AI product photography: what it can and cannot do

The one distinction that decides everything

You have thirty products, a phone, and a kitchen window, and the photos look like exactly that. Then a tool offers two things: turn these into studio shots, or skip the shoot and generate the product from a description.

Those are different offers and only one of them is safe.

Editing a real photograph of the real item — cutting out the background, fixing the light, setting it on a marble counter — changes how the item is presented. Generating an image from text changes what the item appears to be. The first is photography with better tools. The second is a claim about a product, made by a model that has never touched it.

Everything useful about AI product photography follows from that line, so hold onto it while the rest of this post gets specific.

What AI imaging genuinely does well

This part is not hype. The following are reliable enough that paying a retoucher to do them by hand is now hard to justify.

Background removal and replacement. Clean cutouts around hard edges, including moderately fiddly ones like a bag strap or a chair leg. Hair, fur, and fine mesh are still where it fails, but a bottle or a ceramic mug on white is a solved problem.

Relighting and shadow synthesis. Taking a flat phone photo shot under a ceiling bulb and giving it directional light with a believable contact shadow. This is the single biggest quality jump available to a store owner who cannot afford a studio, and because the product pixels are real, it is also the safest.

Scene compositing. Placing a real cutout into a generated setting — a towel on a bathroom shelf, a candle on a side table. The product is photographed; the room is invented. As long as nothing in the invented room implies something you ship, this is honest.

Catalog consistency. Matching crop, margin, background tone, and white balance across hundreds of images that were shot across months by different people. Tedious, mechanical, and now cheap.

Upscaling and cleanup. Dust, sensor noise, a stray thread, a reflection of your own hands in a glossy surface.

Generated human models. Technically good and legally loaded. A generated person wearing a real garment, or a real garment composited onto a generated body, usually distorts fit, drape, and proportion in ways customers notice when the parcel arrives. Several markets are also moving toward requiring disclosure when a body shown in an ad is synthetic or digitally altered. Treat this as the highest-risk use on the list, not the most exciting.

What it still gets wrong

Current models fail in patterns, and the patterns are predictable enough to plan around.

Text and logos. Packaging copy, ingredient panels, a brand name embossed on leather. Generated or regenerated text comes back subtly wrong — a letter swapped, kerning that no font produces. Customers may not read it, but a trademark owner does.

Repeating detail. Knit stitch, weave, basketry, engraving, a printed pattern that has to align at a seam. The model produces something that looks like the idea of the pattern and does not survive a zoom.

Reflective and transparent material. Glass, polished metal, faceted stones, anything with specular highlights. Jewelry is the standing worst case: the stone looks like a stone, and like a different stone from the one in the box.

Exact color. This matters more than anything else in this section, because color is the top driver of returns in apparel and home goods. A model relighting your photo will drift the hue. Dye lots already make color hard to promise; do not add a second source of drift on top.

Hands and interaction. Fingers holding the product, a hand pouring, a strap over a shoulder. Improving, still unreliable, still immediately recognizable.

Angle consistency. Ask for the same product from the back and you get a plausible back, not your product's back. This is the failure that catches people out, because the image looks fine on its own and only breaks in a gallery alongside the real front shot.

The honesty problem, stated plainly

Most countries' advertising rules do not have an AI clause. They have a misleading-representation clause, which is older, wider, and completely indifferent to how the image was made. The question is never "did a model generate this". It is "does this image tell the truth about the thing that arrives in the box".

That standard also happens to be the commercial one. An image that oversells is not a marketing win followed by a neutral outcome. It is a refund, a return shipping cost you pay, a restocking cost, a one-star review that outlives the product, and — if it happens at volume — a payment processor noticing your dispute rate.

Disclosure rules for synthetic media are arriving in several markets, and the direction is toward labeling rather than away from it. So treat "will I have to disclose this later" as a question whose answer is still moving, and prefer images that would survive a label.

The cheap test that replaces all of this

Imagine the customer holding the item, your product page open on their phone beside it.

Would any difference between the two be the kind of thing that makes them start a return? If yes, the image is misrepresenting, no matter how good it looks. If the only difference is that your photo had nicer light than their kitchen does, you are fine.

A shoot you can actually do in an afternoon

The practical answer for a small store is not generated imagery. It is real photographs plus AI editing, and the real photographs take less effort than people assume.

1. One window, indirect daylight, middle of the day. Turn off every indoor light so you only have one color of light.
2. A sheet of white foam board as a reflector on the shadow side. This is the whole lighting kit.
3. Phone on a tripod or propped on books, and do not move it between products — consistency is free if the camera never moves.
4. Lock exposure and white balance. Tap and hold on most phones. Auto settings are where catalog inconsistency comes from.
5. Six to eight frames per product: straight on, three-quarter, back, a detail at the thing people ask about, scale next to a common object, and one in use.
6. Then run AI editing over the set: cut out, relight, match the crop, clean up dust.

The main listing image stays a real photograph. Generated scenes, if you use them at all, belong further down the gallery where they read as mood rather than specification.

Where mAutomate fits, and where it does not

Worth being direct, because this is a post on a commerce platform's blog.

mAutomate does not generate product photography. There is no image generator for your products in it, and anything on this page about generated imagery is about tools you would buy separately.

What the AI Storefront Studio does is build the store around the photos you supply: the storefront, the product pages, the copy, the on-page SEO. It drafts a product description from your product, and whether AI can write product descriptions is a separate question with its own answer. Nothing it drafts publishes until you approve it.

So the division of labor is unglamorous and about right. You own the truthful picture of your product. The platform owns everything that has to be assembled around it.

What to do this week

Pick your five best-selling products. Reshoot them by the window with the phone locked down, then use whatever editor you like to cut out and relight. Compare the new gallery to the old one and to your return reasons.

If returns on those five mention color or size, your imaging problem was never a style problem, and no generator would have fixed it.