Blog / 8 min read · 2026-09-29

Nano Banana Prompt Best Practices: 12 Rules That Actually Improve Your Results

Nano Banana prompt best practices for clearer image edits, stronger compositions, and repeatable results, with copy-and-paste examples you can easily adapt.

The best Nano Banana prompts state what to make or change, what must stay stable, and what the finished image should look like. This is especially useful because Gemini Apps supports editing generated images, uploaded images, and combinations of multiple uploaded images. Google’s current Gemini Apps guidance also recommends naming the subject, action, background, and style.

The short version: write a visual brief, not a bag of keywords

A useful Nano Banana prompt usually follows this order:

[Action] + [main subject] + [setting/composition] + [style and lighting] +
[constraints] + [what must remain unchanged]

For an edit, put the requested change first. For a new image, put the main subject first. If you need a starting structure, use this Nano Banana master prompt template and adapt only the fields that affect the image.

12 Nano Banana prompt best practices

1. Begin with a clear action

Open with a verb that tells Gemini what job it is doing: create, edit, replace, remove, transform, or combine.

Weak:

A coffee shop with a person working

Better:

Create a realistic editorial photo of a freelance designer working in a quiet coffee shop.

For uploaded images, name the operation directly:

Using the uploaded photo, replace the cloudy sky with a pale blue morning sky.

The action prevents the prompt from reading like a loose description.

2. Name one main subject before adding details

Decide what the image is really about. A prompt can contain many objects, but it needs a visual priority.

Create a product photo of a matte black insulated water bottle on a light stone countertop.

Then add supporting elements:

Create a product photo of a matte black insulated water bottle on a light stone countertop, with soft window light from the left and a folded linen napkin in the background.

This is more reliable than asking for “a premium lifestyle scene” and hoping the product remains prominent.

3. Describe the subject doing something

Static nouns leave many decisions open. A pose, interaction, or visible state gives the image a direction.

Create a candid street-style portrait of a woman adjusting the collar of her wool coat while waiting at a crosswalk.

For products:

Create a close-up food photograph of a spoon lifting a ribbon of honey from a glass jar.

Use action only when it helps. A clean catalog image may need the product centered and still.

4. Set the scene, not just the backdrop

“White background” is a backdrop. “A clean studio sweep with a soft shadow beneath the product” is a scene with useful visual information.

Create a full-body fashion photo of a person standing in a sunlit concrete courtyard, with long afternoon shadows and a softly blurred architectural background.

Google’s image-prompt guidance specifically identifies subject, context/background, and style as the core building blocks of a clear prompt. Google AI for Developers

5. State the visual medium and mood

“High quality” is vague. Ask for a recognizable medium or treatment instead: documentary photo, polished e-commerce image, charcoal sketch, retro travel poster, or editorial still life.

Create a polished e-commerce photo of a ceramic desk lamp, warm neutral palette, soft studio lighting, realistic materials, minimal Scandinavian styling.

For people, mood works best when paired with observable details:

Create a natural-light portrait with a calm, thoughtful expression, muted colors, and shallow depth of field.

6. Use composition language when framing matters

Composition terms help when the crop determines whether the image works.

Create a vertical, full-body portrait. The subject stands slightly right of center, leaving clean negative space on the left for a headline. Eye-level camera, soft daylight.

Useful terms include close-up, overhead view, wide shot, profile, low-angle view, centered composition, and negative space. Do not add camera jargon just to make a prompt longer; use it to solve a framing problem.

7. For edits, define the boundary with “change only”

The most practical Nano Banana editing habit is to say both what changes and what does not.

Using the uploaded living-room photo, change only the blue sofa to a warm brown leather sofa. Keep the room layout, cushions, wall art, window light, and camera angle unchanged.

This mirrors the structure of Google’s documented image-editing examples, which specify a local change and explicitly preserve the rest of the scene. Google AI for Developers

Avoid combining five unrelated edits in one instruction when preserving the original matters. Make the first edit, inspect it, then continue from that image.

8. Identify the details that cannot drift

If a face, logo, garment, product shape, or room layout is important, name it and describe it plainly.

Keep the person’s face, hairstyle, skin tone, glasses, and neutral expression unchanged. Replace only the grey hoodie with a navy crew-neck sweater.

For an object:

Keep the bottle’s label text, black cap, proportions, and front-facing orientation unchanged. Replace only the background with a bright kitchen counter scene.

“Keep it the same” is less useful than a short list of the features you genuinely need preserved.

9. Label each uploaded reference image by its role

Multiple-image prompting is a documented Nano Banana workflow in Gemini Apps. Current guidance says Nano Banana 2 can accept multiple reference images and create a new image from uploaded images. Google’s Gemini Apps Help

Give each file a purpose:

Use image 1 as the product reference and image 2 as the background reference. Create a realistic product advertisement: place the exact watch from image 1 on the stone surface from image 2. Match shadows and reflections to the scene.

A role label is far clearer than “combine these.”

10. Ask for a positive result instead of a long negative list

Tell the model what the desired image contains.

Instead of:

No cars, no crowds, no traffic, no signs

Try:

Create an empty residential street just after sunrise, with parked bicycles, quiet sidewalks, and no visible traffic.

Negative constraints still have a place when they protect an essential requirement, such as “no text” or “no extra objects on the table.” Keep them limited and specific.

11. Refine one variable at a time

A good first output is a working draft, not a verdict. Google’s prompt-design guidance describes prompting as iterative and recommends refining based on the observed response. Google AI for Developers

Use follow-ups like these:

Keep everything the same, but make the lighting warmer and softer.

Keep the outfit and background unchanged. Move the camera closer to a chest-up portrait.

Keep the composition unchanged. Remove the coffee cup from the foreground.

One-variable revisions make it easier to see what caused an improvement or a regression.

12. Treat text in images as a separate quality check

If your image needs a headline, label, menu, or poster text, provide the exact wording and keep it short. Current Gemini Apps guidance says Nano Banana 2 has improved text rendering, but you should still inspect every generated word before publishing. Google’s Gemini Apps Help

Create a simple square poster for a weekend flower market. Include the exact headline: “SUNDAY BLOOMS”. Use large cream serif lettering at the top, over a dark green background with illustrated ranunculus flowers.

If the spelling or placement is wrong, revise that single issue rather than rewriting the entire prompt.

Three copy-and-paste Nano Banana prompt examples

Photo edit: replace one item

Using the uploaded portrait, replace only the patterned shirt with a plain charcoal-grey crew-neck sweater. Keep the person’s face, hair, pose, expression, lighting, background, and crop unchanged. Make the sweater fabric look natural and realistic.

For more edit-ready variations, see these Nano Banana prompts for photo editing.

Product image: create a controlled scene

Create a realistic premium product photo of a clear glass skincare serum bottle with a white dropper cap. Place it centered on pale travertine, with a soft beige backdrop, morning window light from the right, and a subtle shadow beneath the bottle. Minimal composition, no extra products, no visible text.

Multiple references: combine a person and an item

Use image 1 as the person reference and image 2 as the jacket reference. Create a realistic full-body street-style photograph of the person wearing the exact jacket from image 2. Preserve the person’s facial features and hairstyle. Overcast daylight, clean city sidewalk, natural fabric folds, vertical composition.

FAQ

How long should a Nano Banana prompt be?

Long enough to remove the decisions that matter. A simple image may need one sentence; a preservation-heavy edit may need several. Add detail for subject, setting, style, and constraints, then stop before details begin to conflict.

Should I use a new prompt or continue editing the generated image?

Continue from the generated image when you like the composition and only need a targeted adjustment. Start again when the core subject, framing, or visual direction is wrong.

How do I preserve a face or product during an edit?

Upload the strongest reference image available, list the critical details to preserve, and state the one requested change. Do not bury the preservation instruction under unrelated styling requests.

Why did Nano Banana change more than I asked?

The edit boundary was probably too broad or unspecified. Use “change only,” identify the target precisely, and list the elements that must remain unchanged.