Ask a small business owner about AI image tools and you will usually get one of two answers: either they tried a free tier once and forgot about it, or they assume it is expensive enterprise software.
Both assumptions are now wrong, and the gap between them is where a genuinely useful and quite cheap capability sits unused.
What Changed
The shift was not in the models — it was in how they are sold. Image generation is now billed per image, in fractions of a cent to a few tens of cents depending on resolution and quality. No monthly commitment, no seats, no minimum spend. A cafe producing a dozen social posts a month spends less than a flat white. A real estate agency generating background and concept imagery across listings still lands well under what a stock photography subscription costs.
That pricing structure matters more than any capability benchmark, because it removes the decision from the budget cycle. There is nothing to approve.
Where It Earns Its Place — And Where It Does Not
Start with the boundary, because getting this wrong is what turns a useful tool into a liability.
Photograph the actual product. A real image of the real item, accurate in colour, texture, and proportion, is not optional — it is the thing your customer is buying. Every gap between the picture and the parcel becomes a return, a refund, or a review you will be reading for years. The same applies to your premises, your staff, and anything a customer could reasonably read as evidence.
Generate everything around the product. The seasonal backdrop. The lifestyle context. The abstract texture behind a flat-lay. The header on the email newsletter. The concept image for a service that has no obvious photograph. In most small businesses that category is the majority of images needed by volume, and historically it was either skipped, or filled with the same stock photo three competitors were also using.
The Cost Nobody Puts in Their Estimate
Here is the number that actually determines the bill: how many attempts precede an image you will use.
Nobody keeps the first one. Realistic usage runs three to eight generations per keeper, so the true cost per finished image is several times the headline rate. Any budget built on the quoted price will be wrong, and wrong in proportion to how particular you are.
The habit that fixes it takes thirty seconds to learn. Generate cheap low-resolution drafts to choose a direction, then regenerate only the winner at full quality. Businesses that adopt this routinely halve their spend with no visible difference in what gets published, because the vast majority of generations are dismissed within seconds.
Anyone comparing tools can check current rates openly rather than guessing. Published pricing for the Nano Banana API on APIMart and competing image models sits on aggregation platforms that expose several models through one account — worth knowing because the model that produces clean product environments is rarely the one that handles illustration, and paying for the wrong fit is the most common way this gets expensive.
Consistency Is the Part That Takes Effort
One good image is easy. The fortieth that still looks like it came from the same business as the first thirty-nine is where most owners quietly give up.
The fix is unglamorous: write down your visual vocabulary once — the same words for lighting, palette, materials, and mood — and reuse that block for every generation, changing only the subject. It reads like paperwork. It is also the entire difference between a coherent brand and a folder of unrelated pictures that happen to feature your products.
Two Practical Limits
Text inside generated images is still unreliable across every model available. If a graphic needs words — a price, a date, a tagline — add them afterwards in a design tool rather than prompting harder.
And keep a simple record of which images were generated and which were photographed. Marketplaces, advertising platforms, and increasingly some regulators ask. Noting it as you go costs nothing; reconstructing it later from a shared drive is genuinely unpleasant.
A Sensible First Month
Pick one real job with a deadline — the seasonal banner set, the category headers, the backgrounds for a product line — and solve only that. Budget the price of a decent lunch. Measure one thing: whether the images actually went live, or stayed in the folder because nobody trusted them.
Most owners who run that test come away with the same two conclusions. The spend was smaller than expected, often by a wide margin. And the real constraint was never the cost or the technology — it was not having decided, before opening the tool, exactly what the picture was supposed to say.
