Nano Banana Prompts: A Copy-Ready Library by Use Case
October 27, 2026
The Nano Banana prompts that work best read like a brief to a photographer, not a list of tags. Say who or what is in the frame, what it's doing, where, in what style, how it's lit and framed, and, if you're editing, what must stay exactly the same. Google's own guidance for its Gemini image models says the same thing: describe the scene in full sentences, because the model reasons about the whole prompt.
Here's a starter you can paste into Gemini right now:
A photorealistic close-up of a ceramic pour-over coffee set on a pale oak kitchen counter, morning light coming from a window on the left, soft shadows, shallow depth of field, steam rising from the cup. Square image.
Below you'll find:
- How to use Nano Banana in the Gemini app, and which model you're actually getting
- A six-part prompt formula you can reuse for any image
- 16 copy-ready prompts across six use cases: product shots, portraits and photo edits, style transfer, text in images, multi-image composition, and consistent characters
- A fix table for the failures people hit most
How to use Nano Banana in Gemini
"Nano Banana" is Google's name for Gemini's native image generation and editing. It's a family now, not one model. As of October 2026, Google's Nano Banana image generation docs list five: Nano Banana 2.1, Nano Banana 2 Lite, Nano Banana 2, Nano Banana Pro, and the original Nano Banana, which Google now calls legacy. If you want the full name-to-model-ID map, our GPT Image 2 vs Nano Banana comparison untangles it.
Most people use it inside the Gemini app. Google's Gemini image generation page describes the flow:
- Open Gemini and start a new chat.
- Choose Create images from the tools menu.
- Type a prompt, or upload a photo and describe the change you want.
- Refine in the same chat, one change per message.
That same page says the no-cost tier runs Nano Banana 2, Google AI plans unlock Nano Banana 2.1 and higher limits, and Pro, Plus and Ultra subscribers can regenerate an image with Nano Banana Pro from the image's menu. The prompts below work on all of them. Where one model is clearly better for a job, it's called out.
The Nano Banana prompt formula
Google's Nano Banana Pro prompting tips break a prompt into subject, composition, action, location, style and editing instructions. Google DeepMind's Nano Banana prompt guide uses nearly the same list. Put together, a reliable shape looks like this:
[Style] of [subject, with specific details] [doing an action] in [location].
[Composition / camera angle / lens]. [Lighting and mood].
[Any exact text in "quotes", plus how it should look]. [Aspect ratio].
For edits, swap the first line for a scope instruction:
Using the provided image, change only [element] to [new description].
Keep everything else exactly the same: [list what must not move].
What each slot does:
| Slot | What to write | Example phrase |
|---|---|---|
| Style | Medium or look | "photorealistic", "flat vector illustration", "watercolor" |
| Subject | Who or what, with concrete detail | "a matte green glass water bottle with a bamboo cap" |
| Action and location | What's happening, and where | "resting on wet river stones at the edge of a stream" |
| Composition | Framing and camera | "low-angle shot", "macro", "wide shot with the subject on the left third" |
| Lighting | Light source and mood | "overcast daylight", "warm golden-hour backlight" |
| Text and format | Exact words, typography, canvas | "OPEN LATE" in a condensed sans-serif; 9:16 |
You don't need every slot every time. DeepMind's guide notes that short prompts still produce good images. Detail is how you take control when the first result drifts. For the general, model-agnostic version of this formula, see how to write a prompt for an AI image.
Nano Banana prompts by use case
Every prompt below is copy-ready. Swap the bracketed or specific details for your own. Brand names in them are made up.
Product shots
A high-resolution studio product photo of a matte black stainless steel water bottle with a brushed silver cap, standing on a slab of light grey concrete. Soft three-point lighting with a large softbox from the left, gentle reflection on the bottle's shoulder. Camera slightly above eye level. Clean light grey background. Square image.
Why it works: surface, light setup and camera height are all named, so the model doesn't invent a busy scene.
Using the provided photo of my candle, place it on a linen tablecloth next to a small bunch of dried lavender and an open hardback book. Keep the candle's label, jar shape and color exactly the same. Warm late-afternoon window light from the right. 4:5 vertical image for an Instagram post.
Why it works: it starts from your real product and locks the label, the detail shoppers will check against the actual item.
Generate three distinct variations of a hero image for a cold brew coffee can: one on a sunny beach towel, one on a marble cafe counter, one held in a hand against a city street. Same can design in all three. 16:9.
Why it works: DeepMind's guide suggests asking for several variations at once, so you can compare directions before you refine one.
Portraits and photo edits
Use these on photos of yourself, or of people who've agreed to it.
Using the provided photo of me, change only the background to a softly blurred bookshop interior with warm string lights. Keep my face, hair, expression, clothing and the camera angle exactly the same. Match the lighting on my face to the new background.
Why it works: "change only" plus a list of what to keep is Google's own inpainting pattern, so the edit stays inside the background.
Turn this photo of me into a professional headshot: navy blazer over a white shirt, plain light grey studio backdrop, soft even lighting, head and shoulders framing. Keep my facial features, skin tone and hairstyle unchanged.
Why it works: it names the outfit, backdrop and framing, and protects the features that make the photo look like you.
Restore this old family photo: remove the scratches and creases, fix the faded areas, and gently sharpen the faces. Keep it black and white and keep everyone's features exactly as they are.
Why it works: it asks for repair, not reinvention, so the model doesn't "improve" faces into strangers.
For the retro instant-photo trend specifically, we keep a separate set of Gemini Polaroid prompts.
Style transfer
Transform the provided photo of my street into a 1950s travel poster illustration. Preserve the layout of the buildings and the road, but render everything with flat color blocks, a limited palette of teal, cream and coral, and visible paper texture.
Why it works: it separates what to keep (the composition) from what to change (the rendering), which is the structure Google's style-transfer template uses.
Recreate the provided photo of my dog as a children's book watercolor illustration: loose washes, soft pencil outlines, lots of white paper showing through. Keep his pose and the red collar.
Why it works: it describes the style in visual terms rather than naming a living artist, and keeps the two details that identify the dog.
Text in images
Create a poster for a neighborhood bake sale. Headline: "Saturday Bake Sale" in a bold rounded sans-serif. Below it: "10am to 2pm, Elm Street Library". Illustrations of cupcakes and loaves around the edges, pastel yellow background. 2:3 vertical.
Why it works: exact words sit in quotes and the font is described, which is what Google's tips ask for.
Design a cafe menu board on a dark green chalkboard with three sections: "Coffee", "Tea" and "Pastries". Under each heading, list three items in neat white handwritten-style lettering. No prices. Landscape 16:9.
Why it works: short headings and a fixed structure leave less text for the model to misspell.
Translate all the English text in the provided infographic into Spanish. Keep the layout, icons, colors and font style exactly the same.
Why it works: Google's API docs use this exact kind of follow-up edit as an example: change the text, leave every other element alone.
For anything text-heavy, Google's docs suggest settling the wording first, then asking for the image. They also recommend Nano Banana Pro for professional assets with text.
Multi-image composition
Combine the provided images. Image 1 is the living room. Image 2 is the armchair. Image 3 is the floor lamp. Place the armchair by the window in the living room with the lamp beside it. Match the room's daylight and shadows, and keep the chair's fabric pattern exactly as in image 2.
Why it works: each upload gets a role, which Google's tips call out as the key to blending references.
Use image 1 for the pose, image 2 for the outfit and image 3 for the background. Create a full-body photo of the person in image 1 wearing the outfit from image 2, standing in the location from image 3. Natural light, eye-level camera.
Why it works: it splits pose, clothing and setting across references so none of them gets averaged away. Use your own photos, or ones you have permission to use.
Consistent characters
This is Pip, a small orange robot with round headlight eyes, a dented left arm and a blue scarf. Create a character sheet of Pip on a plain white background: front view, side view, back view and three-quarter view, same proportions in every view.
Why it works: a name plus fixed identifying details gives the model something stable to carry into later prompts.
Using the character sheet above, show Pip watering a window box of tomatoes on a rainy balcony. Keep Pip's design exactly the same, including the dented left arm and blue scarf. Soft grey daylight, medium shot.
Why it works: it points back at the earlier image and repeats the identifying details, the two habits DeepMind and Google both recommend for consistency.
Make a 3-panel comic strip of Pip learning to ride a bicycle: panel 1 wobbling, panel 2 falling into a hedge, panel 3 riding away proudly. Clean ink outlines, flat colors, Pip looks identical in all three panels.
Why it works: Google's API docs say sequential art like this builds on character consistency, and that it works best on Nano Banana Pro or Nano Banana 2.
Common Nano Banana prompt failures and fixes
Google's tips list its own weak spots: small text and spelling, fact-heavy diagrams, multilingual grammar, artifacts in complex blends, and character consistency that can vary across edits. Here's what to change when you hit one.
| What went wrong | Likely cause | What to add or change |
|---|---|---|
| The whole image changed when you asked for one edit | No scope on the edit | "Change only the [element]. Keep everything else exactly the same." |
| You wrote "no people" and got people | Negative instructions are weak | Describe it positively: "an empty, quiet street at dawn" |
| A face drifted after a few edits | Long edit chain | Start a new chat with the original photo and list the features to keep |
| Text is misspelled or garbled | Too much text, or vague wording | Fewer words, exact text in quotes, describe the font; finalize the copy first |
| The composite looks pasted on | Lighting doesn't match | "Match the lighting, shadows and color temperature of image 1" |
| Wrong shape for where it's going | No canvas set | Name the aspect ratio: "9:16 vertical" or "16:9 landscape" |
| A chart or diagram has wrong facts | The model filled gaps itself | Give it the data in the prompt and check every number before you use it |
The second row comes straight from Google's best-practice list, which calls this a "semantic negative prompt". It's the most common fix people miss.
Two more habits help. Change one thing per message, so you know which instruction caused which result. And when you have a prompt that works, save the whole thing, not just the final image. You'll want it again.
Nano Banana prompt FAQ
What is the best Nano Banana prompt?
There isn't one. The best prompt is the one that names your subject, setting, style, lighting, framing and format in plain sentences. Start with the formula above, generate, then change one slot at a time.
Is Nano Banana free to use?
Google's Gemini image generation page lists Nano Banana 2 as available at no cost in the Gemini app, with paid Google AI plans adding Nano Banana 2.1, Redo with Pro and higher limits. In the Gemini API, each image model is priced separately; Google lists the rates on its Gemini API pricing page.
What's the difference between Nano Banana, Nano Banana 2 and Nano Banana Pro?
Per Google's API docs, Nano Banana 2.1 is the main fast model, Nano Banana 2 is the previous generation, Nano Banana 2 Lite is the fastest and cheapest, and Nano Banana Pro is the premium choice for complex work. Nano Banana Pro is the one Google recommends for professional assets with text. The original Nano Banana is now legacy.
How many images can I upload in one Nano Banana prompt?
Google's API docs say you can mix up to 14 reference images, with lower caps for keeping people's likeness: Nano Banana 2.1 and Nano Banana 2 support up to four characters. Nano Banana 2 Lite isn't optimized for multiple references, so switch models if your blends fall apart.
Can I edit photos of other people with Nano Banana?
Only with their permission, and never to deceive. Google's Generative AI Prohibited Use Policy bars impersonating a person without disclosure in order to deceive, and using personal data or biometrics without legally required consent.
Can people tell an image was made with Nano Banana?
Yes. Google's docs say every generated image carries an invisible SynthID watermark, and Google's Gemini page says images made in the app also carry a visible watermark. You can upload an image to Gemini and ask whether Google AI made it.
Key takeaways
- Write Nano Banana prompts as full sentences: subject, action, location, style, composition, lighting, format.
- For edits, say what to change and what must stay exactly the same.
- Give every uploaded image a job, and give every recurring character a name.
- Put exact text in quotes, keep it short, and describe the typography.
- Describe what you want instead of what you don't, and change one thing per message.
A prompt that finally works is easy to lose in a chat history, and a full image workflow is harder still to rebuild. Taku is an AI-native desktop workspace where you can mirror an AI app someone already got working, run it, edit it and save it as your own, so the next round starts from a setup that works. Browse what's in the Taku marketplace to see what's there today. Taku is in Beta, and the Mac app is available now.