Prompt notebook
Nano Banana prompts for creating and editing images
A useful prompt says what to change, what to preserve, and what a successful result looks like. These examples give you a specific starting point instead of a string of vague quality words.
A prompt is a brief, not a bag of keywords
Use this structure: task + subject + desired change + details to preserve + output context. A short, coherent request is a better starting point than stacking conflicting styles, camera names, and quality adjectives.
Google recommends giving context and intent, then iterating. For images containing text, its developer guide suggests working out the wording first before asking for the image that contains it.
The prompts below are original editorial starting examples. They have not been run as a controlled Nano Banana test, and no output is guaranteed. The separately attributed official video shows Google's published Pro demonstration.

New to the app? Follow how to use Nano Banana first. Brazilian Portuguese examples are available in the Portuguese prompt notebook.
Edit a photo without rewriting the whole scene
Start with a reference image that makes the intended subject easy to identify. Do not combine a change of face, pose, outfit, location, and lighting unless you are comfortable with the result becoming a different scene.
A natural portrait background
Start with: A clear portrait, with the person's permission.
Change only the background of this portrait to a softly lit, neutral studio wall. Preserve the person's identity, facial expression, skin texture, hairstyle, clothing, and pose. Match the original direction of light. Keep the result photographic rather than heavily retouched.
Follow-up correction: The skin looks too smooth. Restore natural texture from the reference and keep the rest of the edit unchanged.
Relight a room
Start with: A room photo with the furniture and windows visible.
Make this room feel like it was photographed in soft late-afternoon daylight. Keep the architecture, window positions, furniture, decor, and camera angle unchanged. Let the light come from the existing window, with consistent shadows. Do not add new objects.
Follow-up correction: The window frame moved. Return to the reference geometry and change only the lighting.
Change one object's color
Start with: A photo with one clearly visible object, such as a chair.
Change the fabric of the chair to a muted forest green. Preserve its shape, seams, texture, position, and shadows. Leave every other object and the background unchanged. The result should look like the same photo with a different fabric color.
Follow-up correction: The chair became glossy. Restore the original matte fabric texture while keeping the new green color.
These are selective-editing instructions, not guarantees of exact masking. If a label, face, or measurement must be preserved exactly, compare at full resolution and use a conventional editor for the final correction.
Create a new image with a clear purpose
A readable event poster
Start with: A new image request with the final wording already decided.
Design a simple portrait-format poster for a community photography walk. Use only this exact text: "PHOTO WALK" and "SATURDAY 10 AM". Use a clean typographic hierarchy, dark lettering on a light background, and a small photographic detail of a camera strap. Leave generous space around the text. Add no other words, logos, or dates.
Follow-up correction: The small line is misspelled. Replace it with exactly 'SATURDAY 10 AM' and keep the layout unchanged.
A restrained logo concept
Start with: A new image request for a fictional project, not an existing brand.
Create three simple black-and-white logo concept sketches for a fictional neighborhood reading group. Explore the idea of an open book and a shared table. Use plain shapes that remain recognizable at small sizes. Place the concepts side by side on white, without words, mockups, shadows, or gradients.
Follow-up correction: Simplify the middle concept to fewer shapes and show it alone, still in black and white.
A consistent character in a new scene
Start with: A reference image of a fictional character you own or are allowed to use.
Place the reference character in a quiet train station, waiting beside a window with a small suitcase. Preserve the character's face, hairstyle, outfit, proportions, and illustration style. Change only the scene and pose needed for the new setting. Keep the composition clear and include no written signs.
Follow-up correction: The jacket color changed. Match the reference outfit and leave the station scene unchanged.
A generated logo concept is not a trademark clearance or necessarily an editable vector file. A generated poster still needs a spelling check. A character reference still needs a consistency check. Name the intended use so you can judge the result against a real goal.
Write the next prompt from what went wrong
- Identify the mismatch: “The title contains an extra word,” not “This is bad.”
- Specify the correction: quote the exact text, color, object, or region to change.
- Protect the successful parts: ask to keep the composition, lighting, or identity unchanged.
- Reset when necessary: return to the original input if successive edits drift too far.
Save the prompt, reference image, model name, and output together when comparing attempts. Without those details, a screenshot of a good result tells you very little about how to reproduce it.
If a usage limit interrupts your work, check the free and paid access comparison. For model-name differences, return to the Nano Banana field guide.
FAQ
Common questions
Are these prompts guaranteed to reproduce the official video examples?
No. Original prompts on this page are editorial starting examples, not a benchmark or a promise of identical output. The separately attributed Google DeepMind video is an official published demonstration, not a run of these prompts. Model versions, input images, and generation settings can change outcomes.
Can I use the same prompt with another model?
You can try the text elsewhere, but supported inputs, reference-image handling, output sizes, and instruction-following differ. Recheck the output and adapt the wording; do not treat results from another model as Nano Banana results.
Should I write prompts in English?
Use the language that lets you describe your goal precisely. For text inside the image, state the exact words and the language, and inspect spelling and accents. The Brazilian Portuguese version includes locally written examples.
How do I keep a person or object consistent?
Use an appropriate reference image when supported, name the identity and features to preserve, and request one targeted change. Review each output against the reference. These steps improve the instruction but do not guarantee perfect consistency.
Sources & editorial notes
Access information and model names were checked against the official sources below. Interfaces, limits, and offers can change.
Independent editorial writing. Original prompt examples are not a model benchmark; external demonstrations carry their own attribution.
Keep exploring