Five honest lessons from hundreds of AI product photos

The lessons start after that. We have now produced product visuals for brands across sanitary ware, furniture, outdoor, fashion, horticulture and construction materials, and the pattern is consistent enough to write down. None of these five lessons is a complaint about the technology. All five are about what you have to put around it.

Lesson 1: a model does not learn your brand, it fills in the gaps

Give a model a thin brief and it does not stop. It completes the picture with whatever is statistically most ordinary, and ordinary is never your brand. That is where the stock-photo feeling comes from: a competent image that could belong to anyone.

What closes the gap is not a longer brief. It is a more specific one. The things that actually decide the look are dull: the exact background value, one light direction for the whole range, a fixed set of angles, how much space sits around the product, what the product colour is measured against. We wrote that list out in rules a model can follow.

Here is the uncomfortable half of this lesson. Most brands discover at this point that those decisions were never written down for the photographer either. They were carried in one person's head, and the photographer absorbed them over three years of working together. A model has no three years.

Lesson 2: the product decides, not the technology

Some products are straightforward and some are a fight, and this has barely changed as the models have improved.

Straightforward: packaging, homeware, furniture, tools, garden and outdoor products, anything with a clear silhouette and a matt or lightly textured surface.

A fight: transparent glass, fine jewellery, high-shine metal, loose knitwear, sheer fabric, and anything where the material itself is the reason someone buys. Also anything where exact dimensions carry the purchase, which is why we work from technical drawings and product references for sanitary ware rather than from a description.

The practical answer is almost never all or nothing. It is a split: AI production for the structural volume, real photography for the few images where the material has to be felt. We say that to prospects before they ask, because finding it out after the invoice is worse.

Lesson 3: consistency is a property of your system, not of the model

A model will happily give you a hundred good images that do not belong to the same catalogue. Write a fresh prompt per product and you get a fresh interpretation per product. The result is a digital version of the patchwork you already had, produced faster.

What produces consistency is a fixed instruction that does not get retyped: the same style decisions applied to product one and product five hundred, with only the product and the setting changing. That is a system, and it has to be built on purpose. We go into how drift creeps in over a growing range in what breaks first when a range grows.

Lesson 4: the buyer judges the image, you carry the cost of the method

The question we get most often is whether customers can tell. For most categories, if the image is on-brand, correctly lit and technically right, the honest answer is that it does not come up. People judge whether the product looks like something they want, not how the file was made.

The cost side is where the real difference sits, and it is not only the price per image. It is the things that quietly surround a traditional shoot: shipping products and getting them back, products that cannot be sold as new afterwards, a studio day that has to be rebooked because one variant arrived late, and the fact that a new colourway means doing it again. For heavy or bulky products that logistics tail is often larger than the photography itself.

We deliberately do not publish a price comparison, because the honest version depends on your range, your variants and your markets. What we can say is which lever moves it: the number of times a product has to physically move.

Lesson 5: AI and photography are not opponents

There are moments where a photographer is simply the better answer. A campaign built around people and a story. A launch where the first hero images set the tone for everything after. A material that has to be felt through a screen.

What AI takes over is the structural work: the long tail of the catalogue, the seasonal refresh, the new colour variant, the same product in a second market. Those are the images you always need and never want to book a studio for.

The lesson that cost us the most: the second batch

This one is not in most articles about AI imagery, including our own earlier ones, and it is the one we would warn you about first.

The first batch goes well. Four months later you add eleven products and produce a second batch. It is subtly off. The light is a shade warmer, a background sits two percent lighter, the camera height moved. Each image is fine in isolation. On a category page the eleven new ones read as a different range.

Two things cause it. The underlying models keep changing, so the same instruction does not give exactly the same result six months later. And the human context changed: a different person ran it, with a slightly different idea of what "the usual" meant.

What fixes it is boring and effective. Keep an approved reference image per product type and compare new output against that image, not against your memory. Keep the full instruction stored, not just the description of it. And keep every approved asset in one place so the comparison takes seconds rather than an afternoon.

There is a sharper reason to care about this when you sell through others. An image on your own site can be swapped in a minute. An image that has already been pulled into dealer portals, retailer product pages and marketplace listings is not something you can recall. The second batch is the one that quietly ends up everywhere.

What we would do in week one

If you are starting now, in this order:

1. Write the eight style decisions down and approve one reference image for each.

2. Split your range into straightforward and difficult, and be honest about the difficult half.

3. Produce twenty images, not two hundred, and have someone who knows the products review every one.

4. Store the approved set and the instruction that produced it in the same place.

5. Only then scale, and check the second batch against the first.

Steps three and four are the ones people skip, and they are the ones that decide whether month six looks like month one. That review step is what we mean by what a person in the loop actually catches.

In closing

The promise of scale is real. It arrives with conditions, and the conditions are mostly about writing down decisions you have been carrying in your head.

If you want to see where your own range sits on the easy-to-difficult scale, we will tell you plainly, including the products we would not recommend us for. You can see how it works first.

Want to try it on your own products? Try LensLab free with five credits.