How to write Nano Banana prompts: creating and editing images with Gemini

“Nano Banana” is the nickname for Google’s Gemini image model, the one you reach through the Gemini app, Google AI Studio and the Gemini API. It behaves differently from tag-based generators. It reads your prompt the way a language model reads any request, which means full sentences, clear instructions and follow-up messages all work in your favour. It also means a pile of comma-separated keywords tends to waste what the model is good at.

This guide covers how to write prompts that create images from scratch, how to edit an existing image in conversation, and how to keep the parts you want unchanged.

Write instructions, not tag lists

Gemini understands grammar, so it can tell which adjective belongs to which object and what you want done with them. Use that.

Before:

cozy cafe, rainy window, woman reading, warm light, bokeh, cinematic, 8k, detailed

After:

Create a photo of a woman reading a paperback at a small table beside a rain-streaked cafe window. Warm tungsten light from a pendant lamp above her, cool grey daylight outside. Shot at eye level with a 50mm lens, shallow depth of field so the street outside blurs. Calm, quiet mood.

The second prompt is longer, but nothing in it is filler. Each sentence settles a decision: what happens, where the light comes from, where the camera is, how it should feel. “8k” and “detailed” are gone because they carry no information about this particular picture.

A useful pattern is to open with a verb that states the task (“Create”, “Generate”, “Make an illustration of”), then describe the scene in the order a viewer would notice it: subject, setting, light, camera, style, mood.

Name the medium early

Gemini will happily produce a photo, a watercolour, a flat vector icon or a 3D render. Say which one in the first sentence, because it changes how every later word is interpreted. “A watercolour illustration of a lighthouse” and “a photo of a lighthouse” followed by the same description will diverge from the first brushstroke.

State the format in words

If your interface has an aspect ratio setting, use it. If not, say what you need: “a wide landscape image”, “a tall portrait format for a phone screen”, “a square image with the subject centred”. Mentioning the intended use (“for a blog header”) also helps composition.

Editing images in conversation

The feature that sets Gemini apart is editing by instruction. Upload an image, or keep working on one you just generated, and describe the change. The trick is to describe it the way you would brief a careful retoucher: what to change, what to change it to, and what must stay as it is.

Vague:

make it better and change the background

Precise:

Replace the background with a softly blurred city street at dusk. Keep the person, their pose, clothing and the lighting on their face exactly as they are. Match the warm colour of the new background to the existing rim light on their hair.

One change per message

When you ask for five changes at once, the model has to decide which ones matter most, and some may be applied loosely or not at all. Ask for one change, check it, then ask for the next. A conversation like this tends to stay on track:

1. “Change the jacket from black to olive green. Keep everything else the same.”
2. “Now add a light dusting of snow on the shoulders.”
3. “Make the background slightly darker so the subject stands out more.”

If an edit drifts too far, say so directly: “Go back to the previous version and only change the jacket colour.”

Say what must stay

Models change more than you asked for when you don’t tell them otherwise. A short “keep” sentence is the most reliable protection you have:

• “Keep the face, expression and hairstyle unchanged.”
• “Keep the camera angle and framing identical.”
• “Keep the product label and its text exactly as it is.”

Identity and composition drift most easily over several edits. If a face starts to change after a few rounds, re-upload the original and restart the edit chain from there rather than piling on corrections.

Point at things unambiguously

“Change the chair” is fine when there is one chair. In a busy scene, locate the object: “the wooden chair on the left, next to the window”, “the smaller of the two mugs”, “the sign above the door”. Position, colour and size are the simplest ways to point.

Text in images

Gemini is generally better at rendering short text than older diffusion models, but it still rewards clear instructions. Put the exact words in quotation marks, say where they go and describe the lettering:

A minimalist poster for a jazz night. At the top, the words "LATE SET" in tall condensed white capitals. At the bottom, smaller text reading "Fridays · 9 pm". Deep navy background with a single brass trumpet drawn in thin gold lines.

Keep on-image text short. Headlines and labels work far more reliably than paragraphs. If a word comes out misspelled, ask for a correction in the next message (“Fix the spelling of the top line to read LATE SET”) instead of regenerating everything.

Combining several images

You can give Gemini more than one image and describe how they relate: put this product in that scene, dress this character in that outfit, apply the colour palette of the second image to the first. Name each image by its role so the model doesn’t have to guess:

Use the first image as the subject: the ceramic teapot. Use the second image as the setting: the sunlit kitchen counter. Place the teapot on the counter near the window, matching the direction and warmth of the existing sunlight and adding a soft shadow.

Only use photos you have the right to use, and don’t use editing to create misleading images of real people.

Create vs edit: worked examples

Create, product shot:

Create a clean studio product photo of a matte black water bottle standing on a pale stone surface. Soft light from a large window on the left, a gentle shadow to the right. Neutral beige background, square format, plenty of empty space above the bottle for text.

Edit, same shot:

Change the bottle colour to sage green. Keep the lighting, shadow, surface and camera position exactly the same.

Create, illustration:

Make a flat vector illustration of a fox sleeping curled up under a pine tree in snow. Limited palette of rust orange, off-white and two shades of blue-grey. No outlines, simple geometric shapes, wide landscape format.

Edit, illustration:

Add a small crescent moon in the upper right corner, in the same off-white and flat style. Don't change anything else.

Notice that the edit prompts are short. Once the image exists, the description has done its job, and the edit only needs to name the change and the constraint.

Starting from a reference

If you already have a picture with the look you want, describe it first and adapt the description, rather than writing from memory. Our Nano Banana prompt generator turns an uploaded image into an instruction-style prompt ready for Gemini, and the reverse prompting guide explains how to refine it one element at a time. If you are starting from a few words instead of an image, turning a short idea into a full prompt walks through the decisions to add.

Prompting habits carry over between models, but formatting doesn’t. The same idea written for Midjourney looks very different, as the Midjourney prompt generator shows, and conversational editing in ChatGPT has its own quirks, covered in our DALL-E prompt tips.

Checklist

• Open with a verb and the medium: “Create a photo of…”, “Make a watercolour of…”.
• Write complete sentences that settle subject, setting, light, camera, style and mood.
• State the format if there is no aspect ratio setting.
• Put exact on-image text in quotes and keep it short.
• Edit one thing per message.
• Always say what must stay unchanged.
• Locate objects by position, colour or size.
• If identity or framing drifts, restart from the original image.