Nano Banana prompt guide: generate and edit images

Nano Banana is the popular name for Google's image generation and editing models in Gemini, and it now covers more than one model variant (such as Nano Banana Pro). What sets it apart from keyword-driven image tools is that it is conversational: you describe what you want in full sentences, upload photos to edit or combine, and refine the result over several turns. This Nano Banana prompt guide collects the guidelines that matter most, based on Google's own prompting tips, with examples you can copy for generation, editing, multi-image composition and text.

The core rule: describe the scene, do not list keywords

Google's guidance for Gemini image generation stresses describing a scene in natural language rather than stacking comma-separated tags. Prompting habits from older image tools often rely on keyword stacks like "8k, masterpiece, trending". Nano Banana is built on a language model, so a descriptive paragraph gives it far more to work with: relationships between objects, the purpose of the image, the mood.

Keyword style (weaker):
cat, window, rain, cozy, cinematic, 8k, bokeh

Descriptive style (stronger):
A ginger cat curled up on a wide wooden windowsill, watching rain run down
the glass. Inside, a warm lamp casts soft amber light across a knitted
blanket. Outside, the street is blurred and blue-grey. Photographed with
a shallow depth of field, cozy and quiet mood. Square format.

Google's tips for the Gemini app suggest thinking in six elements for more control: subject, composition, action, location, style and, for existing images, editing instructions. Be specific about the subject: "a stoic robot barista with glowing blue optics" beats "a robot". Mention the purpose too ("for a café menu header"), because context helps the model make sensible choices about layout and detail.

Photography vocabulary works well: name the shot (close-up, wide shot, low angle, macro), the lens feel (shallow depth of field, wide-angle) and the light (golden hour, soft studio light, neon). Our Nano Banana prompt generator turns those choices into a full descriptive sentence for you.

Editing: "Using the provided image…" and what to keep identical

Editing is where Nano Banana shines. Google's advice is simple: give it a photo and tell it directly what you want to change, for example "change the man's tie to green, remove the car in the background". Plain-language local edits change specific elements without regenerating the entire scene.

The most reliable editing prompts have three parts:

1. Anchor to the image: start with "Using the provided image," so the model knows it is editing, not creating from scratch.
2. One clear change: say exactly what to add, remove or modify, and where.
3. A preservation clause: list what must stay identical, especially faces, pose, framing and lighting.

Using the provided image, replace the plain white wall behind the woman
with a floor-to-ceiling bookshelf full of colorful books. Keep everything
else identical: her face and identity, hairstyle, pose, clothing, the
framing of the shot and the original soft window light. Match the
original photographic style and perspective.

The preservation clause matters because without it the model may "improve" things you did not ask for, like smoothing skin or shifting the camera angle. Naming the elements to protect is more effective than a general "don't change anything else". Our generator has an edit mode that writes exactly this structure: tick the edit option, describe the change, and it adds the "keep identical" clause automatically.

Style transfer

To restyle an image, ask the model to keep the subject and composition while changing only the rendering: "Using the provided photo of my street, redraw it as a watercolor illustration. Keep the layout of buildings, the parked bicycle and the position of the trees exactly as they are."

Multi-image composition

Nano Banana can take several images at once and blend them into a new scene. Google's examples include placing a person into a different setting and applying the colors and textures of one image to an object in another. The key is to say clearly what comes from which image.

Using the two provided images: take the woman from the first image and
place her sitting at the café table in the second image. Keep her face,
hair and outfit exactly as in the first image. Match the warm afternoon
lighting, color palette and camera angle of the second image so she looks
naturally part of the scene. Natural shadows under the table.

• Refer to images by order ("the first image", "the second image") or by content ("the photo of the sofa").
• Decide which image sets the lighting and perspective; mismatched light is the most common giveaway of a composite.
• For product mockups, specify scale: "the mug should be about the size of the laptop's trackpad".

If you want a starting prompt from two reference photos, our two images to prompt tool can draft one. When AI mode is available on this site, it uses a vision model to describe both images; otherwise it analyzes colors, light and framing locally and leaves a [bracketed] subject for you to fill in.

Text in images

Nano Banana is far better at rendering text than older image models, but Google notes it may still occasionally misspell words or struggle with complex typography. Give it the best chance:

• Put the exact text in quotation marks and ask for it to be spelled exactly.
• Describe the typography: "bold sans-serif", "hand-lettered script", "white letters on a dark green banner".
• Say where it goes: "centered at the top", "on the shop sign above the door".
• Keep it short. A headline of a few words is far more reliable than a paragraph.

Create a minimalist poster for a weekend farmers market. A single ripe
tomato on a pale cream background, soft shadow beneath it. At the top,
the text "SATURDAY MARKET" in bold, dark green sans-serif letters, spelled
exactly. At the bottom, smaller text "8 AM – 1 PM". Portrait orientation.

If a word comes out wrong, do not regenerate everything; ask for a targeted fix: "Keep the image identical and correct the word at the top to read SATURDAY."

Iterative editing in conversation

Because Nano Banana is conversational, you can build an image step by step. Google's example: paint an empty room, then add furniture in later turns. Each follow-up builds on the previous result.

Turn 1: Create a photo of an empty, sunlit living room with white walls
and light oak floors, wide-angle, morning light.
Turn 2: Now paint the walls a soft sage green. Keep everything else the same.
Turn 3: Add a low grey linen sofa against the back wall.
Turn 4: Add a stack of three books on a small round coffee table in front
of the sofa.

One change per turn makes it easier to keep what works and undo what does not. Google also notes that character consistency is not always perfect. If a face or key details start to drift after many edits, start a new conversation, upload the best version so far and describe the subject in detail again.

Common mistakes and how to fix them

The negatives point deserves a note: Google recommends what it calls semantic negative prompts, describing the scene you want rather than listing what to exclude. Words like "no cars" still put the idea of cars in the prompt. Our image builder turns the "avoid" field into a short "Avoid: …" sentence, but where you can, phrase the scene positively in the main description as well.

Finally, use these tools responsibly. Edit photos of real people only with their consent, do not create misleading images of real events, and remember that Google marks images created or edited with Gemini with an invisible SynthID watermark. For more image ideas, see the image prompt generator, which supports other models such as Midjourney and Flux, or reverse-engineer a picture you like with image to prompt.