Nano Banana Prompts: Templates for Gemini Images

Google's Gemini API documentation includes a prompting guide for its Nano Banana image models, with a template for each common job. The same principles work in the Gemini app, Google Flow and other tools that host these models. The core idea is simple: these are language models that make images, so they follow a well-written description better than a pile of tags. This guide adapts Google's templates into prompts you can use straight away, with notes on what to change.

The principle: describe, do not list

"Cat, window, sunset, 8k, masterpiece" gives the model fragments. "A ginger cat asleep on a sunny windowsill at sunset, warm light across its fur, soft focus on the garden outside, shot on a 50mm lens" gives it a scene. Google's documentation makes the same point: the more specific you are, the more control you have. Write one or two full sentences, then add a phrase each for light, camera and mood.

Templates for new images

For anything with words in it, Google recommends Nano Banana Pro (Gemini 3 Pro Image) for professional asset production; for comics and storyboards it suggests Pro or Nano Banana 2. Our guide to the Nano Banana models explains the differences.

Templates for edits

Nano Banana's strength is editing by description. Google lists several patterns:

1. Add or remove an element. "Using the provided image of my cat, add a small knitted wizard hat. Keep the cat, the lighting and the background exactly the same." The model matches the original style, lighting and perspective.
2. Change one part (semantic masking). "Using the provided image of a living room, change only the blue sofa to a vintage brown leather chesterfield. Keep everything else unchanged." Naming the object is your mask.
3. Style transfer. "Transform the provided photograph of a city street at night into the style of a watercolour illustration, keeping the composition."
4. Combine images. Provide a product and a person: "A woman wearing the blue floral summer dress from the first image, standing in a sunlit garden."
5. Preserve a critical detail. Describe the face or logo you must keep in detail alongside the edit, so the model treats it as fixed.
6. Bring a sketch to life. Upload a rough drawing and ask for a finished, realistic image of it.
7. Character turnaround. Generate a character, then ask for it "from the side", "from behind" and so on, one view per message.

The editing workflow in the Gemini app is covered in the Gemini image editing guide.

Text that renders correctly

• Put the exact words in quotation marks and keep them short.
• Describe the type: "bold condensed sans-serif", "elegant serif", "hand-lettered".
• Say where the text sits and that no other text should appear.
• Choose Pro, or 2.1, for text-heavy designs, then proofread every letter at full size.

Common prompt mistakes

• Quality tags instead of description. "8k, ultra detailed, masterpiece" adds nothing; describe the light and lens instead.
• Several edits at once. "Change the sofa, add a lamp and make it night" invites the model to redraw everything. Make three edits.
• Not saying what stays. Without "keep everything else the same", the model may tidy up parts you liked.
• Vague text requests. "A sign with a nice message" produces filler; give the exact words.
• Real people and brands. Prompts involving identifiable people or trademarks may be refused; describe a fictional person or an unbranded product.

Grounded images

Most Nano Banana models can use Google Search to base an image on current information, such as a weather forecast, a stock chart or a recent result. In prompts, ask for it explicitly: "Use search to find tomorrow's forecast for Edinburgh and make a clean five-day weather graphic." Nano Banana 2 Lite does not support grounding. Check facts in grounded images before publishing, as you would any chart.

Iterating in one conversation

Google recommends multi-turn conversation as the best way to iterate: generate, then refine with follow-ups in the same thread, such as "change the language on the graphic to Spanish". Change one thing per message, refer to the subject with the same words each time, and go back to the last approved version if a change goes wrong rather than piling further edits on top.

Save the prompts that work in a document of your own, grouped by job, so the next brief starts from a tested template rather than a blank chat.

Try it

A photorealistic close-up of a ceramicist's hands shaping a small bowl on a wheel, wet clay glistening, soft window light from the left, shallow depth of field, warm earthy palette, calm and focused mood, 4:5 format.

Generate it, then edit in the same chat: "Change only the clay to a pale blue porcelain; keep the hands, light and wheel the same." Then try a style transfer: "Turn this into a loose ink and wash illustration, same composition."