AI product photos: cost vs a photographer, marketplace rules and where they fail

How much do AI product photos cost per image?

Vendors publish a monthly fee or a credit pool, not a price per usable image. The table converts each plan into a price per image at full use of its allowance (our arithmetic), with the two marketplace generators we could read.

Per-image prices leave out a clean original of every product, the generations you reject and the time to check each result. Those hours decide whether AI is cheaper.

Where do AI product photos work, and where do they fail?

Image models make convincing pictures, not exact copies. Our rule: let AI change what surrounds a real product, never redraw the product.

• Works: a clean or neutral background behind a cut-out product; Ozon’s editor generates and removes backgrounds.
• Works: lifestyle scenes in secondary slots, which Amazon says may show use or context.
• Works: several backgrounds or crops from one master photo, for testing.

• Fails: the product itself. In Photoroom’s benchmark of 850 clothing, footwear, bag, jewellery and accessory products, the best of four models kept every detail in 29.0% of virtual-model images and the weakest in 16.8%. The test is strict (one reviewer’s flag fails an image) and covers garments on generated models, the hard case, not a plain background swap.
• Fails: logos and text, the most frequent error at 20.1% of generations. Wildberries names distorted text as a reason for a lower rating, and SelSup reports that Ozon’s editor distorts jewellery and packaging text.
• Fails: invented or missing details, altered patterns and colour shifts, all logged in that benchmark.
• Fails: glass, chrome and other reflective or see-through items, where a model must guess what lies behind them. No source we read measures this.
• Fails: scale and contact. A scene can show an item too large, too small or floating; compare it with the spec sheet.

Do Amazon, Wildberries, Ozon and Google allow AI-generated product photos?

None of the pages we could read bans AI images. Each ties them to one test: does the picture show the product the buyer will receive?

Can misleading AI product images get a seller into trouble?

Yes, if the picture changes what a buyer expects to receive. The US Federal Trade Commission reads an advertisement in context, words, phrases and pictures together, and treats a claim as material when it matters to the decision to buy, such as features. An AI image that adds a feature, colour or accessory the product lacks is that kind of claim, and an agency can be held responsible alongside the advertiser.

Amazon’s July 2026 notice ties its tagging rule to recent legislation on AI-generated people in advertisements. This covers the United States only and is not legal advice; ask a local lawyer.

How do you use AI photos without misrepresenting the product?

The safe pattern keeps the product real and lets AI work around it.

1. Photograph the real product firstShoot each item whole and in focus on a neutral background, above the platform’s minimum size. That photo is the source of truth for colour, shape, labels and number of parts.
2. Keep the real pixelsCut the product out and let AI build only the background, shadow and surface; never ask a model to redraw the product, label or logo.
3. Keep the main image plainAmazon wants pure white and Google the exact item without overlays: use a photographed frame there and give AI scenes the secondary slots.
4. Check each result against the itemOpen it beside the original at 100% zoom, product in hand. Check text, logo, colour, count of parts, proportions and contact with the surface.
5. Keep the metadata and the originalDo not strip the IPTC AI tag Google reads, add Amazon’s keyword if a photorealistic AI person appears, and keep the untouched photo with each approved result.
6. Approve per SKU, test firstRecord which frames are AI-made and who approved them; test three to five SKUs before the catalogue.

What would ten SKUs cost with AI, a photographer or both?

Assume ten products with seven frames each: a hero and five secondary frames on white plus one lifestyle frame, 70 in all, no models, basic retouching. Totals are our arithmetic on published prices, not quotes.

The hybrid drops the lifestyle line, $2,000–$5,000 on that list, and swaps it for a person’s hours: roughly $240–$525 per SKU against $440–$1,020 for the photographer alone. If checking costs more time than it saves, the plan fee was never the real price.

What would your catalogue cost with VITON¹³?

Add 3D views for products that do not exist yet or photograph badly, or card design for the slides that carry the pictures.

This package returns a shot map, lighting direction, hero, detail, scale and use frames, a retouched master set and crop exports, in 3–5 working days.

When should you not use AI product photos?

• The product’s look is the claim: jewellery, glass, a cosmetic shade, a fabric sold on weave and drape.
• The label carries information buyers rely on, such as ingredients or certificates; a model can change characters.
• It is the main image on Amazon or the exact-item image on Google.
• You would put garments on generated models, the hardest case in Photoroom’s benchmark.
• You cannot compare every frame with the physical item before publishing.
• The product does not exist yet, or you need exact colour and material variants: a 3D render is built from agreed materials and views, and VITON¹³ prices it separately.