AI product photography: generating the images is the easy part

For most of the history of commerce photography, the constraint was supply. Images were expensive and slow, so having more of them than a competitor was a genuine advantage, and the whole industry organised around lowering that cost. That constraint is gone. Any team can now produce more images in an afternoon than it could previously commission in a year, and so can everyone selling against it.

What follows is predictable. When the cost of an asset falls to roughly nothing, the volume of assets rises until attention, not production, becomes the scarce thing. Feeds fill with technically competent images that were cheap to make and are impossible to tell apart, and the marginal image earns less than the one before it. Buying a faster way to add to that pile is not a strategy, it is a contribution to the problem.

The advantage moved to the other side of publication. Every image you ship is a small experiment that returns a number, and almost nobody collects those numbers in a form that changes what gets made next. A team that does holds something its competitors cannot copy by spending more: a growing, specific, proprietary account of what its own market responds to. That is the asset. The images are just how it gets built.

Every AI image tool demos well. Generate one product on one background and the results are impressive across the board. The problem appears at volume, when a hundred images made from a hundred prompts by four people over three weeks stop looking like one brand — slightly different warmth, slightly different angle discipline, slightly different product scale. Individually fine, collectively a mess.

That drift is a governance problem, not a model problem. Fixing it means the brand rules have to be enforced at generation time and shared across everyone generating, rather than living in a PDF that people consult when they remember to. This is the single clearest dividing line between a point tool and a production platform, and it is worth testing before you commit: generate fifty images with two different people and look at them as a grid.

The second-order effect is what closes the loop. Once generation is consistent and connected to publishing, you can see which visual choices actually perform and feed that back into the rules — so the catalogue does not just get produced faster, it gets measurably better each cycle. Most teams never get here because they are still fighting consistency.

Measure at the level of the visual decision, not the campaign. A campaign result tells you that something worked; it does not tell you that the warm key light worked, or that the in-context shot beat the plain packshot, or that the Arabic-first layout outperformed the mirrored English one in Riyadh but not in Cairo. Tag each generated variant with the choices that produced it, and the numbers start answering questions you can act on.

The metrics that matter are the ones tied to money and attention: conversion rate on the listing, click-through and cost per acquisition on paid placements, ROAS by channel and by market, and how long an asset holds performance before it fatigues. Creative fatigue is the one most teams miss. An image that performed well for three weeks and then decayed is telling you when to regenerate, and that timing is worth more than another thousand fresh assets.

Then do the thing that almost nobody does: write the finding back into the brand definition rather than into a deck. A rule that lives in the system shapes every future generation automatically. A rule that lives in a quarterly review shapes nothing, because the person generating next week never reads it. This is the whole difference between a team that produces content and a team whose content gets better, and it is why the loop matters more than the generator.

The loop is slower to pay off than the generator, and anyone promising otherwise is overselling. You need enough published assets and enough traffic before a pattern separates from noise, which for most catalogues means a quarter rather than a fortnight. Attribution is also imperfect: a listing image and a paid placement influence each other, and no measurement setup untangles that cleanly. Treat the output as a strong signal that improves with volume, not as proof.

Arabic and other non-Latin script rendering fails more often than teams expect. General-purpose image models routinely disconnect Arabic letterforms or produce text that reads as nonsense to a native speaker while looking fine to someone who does not read the script. Treat any label, packaging or in-frame text as a locked asset, and have a native speaker review before publication.

Anything regulated needs a human in the loop. Dosage information, ingredient lists, certification marks, nutritional panels and compliance text must come from the real product, not from a generated approximation. The cost of getting this wrong is not a bad image.

And a real one: if your catalogue is small and static, you do not need a platform. Thirty SKUs refreshed once a year are served perfectly well by a background-removal tool, and anyone telling you otherwise is selling something. The economics turn when you have volume, multiple channels, multiple markets or a seasonal cadence, because that is the point at which there is enough signal to learn from.