How to Prompt Nano Banana Pro

TL;DR: Nano Banana Pro is Google's gemini-3-pro-image model. Prompt it like a creative brief rather than a tag list: name the shot, the subject, the lighting, and the exact text to render, then set aspect_ratio and image_size in response_format. It reasons before it draws and it can ground on Google Search.

Every fact below is checked against Google's own documentation on August 26, 2026, with the URL given. Image model naming has changed twice in nine months, so date-stamped sourcing is the only kind worth having.

How do you prompt Nano Banana Pro?

Write a specific brief and let the parameters do the framing. Google's own best-practice list for these models is short, and it is not the tag-soup style that Midjourney trained a generation of prompters into.

Google's guidance, quoted from the Gemini API image generation docs:

• Be hyper-specific. Their example: instead of "fantasy armor," write "ornate elven plate armor, etched with silver leaf patterns, with a high collar and pauldrons shaped like falcon wings."
• Provide context and intent. "Create a logo for a high-end, minimalist skincare brand" beats "Create a logo," because the model uses purpose as a constraint.
• Iterate and refine conversationally rather than rewriting from scratch.
• Use step-by-step instructions for complex scenes: background first, then the foreground object, then the focal element.
• Use semantic negative prompts. Instead of "no cars," describe the intended state: "an empty, deserted street with no signs of traffic." There is no --no parameter here.
• Control the camera with photographic language: wide-angle shot, macro shot, low-angle perspective.

The practical shape that follows from all six is a sentence-level brief with five slots.

A [shot type] of [subject, described with specific physical detail] in
[setting]. [Lighting description]. Shot from [camera angle] with
[lens type]. [Any text that must appear, in quotes, with font style
described]. [Aspect ratio].

Fill every slot. Vague slots are where the model improvises, and its improvisations are competent and generic.

Which model is Nano Banana Pro, and what is its model ID?

Nano Banana Pro is gemini-3-pro-image. That is the only ID that reaches it.

Google's docs list four models under the Nano Banana name, and the marketing names and API IDs do not line up in any intuitive way:

Three traps live in that table.

The first trap is gemini-3-pro-preview. That was a text model, not an image model, and Google's deprecations page lists it as shut down on March 9, 2026, with gemini-3.1-pro-preview as its replacement. gemini-3-pro-image is a separate, live, stable model with no announced shutdown date. One character of difference between a dead endpoint and a working one.

The second trap is the version number. Pro is gemini-3-pro-image, not gemini-3.1-pro-image. The flash models moved to 3.1; Pro did not. Worth noting: on the current image generation page, the aspect-ratio table for the Pro model is headed "3.1 Pro Image" while the model list and the selection guidance on the same page both call it Gemini 3 Pro Image. No gemini-3.1-pro-image ID appears anywhere in that page or in the deprecations table, so read the heading as a docs slip and use the ID.

The third trap is assuming Pro dominates on every axis. It does not. The flash model takes more object references (10 versus 6), supports style reference images where Pro's row reads N/A, adds Google Image Search grounding, and accepts video as an input. Google positions Pro for "professional asset production and complex instructions" and positions gemini-3.1-flash-image as the "go-to" generalist.

What resolutions and aspect ratios does Nano Banana Pro support?

1K, 2K and 4K, across ten published aspect ratios. Google's tables give exact pixel dimensions for each combination, which is more than most image models publish.

Read that table carefully, because "4K" is not one number. At 1:1 you get 4096x4096. At 21:9 the long edge runs to 6336px. Google's model card describes the 4K tier as roughly 16 megapixels.

Two details that will cost you a request if you miss them:

The K must be uppercase. Google states plainly that lowercase parameters such as 1k will be rejected. Use 1K, 2K, 4K.

2K is free relative to 1K on Pro. The published token table charges 1,120 output tokens for both 1K and 2K on gemini-3-pro-image, and 2,000 for 4K. On gemini-3.1-flash-image the same jump costs 1,680. So on Pro there is no billing reason to request 1K unless you specifically want the smaller file.

Defaults matter too. The docs say the model matches output size to your input image where there is one, and otherwise produces a 1:1 square at 1K. If you never set response_format, you have chosen a square.

{
  "model": "gemini-3-pro-image",
  "input": "A high-resolution studio product photograph of a matte black ceramic mug on polished concrete. Three-point softbox lighting for soft diffused highlights and no harsh shadows. Slightly elevated 45-degree camera angle. Ultra-realistic, sharp focus on the rising steam.",
  "response_format": {
    "type": "image",
    "mime_type": "image/jpeg",
    "aspect_ratio": "3:2",
    "image_size": "4K"
  }
}

How do you get Nano Banana Pro to render text correctly?

Quote the exact string, describe the font in words, and describe the layout. Google calls advanced text rendering one of the headline capabilities of the Gemini 3 image models, specifically for "infographics, menus, diagrams, and marketing assets," and its prompting guide points at Gemini 3 Pro Image for professional asset production.

Google's own template for text work:

Create a [image type] for [brand/concept] with the text "[text to render]"
in a [font style]. The design should be [style description], with a
[color scheme].

Their worked example, which is a good calibration point for how much detail is expected:

Create a modern, minimalist logo for a coffee shop called 'The Daily Grind'.
The text should be in a clean, bold, sans-serif font. The color scheme is
black and white. Put the logo in a circle. Use a coffee bean in a clever way.

There is one non-obvious ordering tip buried in the limitations section, and it is the single most useful line on the page: when generating text for an image, Gemini works best if you first generate the text and then ask for an image with the text. Two turns, not one. Ask the model to write the headline, approve it, then ask for the poster.

Google's own limitations page for the model is candid that text fidelity is not solved. It says the model "can still struggle with small faces, accurate spelling, and fine details," and that for translation work it "may struggle with grammar, spelling, cultural nuances, or idiomatic phrases." If you are shipping a client asset, proofread the pixels.

The vocabulary for describing type and style is the same craft you would use anywhere else, and our Midjourney style modifiers reference transfers cleanly here even though the syntax does not. Nano Banana takes none of Midjourney's double-dash parameters. Everything is prose or a JSON field.

How do reference images work in Nano Banana Pro?

You can mix up to 14 reference images in one request. What matters is the breakdown, because the model treats different reference types differently.

For gemini-3-pro-image, Google documents:

• Up to 6 images of objects to include in the final image with high fidelity
• Up to 5 images of characters to maintain character consistency
• Style references: N/A in Google's table for Pro

That last row is the one people get wrong. Style reference images are documented for gemini-3.1-flash-image only, with up to 3 supported. If your workflow depends on feeding in a mood board and asking for its style, the flash model is the documented path, not Pro.

Google's editing templates are worth copying verbatim because their phrasing is deliberate. For a targeted edit:

Using the provided image, change only the [specific element] to [new
element/description]. Keep everything else in the image exactly the same,
preserving the original style, lighting, and composition.

For protecting a face or a logo through an edit:

Using the provided images, place [element from image 2] onto [element from
image 1]. Ensure that the features of [element from image 1] remain
completely unchanged. The added element should [description of how the
element should integrate].

For a character turnaround, Google recommends iterating and feeding previous outputs back in:

A studio portrait of [person] against [background], [looking forward / in
profile looking right / three-quarter view].

The Gemini 3 Pro Image model card puts hard numbers on the inputs: 14 images per prompt, 7MB per file for inline data, 30MB per file from Cloud Storage, and PNG, JPEG, WEBP, HEIC and HEIF accepted. Each input image consumes 560 input tokens.

If you are working the other direction and want to turn an image you like into a reusable brief, the method in how to reverse-engineer any AI image into a prompt applies directly, and it gets better here because Pro accepts the source image as a reference at the same time.

What does Search grounding change about your prompt?

It lets you prompt for things that happened, rather than things that exist in the weights. Google describes Grounding with Google Search on these models as using Search "as a tool to verify facts and generate imagery based on real-time data (e.g., current weather maps, stock charts, recent events)."

Turn it on as a tool, then write a prompt that assumes fresh facts:

{
  "model": "gemini-3-pro-image",
  "input": "Visualize the current five-day weather forecast for San Francisco as a clean, modern chart. Add a small illustration of what to wear each day.",
  "tools": [{ "type": "google_search" }],
  "response_format": {
    "type": "image",
    "mime_type": "image/jpeg",
    "aspect_ratio": "16:9"
  }
}

Three things to know before you build on it.

It is not free. Google's pricing page gives 5,000 free search requests per month shared across all Gemini 3.x models, then $14 per 1,000 requests.

Image Search grounding is a different feature, and Pro does not have it. Google's note is explicit that grounding with Google Image Search, which pulls web images in as visual context, is available for gemini-3.1-flash-image only. Pro gets Web Search.

Accuracy is still your problem. Google's own limitations text warns that when the model generates infographics or represents complex data it "may misinterpret information or produce factually incorrect results," and tells you to verify data-driven outputs. Grounding narrows hallucination, it does not remove it. The same trade-off we cover in RAG vs fine-tuning vs prompting applies to pixels.

What is "thinking" doing, and should you prompt around it?

Nano Banana Pro is a thinking model. Google states that the reasoning process is enabled by default and cannot be disabled in the API, and that the model generates up to two interim images to test composition and logic before rendering the final one. Those thought images are visible in the response and, per the docs, are not charged.

The practical consequence is that you should write prompts with a logical order the model can reason over, the same way chain-of-thought prompting works for text. Google's step-by-step guidance says as much: build the background, then the mid-ground object, then the focal element, in that order in your prompt.

You can read the reasoning back:

for step in interaction.steps:
    if step.type == "thought":
        for content_block in step.summary:
            if content_block.type == "text":
                print(content_block.text)

When a composition comes out wrong, reading the thought text tells you whether the model misread your brief or read it correctly and made a bad call. Those need different fixes. The same debugging instinct we describe in chain-of-thought prompting applies here.

What is Nano Banana Pro bad at?

We do not build image models, so there is nothing for us to defend here. Google's documented gaps, as of August 26, 2026:

• No structured output and no function calling. Both are listed as not supported on the Gemini 3 Pro Image model card. If you need a machine-readable payload alongside the image, that is a separate call to a text model.
• Video input is not supported for Pro. Video-to-image generation carries an explicit note in the docs restricting it to gemini-3.1-flash-image. The Gemini Enterprise model card's own modality row also lists video as not supported for Pro, although a capabilities row on the same page confusingly reads "Image generation from video input: Supported." Treat it as flash-only until Google's wording settles.
• Output count is not reliable. Google says the model "won't always follow the exact number of image outputs that the user explicitly asks for."
• Complex edits degrade. The limitations page names masked editing, major lighting changes such as day to night, and blending multiple images as sources of "unnatural results, visual artifacts, or disjointed scenes."
• Language coverage is uneven. Google lists 15 recommended languages for best performance, starting with English.
• Rate limits are not published per model. The rate-limits page defers to your AI Studio dashboard rather than listing RPM or RPD for gemini-3-pro-image.
• No free API tier. The pricing table reads "Not available" for the free tier on this model.

One more mismatch worth flagging, because it will bite anyone moving between Google's two clouds. The Gemini API page publishes ten aspect ratios for Pro. The Gemini Enterprise Agent Platform model card for the same model ID lists fifteen, adding 1:4, 4:1, 1:8, 8:1 and 9:21, and marks the 4K resolution tier as "(Preview)" where the Gemini API page does not. Test the ratio you need on the surface you are actually calling.

What does it cost, and where can you use it?

Google's pricing page for gemini-3-pro-image, read August 26, 2026:

On availability, the most recent Google statement we could find is the Nano Banana 2 announcement of February 26, 2026, which says Nano Banana 2 replaces Nano Banana Pro across the Fast, Thinking and Pro models in the Gemini app, and that Google AI Pro and Ultra subscribers keep Pro access by regenerating an image from the three-dot menu. That is six months old at the time of writing and consumer surfaces move fast, so verify in the app rather than trusting a blog, including this one.

For developers, the model card gives a 65,536-token context window, 32,768 max output tokens, GA since May 28, 2026, and a retirement date no earlier than May 28, 2027.

A prompt skeleton worth saving

Everything above collapses into one reusable template. Paste it, fill the brackets, delete what you do not need.

ROLE
You are producing a [asset type] for [audience/purpose].

SUBJECT
[Subject, described with specific physical detail: material, colour,
texture, age, wear, posture.]

SCENE
[Setting. Then the mid-ground. Then the background. In that order.]

LIGHT AND CAMERA
[Lighting setup and direction]. Shot from [angle] with [lens type].
[Depth of field note.]

TEXT
Render exactly: "[string]" in [font style description], positioned
[where], at [relative size].

STATE, NOT NEGATION
The scene is [positive description of what should be absent, e.g.
"an empty street with no signs of traffic"].

OUTPUT
[Aspect ratio] at [1K / 2K / 4K].

That skeleton is also a decent argument for storing prompts somewhere other than a chat scrollback. Ours is the thing we actually build: Prompt Architects turns a one-line idea into a structured brief with role, task, format, constraints and tone, and keeps the result in a library with variables so the same skeleton serves twenty products. We generate the prompt. Google generates the image. Those are different jobs and we are only good at one of them.

Sources

All accessed August 26, 2026:

• Gemini API: image generation and prompting guide
• Gemini API pricing
• Gemini API deprecations
• Gemini 3 Pro Image model card
• Google DeepMind: Nano Banana Pro model page and limitations
• Google blog: Nano Banana 2 announcement, February 26, 2026