AI Product Photography: What Works and What Backfires

You have 400 products in your catalogue and every one of them needs a photo. The studio quote lands, and a question forms: could we just have AI draw all of this?

We hear that question a lot more than we used to. Our answer always starts in the same place, which is that the question is framed slightly wrong. The useful version is narrower: on which image, and how much can I let AI touch it, without losing my marketplace account or watching my return rate climb?

This guide covers what AI product photography tools actually do well, where marketplaces draw the line, the point at which the cost argument flips, and why you should be sceptical of the conversion numbers circulating about all this. For the wider picture of what fits which business, see our industry map.

Do marketplaces allow AI-generated product images?

They do, and the test they apply is not how the image was made but whether it misleads the buyer. Amazon's guidance treats the production method as irrelevant to compliance: a DSLR, a phone, a 3D render, or a generative model are all acceptable inputs. Background removal, colour correction, lighting adjustment and resizing sit comfortably inside the rules.

The technical requirements do not change either. On Amazon the main image still needs a pure white background, the product filling at least 85% of the frame, and a minimum of 500 pixels on the longest side. An AI-assisted image that misses these gets rejected exactly like a badly shot one.

The real boundary sits on secondary images. Lifestyle shots may be AI-generated, but the product shown has to be the product sold, in its real colours, proportions and condition. Drop an accessory into the scene that does not ship in the box and you have created both a policy problem and a wave of "not as described" returns.

One development worth tracking: as of 2026, multiple seller-facing sources report that Amazon requires disclosure when product content has been generated or substantially modified with generative AI, and that AI-image detection and enforcement have widened. We have not verified the exact wording against Amazon's own announcement, so treat the direction as reliable and the specifics as something to confirm in your Seller Central policy pages.

The most expensive outcome here is rarely a suspended account. Accounts come back. The reviews saying "it didn't look like this" stay on the listing.

Where AI product photography genuinely earns its place

The mature capability is narrow and well defined: editing around an existing photograph. Removing and replacing backgrounds, adding shadow, balancing light, cropping to each channel's spec, and producing themed secondary scenes. Tools do this quickly, cheaply and in most cases indistinguishably from manual work.

That sounds modest until you look at where the hours actually go in an ecommerce operation. Even with every product already photographed, standardising 400 items and fitting them to three channels with different cropping rules is days of work. This is the part AI removes.

On the tooling side, roughly what is available:

• Photoroom: free tier available, paid plans starting around $7.50 per month. Strong on mobile background removal.
• Pebblely: around 40 free images a month, paid plans near $19 and $39. Theme-library driven, and the most generous free tier for getting started.
• Flair.ai: a drag-and-drop canvas rather than presets. You place the product and build the scene and lighting yourself, which suits teams that need brand consistency.

Vendor pricing shifts often, so read these as orders of magnitude rather than quotes. Knowing the category runs roughly $20 to $40 a month is enough to make the decision.

Which products break AI images?

Quality degrades as the surface gets complicated. Photoroom's own blog concedes that AI backgrounds look artificial in complex scenes, which is notable coming from the company selling the tool. What we see in practice lines up with that boundary.

The hard categories are consistent: reflective surfaces such as jewellery, watches, glass and chrome; textiles where fabric texture drives the purchase; text and labels on packaging; products held in the hand or worn on a body; and food. In these categories AI can produce something attractive that has quietly stopped being your product.

The risk is commercial rather than aesthetic. An image that makes a fabric a shade glossier sets a different expectation, and the gap shows up when the box is opened. We covered what actually drives returns in our return rate guide, and the distance between expectation and product sits at the top of that list.

A caveat worth stating plainly: there is no measured failure rate for AI imagery in these categories. This is our observation and that of teams doing similar work, not a lab result. Test it on your own catalogue before generalising.

Is studio photography or an AI subscription cheaper?

It depends on volume, and the answer runs opposite to most sellers' intuition. Per-product studio rates vary enormously by market, but the structural fact is stable everywhere: in a studio shoot, the setup cost is paid once. Past the first hundred products, the per-unit cost falls sharply.

Which means the cost advantage of AI is not where people expect it. It is meaningful for small, frequently changing catalogues, not large ones.

Consider a boutique releasing 15 new products a month. Every cycle means booking a session, transporting stock, and paying somebody's minimum package rate. Here a $30 monthly tool that cleans up existing photos is straightforwardly worth it. For a one-off 400-product catalogue shoot, the studio's per-unit cost has already collapsed and the AI saving evaporates.

The second point matters more: an AI subscription does not replace photography, it sits on top of it. Even the best-case workflow needs a source image. Budget for shooting plus editing rather than one instead of the other.

A quick calculation on a 400-product catalogue

Some numbers. The assumptions below will differ for your operation, but the magnitude holds. You have 400 products photographed and you are publishing to three channels: your own site and two marketplaces.

Done by hand, say background removal, white-background conversion and three crops averages 6 minutes per product. That is 40 hours for 400 items, a full week of one person's time. With a batch tool, including review and correction, the same work runs about 1.5 minutes per product, or 10 hours total.

The saving is 30 hours against a tool cost in the $30-a-month range. Whatever your hourly cost, that closes in the first month. The important qualification: all of this saving comes from editing, none of it from the photograph. Any calculation that assumes you have eliminated the shoot will mislead you.

There is also a line item running the other way, which is review time. We have never seen a setup publish AI output unchecked and end well. We folded that check into the 1.5 minutes above; keep it as its own line in your own numbers.

Common questions

Do I have to label AI-generated images? Disclosure requirements are tightening, and Amazon in particular is reported to require it for AI-generated or substantially modified content as of 2026. Rules differ by platform and change quickly, so check your seller policy pages rather than relying on a guide like this one.

Can I generate a model wearing my garment? Technically yes, commercially risky. Fabric drape and texture drive return decisions, and those are precisely what current tools distort most. If you try it, keep it out of the main image and place it alongside a real photograph.

Can I recolour a product with AI to create variant images? This is where marketplaces draw the clearest line. If the variant genuinely exists and the colour matches the real item, it is defensible. If the shade is off, it comes back as returns and complaints.

I am a small seller. Can I skip the photographer entirely? A well-lit phone photo is an adequate source image in most categories, and tools will take it to marketplace standard from there. That path closes for jewellery, watches and high-gloss products, where professional shooting is still the cheapest route.

So what should you actually do?

• Leave the main image alone. Generate it from a real photograph and keep edits to background and lighting. This is where the risk concentrates.
• Test your own category. Run five products through, then put the output next to the physical item and look. Trust your catalogue over anyone's general claims.
• Do the volume maths. Under roughly 20 products a month, an AI subscription pays back fast. Shooting 200 or more at once, price the studio's per-unit cost properly first.
• Track the return rate. Mark the month you changed the workflow and compare the following two months of product-related returns. That is the only measurement that means anything.
• Reread the platform rules annually. Policy here is still settling, and today's permitted practice becoming tomorrow's labelling requirement would surprise nobody.

AI in product photography works as a layer that takes over the tedious work after the shoot, not as a replacement for the shoot. Even in that limited role it gives small teams back hours every week. If you want to work out what that adds up to on your catalogue, get in touch and we will run the numbers with you.