Where generative tools genuinely help a brand, where they still fail, and how we combine them with real photography without the result looking fake.
Conceptual imagery — the surreal, the impossible, the visually arresting — used to require a large budget and a retoucher with a month. Generative tools changed the economics, and the industry is still working out what that means.
Here is our position after using them on real client work.
What they are genuinely good at
Scene and context. Once a product is photographed properly, placing it into twenty different environments is something a model does well and a photographer does slowly.
Variation at volume. A campaign needing the same concept across four seasons, three markets and two orientations is twenty-four images. Generating those from one approved direction is dramatically cheaper than shooting all of them.
Concepting. Showing a client three visual directions before committing a production budget. This alone has saved projects that would otherwise have been shot in the wrong direction.
The genuinely impossible. Products suspended in ways physics does not permit, materials that do not exist yet, sets that could not be built.
What they still fail at
The product itself. A general-purpose model does not know your product. It knows what products like yours look like. For anything a customer will zoom into — stitching, a logo, a control panel, a label — the generated version is plausible and wrong, and plausible-and-wrong is worse than obviously fake.
Text. Packaging copy, labels, signage. Still unreliable.
Consistency across a set. Ask for the same object twice and you get two similar objects. For a catalogue that is fatal.
How we actually combine them
Our default is hybrid: the product is photographed, everything around it is generated.
That gives the customer accuracy where accuracy is checkable, and volume where it is not. The hero and the detail shots are real. The lifestyle contexts, the seasonal variations and the campaign backgrounds are built from those real assets.
Where a brand needs many images of the same product over time, we train a model on that specific product, so the output stays recognisably theirs rather than generically good.
Should you tell people?
Our view: yes where it matters, and stop worrying about it where it does not.
A conceptual campaign visual that is obviously constructed needs no disclaimer — nobody thinks the car is really driving through the clouds. A product photograph a customer will use to decide whether to buy should represent what ships in the box, whatever tool made it. That is not an AI ethics position, it is consumer-protection common sense, and it predates the technology.
What it changes commercially
Less than the hype suggests, more than the sceptics admit. It has not replaced shoots. It has replaced the fourth, fifth and sixth shoot of the same product in different settings — a real saving for brands with large ranges, and almost none for a brand with one hero product.
If someone tells you AI removes the need for photography, ask them to generate your actual product, then compare it to the real thing side by side.
- AI
- conceptual photography
- brand campaigns
- production
