Free Nano Banana Pro Prompt Generator (4K Presets)

TL;DR: A Nano Banana Pro prompt generator is a fill-in-the-blanks brief plus the two API fields that decide framing and resolution. Fill the six slots below, set aspect_ratio and image_size inside response_format, and send it to gemini-3-pro-image. Eight copy-paste 4K presets follow.

Everything numeric on this page was checked against Google's own documentation on August 26, 2026, and each section names the page it came from. If you want the mechanics behind the parameters rather than the fill-in-the-blanks version, our full guide to prompting Nano Banana Pro covers them. This page is the tool.

How do you generate a Nano Banana Pro prompt?

Fill six slots, then set two fields. That is the whole generator.

The reason it works is that Gemini 3 image models parse grammar. They are not tag-matching engines, so a comma-separated pile of adjectives gets averaged into something competent and forgettable. Google's own prompting guidance on the image generation docs says to be "hyper-specific", to "provide context and intent", and to control framing with photographic language rather than parameters. The six slots below are those instructions turned into a form.

NANO BANANA PRO PROMPT GENERATOR: fill every slot, delete the labels

SUBJECT:  [what it is, plus two physical details a stranger could not guess]
SETTING:  [where it sits, what is behind it, what is under it]
LIGHT:    [source, direction, quality, and what it does to the surface]
CAMERA:   [shot type + angle + lens, in photographic words]
TEXT:     ["the exact string", in [font style], placed [where]]
INTENT:   [what this image is FOR: a hero banner, a menu, a thumbnail]

NEGATIVE: [describe the empty state you want; never name the thing you don't]

Delete the labels when you send it. They are scaffolding for you, not syntax for the model.

Here is the same template filled in, so you can see the density Google's examples expect:

A high-resolution studio product photograph of a matte black ceramic pour-over
carafe with a visible hand-thrown ridge and a raw clay foot, standing on polished
grey concrete against an unbroken charcoal paper sweep. Lit by a single large softbox
from camera left at 45 degrees, producing a long soft highlight down the shoulder
of the carafe and one soft shadow to the right. Shot slightly above eye level with
an 85mm lens, shallow depth of field, sharp focus on the pouring lip. The label
reads "SLOW BREW" in a thin uppercase serif, letter-spaced, centred on the lower
third of the body. This is a hero image for an e-commerce product page. The
concrete surface is clean and unmarked, with no props and no reflections of
studio equipment.

Two of those six slots do most of the work, and they are the two people skip.

LIGHT is the one that decides whether the image looks bought or generated. "Studio lighting" is a category, not an instruction. "A single large softbox from camera left at 45 degrees, producing a long soft highlight down the shoulder and one soft shadow to the right" tells the model where every value in the frame comes from. The difference between those two phrasings is larger than the difference between 2K and 4K.

INTENT is the one that costs nothing and changes everything. Google's guidance says explicitly that explaining the purpose of an image influences the output, and gives "Create a logo for a high-end, minimalist skincare brand" against a bare "Create a logo" as its example. Telling the model an image is a hero banner buys you negative space at the edges without asking for it. Telling it the image is a thumbnail buys you a subject large enough to survive being 320px wide.

Notice the last sentence of the filled example. There is no --no parameter and no negativePrompt field on these models. Google's guidance is to use "semantic negative prompts", which means describing the intended state positively. "An empty, deserted street with no signs of traffic" is their example. "No cars" is not.

Then the settings. Every preset on this page uses the same shape:

{
  "model": "gemini-3-pro-image",
  "input": "<your filled-in brief goes here>",
  "response_format": {
    "type": "image",
    "mime_type": "image/jpeg",
    "aspect_ratio": "3:2",
    "image_size": "4K"
  }
}

That is the interactions endpoint shape from Google's image generation docs, which POSTs to https://generativelanguage.googleapis.com/v1beta/interactions. In the Python SDK the same call is client.interactions.create(model=..., input=..., response_format={...}).

If you would rather have a model do the filling in, this is the meta-prompt. Paste it into any chat model, answer its questions, and it hands back a finished brief:

You are a Nano Banana Pro prompt generator. I will describe an image in one
casual sentence. Do not generate anything yet.

First, ask me only for the slots I left empty out of: subject detail, setting,
light source and direction, camera angle and lens, exact text to render, and
what the image is for. Ask them as one numbered list, maximum six questions.

Then output ONE paragraph of 60-110 words in full sentences, in this order:
shot type and subject, setting, lighting, camera and lens, any text in quotes
with its font style and placement, and a final sentence describing the clean
intended state instead of listing things to exclude.

Then output a JSON settings block with model, aspect_ratio and image_size,
choosing the ratio from: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9.
Use uppercase K in image_size. Explain your ratio choice in one line.

Never use the words: 8k, masterpiece, award-winning, ultra-detailed, trending.

That last line matters more than it looks. Quality boosters were an Imagen and Midjourney habit. On a model that reads sentences, they consume attention and return nothing.

What resolution and aspect ratio should you set?

Ten aspect ratios, three sizes, and a full 4K column across every ratio. Google publishes exact pixel dimensions for each combination, which almost no other image model does.

Source: the "3.1 Pro Image" resolution table on Google's Gemini API image generation docs, accessed August 26, 2026.

Two things follow from that table that change how you use the generator.

"4K" is not one number. A lot of coverage in circulation says 4K on this model means 4096x4096. That is only the 1:1 row. At 21:9 the long edge runs to 6336px, and at 16:9 you get 5504x3072. If you are budgeting storage or building a print workflow, size from the row you are actually going to request.

On Pro, 2K is free relative to 1K. Google's token table charges 1,120 output tokens for both the 1K and the 2K row on gemini-3-pro-image, and 2,000 for 4K. There is no billing reason to request 1K on this model unless you specifically want a smaller file. On gemini-3.1-flash-image the same step up costs more, at 1,120 for 1K and 1,680 for 2K.

Choosing the ratio is easier if you pick it by destination rather than by taste. 16:9 and 21:9 are site headers and slide backgrounds, and the ultra-wide row is the one worth knowing about because most models simply do not offer it. 4:5 and 9:16 are the social ratios, and 4:5 in particular is the one that survives being cropped to square without losing the subject's head. 3:2 matches a stills camera and gives you room to crop in two directions afterwards, which makes it the safest default for product work. 1:1 is right for marketplace tiles and app icons and almost nothing else, which is awkward, because it is what you get when you set nothing at all.

Defaults are the other quiet trap. Google says the model matches the output size to your input image where one exists, and otherwise generates a 1:1 square at 1K. Omitting response_format is not neutral. It is a choice, and the choice is a small square.

What are the ready-made 4K presets?

Eight starting points. Each one is a filled brief plus the settings that suit it. Swap the bracketed parts and send.

Read them as a starting shape rather than a finished answer. The parts worth editing first are the light description and the lens, because those are what separate two images of the same object. The parts worth leaving alone are the closing sentences that describe a clean intended state, since those are doing the job a negative prompt used to do and they are easy to delete by accident. Where a preset sets 2K instead of 4K, that is deliberate: reference-driven and editing work rarely benefits from the extra pixels, and on Pro the 2K row costs the same as 1K anyway.

1. Product shot on a paper sweep, 4K. The workhorse. 3:2 gives you crop room for both a wide banner and a square marketplace tile.

{
  "model": "gemini-3-pro-image",
  "input": "A high-resolution studio product photograph of [PRODUCT, with two specific material details] on [SURFACE] against an unbroken [COLOUR] paper sweep. Lit by a single large softbox from camera left at 45 degrees, with one soft shadow falling right and a long specular highlight down the [PART]. Shot slightly above eye level with an 85mm lens, shallow depth of field, sharp focus on [FOCAL DETAIL]. This is a hero image for an e-commerce product page. The surface is clean and unmarked, with no props and no reflections of studio equipment.",
  "response_format": { "type": "image", "mime_type": "image/jpeg", "aspect_ratio": "3:2", "image_size": "4K" }
}

2. Poster with a headline, 4K. Text rendering is the reason to reach for Pro rather than a cheaper model. Quote the string exactly and describe the type in words.

{
  "model": "gemini-3-pro-image",
  "input": "A [STYLE] event poster for [EVENT]. The headline reads \"[EXACT HEADLINE]\" in a [WEIGHT] [SERIF/SANS-SERIF] typeface, set across the upper third with generous letter spacing. Beneath it, smaller text reads \"[DATE AND VENUE]\" in the same family at roughly one third the size. The composition is [LAYOUT DESCRIPTION], on a [COLOUR] ground with [ACCENT COLOUR] accents. Flat print aesthetic, no photographic texture, generous margins on all four sides.",
  "response_format": { "type": "image", "mime_type": "image/png", "aspect_ratio": "4:5", "image_size": "4K" }
}

3. Infographic or diagram, 4K. Google names infographics, menus and diagrams as the target use case for the Gemini 3 text rendering work. Give it the label strings, not a description of labels.

{
  "model": "gemini-3-pro-image",
  "input": "A clean instructional diagram explaining [PROCESS], drawn as [NUMBER] labelled stages flowing left to right, connected by thin arrows. The stage labels read, in order: \"[LABEL 1]\", \"[LABEL 2]\", \"[LABEL 3]\". Each stage has a simple line icon above its label. Flat vector style on a light [COLOUR] ground, one accent colour only, thin consistent stroke weights, a clear title reading \"[TITLE]\" across the top. All type is a geometric sans-serif. Ample white space between stages.",
  "response_format": { "type": "image", "mime_type": "image/png", "aspect_ratio": "16:9", "image_size": "4K" }
}

4. Ultra-wide hero banner, 4K. This is the 6336x2688 row, and it is the reason to care that the 4K column covers every ratio. One call produces a site header no upscaler has to touch.

{
  "model": "gemini-3-pro-image",
  "input": "An ultra-wide cinematic establishing shot of [SCENE]. [TIME OF DAY] light rakes across the frame from the [DIRECTION], throwing long shadows toward the camera and catching [ATMOSPHERIC ELEMENT] in the air. Shot on a 24mm lens from [HEIGHT/ANGLE], deep focus, horizon on the lower third. Muted [PALETTE] colour grade with lifted blacks. Composed with the main subject at the right third and clean negative space across the left half for overlaid text. No lettering rendered in the image itself.",
  "response_format": { "type": "image", "mime_type": "image/jpeg", "aspect_ratio": "21:9", "image_size": "4K" }
}

5. Character consistency across a scene, 2K. Attach up to five character reference images. Pro handles resemblance across them; it does not take style references, which is a real limit covered further down.

{
  "model": "gemini-3-pro-image",
  "input": "Using the attached reference images of [PERSON/PEOPLE], generate a single photograph of them [DOING WHAT] in [SETTING]. Keep facial features, hair and build faithful to the references. [LIGHTING]. Shot at [ANGLE] with a [LENS], [DEPTH OF FIELD]. Everyone is [WEARING/HOLDING WHAT]. Natural expressions, no stiff posing.",
  "response_format": { "type": "image", "mime_type": "image/jpeg", "aspect_ratio": "3:4", "image_size": "2K" }
}

6. Packaging mockup with a legible label, 4K. Two-turn workflow. Google's limitations section says the model works best if you generate the text first and then ask for an image containing it. Ask for the copy, approve it, then ask for the box.

TURN 1: Write the front-of-pack copy for [PRODUCT]. Give me exactly a brand
name, a one-line descriptor under 6 words, and a net weight line. Nothing else.

TURN 2: Generate a studio photograph of a [MATERIAL] [PACKAGE TYPE] standing
three-quarters to camera on [SURFACE]. The front panel carries the copy you
just wrote, set in [FONT STYLE], with the brand name largest and centred. Soft
top-left key light with a subtle fill from the right. 100mm lens, product
photography, sharp on the front panel, background falling gently out of focus.

7. Targeted edit that preserves everything else. Note the missing size fields. When you pass an input image and set nothing, Google says the output matches the input dimensions, which is usually what you want on an edit.

{
  "model": "gemini-3-pro-image",
  "input": "Using the provided image, change only [SPECIFIC ELEMENT] to [NEW ELEMENT]. Keep everything else in the image exactly the same, preserving the original style, lighting, composition and grain.",
  "response_format": { "type": "image", "mime_type": "image/png" }
}

8. Grounded chart or map, 2K. Pro can use Google Search as a tool before it draws, which is how you get an image reflecting something that happened this week. Google's pricing page lists 5,000 free grounding requests a month shared across Gemini 3.x models, then $14 per 1,000.

Search for [CURRENT FACT OR DATA]. Then generate a clean editorial chart
visualising it, with the real figures rendered as labels. Title reads
"[TITLE]". Flat vector style, one accent colour, geometric sans-serif type,
light ground, a small source line at the bottom naming the data source.

If you keep more than a handful of these, they stop being prompts and start being infrastructure. That is the point at which the bracketed slots want to be real variables rather than find-and-replace, which is what a personal prompt library is for.

How do you convert an old Imagen prompt into a Nano Banana prompt?

Delete the quality boosters, turn the descriptor stack into sentences, and remap four config fields. Imagen's generate models were shut down on August 17, 2026, with gemini-3.1-flash-image listed as the recommended replacement on Google's deprecations page, so this is not optional work if you had a pipeline.

The field map, taken from Google's Imagen configuration list and the current image generation docs:

Sources: Google's Imagen documentation and deprecations table, both accessed August 26, 2026.

Two notes on that table, because published sources disagree. Google's own Imagen page does not document a negativePrompt or an addWatermark field on the Imagen 4 generate models, so if your code sets either one, check where it came from before you assume there is something to migrate. And the current Imagen size field is imageSize, not sampleImageSize, which is the name that circulates in older migration write-ups.

Here is the converter itself. It does the first pass; you edit the result.

You are an Imagen-to-Nano-Banana prompt converter.

I will paste an old Imagen prompt, usually a comma-separated list of keywords.

Rewrite it as ONE paragraph of 60-110 words in full sentences, in this order:
shot type and subject, setting, lighting with direction and quality, camera
angle and lens, any text in quotes with typeface and placement, and a closing
sentence describing the clean intended state.

Rules:
- Delete every quality booster (8k, masterpiece, award-winning, ultra-detailed,
  hyperrealistic, trending on artstation). Do not replace them with synonyms.
- Turn every bare noun into a noun with one specific physical property.
- Convert any negative prompt into a positive description of the empty state.
- If the original named an aspect ratio or size, map it to aspect_ratio and
  image_size and put them in a JSON settings block with uppercase K.
- If the original set numberOfImages above 1, tell me it has no equivalent and
  that I need to loop the call.

Output: (1) the rewritten paragraph, (2) the JSON settings block, (3) a two-line
note listing what you deleted and why.

Here is the prompt:
[PASTE]

If you want the wider migration story rather than the prompt half of it, our post on the Imagen retirement covers response parsing, cost re-baselining and the order to do things in. The same descriptor-stack habit is worth unlearning in the other direction too, which is the failure mode behind most of the complaints in why Midjourney prompts don't work.

Which Nano Banana model should the generator target?

Pro when the image has to carry text, reasoning or real-world facts. The flash model for almost everything else. Google itself calls gemini-3.1-flash-image the "go-to" and positions gemini-3-pro-image for "professional asset production and complex instructions".

Three things in that table cost people money.

4K is not a Pro-exclusive feature. Both Pro and gemini-3.1-flash-image publish full 4K columns. The only model capped at 1K is the Lite one. If 4K is your entire reason for choosing Pro, you are paying $0.24 for something available at $0.151.

Pro takes fewer object references, not more. Six, against ten on the flash model. Pro's advantage is character consistency at five faces and its reasoning pass, not raw reference count.

Pro does not take style references at all. Google's table reads N/A for that row. If your workflow is "here is a mood board, match it", the documented path is gemini-3.1-flash-image with up to three style images. This one surprises people who assumed Pro was a superset.

What will this generator not fix?

Four things, and it is worth knowing them before you build a workflow on top.

Spelling is not guaranteed. Text rendering on these models is genuinely good and it is not deterministic. Google's own limitations section recommends generating the text in one turn and then asking for the image in a second turn, which is a workaround, not a fix. Proofread the pixels before a client sees them.

You cannot ask for four options. Google's limitations note that the model "won't always follow the exact number of image outputs that the user explicitly asks for". Where Imagen returned four candidates per call, you now loop. Budget accordingly, because the per-image price is the per-call price.

Thinking is always on. Google states that the reasoning pass is enabled by default and cannot be disabled in the API for Gemini 3 image models. The model produces up to two interim thought images to test composition, and the docs say those are not charged. You cannot trade that reasoning away for latency. Changing model is the only lever.

Every output is watermarked. Google states that all generated images include a SynthID watermark. There is no prompt, no setting and no plan tier that removes it.

And the honest one about us: Prompt Architects does not generate images. We are a prompt layer. We turn a rough idea into a structured brief, store it, version it, and fill its variables, and then you send that brief to Google, Midjourney, or whoever else. If you want the image itself, you still need an account with the model provider. What we save you is the twenty minutes of rewriting the same product-shot brief for the eleventh SKU, and the slow drift where three people on a team each keep a slightly different version in a Notes app.

The image prompt tools sit in the web app and in the Chrome extension, which works directly on Gemini among the platforms it supports, so the generator above runs where you are already typing. Global Variables handle the bracketed slots. The Prompt Library holds the eight presets. The free plan includes 5 prompt enhancements per day, forever, which is enough to test whether the workflow suits you before any money changes hands. Current pricing, including the launch discount, is on the pricing page.

If you would rather start from an existing image than a blank template, the reverse of this workflow is covered in how to reverse-engineer any AI image into a reusable prompt.