AI Ad Generator: Make Static Image Ads Free, Without Faking Proof

1. Pick your free generator

ChatGPT. Sign in at chatgpt.com, start a chat and describe the image you want. Images 2.5 added templates for common formats and comments you place on an image to edit one spot. OpenAI doesn't publish a fixed free image count; ChatGPT shows you when you hit the limit.

Gemini. Sign in at gemini.google.com and ask for an image. Google's plans page says limits may change without notice. New images can carry a small visible watermark; Google's watermark page explains the setting (checked 15 September 2026).

Pick the one you already have an account for and learn it. Switch only if it keeps failing the step 4 test.

Why: both are free to test, and the prompts below work in either.

2. Write your offer sheet before anything gets made

This is the rule that keeps you out of trouble: nothing goes on the image unless it's on this sheet, and nothing goes on the sheet unless you can show where it came from. Your page, your price list, a review screen you can screenshot, a photo with consent.

Mark each price live or not approved. Only live ones go on an ad. If a monthly payment appears, US credit ads have disclosure rules under Regulation Z.

Why: the model will fill every gap with something plausible. A sheet with no gaps leaves it nothing to invent.

OFFER SHEET: [business name]
What I sell in this ad: [one product or service]
Who it's for: [customer, in one line]
The customer's words: [a sentence a real customer said or wrote to me]
One benefit, in my page's wording: [paste]
Price or offer (live only): [paste] | Source: [page URL]
Proof I can show: [review count, rating, award, credential] | Source: [link or screenshot]
Call to action: [Book / Shop / Get a quote]
Where: [city or full address]
Real photos I own, with consent: [list]
Not allowed on this ad: anything not listed above

3. Decide direct or indirect

Swipe structure from ads that already work before you pick. In the Meta Ad Library, the filter I use is simple: still running, and live at least 7 days. Borrow the structure, never the wording. What makes a good Facebook ad goes deeper.

Then paste this with your offer sheet. It picks the type, writes the ad text that carries it, and names the audience.

Why: a direct ad shown to strangers and an indirect ad shown to past visitors both waste money, and the image prompt is different for each.

Here is my offer sheet: [paste].
Audience: [cold, or people who visited my page or engaged with my content].
Decide whether this ad should be direct or indirect and say why in two sentences.
If direct: the image shows the product or service, one benefit, the offer and a call to action.
If indirect: the image only stops the scroll, you can't tell what's sold from it, and the ad text does the selling.
Write the ad text for the version you chose and tell me which audience to run it to.
Use only facts from my offer sheet. No em dashes.

4. Make one small test ad

Before the full layout, run one simple direct ad: a headline, your product, one line and a button. You're testing two things, whether the model spells your words right and whether the picture looks like your business.

Put every word you want on the image in quotes, and tell it there are no other words. If you have a real product photo, attach it and the model will keep it closer to the real thing.

Why: one cheap image shows you the model's habits before you spend an afternoon on a layout.

Make a 4:5 vertical static ad image for [business] in [city].
If a photo is attached, it is my real product. Keep its shape, color, label and branding exactly.
Show [product or service] in a real [setting], natural light, looks like a real photo, not a render.
The only words on the image are, exactly:
Headline: "[SHORT HEADLINE, NO APOSTROPHES]"
Line: "[ONE BENEFIT FROM MY PAGE]"
Button: "[CALL TO ACTION]"
Do not add any other text, logos, star ratings, review counts, prices, badges or statistics.
All text crisp, correctly spelled and fully visible.

5. Build the full layout

Once the test passes, use the layout below. It follows the skeleton of the statics that produced leads across my agency's clinic accounts, generalised for any local business: customer-words headline, the real thing, one benefit, sourced proof, one offer, button, and where you are.

No apostrophes in big headline type. On our render runs, condensed capitals turned ISN'T into IN'T often enough that we write IS NOT, DO NOT, CANNOT.

Change the words, product and person. Keep the layout, so when an ad wins you know why.

Why: a fixed layout turns ad testing into testing one idea at a time.

Premium static ad for [business], 4:5 vertical.
STYLE: [2 brand colors], one accent color only, bold condensed sans headline, clean small sans body, editorial photography, real textures, natural light.
LAYOUT top to bottom:
Headline in capitals, no apostrophes: "[THE CUSTOMER'S SENTENCE FROM MY OFFER SHEET]"
Main image: [real product or service in use, setting, person if any], filling half the frame, looks like a real photo.
One benefit line: "[BENEFIT FROM MY PAGE]"
Proof line, only if on my offer sheet: "[SOURCED PROOF]" (otherwise leave this line out)
Offer: "[ONE LIVE PRICE OR OFFER]"
Button: "[CALL TO ACTION]"
Footer: "[BUSINESS NAME, CITY OR ADDRESS]"
Small print, only if needed: "[TERMS OR DISCLAIMER]"
QUALITY: every quoted string present once, spelled correctly, fully legible, nothing cut off, no extra words.
AVOID: plastic skin, AI sheen, warped hands, extra fingers, stock poses, invented logos, star ratings, review counts, statistics or badges not quoted above, clutter.

6. Use your own photo and real people safely

The most honest ad starts from a real photo: your product, your shop, your team. Upload it and tell the model what to keep and what to change. Autocomplete shows people search for making an ad "from image", and this is how.

A real staff member on an ad needs their consent and this identity lock line. Then compare the render to the photo yourself. If faces come out waxy, make AI images look real has the fixes. For small edits on a render you like, see Nano Banana image editing.

Why: a generated stranger reads as stock, and a changed face on a real person is a misrepresentation.

Use the attached photo of [person or product].
Preserve the exact facial likeness, hairstyle, skin tone, build and clothing as photographed. Do not alter the face.
For a product: keep its shape, color, label and branding exactly.
Re-light and re-compose only: place it [where] at [size] in the layout I gave you.
Change nothing else.

7. Read the batch together, and fix once

Make 4 to 6 versions, then look at all of them side by side before you judge any single one. That's the contact sheet.

Allow one fix per image: keep everything identical, change only the defect. If a text mistake survives one fix, reword the line so it's easier to render. That beats a fifth reroll. A face that looks AI reads dishonest, whatever the copy says, so discard it.

Why: endless rerolls burn your daily limit and drift the image further from what you asked for.

Here are [N] versions of my ad: [attach].
For each: is every quoted string present once and spelled correctly, is anything cut off or duplicated, do hands, faces and the product look real, does it look like my business.
Mark each: ship, fix once (give me the exact one-line fix), or discard.
If the same text mistake appears twice, don't tell me to try again. Rewrite the line so it is easier to render.
No em dashes.

8. Run the check before you publish

This is the step most AI ad tutorials skip. Zoom to 100% and read the image yourself, then paste it back into a chat with the prompt below. It lists every claim and asks where each came from. An invented number gets removed, not swapped for a different number.

In the most-quoted video in this genre, a creator says he thinks the model made up a statistic, then ships the ad anyway. Don't.

Why: once an ad runs, every word on it is yours, legally and with your customers.

CHECK BEFORE PUBLISH
Text on the image
[ ] Every word spelled right at 100% zoom, nothing cut off or repeated
[ ] No words I didn't write (fake badges, stray letters, made-up slogans)
[ ] Phone, address, URL and business name exactly right
Claims
[ ] Every number, rating, review count, award and credential is on my offer sheet with a source
[ ] No result, before and after, or testimonial the model made up
[ ] Nothing says or implies the viewer has a health condition or other personal attribute
Offer accuracy
[ ] Price and offer are live today and match my landing page
[ ] The product or service shown is what the customer receives
[ ] Terms, deposit or financing line present if the offer needs one
People and platform
[ ] Real people have given consent, and their faces weren't changed
[ ] Fits Meta's policies for my category (step 9)

Paste into the chat:
Here is my ad image: [attach] and my offer sheet: [paste].
List every word, number, rating, review count, award, credential and claim on the image, one per line.
For each, point to the line of my offer sheet it came from. If you can't, mark it INVENTED.
Tell me whether anything on the image shows a product, result or service that doesn't exist exactly as shown.

9. Health, beauty and personal attribute rules

If you sell anything health, wellness or beauty related, Meta's health and wellness policy (updated 22 July 2026) applies. It bans statements that attack how someone looks, restricts promises of specific results in a set time without qualifiers, and limits cosmetic procedure ads, including before and afters, to people 18 and over.

Every business is covered by the personal attributes policy: ads can't assert or imply things like a person's health, age, finances or religion. Meta's own disallowed example is "Are you Christian?" A headline built on "you have this problem" gets rejected the same way (checked 15 September 2026).

Never generate a before and after, not even as a placeholder. It's a made-up result, and the FTC's health products compliance guidance expects health claims to have real evidence behind them. Use a real, consented, unretouched photo pair or a different angle entirely.

Why: an image model will happily draw a transformation. The rejection, or worse, lands on you.

10. Make two sizes and launch

Meta's Facebook feed image spec is 4:5 at 1440 by 1800 pixels, JPG or PNG (checked 15 September 2026). Make that your master. Build the 9:16 Stories and Reels version from it with the important parts inside the safe zone, instead of letting it crop automatically, because the headline and the address are what get cut.

In Ads Manager, map each size to its placement, turn standard enhancements off so Meta doesn't rewrite what you checked, and use one primary text and one headline per ad. Judge results on your CRM's lead count, not the platform's. For more angles from a winner, see ad creative variations, and to connect Claude to the account itself, Meta ads MCP.

Why: speed isn't quality. If the offer doesn't work, 50 renders of it won't either.