AI Product Photography: What It Can and Cannot Replace

Artificial Analysis covered the Ideogram 4.5 release this week and led on edit drift, the habit an image has of changing in ways you did not request when you ask for a small adjustment. Their write-up also noted roughly a 27-fold spread in price across current image models. Both details point at the same thing for a retailer: quality and cost have stopped being the deciding factors, and consistency has become the whole question.

What it handles well today

Backgrounds are solved. Removing a cluttered table and placing a product on a clean surface, a marble counter or a studio sweep takes seconds and reads as professional. For a small retailer who has been shooting against a bedsheet, this alone is the difference between a listing that converts and one that does not.

Lifestyle and context shots are the bigger saving. A mug on a desk beside a laptop, a jacket on a trail, a tool in a workshop. These used to mean a location, a model release and a day. Generated from a real product photograph, they are cheap enough to produce a dozen variations and pick the one that performs.

Seasonal and campaign variants follow the same logic. The same product against a summer patio, a holiday table and a back-to-school desk, produced in an afternoon rather than three shoots. So do platform crops: the square, the vertical and the banner, each composed properly instead of cropped badly from one original.

Where it goes wrong

Use	Risk	Rule to adopt
Main listing image	Shows a product that does not exist	Always a real photograph
Colour and finish variants	Shade drifts from the real stock	Shoot one per colourway you hold
Packaging and labels	Text becomes plausible nonsense	Never generate readable copy on a product
Scale and fit shots	Proportions quietly flatter the item	Measure against a known object
Bundles and accessories	Implies items that are not included	Caption what ships, in the image

Every one of these shows up as a return rather than as a complaint about the photo. A customer who receives a slightly different shade does not write to say your image generation drifted; they request a refund and buy elsewhere. The cost lands in logistics, and nobody traces it back to the image.

Consistency across a catalogue

A single generated image can be excellent and the set can still fail. When each of eight listing images was produced in a separate pass, the product gets fractionally longer, the handle shifts, the logo rotates, and the whole listing starts to feel off without any one image being identifiably wrong.

This is why the recent focus on edit drift matters more to retail than to design. A designer iterating on a poster can accept variation. A catalogue cannot, because the buyer compares the images to each other and then to the parcel.

The practical control is to generate every variant from the same source photograph rather than from the previous generation. Chaining edits compounds drift. Going back to the original each time keeps the product anchored to something real, which is the same discipline as working from source records rather than summaries.

The advertising rules have no AI exception

Canada's Competition Act prohibits representations to the public that are false or misleading in a material respect when promoting a product. Nothing in that turns on how the representation was made, so the question for any image is the same question a retoucher has always faced: would a buyer reasonably expect to receive what this shows?

Under that test, a generated kitchen behind a real kettle is fine. A generated kettle with a spout shape you do not sell is not, and neither is a bundle image implying a cable that ships separately. Treat this as general information and get advice on your own listings if the margin is material.

Whether to tell customers an image was generated is a separate judgment from whether it is accurate, and we worked through that calculation in when to disclose AI to customers.

Check the platform rules per channel

Marketplace policies on generated imagery differ and they have been revised repeatedly, so the specific wording is worth reading for each channel you sell on rather than inferring from another one. Most settle in the same place: the main image has to show the actual product accurately against a plain background, and secondary images carry more latitude.

Building your process around that split is the cheap insurance. A workflow where image one is always a photograph and images two onward are generated from it satisfies the strict reading of nearly every policy, and it does not need rework when a platform tightens its terms.

A workable split

Photograph each product once, properly, in daylight against something neutral, with one shot per colourway you actually stock. Generate everything downstream of that: backgrounds, scenes, seasonal sets, platform crops, banner compositions. Review at full size for drift in logos, text and proportions before anything goes live, and keep the source photographs organised, because every future variant starts from them.

For a shop with four hundred SKUs, this converts an impossible content calendar into a manageable one, which is the same leverage that ecommerce automation delivers elsewhere in the operation. The capability side of current image tools is covered in our guide to Nano Banana 2, and the wider marketing use in image and video generation inside ChatGPT.

Frequently Asked Questions

Can AI replace product photography?

It can replace a large part of what surrounds the product shot and very little of the shot itself. Backgrounds, lifestyle scenes, seasonal variations, scale references and platform-specific crops are all work that image models now handle at a quality customers accept. The photograph establishing what the item actually looks like still has to come from the item, because that image is a representation of something the buyer will receive and inspect.

Is it legal to use AI-generated product images in Canada?

There is no rule against the technique itself, and the relevant standard is about accuracy rather than method. The Competition Act prohibits representations to the public that are false or misleading in a material respect when promoting a product, and that applies to an image however it was produced. A generated lifestyle background is not a problem. A generated image showing a finish, a size or an included accessory that differs from what ships is the same problem a retouched photograph would be. This is general information rather than legal advice.

What goes wrong with AI product images?

Drift on details is the common failure. Logos come back subtly wrong, stitch counts change, the number of buttons moves, text on packaging becomes unreadable, and material finish shifts between matte and gloss. These are easy to miss on a phone screen and obvious when the parcel arrives, which is where they turn into returns and chargebacks rather than complaints about photography.

What is edit drift in image generation?

Edit drift is the tendency of an image to change in ways you did not ask for when you request a small adjustment. You ask for a warmer background and the product gets slightly longer. Recent releases have targeted it directly, and the reason it matters commercially is catalogue consistency: a product that changes shape across the eight images in a listing reads as untrustworthy even when no single image looks wrong.

Do marketplaces allow AI-generated images?

Policies vary by platform and they change, so check the current version for each channel you sell on rather than assuming. The durable pattern across most of them is that the main listing image must show the actual product accurately on a plain background, while secondary and lifestyle images have more latitude. Build your process around that split and you will usually stay inside the rules even as the specific wording moves.