An AI ad creative generator that ships finished ads
A creative is not a stock photo with a caption bolted on. It is the picture, the words on the picture, and the post they belong to, decided together. Here is how the generator does that, and how to judge what it hands back.
What an ad creative actually is
In the trade, the creative is everything the reader sees: the image, the words printed on it, the caption underneath, and the one thing you want them to do. People say creative when they mean picture, and that slip is why so many ads are a good photograph carrying a caption that has nothing to do with it.
A generator that only makes pictures leaves you doing the hard part, which is deciding what the picture is arguing. Serious Poster generates the argument first and the picture second, which is the whole difference between the output being a finished ad and being a nice asset you now have to write around.
Words first, picture second
The generator runs the copy stage before the render stage, and everything good about the output comes from that ordering.
The copy model decides the angle for this ad, writes the hook, writes the body, then writes two more things that most pipelines never produce: an array of the exact text to appear on the image, in reading order, and a complete standalone prompt describing the scene, the composition, the styling and where each piece of that text sits.
Only then does the image model run, and it is given both. It is not being asked to imagine an advert for a plumbing business. It is being told to render this scene, with these words, in these positions. That is a far easier job, and it is why the output looks like something with an intention behind it.
| Approach | What the image model is told | What you get |
|---|---|---|
| Picture first | Make an advert image for a plumbing company | A stock-looking scene you must then write a caption for |
| Words first | Render this scene, with these exact words, in these positions | A creative whose image and caption make the same argument |
Why text on AI images used to be a disaster
Anybody who tried this two years ago remembers the result: confident lettering that spelled nothing, words with an extra limb, a headline that read almost like English. It made generated images useless for advertising, because advertising is mostly text.
Two things changed. One is the models. The other is the prompting, and that half is under our control.
- Give the text verbatim. Quoted, exactly, every piece of it. A model asked to write a catchy headline on the image invents the headline and then renders its invention badly.
- Give the reading order. Which line is the big one, what sits under it, what is a small corner label. Layout is information the model cannot infer from a list of strings.
- Keep it short. Six words on a creative reads at a scroll. Thirty does not, whatever the render quality is, because nobody stops for a paragraph in a feed.
- Do not put the offer only on the image. The caption carries it too, because images do not get read by anybody scrolling with data saving on, and they do not get read by search.
Using your own brand imagery
Each business in the studio holds its own imagery: a logo, product shots, and reference images that say what your work looks like. It is stored per business, which matters most for agencies, because the fastest way to lose a client is to put another client's look in their ad.
The point of the reference set is not literal reproduction. It is that the generated creative sits in the same world as the rest of your material, so a person who has seen your van, your storefront or your other posts recognises this as the same operation.
How many creatives a campaign needs
The instinct is either one, which is too few, or forty, which is a waste of money and attention. The number is decided by how the poster uses them rather than by any rule about testing.
A campaign draws its attachment for each post at random from the media sets you attached. So the question is not how many creatives exist, it is how likely two neighbouring groups are to receive the same picture. With four creatives across twenty groups, that is rare enough to stop being a pattern. With one, it is certain.
| Campaign size | Creatives | Why |
|---|---|---|
| Under 10 groups | 2 to 3 | Enough that no two groups in a day match |
| 10 to 30 groups | 4 to 6 | The common case, and where one generation lands you |
| 30 or more groups | 6 to 10 | Plus your own photos, which cost nothing to add |
| Several businesses | 4 to 6 each | Kept separate, because voices should not blur |
One image in forty groups is a signal
This is the reach argument, and it is the reason the media set exists at all. Images are fingerprinted cheaply. Forty posts carrying one identical file group themselves together whether or not the text around them varies, which quietly undoes the work that spintax did on the wording.
You do not get told this has happened. The post stays up, it looks perfectly normal on your own screen, and it is shown to a fraction of the group. The shadowban guide covers how to notice it and what to do about it.
Judging a creative in ten seconds
You will look at twenty of these at once, so a fast rubric beats a careful one. Squint at it on a phone and ask four questions.
- Can I read the biggest line without stopping? If not, nothing else matters. Delete it.
- Does it say something, or just look like an advert? A creative with no claim on it is decoration.
- Would this look out of place in the group? A polished agency-grade banner in a neighbourhood buy-and-sell group announces itself as an intrusion. Slightly plainer performs better there.
- Is anything on it untrue? This is the one to actually check. Generated copy can invent a guarantee, and an invented guarantee on a picture is the version that gets screenshotted.
Mixing generated and real photography
The best-performing media set in a group context is usually mixed. Generated creatives carry the offer and the words. Real photographs carry proof, and nothing renders proof.
| Use | Generated creative | Your own photo |
|---|---|---|
| The offer or price | Yes, this is what it is for | Only if you own a design tool |
| The actual product or property | No | Always, and it beats everything |
| Before and after | No | Yes, and it is the strongest thing you have |
| A concept or an illustrated idea | Yes | Rarely available |
| Faces of real staff or customers | No, and do not fake it | Yes, with permission |
Drop both into the same media set. The poster does not care which is which, and the campaign gains variety from having two visual registers rather than one.
The workflow, end to end
Start to finish, this is about ten minutes of your attention and a few minutes of waiting.
- Describe the business once, with real facts and a couple of brand images.
- Generate. Pick how many audiences and how many ads each, and the studio shows the count against your allowance before it starts.
- Walk away. Rendering takes minutes and runs on the server, so closing the tab is fine.
- Read the batch. Delete the two you dislike, fix the one fact that is wrong.
- Put the keepers in a media set, attach your groups, and start a run with your usual pacing settings.
The AI Facebook ads guide covers the copy and targeting halves of the same pipeline, and AI marketing automation covers running it as a weekly habit rather than a one-off.
Still stuck? Mail us and a person will answer.
Add to Chrome