4 words and 2 or 3 badges of a few words each. A long paragraph can still get distorted.
- Word for word. The caption text goes into the prompt in quotation marks, exactly as it should appear on the card.
- Away from the edges. At the top and bottom of the card the platform places the heart icon, promo labels, rating and price. Captions there will be covered anyway.
The easiest way is the «Marketplace card» mode: upload a product photo, write the name and 2 or 3 benefits or press «Come up with a card», and the AI assembles a 3:4 card with captions. One card at 2K costs 16 tokens, about 25 rubles.
Before publishing, check every letter and digit. If a caption still comes out crooked, regenerate or fix it in a graphics editor.
What it costs
KADR prices, one image, October 2026.
| Model | Price, rubles | Best for |
|---|---|---|
| Flux.2 Pro | 6 | photorealism, cheap, for mass runs |
| Seedream 4.5 | 8 | 4K, detail |
| Nano Banana 2 | 13 | quick editing |
| Nano Banana Pro | 25 | best background and surroundings replacement |
| Recraft v4.1 Pro | 38 | graphics, vector, design tasks |
| GPT-5 Image | 46 | understands the most complex descriptions |
This is not a public offer.
A full set of images for one card, eight pictures, comes to roughly 25 to 370 rubles depending on the model. A 6-second video review costs from 51 rubles on Veo 3.1 Lite.
For comparison: a product shoot with a photographer costs thousands of rubles per item, plus logistics, plus lead time. The video cost breakdown for every model is here.
A workflow for volume
If you have not one product but a hundred, the scheme is this.
- Shoot all products the same way. One background, one light, one angle. It is the only manual stage, and it pays off.
- Debug one prompt on one product. Not on all of them at once. Run drafts on a cheap model.
- Run the batch with the same prompt. All cards will share one style, and that alone looks expensive.
- Check each one. This step is mandatory. The model will spoil the product somewhere, and you should find it, not the buyer.
- Check the captions. Nano Banana Pro writes short text cleanly, but still look at every letter and digit with your own eyes.
Twenty cards in an evening is a realistic figure.
Will the card get banned
The short answer: the issue is not whether the image is generated but whether the product in the picture matches the product in the box.
Allowed: changing the background, adding an interior, placing the product in a usage scenario, generating a model for clothing, making a video review, drawing diagrams.
Not allowed: changing the colour, contents or material, adding features that do not exist, distorting the size.
A simple test: if the buyer opens the parcel and says «this is not what was in the photo», you are in for returns and bad reviews. Everything else is ordinary product presentation.
A separate note on rights: authorship of AI images is not straightforward, and if you sell visuals it is worth understanding the risks. There is a breakdown in the article on copyright for AI images.
The model may refuse: what to do
One more situation to be ready for. You upload a photo, press generate and get a refusal.
It most often happens when there is a person in the frame: models filter realistic photos of people so they cannot be used for deepfakes. For example, Seedance 2.0 is explicitly marked in its description as not suitable for photos of real people. It works great for product shots, but if you try to animate a photo of a seller or a model wearing your clothing, you will most likely hit the filter.
The solution is simple: switch models. Developers have different policies, and what one blocks will pass calmly with another. With product shots there are almost no refusals; they begin where there is a live face in the frame.
Background replacement in detail, with prompts: change a photo background with AI. For sales: banners for 11.11 and Black Friday.
In short
AI does not replace the product photo, it replaces everything around it: background, light, interior, model, video review. One real shot plus a few dozen rubles of generation give a full set of images for a card.
Captions on the card can now be made right in the AI: short, word for word and with every letter checked. And keep the one rule in mind: change the surroundings, not the product.
