Last updated: June 2026
Learning how to use Nano Banana for AI images is mainly an exercise in controlled editing. A first prompt can establish the scene, but the dependable workflow begins when you define what is allowed to change, lock everything else, and make one revision at a time. This approach is more valuable than chasing a perfect one-shot prompt because each accepted image becomes a recoverable checkpoint.
Google's 2026 naming deserves care. Nano Banana 2 is Gemini 3.1 Flash Image, while Nano Banana Pro is Gemini 3 Pro Image. Inside ArWriter, the visible Google choice is labeled Gemini 3 Pro and positioned for image editing and speed. That does not mean ArWriter exposes every model toggle, resolution, or regeneration menu found in Google's own products.
The tutorial below uses one neutral sample project: an editorial visual for a public design workshop. It covers text-only creation, a single-reference edit, an optional multi-reference composition, local changes, text correction, drift recovery, and export checks.
To use Nano Banana effectively, choose a starting mode, provide clean references, define locked and editable elements, then make one localized change per turn. Compare every result with the last accepted image. If identity or composition drifts, return to that checkpoint, reduce the edit scope, and restate the exact location and preservation rules.
What Nano Banana means in 2026
Nano Banana is Google's public name for image-generation and editing models in the Gemini ecosystem. Google identifies Nano Banana 2 as Gemini 3.1 Flash Image, announced on February 26, 2026. It identifies Nano Banana Pro as Gemini 3 Pro Image. The names refer to model families; the interface and available controls depend on where you access them.
Google's official materials say Nano Banana 2 supports multiple aspect ratios and resolutions from 512px to 4K in Google products. Google also reports consistency for up to five characters and fidelity for up to 14 objects in a workflow, plus improved text rendering and localization. These are capability statements, not guarantees for every prompt or every third-party interface.
ArWriter Chat provides a browser route labeled Google with Gemini 3 Pro. It supports generated-image follow-up edits and is positioned for editing and speed. ArWriter also offers a separate ChatGPT route, but this tutorial stays on the Google editing path. Google currently determines size automatically in the ArWriter route; the precise ArWriter size selector applies to its ChatGPT route.
Image creation inside ArWriter requires a paid image-enabled account. Trial and free accounts cannot generate images. Check current image-plan eligibility before starting; this article does not compare plan economics.
For the wider brief, model-routing, size, and export system, use the complete Gemini and ChatGPT image workflow after finishing this focused editing exercise.
Choose one of three starting modes
The starting mode determines how much the model must invent and how much it must preserve. Pick it before writing the prompt.
| Starting mode | Inputs | Use it when | Main risk | First quality check |
|---|---|---|---|---|
| Text-only generation | Written brief | No source image exists | Vague composition or unwanted details | Brief adherence and clean geometry |
| Single-reference edit | One consented image plus instruction | The composition is useful but one element must change | Identity or layout drift | Compare locked details with the source |
| Multi-reference composition | Two or more clearly labeled images | Separate references supply subject, setting, or treatment | Attributes bleed between references | Confirm which source contributed each element |
| Generated-image revision | Last accepted generated output | You need a small improvement | Accumulated drift over many turns | Compare against the accepted checkpoint |
| Text correction | Current image plus exact short copy | Lettering is required inside the visual | Misspelling or added text | Character-by-character proofreading |
| Restart from brief | Clean prompt and best reference | The edit history has become unstable | Losing useful prior decisions | Carry only verified constraints forward |
Text-only is the cleanest way to establish a new scene. Single-reference editing is the safest learning exercise because the source provides an obvious baseline. Multi-reference work is more demanding: label each input, state exactly what to take from it, and prohibit cross-contamination. Do not begin with many references merely because an official model can support them.
Prepare references and permissions before prompting
Reference quality affects both control and risk. Use a well-lit, sufficiently large image with a clear subject and minimal occlusion. Crop irrelevant private information. Avoid screenshots containing names, messages, account details, or unreleased work.
If a real person appears, obtain their permission for the upload and intended edit. Confirm that you have rights to every source image. Do not use the workflow to impersonate someone, create deceptive evidence, or place a person in a sensitive context. A technically successful edit can still be unacceptable or unlawful.
Record the source owner, permitted use, consent date, restricted placements, approval owner, and any crop or conversion in your approved project system. Preserve the original file and avoid compressed messaging-app copies, which make identity and edge changes harder to judge.
Build a preservation contract
A preservation contract tells the model four things: where the edit occurs, what changes, what remains locked, and what must not be added. It is more precise than “make the background nicer.”
Use this grammar:
Location: the upper-right section of the background
Editable element: the plain gray wall
Requested change: replace it with a matte deep-blue wall
Locked elements: central sculpture, table, floor shadows, camera angle, crop, lighting direction
Prohibited changes: people, text, logos, windows, furniture, extra objects
Turn it into a direct instruction:
In the upper-right background, change only the plain gray wall to matte deep blue.
Keep the central paper sculpture, table, floor shadows, camera angle, crop, and lighting direction exactly unchanged.
Do not add people, text, logos, windows, furniture, or other objects.
The word “only” is useful, but it is not a guarantee. The reviewer must still compare the new image against the prior checkpoint. If the model modifies a locked object, reject the version rather than trying to correct several new problems in the same branch.
Step-by-step Nano Banana workflow
The following project is a reproducible exercise, not a reported test result. It creates an editorial workshop visual from scratch, accepts a base composition, edits one region, adds a second reference only if needed, tests a short title, and exports a checked image.
Step 1: Write a six-field base brief
Define purpose, subject, action, environment, treatment, and constraints. The sample avoids real people and commercial products so it can be repeated with low-risk material.
Purpose: editorial hero for a public workshop about practical design systems
Subject: a large folded-paper sculpture surrounded by labeled color swatches
Action: subtle paper strips appear to flow toward the sculpture
Environment: bright studio with a long pale table and uncluttered wall
Treatment: realistic editorial still life, soft side light, restrained blue and amber palette
Constraints: landscape composition, clean left-side space, no people, no brands, no readable text, no watermark
Keep the first brief focused on composition. Do not ask for a final title, several aspect variants, and multiple local edits at once.
Step 2: Open the Google editing route
Inside ArWriter Chat, choose the Create image action, then select Google, labeled Gemini 3 Pro. Paste the base brief. In this route, size is currently selected automatically by Google, so describe the intended landscape composition in the prompt without claiming a precise exposed output size.
If you work directly in a Google product instead, follow its current official help for that specific interface. Google documents generated-image creation, uploaded-image editing, multiple uploaded images, and full-size download. Controls in Google's consumer interface must not be assumed to exist inside ArWriter.
Step 3: Review the base at two zoom levels
At fit-to-screen size, check visual hierarchy, left-side crop space, palette, and whether the sculpture reads clearly. At 100%, check paper edges, shadows, repeated swatches, invented letters, unwanted logos, and geometry.
Accept the base only if its purpose, subject, empty layout space, exclusions, geometry, and shadows all pass review. When multiple structural conditions fail, rewrite the base brief or restart instead of spending several edits repairing an unsuitable composition.
Step 4: Name the accepted checkpoint
Download or otherwise preserve the first acceptable image according to the interface's documented workflow. Name it workshop-editorial-base-v01-accepted. Save the exact prompt next to it.
Every later instruction should start from this accepted checkpoint or an explicitly accepted successor. A checkpoint gives you a clean rollback path when a later edit changes too much.
Step 5: Change the background only
Apply the preservation contract:
In the wall area behind the sculpture, change only the pale wall to a subtle warm-gray plaster texture.
Preserve the paper sculpture, every swatch, table, shadows, light direction, camera position, left-side empty space, and landscape crop.
Add no people, text, logos, shelves, windows, or additional objects.
Compare the output with the accepted base. Inspect sculpture folds, count visible swatches, compare shadow direction, and check the empty region. If any locked item changed materially, return to the base and reduce the instruction to “change wall color only.”
Step 6: Introduce one reference texture if necessary
If text alone cannot communicate the wall treatment, upload a licensed texture as a second reference. Label the roles in plain language:
The first image is the accepted composition and must remain the base.
The second image is a texture reference only.
Apply the second image's subtle warm-gray plaster texture only to the wall behind the sculpture.
Preserve all objects, lighting, shadows, crop, colors, and empty space from the first image.
Do not copy objects, edges, text, or composition from the second image.
This is multi-reference composition at a deliberately small scope. The second file supplies one surface property, not a whole scene. If attributes bleed into the subject, reject the version and return to the single-reference branch.
Step 7: Adjust one color without moving objects
Once the wall edit passes, save a new checkpoint. Then request one color change:
Change only the smallest amber swatch on the right side to muted coral.
Keep its size, position, angle, label area, and shadow unchanged.
Preserve every other swatch, the sculpture, table, wall, lighting, camera, and crop.
Add nothing.
Precise spatial language matters. “Make one swatch coral” leaves the target ambiguous. Location, size, and relationship to nearby objects make the instruction auditable.
Step 8: Add a short title as a separate pass
Text should come after composition and local edits. Use the shortest approved line possible, quote it exactly, specify position and contrast, and prohibit all other copy.
Add the exact title "DESIGN SYSTEMS" once in the empty upper-left area.
Use a clean bold sans-serif style in off-white with high contrast and generous spacing.
Keep the title on one line.
Add no subtitle, date, logo, symbols, or other letters.
Preserve the entire image outside that text area.
Proofread every character. Check that the line appears only once and that no stray marks resemble letters. If exact typography is critical, keep the generated image text-free and add the title in a layout tool. Model improvement does not remove the need for a human proofreader.
Step 9: Create a variant from the accepted image
Start from the latest accepted checkpoint and change the composition only:
Create a portrait variant of this accepted scene. Preserve the sculpture, swatches, wall, lighting, and editorial treatment. Place the sculpture in the lower two-thirds with quiet space above. Remove the title and add nothing.
ArWriter's Google route chooses size automatically, so inspect the result and crop only if the source supports the placement. Precise ArWriter size selection belongs to the separate ChatGPT route.
Step 10: Export and perform final checks
Download the full-size image through the current interface's documented action. Open it outside the chat and check dimensions, compression, text, geometry, edges, and color. Preserve the raw output and the finished file separately.
Use a descriptive name:
design-workshop-editorial-portrait-v04-approved-2026-06.png
Store the source brief, reference permissions, accepted checkpoints, final prompt, provider route, generation date, and reviewer approval. Do not strip or misrepresent provenance information.
Check identity and object consistency after every edit
When a person or recurring character is involved, compare face shape, eyes, hair, skin details, clothing, accessories, proportions, and position rather than calling the result “close enough.” For objects, compare count, location, color, edges, labels, shadows, viewpoint, crop space, and accidental marks. Always use a consented source.
Google reports that Nano Banana 2 can maintain resemblance for up to five characters and fidelity for up to 14 objects in a workflow. Treat those as stated model capabilities, not a promise that any complex scene will remain perfect. Your acceptance test should be stricter than a marketing example because the final image represents your organization.
This is why learning how to use Nano Banana for AI images requires a consistency checklist as well as a prompt. Model capability sets the ceiling; the reviewer decides whether each specific output is safe to accept.
For phone-based comparison and zoom review, the mobile image generator guide explains file handling and small-screen checks.
Troubleshooting drift and failed local edits
Use this ladder in order. Stop as soon as the output returns to an acceptable path.
| Problem | Likely ambiguity | First correction | Recovery point | Reject when |
|---|---|---|---|---|
| Wrong area changed | Location was vague | Name region and adjacent landmark | Last accepted image | Locked region remains altered |
| Subject identity shifts | Edit scope was too broad | List facial, hair, clothing, and pose locks | Clean source or accepted checkpoint | Resemblance is uncertain |
| Extra objects appear | Prompt lacked exclusions | State a short prohibited list | Prior accepted version | New objects persist after one retry |
| Text is misspelled | Copy was long or insufficiently isolated | Quote shorter exact text and prohibit other letters | Text-free accepted image | Any character remains wrong |
| Reference styles bleed | Roles of references were unclear | Define the contribution of each reference | Single-reference branch | Composition changes unexpectedly |
| Quality degrades over turns | Too many stacked edits | Restart from best checkpoint with one change | Earliest clean accepted image | Artifacts compound |
| Crop no longer works | Composition was not locked | Restate subject position and safe area | Accepted placement version | Important content is clipped |
Restate vague locations with a stable landmark, reduce subjective changes to one visible property, and reopen the best checkpoint instead of correcting a bad revision. If the source is compressed or obstructed, replace it with a clean authorized reference. When conversation history contains conflicting requests, start fresh with the accepted image, original brief, and current preservation contract.
The ArWriter image prompt library can help organize starting ideas, but a reusable prompt does not replace reference permissions, checkpoint discipline, or visual review.
Text, provenance, and responsible publication
Google says Nano Banana 2 improves text rendering and can translate or localize text inside images. Even so, a publishing workflow must verify exact characters, word order, punctuation, duplication, and unintended text. For high-stakes information, create the scene without text and typeset approved copy separately.
Google also states that generated content in its current ecosystem uses SynthID and C2PA Content Credentials. Preserve available provenance data and follow organizational disclosure rules. Do not claim that a synthetic scene documents a real event, person, location, or result.
Human approval remains mandatory. Verify the brief, rights, consent, factual and textual accuracy, private data, unwanted brands, crop safety, output quality, provenance, disclosure, filename, and version before publication.
If your workflow involves several reviewers or formal controls, consult the browser image workflow for teams and keep approval evidence in a system that actually supports your governance requirements.
Frequently Asked Questions
What is Nano Banana and where can I use it?
Nano Banana is Google's public image-model branding within the Gemini ecosystem. Nano Banana 2 refers to Gemini 3.1 Flash Image, while Nano Banana Pro refers to Gemini 3 Pro Image. Access and controls vary across Google products and third-party implementations, so follow documentation for the exact interface you use.
Is Nano Banana the same as Gemini image generation?
Nano Banana names Google's image-model family, while Gemini is the broader product and model ecosystem through which image features may be exposed. The terms overlap but are not interchangeable interface labels. A third-party service can provide a Google image route without reproducing every control found in Google's consumer application.
How do I create an image from scratch with Nano Banana?
Start with a six-field brief covering purpose, subject, action, environment, visual treatment, and hard constraints. Generate the base composition without final text or several simultaneous edits. Review it at fit-to-screen and 100% zoom, then preserve the first acceptable version as a named checkpoint before revising.
How do I upload and edit a reference image?
Use a clean image you own or have permission to edit. Upload it through the documented interface, identify one precise location and requested change, list every element that must remain locked, and prohibit additions. Compare the result against the source at 100% zoom and reject unrequested changes.
Can Nano Banana change only the background?
It can perform localized background edits, but “only” is an instruction rather than a guarantee. Name the exact background region, describe the replacement, and lock subject identity, clothing, pose, lighting, camera, crop, foreground objects, and text. Return to the source checkpoint if any locked element changes materially.
How do I keep a face or character consistent?
Use a clear consented reference, limit each turn to one change, and explicitly preserve face shape, eyes, hair, skin details, clothing, accessories, pose, and position. Compare each revision with the last accepted image. If resemblance drifts, return to that checkpoint or restart from the clean reference.
Can Nano Banana combine multiple reference images?
Google documents workflows with multiple uploaded images. Label each reference by role and restrict what it contributes: for example, composition from the first and wall texture from the second. Begin with two simple inputs, inspect for style or object bleed, and return to a single-reference branch if control deteriorates.
How do I download a full-size Nano Banana image?
Use the download action documented for the exact Google or third-party interface, then open the file outside the chat and verify pixel dimensions, compression, text, geometry, and edges. Rename it descriptively and archive the prompt, source permissions, accepted checkpoints, provider route, generation date, and human approval with the output.
Conclusion
The durable answer to how to use Nano Banana for AI images is to treat every accepted image as a controlled checkpoint. Start with one clear mode, define a preservation contract, change one located element per turn, and compare the revision against the prior version. When drift accumulates, roll back instead of adding more corrective prompts.
ArWriter offers a paid browser route labeled Google with Gemini 3 Pro for this editing pattern. It does not expose every control described in Google's own products, so keep platform claims separate. Run the sample project with low-risk material, verify the current interface, and preserve human review from source permission through final export.
Sources
- Google Nano Banana 2 announcement — official current naming, consistency, resolution, text, and provenance claims.
- Google Gemini image help — official creation, generated-image editing, uploaded-image editing, multi-image, and download instructions.
- Google Nano Banana Pro developer announcement — official Gemini 3 Pro Image identity and controls.
- OpenAI Academy image prompting — official incremental revision, reference, constraint, and consent practices applicable to controlled image editing.
Start With One Controlled Edit
Open ArWriter Chat with an image-enabled account, use a consented low-risk reference, and make one localized change. Preserve the source, prompt, checkpoint, and approval record before moving to production.
