From product photo to ad creative: a six-step production line

Turning a product photo into an ad is not a background swap. When the product's label, material and scale are not preserved, the result looks fake in the first second — and the click-through rate says so immediately.

The six steps below are the order our own production line runs in.

Know the product first

Before anything is generated, the pipeline needs to know what it is selling: a physical product, software, a service, or an info product. That branch decides everything after it — a line built on the physical-product assumption produces nonsense for a SaaS product, where a screen matters more than a box on a table.

A pipeline that starts from the wrong assumption amplifies that error at every later step.

Pick a scene, don't invent one

The scene library is made of copyright-free recipes, and each recipe describes a state: where the product sits, where the light comes from, which surface is visible.

• Locked fields — the product itself, its label and proportions. Never changed.
• Slot fields — the surface the product rests on. Left empty by the recipe.
• Adaptive fields — palette and light temperature. Adapted per product.

Without that split, generated scenes either float the product in mid-air or reinvent the label.

Lock the reference

The product photo goes to the model as a reference; the scene image is the second one. Without both, you get "a similar product" instead of yours.

Reference fidelity is the single largest variable in output quality. Swapping models rarely makes this much difference.

Quality control has to be cheap

After every generation a visual review step runs: is the product recognisable, are the proportions intact, is the text mangled? Retries rejected by that review are not billed to the user, because quality control is the line's responsibility, not the customer's.

Produce variants, never a single image

Nothing is tested with one creative. Three angles of the same message — problem, benefit, social proof — go live together and the platform picks the winner. Three angles show which message lands within a week; one image never does.

Write the result back to the brand

Every published creative's performance is written into the brand's memory. The next production does not start from zero: whichever scene, angle and message worked for that brand, the next set is built on top of it.