Marketplace product imagery: where the AI stops and the human eye begins

For most of the past decade, the imagery on a cross-border listing was the single most expensive line item after freight. A sourcing team would brief a studio, wait, receive renders, argue about colour accuracy, and ship. That model has been quietly collapsing for a while now. Image generation tools have reached the point where a competent operator can produce a hero shot, a lifestyle scene and a set of infographic frames in an afternoon.

The library's image tools category is deliberately small, and the reason is that we only score tools against live briefs — the same process described on the about page. What follows is drawn from how those briefs tend to go wrong, and where the reviewer we would previously have called a retoucher still earns their seat.

The render is a draft, not the product

The first and most important distinction is between images that show what a product looks like and images that show that the product exists. Marketplaces, and the buyers on them, treat these differently. A generated lifestyle scene — the kitchen, the model, the soft window light — does real work in a gallery. It sets context and it converts. But when a buyer opens the box and the finish is a shade off, the review you receive names that image.

Our position on the desk is straightforward: generated scenes for context, photographs for evidence. Every listing should carry at least the minimum set of real photographs the marketplace requires, taken against the actual shipment, not the sample from the trade fair. The generated material sits around that core. Midjourney, currently scored 4.3 on the library's scale, is the strongest of the tools we track for scenes and mood, but it is a paid tool that is happiest producing a world, not a specification sheet.

The three failure points

In the briefs we run, unreviewed images fail in roughly three recurring places:

• Policy drift. Marketplaces each have their own rules on overlays, watermarking, promotional text and even the ratio of image to padding. Generation tools optimise for a pleasing picture, not for compliance. A render that passes a human aesthetic check can still bounce at upload.
• Variant accuracy. The generator will happily render the same water bottle in teal when your SKU ships in sage. Colour drift between a prompt and a physical product is the single most common complaint we see in follow-up reviews, and it is invisible unless someone holds the sample next to the screen.
• Cultural context. A background that reads as neutral in one market reads as dated or odd in another. This is where cross-border teams have an edge over pure design teams, and it is precisely the judgement the model does not have.

None of these are exotic. All of them are caught by a person spending a few minutes per listing with the sample and the marketplace's style guide open. The cost of that review is trivial compared with the cost of a delisted listing or a returns spike.

Where each tool earns its place

The pipeline is not one tool. Ideogram, scored 4.0 and freemium, has become the pragmatic choice when the brief involves text inside the image — dimension callouts, feature labels, the small typographic furniture of a good listing. Text rendering is where general-purpose generators historically fell apart, and a tool built around legible type saves the fiddly compositing step that used to eat the afternoon.

Canva, scored 4.4 and also freemium, sits at the end of the line rather than the start. Its value is not generation but localisation at scale: resizing one master layout into the crop ratios and template constraints of several marketplaces, swapping the language on the overlay, keeping brand colour locked. That last part matters more than people expect. When your generated hero shot goes through a template round-trip, colour management drifts, and the reviewer checking the final export against the sample is doing compliance work whether or not anyone calls it that.

The reviewer is now a different job

What has changed is not whether you need a human eye, but what that eye is looking at. The old retoucher spent their day fixing reflections. The new reviewer spends it checking that the generated scene matches the physical product, that the variant mapping is right, that the overlay complies with the target marketplace, and that nothing in the frame would read strangely in the destination market. It is closer to quality control than to design, and sourcing people — the ones who have actually held the product — are often better at it than designers.

There is also a supply-side rhythm worth noting, because imagery sits at the end of a chain that begins with physical goods. Our fuel tracker currently shows Brent at 103.536 and WTI at 96.622, with the next Chinese retail adjustment window due on 2026-11-12, nine days out, and a cut expected rather than a rise. Freight costs shape listing margins, and margin pressure is exactly what pushes teams to generate imagery rather than shoot it. That pressure is rational. It just should not be allowed to also remove the review step that keeps the generated imagery honest.

The teams that do this well treat generation as capacity, not as replacement. They generate the scene, photograph the product, check the variant colours against the sample, read the marketplace rules before upload, and ask someone who knows the destination market to look at the frame before it goes live. The tools in the library — Midjourney, Ideogram, Canva and the rest — can carry most of that weight. The last mile of looking is still yours.

• Generated images work best as drafts for listing layouts and lifestyle scenes, never as substitutes for the photographs that prove the product exists.

• Marketplace policy, cultural context and variant accuracy are the three failure points where an unreviewed image costs more than the review time it saves.

• Midjourney, Ideogram and Canva each occupy a different place in the workflow, and pretending one tool covers the whole pipeline is where teams get caught out.