Nano Banana Pro prompt guide: structure, text, references and edits

A good Nano Banana Pro prompt describes the image in full sentences: who or what is in it, how the shot is framed, what is happening, where it takes place and what style it has. Add the exact words you want rendered in double quotes, say what each reference image is for, and for an edit, say both what should change and what must stay the same. Size and aspect ratio are set as request parameters, not in the prompt. The guidance below comes from Google's own Nano Banana Pro documentation and prompting tips, checked on September 29, 2026, plus the rules of the SeedRouter API.

How do you write a Nano Banana Pro prompt?

Google's prompting tips for Nano Banana Pro list five things to include:

1. Subject. "Who or what is in the image? Be specific."
2. Composition. "How is the shot framed?" For example a close-up, a wide shot or a low angle.
3. Action. "What is happening?"
4. Location. "Where does the scene take place?"
5. Style. "What is the overall aesthetic?" For example photorealistic, watercolor or 3D animation.

Write these as sentences, not as a list of keywords. Nano Banana Pro is a thinking model: it reasons through the prompt before it draws, so a clear description gives it more to work with than a string of tags. Google's image generation guide makes the same point: "The more specific you are, the more control you have over the results."

Two more habits from Google's best practices help on almost every prompt:

• Give the purpose. "Create a logo for a high-end, minimalist skincare brand" works better than "Create a logo", because the model uses the context.
• Describe what you want, not what you don't. Google calls this a "semantic negative prompt": instead of "no cars", write "an empty, deserted street with no signs of traffic."

Here is a prompt built from the five parts:

A close-up photograph of a hand-thrown ceramic bowl filled with ripe cherries, resting on a weathered wooden table in a sunlit farmhouse kitchen. Morning light comes from a window on the left and casts soft shadows to the right. Shallow depth of field, warm and natural colors, editorial food photography.

How do you get accurate text in Nano Banana Pro images?

Text rendering is one of Nano Banana Pro's strengths. Google describes it as "the best model for creating images with correctly rendered and legible text directly in the image". To get the words right:

• Put the exact text in double quotes and say where it goes. Google's example: "The headline 'URBAN EXPLORER' rendered in bold, white, sans-serif font at the top."
• Describe the font in words, such as bold serif, hand-lettered or condensed sans-serif.
• Settle the wording first. Google's guide notes that the model "works best if you first generate the text and then ask for an image with the text." Decide the exact copy before you prompt, rather than asking the model to write and design at once.
• Check small text. Google lists "rendering small text, fine details, and producing accurate spellings" among the areas that may not work perfectly. Keep critical text large and proofread every render.

A minimalist concert poster in a 2:3 vertical layout. The headline "MIDNIGHT ORCHESTRA" sits at the top in tall, condensed, cream-colored serif letters. Below it, a single violin floats against a deep navy background lit by a soft spotlight. At the bottom, the line "Live at the Harbor Hall, October 18" appears in small, clean sans-serif text. No other text.

Nano Banana Pro can also write in other languages and translate text that already appears in an image. Google notes that multilingual text "may make grammar mistakes or miss specific cultural nuances", so have a native speaker check it.

How do you use reference images in a Nano Banana Pro prompt?

A request can carry up to 14 reference images. Google's documentation breaks the Nano Banana Pro limit down like this: up to 6 images of objects kept with high fidelity, up to 5 images of characters for consistency, and up to 3 images used as style references.

The prompt should say what each image is for. Google's tip: "When using uploaded images, clearly define the role of each. (e.g., 'Use Image A for the character's pose, Image B for the art style, and Image C for the background environment.')" Refer to the images in the order you attach them.

On SeedRouter, references are passed as public image URLs in fileData parts. Base64 and file uploads are not accepted. A request with two references looks like this:

{
  "model": "gemini-3-pro-image",
  "contents": [{
    "role": "user",
    "parts": [
      {"text": "Place the lamp from the first image on the desk in the second image. Keep the lamp's shape, color and materials exactly as they are. Match the warm evening light of the second image."},
      {"fileData": {"mimeType": "image/jpeg", "fileUri": "https://example.com/lamp.jpg"}},
      {"fileData": {"mimeType": "image/png", "fileUri": "https://example.com/desk.png"}}
    ]
  }],
  "generationConfig": {
    "responseModalities": ["IMAGE"],
    "imageConfig": {"aspectRatio": "4:3", "imageSize": "2K"}
  }
}

The request goes to POST /v1/images/generations and returns a task ID; the Nano Banana 2 API guide walks through submitting and polling, and the steps are the same for Nano Banana Pro.

How do you edit an image with a Nano Banana Pro prompt?

Attach the image and describe the change directly. Google's tip for edits: "be direct and specific. (e.g., change the man's tie to green, remove the car in the background)".

For a change to one part of the image, Google's guide calls the technique semantic masking: you define the area in words instead of drawing a mask. Its template is "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."

Using the provided photo of a reading nook, change only the gray armchair to a dark green velvet armchair with brass legs. Keep everything else exactly the same, including the bookshelf, the rug, the window and the lighting.

Other edits that follow the same pattern:

Transform the provided photo of a harbor at dusk into a flat vector illustration with bold outlines and a limited palette of teal, orange and cream. Keep the composition of the boats and buildings.

Turn this rough pencil sketch of a desk lamp into a polished product photo of the finished lamp on a white studio background. Keep the angled arm and round base from the sketch, and make the shade brushed aluminum.

To keep refining the same image over several turns, send the earlier turns back with each new request. The Nano Banana Pro API reference explains how to rebuild the model's turn from a finished task.

How do you get 2K and 4K images from Nano Banana Pro?

Set the size in the request, not in the prompt. generationConfig.imageConfig.imageSize takes 1K (the default), 2K or 4K, with an uppercase K. aspectRatio takes one of ten ratios: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9 and 21:9. Without a ratio, the output follows the input image, or is square when there is none.

"generationConfig": {
  "responseModalities": ["IMAGE"],
  "imageConfig": {"aspectRatio": "16:9", "imageSize": "4K"}
}

At 16:9, the three sizes come out at 1376×768, 2752×1536 and 5504×3072. Use 4K when the image will be printed or cropped heavily; for screens, 2K is usually enough. The output size comes from imageSize: leave it out and you get 1K, whatever the prompt says. How size affects the cost is covered in Nano Banana API pricing.

Can one Nano Banana Pro prompt create several images?

One request returns one image: candidateCount accepts only 1. Asking for "four versions" in the prompt does not add images, and Google's guide notes that the model "won't always follow the exact number of image outputs that the user explicitly asks for."

There are two ways around it:

• Send one request per image. Change one detail at a time between requests, so you can compare versions.
• Ask for several designs inside one image. A prompt such as "four icon variations arranged in a 2×2 grid on a white background" produces a single image that shows all four.

Four app icon variations for a weather app, arranged in a 2x2 grid on a plain white background. Each icon is a rounded square with a soft gradient: a sun, a cloud, a raindrop and a snowflake. Flat, clean vector style with gentle shadows. No text.

How long can a Nano Banana Pro prompt be?

Google lists the model's input limit as 65,536 tokens, which covers the text and the reference images together. Most prompts need only a few sentences. Length helps when it adds specific detail; repeating the same instruction in different words does not.

For a complex scene, Google suggests step-by-step instructions instead of one long sentence: "First, create a background of a serene, misty forest at dawn. Then, in the foreground, add a moss-covered ancient stone altar. Finally, place a single, glowing sword on top of the altar."

More Nano Banana Pro prompt examples

Each of these follows the structure above. Swap the subject and keep the rest.

Product shot

A studio product photograph of a matte white wireless speaker on a pale sage-green surface. Soft, diffused light from a large softbox above, with a subtle shadow beneath the speaker. Slightly elevated three-quarter view that shows the fabric grille. Sharp focus, clean background, square format.

Sticker

A cute sticker of a small orange fox wearing a knitted scarf, curled up asleep on a stack of books. Thick white border, bold clean outlines, simple cel shading and bright colors. The background is plain white.

Infographic

A clean infographic titled "How to Brew Pour-Over Coffee" in five numbered steps, arranged from top to bottom: heat the water, rinse the filter, add ground coffee, pour in slow circles, serve. Each step has a simple line icon and a short label. Flat design, warm brown and cream palette, generous white space.

Isometric scene

A 45-degree isometric miniature of a small seaside town with a lighthouse, a harbor with fishing boats and a row of pastel houses. Soft, realistic materials and gentle afternoon light. Clean composition on a solid pale-blue background.

Storyboard

A four-panel storyboard in black-and-white pencil sketch style for a short scene: an establishing shot of a mountain cabin in snow, a medium shot of a hiker opening the door, a close-up of a kettle on a stove, and a wide shot of the cabin window glowing at night.

What does Nano Banana Pro still get wrong?

Google lists these as areas that still need work:

• small text, fine details and exact spellings;
• factual accuracy in data-driven visuals such as diagrams and infographics, which Google says to always verify;
• grammar and cultural nuance in multilingual text;
• natural results in advanced edits such as blending or lighting changes;
• character consistency across edits, which "may vary".

Plan a review step for anything that will be published.

Frequently asked questions

Can a Nano Banana Pro prompt use Google Search on SeedRouter?

No. Google Search grounding is sent as tools, and SeedRouter does not accept it for this model yet. Give the model the facts in the prompt instead.

Do Nano Banana Pro prompts have to be in English?

No. Google lists English and several other languages, including Japanese, Korean, German, Spanish, French, Portuguese, Russian, Vietnamese and Chinese, as the languages with the best performance.

Can I see how Nano Banana Pro interpreted my prompt?

Yes. Set generationConfig.thinkingConfig.includeThoughts to true, and the finished task returns the model's thought summaries in output.thoughts. Reading them helps you see how the model understood the prompt.

Which Nano Banana model should I use for these prompts?

The same prompts work on Nano Banana 2 and Nano Banana Pro. Nano Banana 2 vs Pro vs 2 Lite explains when Pro is worth it.

Try a prompt on Nano Banana Pro

Open the Nano Banana Pro playground to test a prompt before you write code. Every request is billed like an API call and a failed request is not charged. The API reference lists every parameter.