Nano Banana Pro

How Nano Banana Pro reads a prompt

Nano Banana Pro is Gemini 3 Pro Image. Google describes it as a thinking model: before it draws, it reasons through the prompt and can produce interim draft images to work out the composition. That step cannot be turned off, and it is the reason the model handles briefs with many constraints. It also means a short keyword list leaves most of the model's abilities unused. Write in full sentences, the way you would brief a designer, and say what the image is for.

Google's prompting tips name the parts a prompt should cover: the subject, the composition, the action, the location, the style and, for edits, the change you want. Purpose matters as well. "A logo for a high-end, minimalist skincare brand" works better than "a logo", because the model uses the context to make choices you did not spell out.

A structure that works

Put the parts in this order, and drop any that do not apply:

[Format and purpose] A 4:5 product photo for a launch post.
[Subject] A matte green glass bottle of olive oil with a cream paper label.
[Action and setting] Standing on a marble counter next to half a lemon.
[Composition and camera] Eye-level close-up, shallow depth of field.
[Light and style] Soft morning window light from the left, natural color.
[Text] The label reads "Oliva" in a thin serif, nothing else written.

For a complex scene, Google suggests giving instructions step by step: first the background, then the foreground objects, then the detail on top. Describe what you want instead of what you do not want. "An empty street with no traffic" works better than "no cars".

Text, infographics and localization

Text is where this model differs most from the original Nano Banana. Google says it renders legible text from a short tagline to a long paragraph, in several languages, with a range of fonts, textures and calligraphy. To get it right:

• Put the exact words in quotes and say where they go and how they look: "the headline 'Late Set' in tall condensed letters at the top".
• Say how many times the text should appear, and ask for no other text when the layout must stay clean.
• For diagrams and infographics, list every label you need and ask for accuracy ("a scientifically accurate cross-section"). Check the facts yourself; Google lists data accuracy as a known limit.
• To translate an existing image, attach it and ask to translate the text into the target language "while keeping everything else the same". The model keeps the layout and style around the new words.

Google's docs add one practical tip: when the wording itself is not settled, draft the text first and then ask for the image that contains it.

References and editing

The model mixes up to 14 reference images in Google's API: up to 6 objects at high fidelity, up to 5 people kept consistent, and up to 3 style references. Endless accepts up to 10 references for this model. Give each one a role, by number: "use image 1 for the character's pose, image 2 for the art style and image 3 for the background". Without roles, the model has to guess which image controls what.

Edits work in plain language. Name the one thing to change and the things to keep: "change the tie to green, keep the face, pose and lighting the same". Google highlights edits to camera angle, focus, color grading and lighting, such as turning day into night or adding bokeh, and changing the aspect ratio of an existing image by extending or trimming the background. Make one change per turn, then refine the result from there.

Settings that matter

Endless exposes three settings for Nano Banana Pro:

• Resolution: 1K (the default), 2K or 4K. Draft at 1K, and move to 2K or 4K for print, large screens or images with small text.
• Aspect ratio: auto or one of 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3 and 9:16. Pick the ratio here, and also mention the format in the prompt ("a 9:16 vertical poster") so the composition is planned for it.
• Reference images: up to 10, used for editing and multi-image composition.

Common failure modes and fixes

• Small text is misspelled. Google notes that small text and fine detail can still fail. Shorten the copy, make the text larger in the layout, raise the resolution, or spell an unusual word letter by letter.
• A translation reads awkwardly. Multilingual text can contain grammar mistakes or miss cultural detail. Have a native speaker check it before you publish.
• People drift between edits. Restate who must stay the same, attach the original photo again, and keep each edit to one change.
• Added elements do not match the scene. Ask for the light, shadows and color of the target scene to be matched on the added element, and name the light direction.
• The wrong reference controls the result. Number the images and give each a single job in the prompt.
• Facts in a diagram are wrong. Supply the facts in the prompt instead of relying on the model's world knowledge, and verify the result.