Ecommerce is the category where the rules bind hardest, because a product image is a representation a buyer relies on and marketplaces regulate the main image specifically. So the honest answer starts with the line: photograph the product, generate the world around it. Within that line, the picks are Nano Banana Pro at 21 credits when the product must stay exact and Seedream 5 at 6 credits for volume context.
The line
Generate: lifestyle scenes around a real product photo, seasonal variants, backgrounds, environments, model-free context, social crops, ad creative.
Do not generate: the product itself, its finish, its texture, its label text, or anything that changes what a buyer would receive. A generated version of your product is a picture of a thing that does not exist.
Check the marketplace rules before anything goes on a listing. Several require the main image to be an accurate photograph of the actual item on a plain background, and enforcement is real.
The commercial reason, beyond compliance: returns. An image that flatters beyond the product produces a buyer who sends it back and leaves a review about it.
Model picks and per-image costs
Canon 8frame prices, a credit is $0.01 at pack rate:
| Job | Model | Credits |
|---|---|---|
| Volume context and backgrounds | Seedream 5 | 6 |
| Product-accurate scene from a real hero | Nano Banana Pro | 21 |
| Cheap iteration | Nano Banana 2 | 11 |
| Clean cutout | Background removal | 12 |
| Finishing | Image upscale | 13 |
| Instruction edit on a real photo | Flux Kontext Max | 13 |
A twelve-SKU catalogue with four context variants each: roughly 288 credits on Seedream, about $2.88. Against a second shoot day with a stylist and props, that is the argument, and it holds without touching the product.
The workflow that works
1. Photograph one clean hero per SKU. Correctly lit, plain background, accurate colour. This is the irreplaceable asset and the only mandatory capture.
2. Cut it out. 12 credits.
3. Generate the worlds. Kitchen, bathroom, outdoor, seasonal, whatever the listing and ads need, conditioned on the real hero so the product stays itself.
4. Check colour against the physical item. The most common failure: a generated scene shifts the product's colour slightly, and colour is what gets returned.
5. Upscale last.
Save the chain as a reusable workflow and the next product launch runs in an afternoon.
Where this breaks
Materials that are the product. Fabric, jewellery, cosmetics texture. Buyers zoom in, and generated material is wrong in exactly the way that loses the sale.
Colour-critical categories. Paint, cosmetics, textiles. Photograph and colour-manage properly.
Anything with regulated claims. Supplements, medical, food. The imagery has to match the physical item and the compliance surface is not somewhere to experiment.
Scale and fit. Generated scenes get proportion subtly wrong, which matters most in furniture and apparel.
FAQ
Can I use AI images on Amazon or similar marketplaces? Check the specific marketplace's current image policy. Several require an accurate photograph for the main image; supplementary lifestyle images have different rules. This is a compliance question, not a tooling one.
Which model keeps a product most accurate? Nano Banana Pro at 21 credits, when conditioned on a real photograph. Accuracy comes from the reference, not the model.
Do I still need a photographer? For one clean hero per SKU, yes. The saving is in setups and locations, not in the capture. See how much does a photoshoot cost.
Photograph the product, generate the world, check the colour. The 8frame canvas is free and unlimited, and generation is paid from $19/month.