If you want one creator across portraits, outfits and campaign images, a generic ‘make the same person’ instruction is a weak starting point. Nano Banana can be part of a reference-led workflow, but you still need an approved identity, a clear scene brief and a review process.
This guide separates the model names from the production decisions. The prompts are original editorial examples, not benchmark results or a promise that every generation will preserve every detail.
Choosing between providers? Run the same source and prompt with the Grok vs Nano Banana image-edit comparison, then record your own results in the manual scorecard.
Check which Nano Banana model you are using
Google's image-generation documentation now lists Nano Banana 2.1 alongside Nano Banana 2, Pro, Lite and the legacy model. The October update matters when an older tutorial names a model that is no longer selected in your interface. Record the actual name before applying its settings.
| Model label | Google API model code | Check before choosing |
|---|---|---|
| Nano Banana 2.1 | gemini-nano-banana-2.1 | Current image-generation and conversational-editing model |
| Nano Banana 2 | gemini-3.1-flash-image | Earlier model; do not silently substitute it for the newer one |
| Nano Banana Pro | gemini-3-pro-image | Separate model for complex visual work |
| Nano Banana 2 Lite | gemini-3.1-flash-lite-image | Different speed and reference tradeoffs |
| Legacy Nano Banana | gemini-2.5-flash-image | An older tutorial may refer to this model |
The current model page documents up to 14 reference images and image output at 1K, 2K or 4K. Those are documented capabilities, not a guarantee that every attached object will remain exact. A third-party workspace may expose fewer controls. Compare accepted exports for your task rather than declaring a winner from the model name.
| Decision | What to record |
|---|---|
| Model | The exact visible model name, not just the Nano Banana family label |
| Input | Text-only, identity reference, product reference or an existing image edit |
| Output | Aspect ratio, requested resolution and downloadable file |
| Revision | Which details are allowed to change and which stay fixed |
Use Nano Banana in Higgsfield with the right controls
Open the Higgsfield image workspace with Nano Banana selected. In our signed-out October 17 check, the loaded model button read Nano Banana 2.1. The surrounding introductory banner still promoted Soul Cinema, so the selected model control was the useful signal.
- Confirm the model in the composer, then choose the shape of the finished image
- For an edit, attach the accepted source through References before entering the change
- Select an available resolution and record the chosen Thinking setting
- Decide whether the task needs the available web or image search controls
- Check the displayed charge and account access, generate one test and download it for review
The public Quality menu offered 1K, 2K and 4K; Thinking offered Minimal, Medium and High. Search settings exposed Web search and Image search. We inspected those menus without uploading a reference or submitting a generation. A visible control does not establish a free allowance or a tested result. Start with the settings needed for one deliverable, then change one decision at a time.
For a closed product-color revision, the accepted reference defines the object. For a factual subject that needs external references, search can be a different part of the brief. Name the reference role so a reviewer knows which input establishes identity, which supplies the product and which only suggests a visual style.
How to write prompts for Nano Banana
Write a scene the model can depict: who is visible, what they are doing, where they are, how the picture is framed and where the light comes from. Put fixed identity details in a separate instruction when using a reference. Replace vague requests such as “make it professional” with concrete decisions such as a waist-up portrait, a plain background and soft window light.
The Google Nano Banana prompting guide discusses subject, composition, action, location and style. Use it alongside the API documentation: a clear brief describes the image, while the interface controls determine which inputs and output settings the model receives.
Turn a vague request into a reviewable brief
| Vague instruction | Specific replacement | What you can inspect |
|---|---|---|
| Make a good creator photo | An original adult creator seated at a cafe table, with the head and both hands visible | The crop and visible anatomy |
| Make it realistic | Soft window light, cotton texture and natural skin detail | Lighting direction and material consistency |
| Keep everything the same | Change the notebook cover only; preserve the face, pose, mug and background | The requested change and unintended revisions |
Use the builder above to generate a prompt from those decisions. It assembles your fields in your browser, requires a subject and scene, and provides a selectable result. It does not call an image model or charge generation credits. Copy the output into your chosen image workflow and attach any required reference there.
Nano Banana JSON prompt versus an API request
A Nano Banana JSON prompt can mean two different things: a creative brief saved as JSON, or the actual request body a developer sends to a provider. The JSON download in this guide is a creative brief. Its subject, scene and preservation fields organize your instructions; they are not special model controls.
{
"format": "Clout editable image brief",
"subject": "An original adult creator in a mustard overshirt",
"scene": "Seated at a cafe table with a closed green notebook",
"framing": "Medium-wide photograph with both hands visible",
"lighting": "Soft window light and natural cotton texture",
"preserve": "Approved face and hairstyle when using a reference",
"avoid": "No lettering or additional people"
}
Use the plain-language output in a prompt box. For an API integration, put that prompt text inside the documented request field, supply image references in the documented format and set only supported model options. Do not send the custom brief object as if its keys were Google's request schema. Downloading JSON does not attach a reference image, set an aspect ratio or select a model.
Structured briefs help a team keep decisions consistent and revise one field at a time. They do not establish that JSON produces better images than ordinary language. Evaluate the exported result, not the appearance of the prompt format.
A small revision prompt with an actual output
Change only the forest-green notebook cover to cobalt blue. Keep the creator's face, hairstyle, hands, mustard overshirt, cream shirt, mug, table, lighting and composition unchanged. Do not add any writing to the notebook.
Open the source-and-edit example in our picture guide to download both images and review this instruction against an actual result. That pair was generated with the built-in image tool and is not a Nano Banana test. It shows the review process you should apply to your own Nano Banana exports.
Inspect an original lamp source and color revision
This exercise turns a vague “keep the product the same” instruction into observable checks. The source is a fictional ceramic lamp on an oak table, with a cobalt sketchbook and an ivory cup. The requested change is the shade color. Both images were made with the built-in image tool; they are not Nano Banana or Higgsfield results.

Download the native source PNG
Change only the terracotta-orange lamp shade to a muted sage green glaze. Preserve the shade silhouette, rim and texture, ivory base, brass switch, cobalt sketchbook, ivory cup, table, background, camera angle and daylight. Keep the lamp switched off.

Download the native revision PNG
At native size, the shade reads as green rather than orange, and the base, book and cup remain recognizable in the same general arrangement. The revised shade also has a different sheen and texture rendering. That is a useful concept revision, but it does not prove that the original material or every surrounding pixel was preserved. A real product brief may reject that surface change.
| Review area | What this pair lets you inspect |
|---|---|
| Requested change | Does the shade read as the intended green |
| Product shape | Compare the rim, slope, base proportions and switch position |
| Material | Check whether glaze texture and highlights changed beyond the brief |
| Unrequested changes | Compare the book, cup, wall shadows and pottery shelves |
| Delivery | Both native files are 1672 by 941; check your actual provider export separately |
Download the complete lamp exercise for both native PNGs, optimized previews, reusable prompts and a blank review CSV. Or use the review sheet alone. The prompt file organizes a creative task; it is not a model request, uploaded reference or executed preset.
Choose Load lamp example in the prompt builder near the top of this page to adapt the revision. Change the target color or a preservation instruction, copy the prompt, then attach the original source in your selected image editor. The local builder does not attach files, select Nano Banana or generate an image.
Keep prompts and references aligned
If the prompt says to preserve a face but no identity image was attached, the model lacks that specific baseline. If the prompt asks for a waist-up scene while the reference crops tightly around the face, check whether the selected operation can create the new framing. If your revision changes too many details, return to the accepted image and request one change at a time.
Build an identity reference you can review
Use a clear approved portrait of an original adult character or a person whose likeness you have permission to use. Avoid beginning with a blurry screenshot or a reference where the face is obscured. Save that baseline separately from later scene outputs.
Create a short identity note: facial landmarks, hair shape, apparent age and proportions. The note helps reviewers agree on what makes the character recognizable. It complements the reference image; it does not replace it.
Start with one scene transformation
Use the creator reference for the same adult character. Make a waist-up editorial photograph in a bright pottery studio. The creator wears a mustard linen blouse and holds a plain coral ceramic mug with relaxed natural hands. Keep the face, hairstyle and proportions consistent. Soft window light, a simple pale wall and realistic fabric texture.
Approve the face before approving the background. Then check the hands, mug and clothing. If the identity is wrong, return to the original reference instead of using the failed scene as the next reference.
Do not supply a collection of contradictory portraits and expect the model to know which one defines the person. Label the role of each permitted input in your brief: identity, product, scene or style. Interface syntax for reference labels varies, so use its documented controls.
Use a precise edit for a small change
Change only the blouse color from mustard to deep blue. Preserve the creator's face, hairstyle, hand position, mug, studio background and lighting. Keep the fabric weave and natural folds.
Compare the revised image with the accepted baseline. Check the mouth, eyes, fingers and object edges even when you asked only for a color change. If those areas changed, the result needs another review or a more controlled local edit.
For an outfit campaign, use a consistent framing brief and approved garment references. Exact seams, labels and fabric patterns remain approval requirements. A generated image should not be treated as proof of physical fit or product performance.
Test continuity with a small sequence
- Neutral portrait: establish a clear identity baseline.
- Lifestyle scene: move the creator into a plausible environment.
- Wardrobe variation: change one outfit while keeping the identity.
- Product scene: add a simple interaction with a reference object.
- Local revision: change one accepted detail without restarting the whole scene.
Review the five images together. Record where the identity changed, where the prompt was ignored and how much repair was necessary. This is a practical trial design, not a success-rate claim.
Diagnose the failure instead of lengthening every prompt
| Problem | Next test |
|---|---|
| The face changes | Return to one clear identity reference and simplify the new scene |
| The edit changes unrelated details | Use a more specific revision and supported local editing controls |
| Hands or props look wrong | Choose a simpler interaction with visible contact points |
| The image is too small | Check the selected model and export settings; download the master |
| Text on a product is inaccurate | Compare the reference and use controlled finishing or a real detail image |
Keep the accepted full-resolution source, prompt and model name together. This makes it easier to reproduce the campaign direction even if a later interface update changes a control or default.
For a larger still-image deliverable, use the Nano Banana Pro upscaler guide and local export checker to inspect the downloaded dimensions and details before approving it.
Understand image output and video handoff
A static image generator and an animation workflow are separate steps. Do not assume an image export is a GIF or that an interface offering Nano Banana also supports the exact video operation you want. Approve the image, then use an appropriate image-to-video or editing workflow.
Google states that generated images include SynthID. Follow the selected provider's current usage conditions and the publishing channel's disclosure requirements. A clean-looking export does not remove those obligations.
Build the creator and campaign together
In Clout, create your character and produce photos and videos around that recognizable identity. Keep your approved reference and campaign brief central to the workflow. Build your AI photo creator and turn the first accepted image into a reusable campaign direction.



