Nano Banana Pro prompts: Google's own prompt guide, boiled down, plus the one prompt we actually ran

The best Nano Banana Pro prompts, per Google's own guides read on November 8, 2026, are full sentences that name the subject, framing, action, place and style, put any words to render in quotes, and give every reference image a job. Nano Banana Pro is Google's Gemini 3 Pro Image model. Below is what Google's four official guides tell you to write, Google's own templates word for word, how to run Pro in the Gemini app, Google Flow and Higgsfield, and the one character-consistency prompt we have run ourselves, on Nano Banana 2 rather than Pro. We don't list "prompts that work" we haven't run.

What should a Nano Banana Pro prompt include?

Google's Nano Banana Pro tips post (Google blog, November 20, 2025) lists the elements, each with an example. Here they are with Google's own examples:

Google's Cloud team turns that into one formula in its Nano Banana prompting guide (March 5, 2026, written for Nano Banana Pro and Nano Banana 2): [Subject] + [Action] + [Location/context] + [Composition] + [Style]. Its advice on how to write it: "A simple list of keywords won't cut it; you need to describe the scene narratively." The same guide says to "start a prompt with a strong verb" that names the operation you want.

Every other Nano Banana Pro question, from the free allowance to errors, is answered on our Nano Banana Pro hub.

What are Google's best practices for Nano Banana prompts?

Google's developer docs for the Gemini image models list six, and its other guides repeat them:

1. Be hyper-specific. "The more detail you provide, the more control you have." Google's example swaps "fantasy armor" for a description of the armor's material, pattern and shape.
2. Give the context and intent. "Explain the purpose of the image." Google's example: "Create a logo for a high-end, minimalist skincare brand" does better than "Create a logo."
3. Iterate. "Don't expect a perfect image on the first try." Follow up in the same conversation with small changes.
4. Break complex scenes into steps. "For complex scenes with many elements, break your prompt into steps": first the background, then the foreground object, then the detail.
5. Use positive wording. Google calls it a semantic negative prompt: "Instead of saying "no cars," describe the intended scene positively", such as an empty, deserted street.
6. Use camera language. Terms like wide-angle shot, macro shot and low-angle perspective control the composition.

Those are from the AI Studio image-generation docs, read November 8, 2026. The same page gives Google's fill-in templates; here are the two most people need, copied exactly:

These are Google's templates, not prompts we've run. Fill the brackets with your own subject.

How do you get readable text in a Nano Banana Pro image?

Put the exact words in quotes, describe the font, and say where it goes. Google's DeepMind guide puts it in one line: "Enclose your desired words in quotes", then describe the typography, "like a "bold sans-serif font" or "neon cursive signage"". Google's AI Studio docs recommend Gemini 3 Pro Image "for professional asset production" with text, and give this template:

Three more text rules from Google:

• Write the words first. Google's Cloud guide says the image models "work best if you first converse with it to generate the text concepts, and then ask for an image with that text." Agree the wording in chat, then ask for the image.
• Translate in the prompt. Write the prompt in one language and name the target language for the text; if the phrasing matters, give the exact translation in quotes. Google's own example: "translate all the English text on the three yellow and blue cans into Korean, while keeping everything else the same".
• Expect small text to slip. Google's limitations note says "Rendering small text, fine details, and producing accurate spellings may not work perfectly." Zoom in before you publish.

Higgsfield's help article gives the same fix for misspelled text on its version of the model: exact text in quotes, plus the font style, size and placement.

How do you edit an image with Nano Banana Pro?

Name the change, then name what stays. Google's Cloud guide: "Be explicit about what to keep exactly the same." Its blog post shows how short an edit prompt can be: "Turn this scene into nighttime" and "Focus on the flowers". For targeted edits, Google's docs give two templates:

Google calls the second one semantic masking: you mark the area with words instead of a brush. Its limitations note warns that "blending or lighting changes can sometimes produce unnatural artifacts", so check edges and shadows after a lighting edit.

How do you use several reference images?

Give each image a role, and know the limit. Google's API docs say gemini-3-pro-image "supports 5 images with high fidelity, and up to 14 images in total". Google's blog adds that the 14 "varies by surface", and Higgsfield lists up to 14 references for Nano Banana Pro.

• Say what each image is for. Google's example: "Use Image A for the character's pose, Image B for the art style, and Image C for the background environment."
• Name your characters. DeepMind's guide says to upload clear references and "assign a distinct name to each character or object in your prompt."
• Describe what must survive. For a face or a logo, Google's docs say to "describe them in great detail along with your edit request."

What's the one prompt we ran ourselves?

A character-consistency prompt, on Nano Banana 2, not Pro. On October 27, 2026 we made one fictional woman with Higgsfield's Soul 2.0, then gave that single reference photo to Nano Banana 2 and Seedream 5.0 Pro with eight scene prompts. Every scene prompt repeated her details in this exact clause, which is what our same-face test published:

It is Google's own advice in practice: describe what must stay, in detail. The result on Nano Banana 2, at 2 credits a 2K image: 4 of 8 shots kept all five details we scored, and all 16 shots across both models read as the same woman. The misses were small: freckles faded in bright window light, side views added earrings nobody asked for, and Nano Banana 2 turned her the wrong way when we asked for her left profile. Seedream 5.0 Pro kept 6 of 8.

What we'd change, from our own misses: a reference that shows the angles you'll use (every jewellery miss came in a side view), and a name for the character, which Google recommends and we didn't use. We haven't rerun the clause on Nano Banana Pro, so we can't say whether Pro holds the details better. For a trained face instead of a reference photo, see Higgsfield Soul vs Nano Banana.

How do you use Nano Banana Pro in the Gemini app?

As a redo, on a paid plan. Google's Gemini help says Nano Banana Pro is "Available with a Google AI Plan" when the Gemini model is set to Pro, and that it regenerates an image the app first made with Nano Banana 2:

1. Be on Google AI Plus, Pro or Ultra.
2. Set the Gemini model to Pro, then type your prompt. Google's help suggests starting with a verb such as draw, generate or create, naming the style and describing the subject, action and setting.
3. Under the image, click More, then Redo with Pro.

Google's help adds that image editing isn't available to users under 18, and that downloads are 2K with a Google AI plan. If Pro doesn't appear or stops, our Nano Banana Pro not working guide walks through each cause.

How do you use Nano Banana Pro in Google Flow?

Pick it in the prompt box. Google's Flow help, read November 8: open a project, click the model name in the prompt box, choose Image, set the aspect ratio, number of outputs and model, write the prompt, add any ingredients, and click Generate. Ingredients are your reference images: drag media into the prompt, or use More then Add to Prompt on an image; typing @ and a character name pulls in a saved character. Edits don't overwrite the original; earlier versions and their prompts stay in the History panel.

Flow is also the free route. Without a Google AI plan it allows up to 20 Nano Banana Pro images a day, without spending Flow credits:

One trap: Flow's pages disagree on the default model, and its Agent mode picks a model for you. Turn the Agent off and choose Nano Banana Pro yourself to be sure. Limits, watermarks and age rules are in is Nano Banana Pro free. Google's Cloud guide also suggests making keyframes with Nano Banana and animating between them with Veo; our image-to-video guide compares the video side, and the best AI video generator guide ranks Veo against the rest.

How do you use Nano Banana Pro on Higgsfield?

Higgsfield's help article (last modified September 9, 2026) gives the steps:

Higgsfield's pricing page listed Nano Banana Pro at 2 credits an image and 4 at 4K on November 3. At the monthly Plus rate ($59 for 1,200 credits) that's about $0.10 an image, our arithmetic. Google's own content rules still apply there: Higgsfield's help says the model provider runs its own safety check on top of Higgsfield's.

Is there a Nano Banana prompt generator?

Google's answer is Gemini itself. Its Cloud guide says "Gemini 3 can help you create prompts and with creative direction", and its text-first advice works the same way: ask Gemini for the wording, then for the image. Prompt-generator sites with "Nano Banana" in the name aren't Google, and we haven't tested any. If a model keeps failing your prompt, Nano Banana Pro alternatives lists what to try instead, priced per image.

What does Google say Nano Banana Pro still gets wrong?

Five things, from Google's own limitations note on the Pro tips post:

• Small text, fine details and spelling "may not work perfectly".
• Data in diagrams and infographics needs checking for facts.
• Translated text "may make grammar mistakes or miss specific cultural nuances".
• Blending and lighting edits can produce artifacts.
• "character consistency across edits may vary". Our test saw the same on Nano Banana 2: the face held, small details drifted.

Google's API docs add that the model "won't always follow the exact number of image outputs" you ask for, and list the languages it works best in, from English to Hindi, Japanese and Vietnamese. Every image carries Google's invisible SynthID watermark.

What we couldn't verify

• Our clause on Nano Banana Pro. We ran it on Nano Banana 2 and Seedream 5.0 Pro only.
• Any prompt as "proven" on Pro. Every template above is Google's; we haven't run them.
• The full scene prompts from our test. Our test page published the shared clause and the scene list, not each full prompt, so we quote only the clause.
• The 14-reference limit in each app. Google says the limit varies by surface; Flow's help doesn't give a number.
• Third-party prompt generators. We haven't tested any, and they aren't Google.
• The Gemini app's daily image count. Google doesn't publish one for personal plans.

The short version: Google's own guides are the best Nano Banana Pro prompt guide there is; describe the scene, quote the text, protect what must not change, give references a role, and check the result at full size.