Production guide

The best AI image models for ads, judged on what an ad needs

Nano Banana, Seedream, GPT Image, FLUX.2 and Ideogram judged on what an ad needs: your product kept intact, text spelled right, price per usable image and speed.

Guide · No. 31

Reading time
11 minutes
Updated
Topic
Production

The short version

  • Judge a model on four things an ad needs: it keeps your real product intact, spells on-image text right, costs little per usable image and returns a batch quickly.
  • For ads built from a product photo, use an editing model that takes reference images, such as Nano Banana 2 or 2.1, Nano Banana Pro, Seedream 5.0 Flash edit, FLUX.2 or GPT Image 2.
  • List prices on the providers' own pages ran from under a cent to about 24 cents per image in October 2026, which is small next to what it costs to test an ad.
  • Compare price per usable image, not price per image: a cheap model that mangles labels costs more in review time and risk.
  • Models change and retire quickly, so run a short test on your own packshot before you commit, and repeat it when a new version ships.

The short answer

There's no single best AI image model for ads. For ads built from your real product photo, use an editing model that takes the photo as a reference: as of October 2026 that means Google's Nano Banana 2 family or Nano Banana Pro, ByteDance's Seedream 5.0 Flash edit, Black Forest Labs' FLUX.2 or OpenAI's GPT Image 2. For designs where the headline does the work, start with a model whose maker puts text rendering first, such as Ideogram V4 or Seedream 5.0 Pro.

Then choose on your own product, not on a demo reel. A model that paints lovely landscapes can still redraw your label.

What an ad image has to get right

Winning an art leaderboard is a different skill from making a usable ad. An ad image has four jobs, roughly in this order:

  1. Keep the product true. Same shape, label and colours as the thing that arrives in the box. A redrawn jar is a promise you can't keep.
  2. Spell the words. If the headline sits inside the picture, every letter has to be right. One wrong character in a price and the ad goes in the bin.
  3. Cost little per usable image. Not per image. A model at half the price that needs three tries to get a clean label is the dearer one.
  4. Come back fast. Speed matters when you make a dozen ads at a time and review them in one sitting.

Also check that it makes 1:1, 4:5 and 9:16 without awkward crops, and whether it adds a watermark (more on that below).

How the main models compare

This is what each provider said on its own pages in October 2026. "Reference images" means photos you can hand the model to keep in the result, such as your packshot.

Model (maker) Built for Reference images List price per image What its page says about text
Nano Banana 2.1 (Google) Recommended by Google for all new projects Up to 10 object images $0.034 at 1K, $0.050 at 2K Legible, stylised text for marketing assets
Nano Banana 2 (Google) The previous-generation workhorse Up to 10 object images $0.067 at 1K, $0.101 at 2K Same claim as above
Nano Banana Pro (Google) The most complex, professional asset work Up to 6 object images, plus 3 style images $0.134 at 1K or 2K, $0.24 at 4K Same claim as above
GPT Image 2 (OpenAI) Generation and editing, reads every input at high fidelity One or more, no limit stated $0.005 low, $0.041 medium, $0.165 high at 1024×1536 OpenAI's guide says precise text placement can still be a struggle
Seedream 5.0 Pro (ByteDance, on fal) Flagship text-to-image Not stated $0.0675 up to 1536×1536, marked tentative Native text in 14 languages, dense layouts
Seedream 5.0 Flash edit (ByteDance, on fal) Editing where speed and budget matter Several, the example uses three $0.027 No claim made
FLUX.2 pro (Black Forest Labs) Production at scale Up to 8 on its own API, 9 on fal On fal, $0.03 for the first megapixel plus $0.015 per extra megapixel in or out Reliable text for marketing materials, with the flex model tuned for typography
Ideogram V4 (Ideogram, on fal) Posters, logos and text-led design Not stated on the fal page $0.0075 to $0.025 per megapixel, by speed setting Accurate text rendering

Sources: Google's Gemini API pricing (opens in a new tab) and image generation docs (opens in a new tab), OpenAI's image generation guide (opens in a new tab), Black Forest Labs' FLUX.2 overview (opens in a new tab), and fal's model pages for Seedream 5.0 Pro (opens in a new tab), Seedream 5.0 Flash edit (opens in a new tab), FLUX.2 pro edit (opens in a new tab) and Ideogram V4 (opens in a new tab). Google's prices are its standard API rate.

Two rows deserve a second look. Google lists the newer Nano Banana 2.1 at about half the price of Nano Banana 2. And the same model can cost more through a reseller: fal's page for Nano Banana 2 edit (opens in a new tab) charges $0.08 per image at 1K, against $0.067 on Google's own API.

OpenAI also lists two GPT Image 2.5 models, Sunburst and Flare, at the same token rates as GPT Image 2 on its pricing page (opens in a new tab), with no per-image price yet. Its guide suggests Sunburst where editing precision matters most and Flare for fast everyday images, so Sunburst is the one to try on a packshot.

Per-megapixel prices are easier to read at a real ad size. A 4:5 feed ad at 1080×1350 is about 1.4 megapixels by fal's count. FLUX.2 pro rounds that up to two, so it costs $0.045, or $0.06 with a one-megapixel reference photo. Ideogram V4 comes to roughly $0.01 to $0.035, by speed setting.

Two cautions. "Up to 10 reference images" is a ceiling, not a promise that all ten survive intact. And a maker's claim about text says nothing about your product's name.

Old versions also disappear. Google's pricing page marks the original Nano Banana as deprecated, with a shutdown date of 2 October 2026, and its docs list the older Imagen models as shut down. Plan to retest when versions change.

Product faithfulness: start from your photo

There are two ways to get your product into an AI ad.

Text-to-image describes the product in words and lets the model draw it. That's fine for a scene with no product, or a mood board. For a product ad it fails predictably: the model draws a typical bottle, not yours, with invented label text and a generic cap.

Editing with a reference hands the model your actual packshot and asks it to build the scene around it. It's the only approach that holds up for ads, because the product starts from your pixels. Even so, models drift on small things. Check these five spots on every image:

  • Small label text such as ingredients and weights, which blurs first.
  • Logos with thin strokes or unusual letterforms.
  • Caps, pumps and lids, which models like to tidy into a standard shape.
  • Muted brand colours, which drift towards brighter versions.
  • Glass and shiny packaging, where the model has to invent the reflections.

The input matters as much as the model. A sharp, evenly lit photo with the label readable at full zoom does more for faithfulness than any upgrade. Our guide to product photos for AI ads covers shooting one on a phone, and the white background product prompts show what a clean packshot prompt looks like.

On-image text: what to ask for and what to add later

One wrong letter makes an ad unusable, so treat on-image text as something you check, not something you trust:

  • Keep it short. Every extra word is one more chance of a wrong letter, and five words take seconds to check. Long copy belongs in the primary text field.
  • Put the exact words in quotation marks in the prompt, and describe the type plainly, such as "heavy white sans-serif".
  • Watch made-up words. An invented product name is often the hardest thing for a model to spell, because it has never seen it.
  • Check every number twice. Prices and pack sizes are where one wrong character costs most.
  • Leave room instead. If a model keeps fumbling, ask for clear space in the top third and add the headline as a text overlay afterwards, where the type is exact every time.
  • Keep legal lines and prices out of the image when you can, so a wrong character never reaches the picture.

Price per usable ad, not price per image

The list prices above are tiny. What decides your real cost is how many images you make, and look at, before you get one you'd put money behind.

The keep rates in that example are invented to show the sum. Yours depend on your product, and they're the number worth measuring. Two lessons hold either way:

  • The model is rarely the expensive part. Review time and test spend are. Saving ten minutes of checking per batch beats saving a cent per image.
  • A bad image that slips through costs most. A mangled label that reaches buyers outweighs any saving on generation.

For the full picture with people and time counted, see the cost per ad creative.

Speed: what the providers say

Speed is the least comparable line on any model page, because it depends on size, quality settings and load. Here's what the providers said in October 2026:

  • Black Forest Labs describes its FLUX.2 klein models as sub-second, built for real-time and high-volume work.
  • fal describes Nano Banana 2 edit as running at Flash-tier speed, with edits back in seconds.
  • OpenAI's guide says complex prompts can take up to two minutes.
  • Google's batch mode charges about half the standard rate in return for a turnaround of up to 24 hours.

For ad work, the question is whether a batch of a dozen comes back while you're still at your desk. The half-price overnight mode suits big catalogue jobs where nobody is waiting.

A short test to pick your model

Run the same three prompts through two or three models with your own packshot, all at 4:5, and score what comes back.

Prompt 1, the product test (editing model, packshot attached):

Place the product from the reference photo on a pale oak kitchen counter beside a glass teapot, soft morning light from a window on the left. Vertical 4:5. Keep the product exactly as in the reference: same tin shape, same lid, same label text and colours. Leave the top third of the frame clear.

Look for: every changed letter, colour or shape on the label at full zoom.

Prompt 2, the text test (packshot attached):

Bold flat-colour ad on a deep green background. The product from the reference photo on the right third. On the left, the headline in heavy white sans-serif letters: "Smoky lapsang, zero bitterness". Vertical 4:5.

Look for: every letter of the headline, and whether the product survived the bolder design.

Prompt 3, the scene test (no reference):

Phone-camera photo of a cluttered home desk at 4 pm, a laptop, a notebook and an empty mug, warm lamp light, slightly grainy, no people, no text, no logos. Vertical 4:5.

Look for: a real phone photo, not an advert pretending to be one. This tests the lo-fi creative look many ads rely on.

Make four images per prompt on each model and score each one:

Check Passes if
Product shape and lid Identical to the packshot
Label text Every word readable and unchanged
Colours Match the real product in daylight
Headline Every letter right, no stray characters
Composition Clear space where you asked, product not cropped
Realism No warped hands, melted objects or impossible shadows

Divide each model's price by the share of images that passed every line. That price per usable image is the figure to compare. Keep the prompts and scores for the next version.

Watermarks, metadata and AI labels

Know how your images may be marked before you launch:

  • Google's image generation docs (opens in a new tab) say every image from its Gemini image models carries an invisible SynthID watermark.
  • Meta's post on labelling AI content (opens in a new tab), first published on 5 April 2024 and last updated on 23 October 2025, describes an "AI info" label for content where Meta detects industry-standard AI signals or people disclose AI. An update dated 10 May 2024 says that labelling applies to organic content. The same post says advertisers must disclose digitally created or altered political or social issue ads in certain cases.
  • For ads, Meta's Marketing API documentation (opens in a new tab) says images created or materially edited with certain of Meta's own generative AI features may show "AI info" in the ad's menu or next to the Sponsored label (as of October 2026). That page is about Meta's own tools, not images you make elsewhere.

For an ordinary product ad, the rule is older than any model: don't show what the product can't do. An AI kitchen behind your tea tin is a setting. An AI "result" your product doesn't deliver is a misleading claim, whatever made it. Our guide to Meta's generative AI ad tools covers those features one by one.

How BulkAds uses image models

You don't pick a model in BulkAds. The BulkAds team sets one per job, and every result is checked:

  • Ads made from your product photo use an editing model that takes your packshot as the reference. By default that's Google's Nano Banana 2 edit, run through fal.
  • Scenes made without a product photo use a text-to-image model, Seedream 5.0 Pro by default.
  • Every image is checked by a separate AI review that compares the product with your original photo, reads the on-image text against what was asked for, and checks brand colours, quality and likely policy risks. Ads that fail are flagged before you see them. See product photo ads.
  • Your price doesn't follow the model. An image ad costs 2 credits whichever model made it, so a model switch doesn't change what you pay per ad.

To skip the setup, a batch of up to 12 ads from your packshot in the AI ad generator shows how today's models handle your product.

Questions people ask

Which AI image model is best for product ads?

The one that keeps your product unchanged in a test on your own photo. As of October 2026, editing models that take a reference image, such as Google's Nano Banana 2 and Nano Banana Pro, ByteDance's Seedream 5.0 Flash edit, FLUX.2 and GPT Image 2, are the ones to try. Text-to-image models draw a typical product, not yours.

Which AI model is best at text in images?

As of October 2026, Ideogram, Google and Black Forest Labs all put text rendering near the top of their model pages, and fal's page for Seedream 5.0 Pro claims native text in 14 languages. OpenAI's guide says its model can still struggle with precise text placement. Test your own headline, keep it short, and add long or legal text as an overlay.

How much does it cost to make an AI ad image?

On the providers' pages in October 2026, list prices ran from under a cent per image at the cheapest settings to about 13 to 24 cents for Nano Banana Pro. Your real cost is the price divided by the share of images you can use, plus the time it takes to review them.

Can AI keep my product's label exactly the same?

Often, not always. Editing models start from your photo, so the product usually survives, but small text, thin logos, lids and exact colours can drift. Check every ad at full zoom before it runs, or use a tool that compares each ad with your original photo.

Will Meta label my ad as made with AI?

It can if you used Meta's own generative AI features. As of October 2026, Meta's Marketing API documentation says ad images made or materially edited with certain of them may show an AI info label. Meta's wider AI labelling post covers organic posts, and political or social issue ads have their own disclosure rule. Some models also add invisible marks, such as Google's SynthID.

Which image models does BulkAds use?

By default, Google's Nano Banana 2 edit for ads built from your product photo and Seedream 5.0 Pro for scenes without one. Every ad is checked against your original photo, and an image ad costs 2 credits whichever model made it.

Keep going

Written by the BulkAds team. Published , last checked . Meta changes its tools often — if something here no longer matches Ads Manager, tell us at contact@bulkadsai.com.

BulkAds AI is an independent product and is not affiliated with, endorsed or sponsored by Meta Platforms, Inc. Facebook and Instagram are trademarks of Meta Platforms, Inc.

Your next fifty ads are one link away.

Paste your website, meet your brand kit and see your first batch in minutes. 7 days for $9.99 — cancel anytime.

yourbrand.comMake my ads

See pricingExplore featuresBrowse the library