Nano Banana 21
Image generation and reference editing in your browser
Open the image workspace · Explore prompt ideas · Check current pricing
Nano Banana 21 is a browser-based workspace for generating images from written ideas and editing reference images. Its shared interface offers model-specific settings, credit estimates and saved creations. Decide whether you need a new composition or a change to an existing picture, then choose the appropriate input route. Visit Nano Banana 21.
This repository documents the independently operated website and includes original prompts. It provides no model weights, training code or Hugging Face inference endpoint. Nano Banana 21 is separate from Google and other underlying model developers. This is not an official model release; the site's Terms of Service describe its use of third-party AI providers.
What you can explore
Explore campaign concepts, product scenes, illustrations, social graphics and reference-led edits. Give the first project one clear goal: replace a backdrop, test a composition or explore a poster layout. Review an appealing draft against production requirements before using it.
The public site offers a prompt gallery and links to My Creations. Its documented workflow treats each submission as a separate task; further edits can use a completed image as a new reference. Downloads are available from results or saved creations. See the website workflow.
Choose a model for the input
These are the website's product labels, not independently verified upstream model identities. Check live controls before submitting because supported inputs and settings can change.
- Nano Banana 2 Lite: the page's input guidance specifies text-to-image creation. The inspected settings included square, portrait and landscape aspect ratios, with no resolution selector. Start here for a written scene brief; choose an editing model for uploaded-reference changes.
- Nano Banana Pro: its dedicated page requires a reference image and presents an image-to-image workflow with aspect-ratio controls and 1K, 2K and 4K resolution choices. This is a relevant starting point for exploring changes to a supplied subject or product.
- Qwen Image 2.1: the page describes text creation, image editing and multiple-reference compositions. Visible controls include 1K/2K resolution and opaque/transparent backgrounds. Upload guidance allows up to 10 PNG, JPEG or WebP images, with a 20 MB and 25 MP limit per image.
- GPT Image 2.5: the page presents text-to-image and photo-editing uses. Its inspected settings expose Flare/Sunburst choices, aspect ratios, low/medium/high quality and automatic, transparent or opaque backgrounds. Treat these as interface labels when recording a task.
Choose by required input and output controls rather than assuming one option is universally best. This guide does not rank image quality or claim a measured speed advantage.
Quickstart
- Open the creation workspace. Review the account requirements, available credits and current access conditions before beginning.
- Choose the input route. For a new image, prepare a text brief. For an edit, select an image-capable model and use only reference material you have permission to upload.
- Write the main request. Describe the subject, setting, composition and lighting. For an edit, distinguish the requested change from details that must remain recognizable.
- Inspect the controls. Pick the intended aspect ratio and any available quality, resolution or background setting. A setting on one model should not be assumed to exist on another.
- Review the displayed credit quote before pressing Generate. Keep a copy of the prompt and settings so comparisons remain understandable.
- Examine the result at full size. Download a useful version, then submit a focused revision if necessary. The documented workflow explains reusing an output as an input for another supported editing task.
For an efficient first session, choose one brief and one target format. Decide what would make the image acceptable before spending credits on variations. Three attractive pictures are less useful than one that leaves enough room for the actual headline or preserves the right product silhouette.
Write a clear visual brief
A reusable prompt structure is: subject, requested action, environment, composition, light, materials, wording and preservation instructions. Not every prompt needs every part. Specific relationships often help more than long lists of adjectives: place the object on the right, leave the left third empty, or retain the original camera angle.
For text in an image, quote the exact words and state where they belong. Keep the first attempt short enough to proofread. For several references, explain each image's role and avoid mixing competing instructions about identity, shape and style.
Original prompt examples
These prompts are original, untested starting points. They have not been run on Nano Banana 21 and are not demonstrated outputs, performance tests or quality guarantees. Replace the subjects and requirements with your own brief.
Product concept from text
Suggested route: a text-to-image option such as Lite.
Create a studio concept photograph of an unbranded matte ceramic water bottle on a pale limestone plinth. Place the bottle in the right third of a wide composition. Use soft daylight from the upper left and a restrained sand and sage palette. Leave the left third clear for a headline added later. Show a believable contact shadow and subtle surface texture. No logo, lettering or extra objects.
Review the bottle outline, shadow direction and usable empty space before adding campaign copy.
Background change with a reference
Suggested route: Pro or another reference-editing option.
Use the uploaded chair photograph as the product reference. Keep the chair's frame, woven seat, proportions and viewing angle recognizable. Replace the surrounding room with a quiet gallery featuring warm white walls and a light oak floor. Match the light direction to the chair and add a grounded shadow. Do not introduce decorative objects or redesign the chair.
Compare against the original at the same zoom level, especially thin legs, woven edges and floor contact.
Poster with a short headline
Suggested route: Qwen Image 2.1 or GPT Image 2.5.
Design a vertical poster for a fictional community ceramics exhibition. Use a close-up photograph-style composition of three handmade bowls with warm side lighting. Put the exact headline "FORM AND FIRE" in the upper quarter, in clean uppercase lettering. Use muted clay red, charcoal and cream. Leave the lower quarter uncluttered for details added later. No additional words, logos or invented dates.
Check every letter and reserve final event details for a controlled layout pass.
Transparent illustration asset
Suggested route: Qwen with its transparent-background setting, or another option exposing that control.
Create a small original illustration of a cheerful terracotta watering can with two green leaves emerging from its spout. Use rounded shapes, simple shading and a clean silhouette. Center the complete object with generous clear margins. No words, frame, floor or surrounding scene. The illustration should read clearly at a small size on both a light and a dark background.
Select transparency in the interface as well as describing the asset. Inspect the downloaded file on contrasting backgrounds for edge halos and unwanted opaque areas.
Practical editing workflows
Refresh a product setting
Start with a clear product photograph and one environmental change. Preserve shape, material, distinguishing details and camera angle in the prompt. First review whether the product remains accurate; only then evaluate styling. If a label changes, treat that version as a concept rather than a factual product listing image.
Build a reference combination
For the Qwen multiple-reference workflow, organize images in a deliberate order and assign each a role. A useful brief might designate image one as the object, image two as the background inspiration and image three as the color palette. Avoid asking every reference to control the entire result. Inspect whether an unwanted feature migrated between references.
Develop a consistent series
Keep a small record containing the source image, prompt, model label, settings and what changed. Reuse an approved reference when the selected editing model supports it. Change one dimension per iteration, such as background or lighting, and compare identity or object details across the series. Consistency remains something to review, even when the instructions request preservation.
Review an image before using it
Use this practical checklist for your own outputs. It is an evaluation aid, not a benchmark result.
- Brief fit: Does the picture communicate the intended subject, action and mood?
- Reference fidelity: Are recognizable proportions, faces, materials and important details retained?
- Local accuracy: Look closely at hands, edges, reflections, shadows, repeated patterns and object joins.
- Typography: Proofread spelling, punctuation, numbers and placement at actual delivery size.
- Layout readiness: Confirm crop, safe margins and sufficient space for captions or interface overlays.
- File suitability: Check the downloaded dimensions and format. For transparency, inspect actual alpha behavior rather than relying on a preview.
- Rights and context: Confirm permission for references, likenesses and brand elements, and disclose synthetic imagery where appropriate.
Record why a version passed or failed. A short note such as "good layout, incorrect cap shape" is more actionable than a general quality score.
Access and credits
The service uses credits, with a quote shown before generation. Its pricing page describes subscription allowances, one-time credit packs and plan-specific concurrent-task limits. Advertised image counts depend on the model and settings; they are not a combined allowance for every listed model. Review live prices, signup offers, purchase availability and billing terms directly on the site.
Budget for revisions as well as initial attempts. Compare the cost of reaching an acceptable image, not just the lowest displayed starting quote. Do not assume that a free signup promotion provides unlimited generation or every paid feature.
For commercial work, review the current Terms of Service, Privacy Policy and Acceptable Use Policy. Service permissions do not remove your responsibility for third-party rights. Avoid uploading confidential client material until its permitted handling is clear.
Frequently asked questions
Can I run this repository locally?
This repository contains documentation and prompts. Use the linked web application for image creation; there is no local inference package here.
Can Lite edit my uploaded photograph?
Its public input guidance specifies text generation. Select a reference-editing option and check its current upload requirements instead. Lite input guidance.
Is there an API or SDK in this repository?
No. This card documents the browser workflow and does not provide API endpoints, integration code or package-installation instructions.
Will an edit preserve everything outside my request?
Preservation instructions are useful, but inspect the complete result. Unrequested details can change, particularly across repeated edits.
Are the examples tested?
No. The prompts above are original suggestions for readers to adapt. No generated examples, benchmark scores or timing measurements are reported here.
Documentation license
The MIT license declared in this repository applies only to the original README text and original prompt examples. It does not license the Nano Banana 21 service, third-party models, model weights, trademarks, website images, user inputs or generated outputs. Those remain subject to their respective rights and applicable terms.
Public pages and interface controls reviewed on October 7, 2026. Availability and settings may change.