Nano Banana 2.1 Prompt Guide

The prompt formula

Google's own Gemini guidance starts from a simple structure: action, subject, setting, atmosphere. For images you will actually use, fill in a few more slots. You rarely need all of them — you always need the subject.

Words like "4k", "masterpiece" or "high quality" don't change the output size — resolution is a setting, not a prompt word. Spend those words on the subject and the light instead.

Getting text to render exactly

Text rendering and infographic layouts are among the improvements Google lists for 2.1. You still get the best results by being explicit.

• Quote every exact string. the sign reads "OPEN LATE", not a sign saying open late.
• Say where it goes and how it looks. "Centered at the top, bold condensed sans-serif, white on red."
• Keep strings short. Headlines, labels and short lines render far more reliably than paragraphs. Set body copy in a design tool.
• Limit the number of text elements. Two or three per image is comfortable; a dozen invites a typo.
• Name the language for non-English text. the menu title reads "早餐" in Traditional Chinese.
• Fix typos with a targeted edit. Change the headline to "SUMMER SALE". Keep everything else exactly the same.

Editing a photo

Upload the image, then describe the change — and what must stay the same. The pattern is: the change, what to keep, and how it should blend.

• Name what to preserve. Face, expression, pose, product shape, logo, existing text — anything you don't list is fair game.
• Point at the area. "The mug on the left", "only the sky", "the text on the sign".
• One precise change per request. Combine changes only when each one is simple.
• Restyle, don't replace. For a new look on the same photo, say "restyle" and list what stays: "Restyle as a watercolor painting; keep the composition and everyone's position."

Multi-turn edits

Multi-turn character consistency is one of the four areas Google improved in 2.1. To get the most out of it:

• Take small steps. Several small, sequential edits beat one giant rewrite.
• Refer back. "Same character as before, now sitting at a desk."
• Re-anchor when something drifts. "Keep her face, red hair and freckles exactly as in the first image."
• Start fresh when a chain gets long. Begin a new session from the best image so far rather than stacking edit after edit.
• Upscale last. Ask for the final image at 4K as one more follow-up edit. Re-running the original prompt at 4K produces a new composition — the model has no seed.

Reference images

2.1 accepts up to 14 reference images in one request — up to 4 for keeping characters consistent and up to 10 for objects you want reproduced faithfully.

• Refer to images by order. "The woman in image 1, wearing the jacket from image 2, holding the bottle from image 3."
• Use clean references. One subject per image, well lit, plain background.
• Style references are a Pro feature. Google's docs list style-reference slots only for Nano Banana Pro (up to 3). On 2.1, describe the style in words — medium, palette, lighting, era — or use Nano Banana Pro when matching a look matters.

Aspect ratio and resolution

Pick the ratio for where the image will be used, and set it in the controls rather than in the prompt. 2.1 supports 14 ratios: 1:1, 1:4, 1:8, 2:3, 3:2, 3:4, 4:1, 4:3, 4:5, 5:4, 8:1, 9:16, 16:9 and 21:9.

1K is the default and there is no 512px option. Iterate at 1K, then ask for the keeper at 4K as a follow-up edit.

Grounded prompts

2.1 can ground a prompt in Google Search: web results, and with Image Search, web images used as visual context. It helps most with real places, current products and recent events.

• Ask for the real thing explicitly. "The Taipei 101 observatory entrance as it looks today", "this year's packaging".
• Ask for details grounding can pin down. Signage, season, materials, colors.
• Not for real people. Google says Image Search grounding can't use real-world images of people.
• Still check the result. Grounding reduces guessing; it does not turn the output into a photograph of the real place.

In the API, grounding stays off until you add the google_search tool — see the API guide.

Thinking level (API)

2.1 can reason through a prompt before it renders, producing up to two interim "thought images" to test the composition. You aren't charged for those images, but thinking adds latency and thinking tokens.

Mistakes to avoid

• Keyword soup. Write sentences. The model reads descriptions better than tag lists.
• Negative lists. "No cars, no people" works worse than describing the scene you want: "an empty street at dawn". Google's own prompting guidance recommends this positive phrasing.
• Clashing styles. "Photorealistic watercolor anime" — pick one medium.
• Paragraphs of text. Keep strings short and add long copy in a design tool.
• Upscaling by re-prompting. The same prompt at 4K gives a new image. Ask for 4K as a follow-up edit instead.
• Real people through grounding. Image Search grounding doesn't use real-world images of people.

24 Nano Banana 2.1 prompts to copy

Each card lists the aspect ratio and size it was written for. Swap in your own product, place or text.

Product and e-commerce

Posters and text-heavy designs

Photo editing

Characters and mascots

Infographics and diagrams

Banners and social formats

FAQ

What is the best prompt format for Nano Banana 2.1?

Full sentences that describe the scene: subject, action, setting, composition, lighting and style, with any exact text in double quotes. Set the aspect ratio and size in the controls, not in the prompt.

How do I get Nano Banana 2.1 to spell text correctly?

Put the exact words in double quotes, say where and how they should appear, keep each string short and limit the number of text elements. If one word comes out wrong, fix it with a follow-up edit that changes only that text.

Does Nano Banana 2.1 support negative prompts?

There is no separate negative prompt field. Describe what you want instead: "an empty street at dawn" works better than "no cars, no people".

How do I keep the same character across images?

Edit in one multi-turn session, restate the identity details such as face, hair and outfit when something drifts, and pass up to 4 character reference images when you start a new session.

Do Nano Banana 2 prompts work on 2.1?

Yes. 2.1 is an update to Nano Banana 2 and reads the same prompts. Text-heavy prompts and long edit chains are where Google says 2.1 improved, so re-test those first. What changed in 2.1.

Sources

1. Gemini API — Nano Banana image generation guide
2. Gemini API — Gemini Nano Banana 2.1 model page
3. Gemini app — Image generation overview and prompting tips

Prompt advice beyond Google's documentation is our own, based on how the model responds in practice. Google's pages are the authority on what the model supports.