How to Write Prompts for Nano Banana Pro: Structure, Tips and Techniques

Published: July 25, 2026

Most disappointing Nano Banana Pro images come from prompts that were too short, not prompts that were too ambitious. "A professional photo of me" gives the model five words to work with and total freedom over the other forty decisions — lens, light, wardrobe, background, mood. It fills those gaps with an average of everything it has seen.

This guide covers the structure that works, three prompts rewritten from vague to specific, the mistakes worth knowing before you spend a generation, and a checklist you can copy. If you would rather start from finished prompts, the full Nano Banana Pro prompt library has more than 20 organized by use case.

What makes a good Nano Banana Pro prompt

A good prompt fills five slots: subject, style, composition, lighting, and constraints. Cover all five and the model has almost nothing left to guess; skip one and it improvises there. As of July 25, 2026 Nano Banana Pro carries no length penalty, and a detailed paragraph reliably beats a clever one-liner.

Most people skip the constraints slot, and it is what separates a usable image from a near-miss. Constraints are where you say "keep this person's face exactly as it is", "the text must read GRAND OPENING", or "do not alter the background".

You do not need to label the slots in your prompt. Writing them as short labeled blocks does help on complex prompts, though — the model handles a structured brief well, and you can edit one block without rewriting the whole thing.

How to write the prompt, slot by slot

Start with the subject in plain words, then add one slot at a time and read the prompt back after each. Building in that order stops you from producing a wall of adjectives that never says what the picture is.

Here is the same idea at three stages.

Subject only:

1A ceramic coffee dripper.

Subject + style + composition:

1Commercial product photograph of a matte white ceramic pour-over coffee dripper, three-quarter view, centered in frame, slight low angle so the cone reads as tall.

All five slots:

Settings: 1:1 · 1K (draft run — run the finished prompt at 4K)

1Subject: A matte white ceramic pour-over coffee dripper on a pale oak surface.
2
3Style: Commercial product photography, photorealistic, minimal Scandinavian styling.
4
5Composition: Three-quarter view, centered, slight low angle so the cone reads as tall. Generous negative space above the product.
6
7Lighting: Large softbox from the left at 45 degrees, white bounce card on the right to open the shadows, soft contact shadow anchoring the base.
8
9Constraints: Keep the ceramic surface matte with no blown highlights. Seamless off-white background, no props competing with the product, no visible reflections of studio equipment.

The third version takes a minute to write and removes almost every decision the model would otherwise make for you.

Before and after: three prompts rewritten

Each rewrite below keeps the original intent and fills the empty slots. Read the "after" prompts as templates — swap the nouns for your own subject and the structure still holds.

1. Portrait for a professional profile

Before:

1A professional photo of me for LinkedIn.

Nothing here says what "professional" means, so the model picks a look, a light, and an outfit at random. Nor does anything in it say your face has to survive the edit.

After — attach your photo as a reference.

Settings: 1:1 · 4K (final-run settings — draft at 1K)

1Preserve the subject's facial identity from the attached image exactly — bone structure, skin tone, hairline and expression lines unchanged.
2
3Subject: The same person, wearing a well-fitted navy blazer over a plain white shirt.
4
5Style: Corporate headshot, photorealistic, natural skin texture with visible pores.
6
7Composition: Head-and-shoulders crop, subject facing camera, eyes level with the lens, centered.
8
9Lighting: Soft key light at 45 degrees from the left, subtle fill on the right, gentle rim light separating the head from the background.
10
11Constraints: Mid-gray seamless background with a soft falloff. No retouching that smooths skin into plastic. Natural catchlights in both eyes.

2. Product shot for a store listing

Before:

1Make my candle look nice and professional on a white background.

"Nice" is the whole problem. The model has no material, no light shape, and no instruction about the label — which is usually the part that comes back garbled.

After — attach the product photo.

Settings: 1:1 · 4K

1Isolate the candle from the attached photo and place it on a seamless white studio background.
2
3Subject: A single amber glass candle jar with its paper label intact.
4
5Style: E-commerce catalog photography, photorealistic, clean and neutral.
6
7Composition: Straight-on eye-level view, product centered, occupying roughly 70 percent of the frame height.
8
9Lighting: Wrap-around softbox lighting that reveals the glass without harsh specular hotspots, plus a soft contact shadow under the jar.
10
11Constraints: Reproduce the label text and artwork exactly as they appear in the reference. Neutral white balance. No hands, surfaces or props from the original photo.

3. Poster with text

Before:

1A cool poster for my bakery's opening with some nice typography.

Ask for "some text" and you get text-shaped decoration. Nano Banana Pro renders legible type when you tell it the exact words, so write them out verbatim.

After:

Settings: 4:5 · 4K

1Subject: A promotional poster for an artisan bakery opening.
2
3Style: Warm editorial print design, hand-printed riso texture, two-color palette of terracotta and cream.
4
5Composition: A dramatic close-up of a scored sourdough loaf occupying the lower two thirds, type stacked in the upper third.
6
7Lighting: Warm side light raking across the crust to bring out the scoring and the crumb texture.
8
9Text, exactly as written:
10- Headline: "GRAND OPENING" in bold condensed sans-serif.
11- Subhead: "Saturday, September 12 · 8 a.m." in a lighter weight below.
12- Footer: "412 Grand Street, Brooklyn" in small caps.
13
14Constraints: All three text lines must be fully legible and correctly spelled. Keep the type clear of the loaf's edges. Textured paper background.

Nano Banana Pro prompting tips that raise your hit rate

Five habits do most of the work when you want the best results from a Nano Banana Pro prompt: name real photographic gear and light shapes, write text verbatim, protect what you want kept, change one variable per iteration, and draft cheap before you finish expensive.

Borrow the vocabulary of photography. "85mm, shallow depth of field, softbox at 45 degrees" carries far more information than "beautiful bokeh". Describe the scene the way a director of photography would brief a crew.

Write text verbatim, in quotes. Spell out every word you want rendered and say where it goes. Vague instructions like "add a catchy headline" invite the model to invent one — and invented text is where garbling usually creeps in.

Say what must not change. On any edit, name the elements to preserve before you name the change. "Keep the face, pose and lighting direction exactly as they are; replace only the background" gives the model a boundary it can respect.

Change one thing at a time. When an image comes back almost right, resist rewriting the prompt. Adjust a single slot — the lighting, or the angle — and rerun. Rewriting everything means you never learn which sentence was doing the work.

Draft cheap, finish expensive. Iterate at 1K until the prompt is right, then run that prompt at 4K. The prompt is what carries over — a 4K run is a fresh generation, so if you want that 1K composition back, select it as a reference and add a preservation clause before you raise the quality.

Common Nano Banana Pro prompt mistakes

Most bad generations trace back to four habits — adjective-stacking, silent aspect ratios, unquoted text and unbounded edits — plus a few lesser mistakes below. Each one hands a decision back to the model.

• Stacking quality words. "Stunning, beautiful, ultra-detailed, masterpiece" describes no light, no lens and no framing. Replace the pile with one concrete setup.
• Ignoring the aspect ratio. A full-body fashion shot cropped to 1:1 fights itself. Pick the ratio for the destination first, then compose the prompt to suit it.
• Asking for text without writing it. If you do not supply the exact string, the model supplies one.
• Editing without a preservation clause. "Change the background" without "keep the subject unchanged" invites the model to redraw the person too.
• Keeping the ratio loose on an edit. When you modify an existing image, hold its original aspect ratio, or the crop will surprise you.
• Changing five things per iteration. You lose the ability to attribute the improvement, and usually the good parts along with it.
• Drafting at maximum quality. Composition problems are visible at low resolution. Save the high-resolution run for the prompt you have already approved.

Quick-reference checklist

Copy the template below, fill the five slots, and add your settings before you generate. It is the same skeleton every prompt in this guide uses.

1Subject: [who or what is in the frame, in plain words]
2
3Style: [type of image - photorealistic, editorial, flat vector, riso print...]
4
5Composition: [angle, crop, where the subject sits, negative space]
6
7Lighting: [source, direction, quality, shadows]
8
9Text, exactly as written: [every word to render, in quotes, with placement]
10
11Constraints: [what must not change, background, hard requirements, what to exclude]

Before you hit generate, check:

• Does the prompt name a subject, not just a mood?
• Is the style a specific genre rather than an adjective?
• Have you set the angle and the crop?
• Have you described the light source and its direction?
• If the image contains text, is every word written out verbatim?
• Have you said what must stay untouched?
• Does the aspect ratio match where the image will be published?
• Are you drafting at 1K and saving 4K for the final run?

Where to run these prompts

You can run every prompt on this page in your browser: free account, no API key.

1. Sign in with a Google or Microsoft account.
2. Open the image generator.
3. Paste the prompt.
4. Pick the model (Nano Banana Pro), an aspect ratio and a quality.
5. Generate.

Iterate at low quality — a draft run fits the credits your free plan includes — then spend on the run worth keeping: a 4K Nano Banana Pro render draws from your credit wallet, which you top up as you go. The guide to running Nano Banana Pro prompts free covers the available models and what each image costs in credits.

Frequently asked questions

How do I write a prompt for Nano Banana Pro?

Fill five slots: subject, style, composition, lighting and constraints. Write the subject in plain words first, then add the other four one at a time. Choose the aspect ratio for wherever the image will be published, and write any text you want rendered exactly as it should appear.

How long should a Nano Banana Pro prompt be?

As long as it needs to be. There is no length penalty, and detailed prompts consistently outperform short generic ones. A structured paragraph covering all five slots is a reasonable target; a single sentence rarely is.

Do prompts have to be written in English?

No. Nano Banana Pro handles prompts in Spanish, French and other languages with comparable results. Specificity matters far more than the language you choose.

How do I stop the model changing someone's face?

Put the preservation instruction first. Open with a line such as "Preserve the subject's facial identity from the attached image exactly" before you describe anything you want changed, and attach the reference photo.

Why does the text in my images come out wrong?

Usually because the prompt did not spell it out. Write each string verbatim in quotes, say where it sits in the layout, and add a constraint that every line must be legible and correctly spelled.

What should I do when an image is almost right?

Change one slot and rerun. Adjust the lighting, or the angle, or a single constraint — never all of them at once — so you can tell which edit produced the improvement.