Most Nano Banana Pro tutorials open with a sign-up link. Your output is decided elsewhere, in the first fifteen words you type.
You will not need a paid plan here, nor Google’s documentation: what follows is built on it. If you have never generated an image, start with the getting-started walkthrough.
Nano Banana Pro tutorial: the short answer
Fill six slots and the first image is usually usable: subject, composition, action, location, style, and an explicit edit instruction when you are editing rather than generating.
Five are Google’s own list. The sixth is where beginners leave a blank, and a blank lets the model decide what may change.
Six lines, one screen. The rest of this page fills them in.
The six slots Google tells you to fill
Google’s tips post names each element with a worked example attached, and the examples teach more than the labels.
The elements Google names
Subject is who or what is in the image. Google’s example is a stoic robot barista with glowing blue optics: an occupation, an expression and a light source, not the word robot. Composition is how the shot is framed, action is what happens, location is where it takes place, style is the aesthetic.
Every slot wants a noun someone could photograph; Cinematic belongs in the camera or lighting slot, unpacked.
Subject: a stoic robot barista with glowing blue optics
Action: brewing a cup of coffee, steam rising
Location: a futuristic cafe on Mars, dust light through a wide window
Composition: low angle shot, portrait framing, subject slightly left of centre
Style: 3D animation with 1990s product photography lighting
Editing instruction: leave blank when generating from scratchFour of those lines are Google’s, reordered. Copy the shape, not the subject.
Where the six slots go wrong
Two failures explain most weak first attempts: a style slot holding an adjective instead of a reference, and an action slot describing a state. Beautiful is not a style, and a cat sitting is not an action.
Where to run it, and why prices disagree
Nano Banana Pro is built on Gemini 3. Google names the Gemini app first, with AI Studio and Vertex rolling out after it; every other interface is a third-party wrapper with its own billing.
Which is why the prices disagree. Google’s pages publish no per-image price, no tier and no free quota, while wrapper sites quote plans and credit tables belonging to the site quoting them.
What the vendor publishes, and what it does not
Published: the model runs on Gemini 3, accepts up to 14 images in one composition depending on surface, outputs at 1K, 2K and 4K, and stamps a SynthID watermark. Not published by the vendor: price per image, free quota, the font list, the full aspect-ratio set, and safe-area figures.
The API guide shows where a request is configured; the free-access write-up collects what is documented rather than quoted second-hand.
Camera, light and aspect ratio: what to add
Three lines turn a description into a shot, and Google supplies an example of each.
Composition and aspect ratio: A 9:16 vertical poster, subject centred, copy space across the lower third.
Camera: A low-angle shot with a shallow depth of field (f/1.8), 35mm.
Lighting: Golden hour backlighting creating long shadows, cinematic colour grading with muted teal tones.Where the detail changes the image
Put the aspect ratio inside the sentence, not in a setting you hope gets applied, and name the light direction: soft light and light from the upper left are visibly different pictures. Then move one element per round, because a round that shifts two variables teaches you nothing.
Three worked examples you can copy
Each runs in one turn: the first generates from nothing, the second composes three references, and the third edits a photo you already own.
Example one: a poster whose headline must be spelled right
Start from the constraint you cannot repair later: Google’s text rule asks for the string and how it looks.
Create a poster for a fictional urban photography festival. A 9:16 vertical poster, subject centred. The headline "URBAN EXPLORER" rendered in bold, white, sans-serif font at the top. Below it, a low-angle shot of a person walking through wet city streets at dusk, shallow depth of field (f/1.8). Golden hour backlighting creating long shadows, cinematic colour grading with muted teal tones. Copy space across the lower third. No watermark, no added logos.Short and quoted, both deliberately. Google lists accurate spelling as an unresolved weakness, and a wrong letter in a headline cannot be fixed afterwards.
Example two: three references, three named roles
Google’s rule is to define the role of every uploaded image. The failure is rarely the model but an unlabelled second file.
Use Image A for the character's pose, Image B for the art style, and Image C for the background environment. Keep the character's face, hair and build consistent with Image A. Render the result as a cinematic 21:9 wide shot with soft rim lighting from the right. No text, no watermark.Drop the role list and the model follows one image and quietly ignores the others.
Example three: an edit that leaves everything else alone
Google’s editing examples are blunt: a keep-list followed by one change.
Keep the subject's face, skin texture and expression exactly as they are. Change only two things: change the man's tie to green, and remove the car in the background. Fill the space where the car was with the existing pavement and shopfront, matching the original perspective and light direction.Two changes in one round is a practical ceiling, past which you cannot tell which instruction produced the artefact. Google warns that blending and lighting edits are where artefacts appear.
What still breaks, and what to check first
The same page lists five limitations, which makes it the most useful checklist the vendor publishes.
The five documented failure modes
Small text, fine details and accurate spelling may not render perfectly. Data-driven visuals may be factually wrong. Multilingual text may carry grammar mistakes. Complex edits can produce artefacts, and character consistency is reliable but not guaranteed.
Read that list against your own image. A face that must stay readable at thumbnail size, a licence plate and a legal line are where to look at full resolution. For cut-out edits the photo editor walkthrough goes deeper; for charts and headlines, the poster and infographic notes are the closer match.
Frequently asked questions
Do I have to write the prompt in English? No. Google treats localised text and translation inside images as a first-class task; write the instruction in whichever language you are precise in.
How many reference images can I use? Up to 14 in one composition, varying by surface. Google’s guidance is to name the role of each.
Is there a free quota? Not published by the vendor, so any number you read elsewhere belongs to the site quoting it.
How long should the prompt be? Google says short prompts still produce good images, and that control comes from adding detail one element at a time.
Fill the sixth slot, then change one thing per round. That is the whole method, and the part the wrapper sites leave out.