Nano Banana prompts for image generation and precise editing

Choose generation, editing, or reference-based composition

The first decision is whether you need a new image or a controlled change to an existing one. A generation prompt describes the whole scene. An editing prompt identifies the change and protects everything else. A reference-based prompt assigns a specific purpose to each uploaded image instead of asking the model to combine them indiscriminately.

The current model commonly called nano banana 2 is identified in Google’s documentation as Gemini 3.1 Flash Image. The earlier Nano Banana model is Gemini 2.5 Flash Image. Select the model deliberately: a familiar nickname in a third-party interface does not tell you which endpoint or output settings it uses.

• Generate from text when the scene does not need to match an existing person, product, or photograph.
• Edit an uploaded image when framing, identity, or product geometry already matters.
• Use multiple references when identity, product details, and visual style come from different images.

Gemini suits conversational work; Google AI Studio exposes direct model experimentation; the API and Vertex AI suit automated workflows. Krea offers a credit-based route. For simpler photo adjustments, Pict.AI is another option: an AI photo editor app for iPhone/Android and a website with guides and free image tools. A different editor is not necessarily running the same Gemini model.

Build a Nano Banana prompt in seven steps

Write instructions in priority order. A clear task and a short list of protected details are more useful than a long string of photographic adjectives. Use this workflow for both new images and edits.

1. Name the action. Begin with “Create,” “Edit,” “Replace,” or “Recompose.” Say whether the attached image is the starting point or only a reference.
2. Describe the subject. Specify the number of people or objects, their pose, materials, clothing, and important distinguishing features.
3. Set the composition. Define camera angle, crop, subject placement, background, and empty space needed for a headline or interface overlay.
4. Explain the lighting. State direction, softness, and contrast. “Soft window light from the left” gives a clearer target than “beautiful lighting.”
5. Assign reference roles. For example: image 1 supplies identity, image 2 supplies the product, and image 3 supplies the color palette. State which reference takes priority if they conflict.
6. Add output constraints. Request an aspect ratio, available resolution tier, exact wording, and protected details. Set resolution in the interface or API controls where available, rather than relying only on prompt text.
7. Revise one variable. Ask for a warmer background, a wider crop, or a corrected word separately. Compare each revision against the original reference.

Reusable structure: “Create/edit [asset]. Subject: [description]. Composition: [framing]. Environment: [setting]. Lighting: [direction and softness]. Text: [exact wording and placement]. References: [role of each image]. Output: [ratio and resolution]. Do not change: [protected details].”

Not every template field is necessary. For a background replacement, the existing photograph already supplies most of the subject and composition information.

Six copy-ready prompts for common image tasks

These Nano Banana prompts are starting instructions, not guarantees of a finished asset. Replace the bracketed details, attach the specified references, and choose the output settings supported by your interface.

1. Product hero image with protected packaging

“Edit the attached product photo into a clean ecommerce hero image. Keep the bottle shape, cap, label artwork, and printed wording unchanged. Place it upright on a warm off-white surface with soft light from the upper left and a subtle contact shadow. Use a straight-on camera angle. Center the product with generous margins. Output a square image. Do not add props or extra text.”

2. Portrait background replacement

“Replace only the background of the attached portrait with a softly blurred studio backdrop in muted gray. Preserve the person’s facial features, expression, hairstyle, clothing, pose, and skin texture. Match the background light to the existing light on the face. Keep fine hair edges natural. Do not smooth the skin or change the crop.”

3. Poster with an exact headline

“Create a portrait-format poster for a neighborhood plant sale. Show three potted plants across the lower half on a pale cream background. Put the exact headline ‘SATURDAY PLANT SALE’ in large, dark-green, bold sans-serif lettering at the top. Below it, write exactly ‘10 AM, 2 PM’. Leave clear margins and strong contrast. Include no other words, logos, or decorative lettering.”

4. Combine a product and a room reference

“Use image 1 as the exact chair reference and image 2 as the room reference. Place one chair beside the window in the room. Preserve the chair’s upholstery color, arm shape, and leg design. Match its scale, perspective, shadows, and lighting to the room. Keep the room layout unchanged. Do not duplicate the chair.”

5. Consistent three-shot storyboard

“Using the attached character reference, create a wide shot of this person entering a quiet train station at dawn. Preserve their face, hairstyle, navy coat, and brown shoulder bag. Use cool ambient light and restrained colors. Output a 16:9 image.” For subsequent shots, request a medium shot and then a close-up while explicitly retaining those same identity and wardrobe details.

6. Wide banner with space for later typography

“Create a 4:1 banner showing a ceramic coffee cup on a light wooden desk. Place the cup in the right third. Keep the left half uncluttered with a plain, softly lit background for text to be added later. Use a slightly elevated camera angle, realistic ceramic texture, and soft morning light. Do not generate lettering.”

Check identity, typography, and edges before export

A good-looking thumbnail can hide errors that matter in a product listing or printed poster. Inspect the image at its intended display size, then zoom in on details that carry identity or meaning. Higher resolution does not correct a misspelled label or an altered face.

• Identity: compare facial proportions, hairstyle, distinguishing marks, and clothing against the reference. For a sequence, compare every frame against the same source image.
• Product fidelity: check silhouette, seams, controls, cap shape, label layout, and color. A plausible-looking substitute is still the wrong product.
• Text: read every word, number, punctuation mark, and repeated letter. Check line breaks and margins, not just spelling.
• Compositing: inspect hair, transparent materials, object boundaries, contact shadows, and reflections. Look for mismatched perspective or a floating object.
• Composition: verify that the final crop leaves sufficient space for headlines, navigation, or platform overlays.
• Instruction compliance: count subjects and objects, and check every detail listed under “Do not change.”

When an output fails, name the defect precisely: “Restore the original cap shape; keep the background and lighting unchanged.” Avoid asking the model to “make it better,” which leaves the direction of the revision undefined.

For text that must be exact, generating the illustration without lettering and adding typography afterward is often the more controllable workflow. Keep a copy of the original input so you can restart an edit if successive revisions drift.

Compare interfaces, output sizes, and API costs

The same prompt can encounter different upload limits, resolution controls, and billing rules across interfaces. Choose the access route before planning a batch of images. Google’s image-generation documentation identifies the model options and supported generation settings.

As of September 2026, Gemini 3.1 Flash Image has 512px, 1K, 2K, and 4K output tiers. The 1K and 2K tiers are generally available; 4K remains in preview. Supported aspect ratios include 4:1, 1:4, 8:1, and 1:8, although individual interfaces may not expose every option.

The image-output rate is $60 per million output tokens. These amounts cover image output, not necessarily the entire request. At 1K, 100 image outputs cost approximately $6.70 for that component. Consumer subscriptions and third-party credits are separate from Gemini API pricing.

Prompt myths that cause avoidable revisions

Myth: a longer prompt always produces a better image.

Length helps only when it adds relevant instructions. Conflicting lighting, camera, and style requests make the target less clear. Put the task first, then separate visual requirements from protected details.

Myth: a style reference also preserves identity.

A reference can serve several purposes, but you should identify its intended role. If one image supplies a face and another supplies the art direction, say so explicitly rather than leaving that distinction implicit.

Myth: uploading more references always improves consistency.

The current endpoint supports up to 14 reference images in relevant interfaces. That is a capacity limit, not a recommendation. Use only images that contribute a clear identity, object, composition, or style constraint.

Myth: requesting 4K fixes visual mistakes.

Resolution specifies output size, not factual correctness. Correct the subject, text, and composition before requesting a larger final output.

Myth: “Nano Banana AI” is a separate consumer product.

The name informally describes access to Google’s image models through Gemini, development tools, or third-party services. The interface determines the controls and commercial terms you actually receive.

For reusable Gemini Nano Banana prompts, save the successful instruction alongside its reference images, model identifier, aspect ratio, and resolution. A prompt without those settings is an incomplete recipe, especially when moving between interfaces.