Turn a written description into an image with the three Nano Banana tiers this site runs: Nano Banana, Nano Banana 2 and Nano Banana Pro. You pick the tier yourself, choose how large the output should be, set the frame, and attach reference images when a series has to stay consistent.
Three tiers, all from Google · 1K, 2K and 4K on two of them · Up to 14 reference images
Every number below is read from this site’s own model configuration, not from a marketing page. Google registers a fourth tier, Nano Banana 2 Lite, which this workbench does not run, and one officially published output size that this workbench does not expose. Both gaps are named rather than folded into the list.
The base tier. Output is fixed at 2K with no resolution selector. Eight aspect ratios are available, covering 1:1, 2:3, 3:2, 3:4, 4:3, 16:9 and 9:16 alongside the automatic frame. Reference images are accepted up to a limit of 5 per prompt.
Reach for it when: you are testing a prompt and want the cheapest way to see whether the idea works.
Three output sizes, 1K, 2K and 4K, with 1K as the default. Fifteen aspect ratios, the widest set on the site, including the tall and wide frames 1:4, 4:1, 1:8 and 8:1. Reference images are accepted up to 14 per prompt.
Reach for it when: you need 4K, an extreme frame, or more reference images than the base tier accepts.
Three output sizes, 1K, 2K and 4K, with 1K as the default. Eleven aspect ratios, including the cinema frame 21:9, 4:5 and 5:4. Reference images are accepted up to 14 per prompt.
Reach for it when: the composition has to hold together while small details inside it stay readable.
This page is the entry point to image generation on this site. You write a description, pick one of the three Nano Banana tiers registered here, decide how large the output should be and how it should be framed, and generate. Nothing is routed on your behalf. The tier you select is the tier that runs, which matters because the tiers do not accept the same settings.
Nano Banana is Google’s image generation and editing family, and all three tiers here come from that one provider. They differ in the controls this site exposes rather than in the interface you use: the prompt field, the aspect ratio selector and the reference image area are shared, while the resolution options, the number of available frames and the reference image limit change from tier to tier. The comparison above lists those differences one tier at a time.
Every tier accepts a prompt written in ordinary language, and reference images stay optional. Aspect ratio is a first-class control rather than something applied after the fact, so the frame you choose is the frame that is generated. Two of the three tiers let you choose between 1K, 2K and 4K output; the base tier produces 2K.
Generation is billed in credits rather than being capped by a plan tier, and the amount comes from the live tool catalog rather than from a number hard-coded on this page. Current figures belong on the pricing page, because they change and this landing page does not.
The rest of this page is deliberately specific about the line between what Google publishes about these models and what this workbench actually exposes. Where the official line-up is wider than the site, the gap is stated. Capabilities that the official API accepts but this workbench does not offer, such as document input, are described as official behaviour only and are never presented as features of this site.
The five frameworks and four rules below are Google’s published guidance for the Nano Banana image models. They are reproduced here because they apply to every tier in this workbench, and because one of them contradicts a habit that third-party tutorials repeat constantly.
Google lists starting with the main action as a best practice, which means a strong verb describing what happens in the frame rather than a stack of adjectives. Staging detail written first tends to be read as the subject itself, and the subject is the one thing a prompt cannot afford to lose.
The published order is subject, then action, then location or context, then composition, then style, written as [Subject] + [Action] + [Location/context] + [Composition] + [Style]. Keeping style last stops it from leaking into the description of the subject.
The order becomes [Reference images] + [Relationship instruction] + [New scenario]. State what each uploaded image contributes, state how the references relate to one another, and only then describe the new scene you want built from them.
Generation and editing are different jobs, and an edit prompt says what changes and just as explicitly which parts must stay exactly as they are. Text alone is enough to mark the region to protect, and one change per pass is what makes two versions comparable.
The formula is [Source or search request] + [Analytical task] + [Visual translation]. Ask for the fact, work out the answer as text, and only afterwards describe how that fact should look as a picture. Google records a knowledge cutoff of January 2025 for these models and reaches anything more recent through a live web search.
Put the exact characters in quotation marks, name a font style, state the target language, and describe the text before describing the scene around it. Google also documents a text-first route in which you settle the wording in conversation before asking for the image.
Four lever groups carry most of the art direction: how the scene is lit, the camera and lens with its focus behaviour, the colour grade and film stock, and the material texture of the objects in frame. Reuse the same vocabulary each time you iterate, so a difference in the output is attributable to one change rather than to reworded styling.
Google’s fourth rule is positive framing, and its own example swaps an exclusion for a description: write an empty street rather than no cars. This is the rule that third-party prompt lists most often get backwards, and on this site it is also the only option, because the image workbench has no field dedicated to exclusions.
Every statement above is checkable against either Google’s published material for these models or this site’s own configuration. Where the two disagree, the disagreement is written down rather than smoothed over.
Read the service termsText to image is the workflow that starts from nothing but a written description. You supply words and the model returns an image. There is no source photo to preserve, which is exactly what separates it from image to image, where the starting point is a file you upload and the job is to change it without losing its structure.
The Create workbench is one surface for both jobs. The tier selector, the prompt field, the aspect ratio control, the resolution control and the reference image area all sit in the same place, and moving from generating a new image to editing an existing one changes which uploads are expected rather than which page you are on.
Reference images sit between the two modes. Attaching an image to a text prompt does not turn the task into an edit; it narrows the generation. The prompt still describes the scene, while the reference supplies identity, palette, texture or layout. Because the limit differs by tier, the same set of references can be enough on one tier and too many on another.
The result is an ordinary image file that you download and keep. What you may do with it afterwards is governed by the service terms rather than by anything written on this landing page.
Three steps from a written description to a file on your device.
Name the subject and what it is doing, then add setting, composition and style. Google’s guidance is to describe what you want in the frame rather than what you want removed, so write the state you are aiming for instead of a list of exclusions. This workbench has no separate field for exclusions, so the prompt is the only place a constraint can live.
Choose Nano Banana, Nano Banana 2 or Nano Banana Pro, set the aspect ratio, and choose 1K, 2K or 4K where the tier offers a choice. Attach reference images if the result has to match something you already have, keeping inside the limit for that tier.
Run the prompt, then adjust a single variable before running it again: the wording, the frame, the tier or the output size. Holding everything else fixed is what makes the difference between two results readable, and it is also how the official guidance describes iteration. Download the results that work.
These are the controls this workbench actually exposes. Google’s own specifications are wider in two places, and both are named in the list rather than quietly dropped.
There is no automatic routing between Nano Banana tiers. You select the tier before generating, and the same prompt can be run on a different tier simply by changing the selection, which puts the same description in front of two tiers and makes the difference visible.
Nano Banana 2 and Nano Banana Pro offer 1K, 2K and 4K, with 1K as the default. The base Nano Banana tier produces 2K output and has no size selector. Google also publishes a 512px size for Nano Banana 2 that this workbench does not expose.
Fifteen frames on Nano Banana 2, including 1:4, 4:1, 1:8 and 8:1; eleven on Nano Banana Pro, including 21:9, 4:5 and 5:4; eight on the base tier, including 16:9 and 9:16. Choosing the frame before generating avoids losing composition to a later crop.
Up to 14 reference images per prompt on Nano Banana 2 and Nano Banana Pro, and up to 5 on the base tier. Google’s published figure for a single prompt on these models is 14, with PNG, JPEG, WEBP, HEIC and HEIF accepted as input formats.
There is no negative prompt field, which happens to match Google’s published guidance: describe what you want rather than what you want to avoid. Constraints belong in the sentence describing the scene, which keeps one description of the image instead of two competing ones.
Generation is billed in credits, and the figure is read from the tool catalog at the time of use rather than hard-coded into this page. The pricing page holds the current numbers, including the difference between the base tier and the two higher tiers.
The most common shape of work is a short series rather than a single image. A campaign needs six frames that look like one set, a product page needs the same object against several backgrounds, a set of article headers needs a consistent palette. That is reference image work, and it is why the reference limit per tier matters more in practice than the resolution selector for many people.
The second group works frame-first. Poster layouts, vertical social crops, cinema-width headers and the very tall 1:8 and 1:4 frames on Nano Banana 2 are chosen before the prompt is written, because a composition that has to survive a crop is already a different composition. Generating the intended frame directly is the reason the aspect ratio control sits next to the prompt rather than behind an export step.
The third group is iterating on wording. Prompt testing needs volume more than size, and the base tier is the cheaper way to find out which description produces the picture you had in mind before moving that same prompt to a tier with 4K output and a wider reference limit.
A fourth, smaller group arrives with an image already in hand and a change to make. That work belongs to image to image rather than to text to image, and the same workbench handles it, so nothing has to be rebuilt in a second tool.
Open the Create workbench and turn your next description into an image, at 1K, 2K or 4K.
Go to the image workbench