Image generation moves into the API. When that is worth using
OpenAI opened its image model to developers on Wednesday. For most SMEs the interesting case is not marketing — it is the catalogue.
On 23 April 2025 OpenAI made its image generation model available through the API, so it can be built into other products rather than used only through a chat window.
A month ago this blog covered the chat version and drew a line: generated images where you would have used a stock photograph, never where a customer could think they are seeing your actual work. That line still holds. The API version raises a different question: when is it worth putting this inside your own systems?
Three cases where it genuinely is
Catalogue placeholders. A wholesaler with eleven thousand articles has photographs for perhaps four thousand. The rest show a grey box. A generated, clearly-labelled category illustration is better than a grey box and costs a fraction of a photo shoot. Label it as an illustration, and never for an article where appearance affects the purchase.
Internal material at volume. Training slides, process diagrams, intranet headers. Nobody is buying anything based on these, and the alternative is usually clip art from 2011.
Concept visuals before a commission. Showing a client three directions before paying a photographer for one. This is where the money actually gets saved, because the expensive step becomes better targeted rather than replaced.
Two cases where it is not
Anything a customer might mistake for documentation. Your machine, your installation, your finished work. Covered last month, still the most expensive mistake available here.
People. Generated faces in a company context imply staff or customers who do not exist. It reads badly when noticed, and it is noticed more often than people expect.
The practical control
If you build this into a system, store one extra field with every generated image: that it was generated, when, and from what instruction.
Two years from now somebody will ask whether a particular picture is a photograph. That field is the difference between answering in ten seconds and not being able to answer at all — which is the same argument as February’s logging post, applied to pixels.