Nano Banana Pro for Ad Creative: Legible Headlines, Price Badges, and the 2K Sweet Spot

Your headline has to survive the render

A skincare brand needs twenty-four statics by Friday: three headlines, two price badges, four backgrounds, in both the feed shape and the story shape. The product photography is already done. The blocker is the four words sitting on top of it, because the last batch came back with the product name spelled two different ways inside the same set, and nobody caught it until the second review. Resolution was never the hard part.

Nano Banana Pro, which Google also calls Gemini 3 Pro Image, was announced on November 20, 2025 and is, in Google's own words, "Built on Gemini 3 Pro." The launch line that traveled furthest was about pixels: output ready for any platform, "thanks to a range of available aspect ratios and available 2K and 4K resolution."

Nine months on, that line has aged. Google's own developer docs now describe four Nano Banana models, and the generalist one sitting below Pro "balances speed with state-of-the-art 4K generation" as well. So this is not a story about a new 4K model, and if you came looking for one, the number stopped being a differentiator a while ago. What did not age is the part of that announcement an advertiser should actually care about: the words inside the picture, the consistency of a set, and the resolution you deliberately choose to pay for.

Where Nano Banana Pro sits in Google's lineup today

Before deciding whether to reach for it, place it. A model's position in its own family changes what it is for, and Google restructured this family after the launch that made the model famous.

Four models, and Pro is the premium slot

Google's Gemini API documentation is blunt about the shape of the line: "Nano Banana refers to four distinct models available in the Gemini API." In the docs' own descriptions, they stack like this:

• Nano Banana Pro (Gemini 3 Pro Image), described as "The premium choice for the most complex visual tasks, offering the highest level of world knowledge, advanced localization, accurate brand consistency, and precision creative control."
• Nano Banana 2 (Gemini 3.1 Flash Image), the generalist, which "balances speed with state-of-the-art 4K generation, world knowledge, and reliable text rendering."
• Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image), "Our fastest and cheapest Gemini image model," built for velocity and scale.
• Nano Banana (Gemini 2.5 Flash Image), the original, which Google now describes as "The legacy pioneer of the Nano Banana series."

So Pro is the top of that list by capability, not by recency. Read its four listed strengths as a job description rather than as marketing: world knowledge, localization, brand consistency and creative control are, in order, most of what a static ad needs from a renderer.

The resolution ladder flattened

The launch framing put 2K and 4K on the Pro tier. That is no longer where the line is drawn, and the giveaway is sitting in the list above: the phrase "state-of-the-art 4K generation" belongs to the generalist model now, not to Pro. At the bottom of the range, "Gemini 3.1 Flash Image adds the smaller 512px (0.5K) resolution."

So the family runs from half a thousand pixels to four thousand, and the top of that range is reachable on more than one rung. If your reason for choosing Pro is the 4K number, you no longer have a reason. We looked at the fast end of the same family when Nano Banana 2 Lite arrived built for volume, and the split holds: cheap tiers for exploring, premium tiers for the asset that has to be exactly right.

What is still Pro's argument

Strip the resolution claim out and something specific is left. Three of the four strengths Google lists for Pro are things a media buyer can act on directly: localization, brand consistency, creative control. None of them are about sharpness. All of them are about whether a set of twenty-four images still reads as one brand, in the language the market speaks, with the label saying what the real label says.

The in-image text problem is the ad-creative problem

Here is the thing about static ads that people who do not make them tend to miss. Most of the surface area is not the photograph.

A static ad is mostly typography

Open any high-performing feed static and count what the render has to get right:

• the headline, which is the whole hook
• the price or discount badge, which is usually the reason to stop
• the CTA, which has to read at a glance
• the pack copy on the product itself, which has to match the real pack
• the legal line, which is small and unforgiving

Five text elements, one image. A model that produces gorgeous lighting and garbles two of them has produced a mood board, not an ad.

What Google actually claims here

This is where Nano Banana Pro's pitch is strongest and most specific. Google says it "is the best model for creating images with correctly rendered and legible text directly in the image," for a short tagline or a long paragraph alike. The multilingual half matters just as much for anyone running one offer across several markets: "you can generate text in multiple languages, or localize and translate your content."

That is a claim about the exact failure mode that made image models unusable for statics for years. It also is not unique to Google. Ideogram built an entire release around in-image text with explicit layout control, and Seedream 5 Pro leans on multilingual rendering for the same reason. Text rendering has quietly become the axis image models compete on, because it is the axis commercial work actually needs.

Read it back before you ship it

A model that renders text well will still render text you never wrote. Given a bottle and a vague brief, it fills the empty label space with something that looks like marketing, and the invention is always plausible enough to survive a glance.

The discipline that fixes it is boring and it works: write out every word that is allowed to appear, then proofread the picture instead of the brief. Four checks, on the first render of the set:

• Zoom to 100 percent on the label and read the pack copy word by word against the real pack, including the size and the unit.
• Read the headline out loud. A dropped letter that the eye smooths over at thumbnail size does not survive being spoken.
• Check the discount figure against the offer that is actually configured in the ad account, not the one in the brief.
• Confirm the legal line is present and complete, because it is the smallest text in the frame and the first thing a renderer abbreviates.

Do that on render one, not render twenty-four, because a wrong word cloned across a variant set is twenty-four corrections instead of one. Most weak AI ad creative is not weak because the renderer lost a benchmark, which is the quality gap that gets blamed on models every time a new one launches.

Resolution is a budget decision, not a quality dial

Now the part that changes what you spend. The resolution question looks like a quality question and behaves like a line item.

1K is the default, and 4K is a request

Google's docs are explicit: "Gemini 3 image models generate 1K images by default but can also output 2K, 4K, and 512px." Higher resolutions are opt-in, and the mechanism is a parameter: "To generate higher resolution assets, specify the image_size in the response_format." The docs even warn that "You must use an uppercase 'K'", because "Lowercase parameters (e.g., 1k) will be rejected."

Google DeepMind's model page says the same thing from the product side: "Generate crisp visuals at 1k, 2k or 4k resolution." Three rungs, chosen deliberately. Nobody is handed 4K by default, and nobody is billed for it by accident either.

What each rung costs in credits

Inside Novoads, Nano Banana Pro's three resolutions map to two prices:

Resolution	Long edge	Credits per image
1K	1,024 px	0.5
2K	2,048 px	0.5
4K	4,096 px	1.0

The shape of that table is the whole recommendation. 1K and 2K cost exactly the same, which means 1K has no reason to exist in an ad workflow: you would be choosing fewer pixels for the same money. 4K costs double. So the real decision is not "how good do I want this to be," it is "is this asset going somewhere that can use four thousand pixels."

Stay at 2K when:

• the asset is a feed static, a story frame, a marketplace tile or a display banner
• the platform is going to recompress it anyway on upload
• you are generating a variant set and want the credits to go into more angles

Pay for 4K when:

• the asset will be printed, or mocked up at physical size
• somebody will crop hard into a corner of it and still need detail
• it is a single hero image, not one of twenty

Check what your placement actually requires before you pay for pixels a compressor is going to throw away.

A worked example: the twenty-four-variant week

Take the brief from the top of this post. Three headlines, two price badges, four backgrounds is twenty-four combinations once you produce both shapes.

At 2K that set costs 12 credits. At 4K the identical set costs 24 credits. On the entry plan's monthly grant of 50 credits, the same week of creative is either under a quarter of the month or nearly half of it, and the only difference is a setting almost nobody thinks about. Choose 2K and you can run the week twice and still have room to react to what wins.

That arithmetic is the actual argument for caring about the resolution parameter, because creative testing is a volume activity and volume is where per-asset prices compound. A doubled unit price does not cost you one expensive image. It halves how many angles you get to try.

Consistency is what turns one image into an ad set

One good image is a lucky render. An ad set is twenty images that look like they came from the same brand on the same day, and that is a much harder problem than it sounds.

Fourteen inputs, five faces

Google's stated ceiling here is unusually concrete. Nano Banana Pro blends more elements than its predecessors, "using up to 14 images and maintaining the consistency and resemblance of up to 5 people."

For a brand set that reads as two capabilities. Fourteen inputs is enough to hand the model the whole brand kit at once instead of describing it and hoping:

• the packshot, from the angle the pack is actually recognised at
• the palette, as a swatch plate rather than a list of hex codes
• a typography sample, so the headline inherits a real typeface
• the background and prop references the set is supposed to share
• two or three lifestyle plates that set the lighting

Five people is enough for a family of testimonial statics where the same faces recur across the set. The same instinct drives keeping an AI actor consistent across ad variations: variation should happen in the message, not in who or what is on screen.

Edit the region, do not re-roll the image

The other half is repair. Google describes localized editing as being able to "Select, refine and transform any part of an image with improved localized editing."

That is worth more than it sounds to anyone maintaining a set. When one variant comes back with the badge in the wrong corner, a full re-roll gives you a different image, not a fixed one. The lighting moves, the shadow moves, the bottle rotates two degrees, and now that variant no longer belongs to the set. Editing the region keeps the twenty-three good ones good.

Google routed it into Google Ads, and the file still carries a tag

The most advertiser-specific fact in the announcement had nothing to do with the model's specs.

The line worth reading twice

Google wrote that it was "upgrading image generation in Google Ads to Nano Banana Pro to put cutting-edge creative and editing power directly into the hands of advertisers globally." That is a platform putting a specific frontier model behind the generate button inside its own ad manager, worldwide.

Two practical reads follow from that one sentence:

• You may already be using this model without choosing it. If your team generates images inside Google Ads, the renderer behind that button is this one, and the output inherits its strengths and its habits.
• The platform's rules apply to the output, not to your intent. Which is why the rules for AI-generated ad images are worth knowing before the assets exist rather than after a disapproval.

In-platform convenience is not a pipeline

Here is the unhedged version. A generator that lives inside one ad account produces assets for that ad account, and that is the ceiling of its usefulness.

A creative operation needs the file itself, which means four things an in-account generator does not give you:

• the same asset in every placement shape, not only the ones that platform sells
• a library you can search six months later when a winner needs a refresh
• a handoff into the video pass, because the still is usually the first frame of something
• reuse on Meta and TikTok without regenerating the concept somewhere else

Convenience inside one platform is genuinely nice. It is not the same thing as owning your creative assets.

SynthID rides along, and so does disclosure

There is a compliance layer the resolution conversation hides. Google states that all media generated by its tools is "embedded with our imperceptible SynthID digital watermark," and the Gemini API docs repeat it flatly: "All generated images include a SynthID watermark." Google adds that "we will maintain a visible watermark (the Gemini sparkle) on images generated by free and Google AI Pro tier users."

Two practical consequences:

• The invisible tag travels with the file regardless of which surface you generated it on, so plan on the provenance signal existing rather than on stripping it.
• A visible mark is tier-dependent, so check what your own output actually carries before it goes near an ad account, instead of assuming from someone else's screenshot.

Either way, platform-side disclosure is a separate obligation from the watermark, and a workflow that already handles labeling AI-generated ads is doing work the model call does not.

How Novoads solves the resolution-and-consistency problem

The reason to write a post like this at all is that a marketer should not have to hold a provider's parameter table in their head to know what an image costs.

The model is already in the picker

Nano Banana Pro is one of the image models in Novoads, which means the resolution question arrives as a dropdown with a price next to it rather than as a parameter in someone's request body. What that looks like in practice:

• 2K by default, at 0.5 credits per image, with 4K available at 1.0 credits when an asset genuinely needs it
• Eleven aspect-ratio options, including 4:5, 9:16, 1:1, 16:9 and 21:9, so a placement change is a dropdown rather than a re-render at the wrong size
• The rest of the image catalog on the same balance, including GPT Image 2, Seedream 5 Pro, Seedream 5 Lite and Reve 2.1, when a job wants a different engine

One number per image, one balance, and the resolution decision made explicit instead of buried in a request body.

Where the still goes next

A static is where an angle is born, and it is rarely where the campaign ends. In the same workspace the approved image becomes the opening frame of a vertical clip: write or auto-generate a script, pick an AI actor whose age, gender and accent match the audience, and get synthetic voice, lip-sync and captions on top. Our guide to making product videos with AI walks that handoff end to end.

Trying it costs $1 for three days of access, which then becomes the $49-a-month Inicial plan. That first charge grants 10 credits, enough for about one video. Cancel anytime.

Legibility is the feature. Resolution is the invoice.

Nano Banana Pro earned its reputation on a number that has since spread across its own family, which is the normal life cycle of a spec. The thing that did not spread is the reason it was interesting for advertising in the first place: it renders the words inside the picture, in the language the market reads, across a set that stays recognisably the same brand.

So judge an image model the way you judge a printer. Not by the largest size it can output, but by whether the text comes out right the first time, and by what a hundred of them cost. One of those questions decides whether the ad ships. The other decides how many you get to try.