Nano Banana prompts for designers: seven examples to adapt

The best Nano Banana prompts read like a brief to a photographer or a retoucher. They say what is in the frame, where the light comes from, what the image is for and, most important, what must stay exactly as it is. Nano Banana, the name Google gives the image models in Gemini, is at its best when you hand it a real image and change one thing with words.

Below: what Nano Banana is in October 2026, what Google says it does well, and seven example prompts for everyday design jobs. Each is a labelled starting point with notes on what to adjust, not a result I am showing. Then Arabic text, SynthID and responsible use. The general method is in prompt writing for designers, and the wider picture in AI in graphic design.

What Nano Banana is, as of October 2026

Google first used the name in August 2025, when it introduced Gemini 2.5 Flash Image "aka nano-banana". The nickname stuck, and Google now uses it for a family. Its developer documentation, last updated on 23 September 2026, lists four models:

In the Gemini app, Nano Banana 2 has been the main image model since February 2026, and paid Google AI plans add Nano Banana Pro and higher limits. Developers use AI Studio, the Gemini API and Vertex AI, and Photoshop offers Nano Banana as a partner model in Generative Fill. Treat the table as a snapshot.

What Google says it is good at

Three strengths matter most to a designer.

1. Editing with words. Upload an image and describe the change: add or remove an element, or change only one part and keep the rest, which Google calls semantic masking.
2. Keeping a subject consistent. Google says the models keep people and objects looking the same across edits and scenes, and adds that this "may not always get it right". Check every image in a series.
3. Combining images. Give it several images and ask for a new scene: the product from one, the setting from another, the style from a third. Google's API docs say the Gemini 3 image models can mix up to 14 reference images.

Its stated limits matter as much: the model "can still struggle with small faces, accurate spelling, and fine details", and masked edits or major lighting changes "may produce unnatural results".

How to use the examples

Each example follows the shape of Google's own templates (subject, setting, light, camera, purpose, format) and adds one habit that pays off with editing models: say what must not change. Attach your own images where the prompt mentions them, change one thing per round, and if a product starts to drift after many edits, start a fresh conversation with the original image.

Example 1: a product shot on a set (starting point)

Use it when you have a clean photo of the real product and need it in a styled scene.

Using the provided photo of the coffee bag, place the bag on a set: a warm travertine plinth in front of a plain sand-coloured wall, with a few green coffee cherries and a small brass scoop beside it. Soft window light from the left, a gentle shadow falling to the right. Keep the bag's shape, colours, label and logo exactly as they are in the photo. Eye-level camera, the look of an 85mm lens, shallow depth of field. This is a product photo for an online shop. Vertical 4:5 format.

Adjust: the set, using materials and props from the brand's world, and the light direction, to match the series. If the label changes at all, put the real artwork back in Photoshop.

Example 2: changing a background (starting point)

Use it when the product photo is good but the background is wrong for the channel.

Using the provided image, change only the background to a smooth gradient from deep olive green at the bottom to a lighter olive at the top. Keep the product, its position, its reflections and its shadow exactly the same, and match the new background to the existing light. Do not add any other objects.

Adjust: the colours, to the brand palette; an exact match is not guaranteed, so check and correct afterwards. If the background implies a different surface, describe it, or the shadow will look pasted on.

Example 3: colour variations of a pack (starting point)

Use it when you want to preview colour directions for a range before you build the artwork.

Using the provided image of the medicine box, create three versions side by side. Change only the colour of the large main panel: version one in deep teal, version two in warm terracotta, version three in soft lavender. Keep the box shape, the white areas, the layout, the logo and every line of text exactly as they are. Same camera angle and lighting in all three. Plain light grey background. Landscape 3:2 format.

Adjust: the colours, and how many versions you ask for at once; fewer usually stay cleaner. This is for direction only: final colours are specified in the artwork with print references. The Twins Pharma range, seven products each with its own colour and a clear bilingual label, is the kind of system where a quick preview helps a client conversation, while the colours and labels belong in the real files.

Example 4: a moodboard grid (starting point)

Use it when you need a first visual direction to discuss with a client.

A moodboard for a modern Arabic bakery brand: a clean three-by-three grid of square photographs with thin white gutters. The tiles: a stone oven glowing at dawn, sesame ka'ak on linen, flour on dark wood, hand-glazed green tiles, morning light through a carved wooden screen, kraft paper bags, crust texture in macro, a brass tray of tea glasses, a quiet street at sunrise. Natural, warm light, muted greens and wheat tones. Photographs only, with no lettering or signs. 16:9 format.

Adjust: the tiles, so they come from the brand's real world, and the palette. You can also upload two or three of your own reference photos and say what to take from each. Treat the grid as a direction, not a finished design.

Example 5: a social post background with room for Arabic text (starting point)

Use it when the post will carry an Arabic headline that you will set yourself.

A calm background for a social media post: a soft morning sky over low sand dunes, the dunes filling only the bottom fifth of the frame, with one small brass lantern standing on the sand at the lower left. The right side and the upper half of the image stay empty, smooth and evenly lit, leaving clear space for a headline. No text, no letters, no logos and no signs anywhere in the image. Warm, quiet mood, gentle colours. Portrait 4:5 format.

Adjust: where the empty space sits. Arabic reads from right to left, so a headline usually starts on the right. Then set the Arabic as live type, and remove any letter-like marks the model adds anyway.

Example 6: restyling a sketch (starting point)

Use it when you want to show a client how a rough idea might look before you commit to drawing it properly.

Turn this rough pencil sketch of a perfume bottle and its box into a clean product concept render. Keep the proportions, the curve of the bottle's shoulders and the position of the label panel exactly as in the sketch. Render the bottle in frosted glass with a matte black cap, and the box in uncoated cream board. Neutral studio light, plain light grey background, three-quarter view. Leave the label panels blank.

Adjust: the materials, the angle, and how closely to follow the lines ("keep the proportions exactly" or "use the sketch as a loose guide"). The Pupal identity shows the stage this suits: the mark grew from iPad sketches, through three ideas presented to the client, to the final mark. A render can help a client picture an idea at the sketch stage; a logo itself still has to be drawn properly.

Example 7: a logo mockup on stationery (starting point)

Use it when you need a quick presentation scene for a logo you own or your client has approved.

Using the provided logo, create a photo of a stationery set on a light oak desk: a business card, a letterhead and an envelope, arranged at a slight angle. Print the logo from the image on each item: centred on the card, small in the top corner of the letterhead and on the envelope flap. Keep the logo's shapes, proportions and colours exactly as in the provided file, and do not redraw it. Soft daylight from a window on the right, gentle shadows. Uncoated off-white paper with a fine texture. Landscape 3:2 format.

Adjust: the paper, the desk and the items. Then compare the logo with the original at full size; if anything moved, place the real vector over it as a smart object, or use a classic mockup for the final presentation.

Arabic text: generate without it

The safest Arabic workflow with Nano Banana is the same as with any image model in 2026: prompt for space, not letters. Ask for "no text, no letters, no signs", say where the empty area goes, and set the Arabic yourself in a tool that shapes the script properly. I explain why the letters break in Arabic text in AI images.

Two details are specific to Gemini:

• Writing the prompt in Arabic is fine. Google's Gemini API documentation lists Arabic (ar-EG) among the languages with the best performance. That is about understanding your prompt, not drawing Arabic letters correctly.
• If you must test Arabic inside the image, keep it to a word or two and follow Google's tip: settle the exact text first, then ask for an image with that text. Check it letter by letter with a native reader, and never use it for a logo, a label or a price.

SynthID and watermarks, as Google states them

• SynthID. Google says every image from its Nano Banana models carries an invisible SynthID watermark marking it as AI-generated or AI-edited.
• Content Credentials. The Gemini app also adds C2PA metadata about the image's origin.
• The visible watermark. In the Gemini app it is a setting (Settings, then Media Watermark); in India, South Korea and Vietnam only AI Ultra subscribers see it. Turning it off leaves SynthID and Content Credentials in place.
• Checking an image. Ask Gemini whether an uploaded image was created or edited by Google AI. It only recognises Google's own AI content.

Don't try to strip these marks. Google's policy forbids presenting generated content as made only by a human in order to deceive.

Responsible use

• Real people need consent. Google's prohibited use policy rules out impersonating someone to deceive, and using personal data or biometrics without the consent the law requires.
• Logos you don't own stay out. Use your own marks or ones your client has approved, and don't ask for a copy of a competitor's pack.
• Client material needs care. On personal accounts, Google says human reviewers may read some Gemini conversations, and advises against entering confidential information. For unreleased client work, use a work account or a paid setup with clearer data terms, and ask the client first.
• Label what you generate, so people know what was photographed, rendered or generated.

From my work

In the video series I direct at SkillUp MENA (AI Video Studio), I direct AI tools for stills, motion and voice. The lesson that carries over is about order: stills must be consistent before anything moves, because a product that drifts between frames drifts more in motion. My guide to turning images into video picks up from there, and if you are still choosing a tool, see Midjourney vs ChatGPT vs Nano Banana.

Conclusion

Nano Banana is strongest when you treat it as an editor rather than a slot machine. Start from a real image, describe the change like a brief, say what must stay the same, and change one thing per round. Swap your own brand's materials and colours into the seven examples, keep Arabic type and real logos in your own hands, and label what you generate.

If you want product visuals, packaging or a campaign where AI speeds up the work and every decision is still designed, start a project here.