How many AI images does it take to get one you can use?
About eight, in our own work. Across four real campaigns made with AI image models, 361 renders gave us 45 photos we chose to use. The spread was wide: a supplement ad took about 3 renders per keeper, and a surf apparel shoot on one model took about 14. Budget for several tries per photo, and for more when a person, motion or text is in the frame.
The numbers from four real campaigns
Our agency made these four campaigns in 2026, by hand, in Figma Weave, a tool where you wire AI models together on a canvas. Every image the model returned was saved, and an art director marked the one picked for each shot. That makes the count simple: every file is a render, and every file marked as picked is a keeper.
| Campaign | What we made | Renders | Keepers | Renders per keeper |
|---|---|---|---|---|
| Gummy supplement | Static ads of the jar, with the words added in a second pass | 38 | 12 | 3.2 |
| Men’s trousers | Store photos on a model, plus editorial shots | 40 | 11 | 3.6 |
| Brownies | Food ads, with brownies breaking apart and cocoa in the air | 120 | 10 | 12.0 |
| Surf apparel | A character sheet of one model, then studio, cliff and ocean shots | 163 | 12 | 13.6 |
| All four | 361 | 45 | 8.0 |
Each campaign used a different image model and a different kind of shot, so these numbers cannot tell you which model is best. They do show how much the number of tries changes from one job to the next.
Why some jobs took four times as many tries
We cannot prove the causes, but the two slow jobs and the two fast ones differed in plain ways.
- Slow: a person who has to stay the same person. The surf job needed one model to look identical in every photo, in a studio, on a cliff and waist-deep in the sea. The character sheet alone, the same person shown from every side, took many attempts before it was right, and every later shot had to match it. See how to keep the same model in every AI photo.
- Slow: motion and physics. The brownie ads showed brownies breaking apart mid-air and cocoa scattering through the frame. Image models often get the product right and the physics wrong, or the physics right and the product wrong.
- Fast: the product alone, with no words in the render. The gummy ads showed the jar by itself, and the headline was set in a separate pass, so no render had to spell anything.
- Fast: a settled process. The trouser job ran on our most developed workflow: a written shot plan, fixed reference photos for each shot, and a locked frame so every store photo lined up. It had a person in it too, and it still needed only 3.6 renders per keeper.
What it means for your budget
The cost of one usable photo is the price of one render times the number of renders it takes. Using our range of 3 to 14, and the prices Google and OpenAI publish for one image:
| Image model | One render | At 3 renders | At 8 renders | At 14 renders |
|---|---|---|---|---|
| Nano Banana 2 Lite (Google) | $0.034 | $0.10 | $0.27 | $0.47 |
| Nano Banana 2 (Google, 1K) | $0.067 | $0.20 | $0.54 | $0.94 |
| Nano Banana Pro (Google, up to 2K) | $0.134 | $0.40 | $1.07 | $1.88 |
| GPT Image 2 (OpenAI, high quality) | $0.211 | $0.63 | $1.69 | $2.95 |
The review takes more time than the model bill takes money. At eight renders per keeper, a 30-photo campaign means looking at about 240 images to choose 30, so plan the reviewer’s hours as well as the budget. For a full breakdown of subscriptions and credits, see how much AI product photos cost.
How to get more keepers from each render
- Choose the reference photos for each shot before you render. A back view gets the back photo, a close-up gets the texture photo, and nothing else. Extra references give the model more to mix up. See why AI changes your product.
- Approve the person first. If a model appears in the photos, make a character sheet and approve it before any scene. A wrong face there costs you every photo after it.
- Draft on a cheap model, finish on a good one. Try ideas on a fast, low-cost model, then remake only the photos you like on the better one.
- Fix one photo at a time. When one shot is wrong, remake that shot with a note about what to change, and ask for two to four versions side by side. Re-running a whole batch to fix one frame pays for every frame again.
- Write down what each photo is for before you render: store page, feed ad, banner or billboard. A model that knows the job guesses less.
- Save the prompts of the photos you keep, and start the next similar shot from them.
How to count your own
Save every attempt, including the bad ones, and mark the one you keep. At the end of a job, divide all renders by keepers. Count packshots, photos with people, and photos with motion separately, because they behave differently. After two or three jobs you will know your own number, and you can quote cost and time from it instead of from a price list that assumes every render is a keeper.
Questions people also ask
- Does a more expensive model mean fewer tries?
- Sometimes, but our numbers cannot show it, because each campaign changed the model and the kind of shot at the same time. To find out for your own product, run the same shot on two models ten times each and count the keepers.
- What counts as a keeper?
- In our count, a keeper is the version an art director picked for use: the product looks like the real product, nothing is warped, and it does the job it was made for. A render that would need heavy retouching to get there did not count.
- Is eight renders per photo a lot?
- Not in money. At a few cents to about twenty cents a render, eight tries costs less than two dollars per photo on the models above. It is a lot of looking, so the real cost is the time someone spends reviewing.
Where Overs fits
Overs follows the workflows behind these four campaigns. It shows the shot plan and an estimated cost before any photo is made, gives each photo only the reference pictures it asked for, and lets you remake one photo with a note, up to four versions at once. Its estimate counts one render per photo, so for a finished set, budget for the tries above.
Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.