Nano Banana Prompt Guide: How to Write Better Prompts

Nano Banana reads prompts like instructions from a person, not like a list of tags. The single most useful rule, which Google repeats in its own prompting guidance, is to describe the scene in full sentences instead of listing keywords. This guide covers the formula for new images, the rules for edits, text, several photos, JSON prompts and the mistakes that cost you retries.

The formula for new images

Subject, then setting, then light, then camera or style.

• Weak: “woman, cafe, cozy, 4k, cinematic”
• Strong: “A woman in her thirties reading by the window of a small Paris café on a rainy afternoon, soft grey daylight on her face, steam rising from a cup, shallow depth of field, 50mm photo.”

Add the details that matter to you and leave out the rest. Camera words (85mm, macro, wide shot), light words (golden hour, softbox, neon) and medium words (oil painting, watercolor, 3D render) all change the result reliably.

The rule for edits: change, then keep

When you upload a photo, name the change first and then say what must stay:

“Replace the background with a snowy forest at dusk. Keep the person, their face, hair and clothing exactly the same.”

• One change per prompt. Chain several short edits rather than one long one.
• Name what to keep, especially faces. The keep list is what protects your subject.
• Describe the light when you move a subject, so the result looks like one photo.

More edit prompts are in Nano Banana photo editing prompts.

Text in images

• Put the exact words in quotes: headline “SUMMER SALE”.
• Keep each text element short, and say where it goes and what the font looks like (bold condensed sans-serif, elegant serif).
• For more than a headline, use Nano Banana Pro; it handles several lines far better than the Flash models.

Several photos in one prompt

Refer to images by order and say how they combine: “The person from the first photo wearing the jacket from the second photo, standing in the street from the third photo, matching the light.” Google’s documentation lists how many references each model takes; on Nano Banana AI each model accepts up to 4 images per request. See Nano Banana character consistency for keeping the same face across a series.

Say what you want, not what you don’t

Nano Banana has no separate negative prompt field. Instead of “no cars”, describe the result you want: “an empty street at dawn”. Google calls this a semantic negative prompt, and it works better than a list of things to avoid.

JSON prompts

Some people write prompts as JSON, with fields such as subject, style, lighting and camera. There is no special JSON mode: the model simply reads the structure as text, and it helps mainly as a checklist so you do not forget a field. For example:

{
  "subject": "a ceramic coffee mug with a hand-painted lemon",
  "setting": "rustic wooden table by a window",
  "lighting": "soft morning light from the left",
  "camera": "close-up product photo, shallow depth of field",
  "style": "warm, natural, editorial"
}

A well-written sentence with the same details works just as well.

Aspect ratio and size

In the Gemini app, ask for the shape in the prompt (“wide 16:9 image”). In AI Studio and the API, and on Nano Banana AI, you pick the aspect ratio and resolution as settings, which is more reliable than asking in words. Details in Nano Banana 4K and aspect ratios.

Iterate like a conversation

Treat the first image as a draft. Follow up with small, specific changes (“warmer light”, “move the title to the top”, “make her smile slightly”) rather than rewriting the whole prompt.

Common mistakes

• Asking for many changes at once, which makes faces drift.
• Long text in a Flash model instead of Pro.
• Vague style words (“beautiful”, “amazing”) instead of concrete ones.
• Forgetting to say what to keep in an edit.

For copy-paste starting points, browse the Nano Banana prompts library or build one with the Nano Banana prompt generator.