VidoGridWhat good Nano Banana prompts contain
The models sit on top of Gemini, so they read language rather than match keywords. A short prompt still returns a usable picture, but detail is how you take control. Say what the image is for, too: "a hero image for a refill shop's product page" gives the model more to work with than "a bottle photo".
For busy scenes, give the order of work: background first, then the main subject, then the detail that matters most. Google DeepMind suggests deciding five things before you write:
- Style: photograph, watercolor, flat vector, clay render.
- Subject: who or what is in the frame, what they wear, how they are posed.
- Setting: the place, the time of day, the weather.
- Action: what is happening at that moment.
- Composition: portrait or landscape, close-up or wide, camera height.
A studio photograph of a refillable glass perfume bottle with a brushed brass cap on wet black slate. One soft light from behind makes the liquid glow amber. Eye-level camera, 100mm macro lens, shallow depth of field. For an online shop listing, square.
Composition, camera and negative space
Photographic vocabulary works well in Nano Banana prompts. Shot types (wide shot, macro), lens lengths (24mm, 85mm), camera angles (low angle, overhead) and a named light source give the model concrete choices.
If text will go over the image later, say where the subject sits and which part of the frame stays empty. To leave something out, Google recommends semantic negative prompts: instead of "no people", describe the empty version of the scene, such as a deserted plaza at dawn.
A wide photograph of a salt flat after rain, its water mirroring a pink dawn sky. A single cyclist rides along the horizon in the lower third, small in the frame. The upper two thirds stay calm and empty for a headline. Ground-level camera, 24mm lens, 21:9.
Text, diagrams and real-world facts
Legible text is one of the family's strengths, and Google points to Nano Banana Pro for professional assets with lettering. Put the exact words in quotation marks, describe the type (bold condensed sans-serif, neon script) and say where each line goes. Settle the wording in the chat first; the documentation says text comes out better that way.
For another language, name the target language and any regional cues; if the phrasing must be exact, quote the translated text itself. The same Nano Banana prompts can ask for timelines, flowcharts or infographics, and with Google Search grounding on, the model can check current facts such as a forecast (not on the Lite model).
A vintage screen-printed poster for a neighborhood jazz night. At the top, "LATE SET" in tall condensed sans-serif letters, cream on deep teal. In the middle, an upright bass in three flat colors with paper grain. At the bottom, in small italic serif: "Fridays, 9 pm, Harbor Street Hall". Portrait, 2:3.
Nano Banana prompts for edits and reference images
For edits, name what changes and what stays: "change only the jacket to olive green; keep the face, pose, lighting and framing" beats "make the jacket green". The model matches the original style, lighting and perspective; to protect a face or logo, describe it in detail in the edit request.
When combining images, say which element comes from which file and give each person or object a distinct name. Nano Banana 2, 2.1 and Pro accept up to 14 references: up to 10 objects and 4 characters on Nano Banana 2 and 2.1, or 6 objects, 5 characters and 3 style images on Pro. The original model works best with three inputs or fewer, and 2 Lite is not tuned for many references. The Nano Banana 2 models also accept a video, such as a YouTube link, as the reference for a thumbnail.
Image 1 shows Mira, a woman with a silver bob and round tortoiseshell glasses. Image 2 shows her spaniel, Pepper. Place both in a narrow street at night in light rain, Mira holding a clear umbrella while Pepper looks up at a glowing noodle-shop sign. Keep Mira's face and glasses exactly as in image 1. 16:9.
Settings that change the result
Some choices work better as settings than as words. In the Gemini API they sit in response_format; in the Gemini app, name the ratio and format in the prompt, such as "a wide website banner". Write Nano Banana prompts in a language Google lists for best performance, such as English, Spanish, French, Japanese or Chinese.
- aspect_ratio: ten ratios from 21:9 to 9:16, plus 1:4, 4:1, 1:8 and 8:1 on Nano Banana 2 and 2.1. Unset, the output follows the input image or comes out square.
- image_size: 1K by default on the Gemini 3 models, or 2K and 4K, written with a capital K.
- Thinking level: minimal, medium (default) or high on Nano Banana 2.1; minimal (default) or high on Nano Banana 2 and 2 Lite. Thinking cannot be switched off and its tokens are billed.
- Output: text and image by default; ask for image only through response_format.
Common mistakes and fixes
Most weak Nano Banana prompts fail in a few predictable ways.
- A list of tags: rewrite it as two or three sentences covering style, subject, setting, action and composition.
- Expecting a perfect first image: ask for small changes in the same conversation and say what stays.
- A character drifting after many edits: open a new conversation with a full description and the reference images again.
- Asking for an exact number of images: the model does not always return that count, so check what came back.
- An edit that changes the frame shape: set aspect_ratio, or tell the model to keep the input aspect ratio.