Craft

How to make AI image ads with Nano Banana and GPT Image

By James Tumbucon
Updated · 6 min read

To make AI image ads, give the image model your real product photo and brand colours, describe one format at a time (flat lay, splash, infographic, poster, social post), generate a grid of options, pick and upscale the best, then check text, logo and safe zones. Most creators use Nano Banana or GPT Image.
On this page
  1. What is an AI image ad?
  2. Tools creators use for it
  3. How do you make AI image ads step by step?
  4. Prompt template for a static ad
  5. Common problems and fixes
  6. What creators report

What is an AI image ad?

An AI image ad is a single static image, made with an image model, that runs as a feed, Story or carousel ad. It shows the product in a designed scene with a headline, a short benefit and often a price or offer.

Static ads are the cheapest format to test, because one generation is one finished candidate. That is why creators often test an angle as a static first. The common static formats in the tutorials are product photography (flat lay, splash, levitating product), infographics, posters, testimonial or review cards, and native-looking social posts.

Tools creators use for it

Two image models dominate this workflow. The counts show how many of the reviewed static ad tutorials used each tool.

Based on 28 public tutorials posted between 1 Jan 2026 and 11 Oct 2026, reviewed by Anglr.

Tool What it does in this workflow Tutorials that used it
Nano Banana (Pro, 2) Product photography, grids, flat lays, splashes, posters 23
GPT Image (1.5, 2) Text-heavy layouts, handwriting, social interface mock-ups 13
Higgsfield Runs GPT Image and other models; its Ads Studio builds statics from brand inputs 3
Figma Holds real headline, copy and logo layers over AI images 1
Claude Writes angles, copy and prompts and checks outputs 1

Nano Banana and GPT Image have no page in our AI video pricing set, because they are image models; check the price in the app you use them through.

How do you make AI image ads step by step?

You make them in six steps: brand inputs, product reference, format choice, prompt, grid and pick, then text and checks. The steps are our synthesis; each notes where the sources agree and differ.

1. Collect the brand inputs once

Put the logo, brand colours as hex codes, fonts, product photos and tone rules in one place before you prompt. Estefania Freites builds one brand research document that every ad inherits. Tim Davidson keeps one agent per client holding the brand rules. Mike Futia feeds in the logo file, a product photo and hex codes. Ruizhe (Richard) Ma had to correct the colours to the brand palette after the first try, which is what a fixed input sheet prevents.

2. Attach the real product photo

Upload the product photo and tell the model to use it as the product. @oggii_0 builds a flat lay around an uploaded product. Richard Ma asked GPT Image 2 to use the actual product image. Mike Futia and Estefania Freites also use real product references. Google's Nano Banana guide describes this as reference images plus a relationship instruction plus the new scene.

3. Choose one format per prompt

Pick the format before you write the prompt; each needs different fields. These formats appear in three or more tutorials each:

Format What the prompt specifies Seen in
Flat lay Top-down view, product centred, matching props or ingredients @oggii_0, @ChillaiKalan__, @Kerroudjm
Splash or levitation Product in mid-air, frozen liquid, droplets, condensation @AIwithSynthia, @Fujimoto_hina, @Naiknelofar788
Infographic Product plus ingredient icons, steps or callouts @AIwithkhan, @AIwithSynthia, @AmirMushich
Poster Oversized headline, brand colour blocks, illustrated or photo scene @AmirMushich, @AIwithkhan, Richard Ma

Will Sartorius adds native social formats: a post that looks like a Story or a feed post, and handwritten notes. Estefania Freites shortlists five or six formats that are currently working in the category rather than using every template.

4. Write the prompt with fixed fields

Write the prompt as a set of fields you can change one at a time: product, scene, composition, camera, light, style, text and exclusions. Many creators share these as JSON (@mehvishs25, @Fujimoto_hina, @john_my07, @Just_sharon7). Others use fill-in templates with placeholders (@Kerroudjm, @ShamsAmin56, @ChillaiKalan__).

The sources differ on format, not on content. Google's guide writes prompts as sentences and says to state the purpose of the image, for example an ad for a minimal skincare brand. The template below uses sentences.

Style words matter too. Wayne St Ledger recommends restrained, editorial words over hype words, and one subject per image: "Pick one hero, one product or one person." He also lists what the model must not add.

5. Generate a grid, pick, then upscale

Ask for several options in one image, pick the best, then upscale it. @hellorob generates a nine-image grid, extracts the favourite and upscales it to 4K using the grid as the reference. @gizakdag asks for a nine-image feed. @Gdgtify uses a 2x2 grid with one style instruction applied to each cell.

Run the same prompt in both models when you are unsure. At least five authors compare Nano Banana and GPT Image side by side with one prompt. Their verdicts differ by format: Will Sartorius and Richard Ma prefer GPT Image 2 for text and handwriting; Will Sartorius keeps Nano Banana 2 for products that photograph like a retail shelf.

6. Add text, check safe zones and export sizes

Decide whether the model writes the text or you add it after. GPT Image 2 renders short headlines legibly according to Richard Ma and Will Sartorius. Mike Futia takes the other route and recommends generating only the background, with headline, copy, call to action and logo as live layers in Figma, so a typo is a retype rather than a regeneration.

Keep the headline out of the areas the platform covers. Will Sartorius reserves the top and bottom of a 9:16 image as empty. Check yours with the safe zone checker. Export 1:1, 4:5 and 9:16 versions, as Mike Futia and Estefania Freites do.

Before it runs, check the logo is the real one, the product matches the photo, the text has no errors and every claim is one the brand can make. Meta labels ad images made or significantly edited with third-party AI tools with AI info (Meta).

Prompt template for a static ad

This template is ours, written in the order Google documents for Nano Banana: product and reference, subject and action, setting, composition, camera, light, style, then text in quotation marks and the aspect ratio. The same fields work as a starting point in GPT Image; OpenAI's structure is not in our prompt notes, so test it. More formats are on Nano Banana prompts, and the AI ad prompt builder fills the fields for you.

Use the attached photo of [product] as the product: same shape, colours and label. A static ad image for [brand type], made to sell [benefit] to [audience].
Scene: [product] [action: rests on / rises out of / is held above] [surface or setting], with [one or two props].
Composition: product in the lower centre, empty space in the top third for the headline, nothing in the top and bottom bands.
Camera: [angle], [lens], sharp focus on the label.
Lighting: [light source and direction], [mood].
Style: restrained, editorial, natural colours in [brand colours]. One product only. No badges, icons, stickers or extra text.
Text: the headline "[headline]" in a bold sans-serif in [colour], top centre.
4:5.

Common problems and fixes

Problem Fix Source
Grid cells are too small, so product text is lost Extract the chosen cell and upscale it with the original as reference @hellorob
The headline lands inside the platform's interface Mark the top and bottom bands as must-be-empty and over-specify it Will Sartorius
Nano Banana 2 uses a handwriting font instead of real handwriting Use GPT Image 2 for handwritten formats Will Sartorius
A prompt from an older model gives flat, repetitive results Remove phrases such as same aesthetic that the newer model follows too literally @gizakdag
Baked-in typos and a logo that is almost right Generate the background only; add text and logo as real layers Mike Futia
The model redraws the uploaded logo instead of using it No reliable fix stated; a later prompt worked for the author Tim Davidson
The image fills with clutter Say what the model must not add Wayne St Ledger

What creators report

These are the authors' own claims.

  • Will Sartorius reports running one brand through 23 ad formats in more than 50 generations, with most usable on the first or second pass.
  • Estefania Freites reports shipping 40 static ads in a week and reading the first 24 hours of results to pick the next formats.

If your static and video ads already perform, see how Anglr pays creators on performance on Anglr for creators.

Who this is for

  • Independent AI creators and editors. You make the ads yourself. You know your work performs, so you'd rather be paid on what it does than on a flat fee.
  • AI agencies and studios. You run a team of editors. You write the brief, your team makes the versions, and you're paid on performance because you know your team's ads scale.

Not for you if

  • You only want a flat fee.
  • You're new to editing and not yet confident your ads will perform.
  • You want every shot spelled out for you.
  • You make one style and won't try others.

FAQ

Is Nano Banana or GPT Image better for ad images?

The tutorials we reviewed do not agree. Several found GPT Image 2 better at legible text, handwriting and social interface layouts, while one author keeps Nano Banana 2 for clean retail product photography. Run the same prompt in both and compare.

Can AI image models write the headline on the ad?

Newer models render short headlines much better than older ones, and Google's guide says to put the exact words in quotation marks. Many creators still add the headline, price and logo in a design tool so they can fix typos and change copy without regenerating.

What sizes do AI static ads need?

Creators in the tutorials export 1:1 and 4:5 for feeds and 9:16 for Stories and Reels. Exact sizes and safe zones per platform are on the ad specs pages, and you can check a finished image with the safe zone checker.

Do AI image ads get a label on Meta?

Meta says it labels ad images created or significantly edited with third-party AI tools with AI info, shown in the About this ad menu (Meta). Details for both platforms are in AI ad labels on Meta and TikTok.

Do I need JSON prompts for Nano Banana ads?

No. Many creators share JSON prompts because the fields make it easy to swap one thing at a time, but Google's own guide uses full sentences. The fields matter more than the format.

Sources

Checked: 12 Oct 2026