Unnamed references
“Combine these images” does not explain which file supplies the person, room, clothing, or style. Label every source and state its role. A Nano Banana prompt becomes easier to audit when each borrowed element has one origin.
Gemini reference edit desk
Nano Banana Prompt Generator — Reference Image Editing
Create a Nano Banana prompt that treats the uploaded image as an active reference. Define the subject, setting, action, composition, style, preservation rules, requested edits, and production format in natural language, then copy the result into Gemini.
- Strength
- Reference editing
- Method
- Natural language
- Output
- 6 prompt versions
Reference translated
Reference / role / editUse the uploaded watch image as the composition reference. Preserve the exact overhead placement, white strap, silver case, and soft studio reflections. Change the surface to pale limestone and keep the watch face blank.
Reference role
composition and product
Preserve
shape, angle, reflections
Change
surface material only
Deliverable
vertical catalog image
Gemini reference input
Assign each reference a role before requesting the edit
Generate the six outputs, open Nano Banana, and verify which source supplies the subject, composition, style, or environment before you ask for a bounded change.
Drop images herePaste an image, or choose how you want to add one.JPG, PNG, or WebP · up to 10 MB each
No image selected
Add an image to create all six prompt formats.General · Midjourney · FLUX · Stable Diffusion · ChatGPT Image · Nano Banana
How the reference participates
A Nano Banana prompt should explain how the reference participates
The instruction can direct image generation or editing in Gemini. Google guidance asks users to think about style, subject, setting, action, and composition, then add details gradually. With an uploaded image, the instruction can also assign a role to the reference: preserve a person, borrow a room layout, transfer a style, replace an object, or build a new scene around the same subject.
This generator reads the visible decisions in your image and writes a reference-aware instruction. It identifies subject, environment, composition, lighting, color, camera feel, materials, style, and mood. Then it says which parts should remain stable and which can change. That boundary is more useful for editing than a flat list of visual tags.
The generated text is a reconstruction, not the original hidden instruction. A finished image cannot disclose the exact Gemini conversation, model mode, uploaded references, prior edits, source files, or generation history. The tool infers creative direction from pixels and gives you an editable starting point. Review details that could be ambiguous or sensitive.
Nano Banana is a public name used for Gemini image capabilities, and Google offers different modes and model generations over time. The exact product controls can change. Keep the descriptive prompt readable and place resolution, aspect ratio, mode selection, or other application controls in Gemini or Google AI Studio as appropriate.
Source-role map
Assign every image a role in the Nano Banana prompt
Reference-aware editing works when the instruction distinguishes source material from the requested output. The model needs to know what each image contributes and which relationships should survive the change.
- 01
Name the reference
Start by identifying the uploaded image and its role. Say “use the uploaded portrait as the subject reference” or “use image one for the product and image two for the room style.” Clear labels become more important when several references enter the same request.
- 02
Define the subject
Describe the person, object, animal, or place that must remain recognizable. Google suggests assigning distinct names to subjects when maintaining consistency across images. Use a simple label such as “Mara” or “the red chair” and repeat it in later instructions.
- 03
Map setting and action
State where the scene takes place, what the subject does, and how objects relate. Keep camera angle, crop, or pose when those define the reference. Say which geometry remains fixed before requesting a new season, outfit, material, or environment.
- 04
Separate preserve from change
Use direct clauses: preserve the face, pose, product proportions, and composition; replace the background and wardrobe. This gives the edit a visible success test. Avoid asking for a complete style overhaul while insisting that every surface, shadow, and color stay identical.
- 05
Specify the deliverable
Finish with the intended format, aspect ratio, resolution option, text requirement, number of variations, or series continuity. Google guidance supports asking for multiple images, production-ready aspect ratios, and subject consistency. Check the controls available in your current Gemini interface.
Reference identity
Reference labels prevent the wrong image from controlling the wrong detail
A multi-reference instruction should state the role of each file in plain language. Image one can supply the subject, image two the environment, and image three the color treatment. Without that map, several reasonable combinations are possible. The model may borrow clothing from the style reference or move the product into the wrong scene.
Preservation rules need priorities. “Keep the same person” is broader than “preserve facial features, hairstyle, skin tone, and age while changing the jacket.” The second version gives you specific review points. For products, preserve silhouette, proportions, branding where permitted, material, and camera angle before changing the backdrop.
Google recommends adding detail bit by bit. Use that as an editing discipline. Start with one structural change, inspect the image, then adjust style, typography, or small props. A long instruction can still work, but one crowded request makes it hard to identify which clause caused an unwanted change.
Generation and editing are probabilistic. Reference consistency can improve without becoming identical. Review faces, hands, product geometry, visible text, logos, and small repeated details after every pass. If exact reproduction is required for legal, catalog, or identity-sensitive work, keep the original asset and use standard editing tools where precision matters.
Controlled edit passes
Use a Nano Banana prompt to edit in controlled passes
A reference image gives the model more evidence, but evidence still needs direction. The workflow becomes predictable when each pass has one job and an explicit preservation boundary.
- PASS 01
Choose what the image contributes
Before uploading, decide whether the reference supplies subject identity, composition, style, product geometry, environment, or color. Write that role into the generated brief. If the image should only inspire lighting, do not accidentally ask the model to preserve every object in it.
- PASS 02
Review the generated instruction
Check the generated instruction against the source. Remove invented text and guessed personal details. Correct the subject, camera angle, setting, and preserve list. If a detail is optional, do not phrase it as fixed. The edit needs room to produce the requested change.
- PASS 03
Run the structural edit
Change the background, action, format, or composition first while preserving the required subject features. Review the whole output. If identity or product shape drifts, narrow the change and restate those constraints in the next turn.
- PASS 04
Refine surfaces and text
After structure holds, adjust color, lighting, material, typography, or small props. Save the edited text and the related visual analysis in History. A short sequence of clear passes is easier to reuse than one giant instruction with several competing goals.
Consistency failures
Why a Nano Banana prompt loses subject consistency
Consistency problems often begin before generation. The reference role is vague, the requested change conflicts with preservation, or too many subjects compete for attention.
Conflicting instructions
“Keep the lighting identical” conflicts with moving a subject from a studio into a sunny beach scene. Choose the relationship that matters. The instruction can preserve face and pose while allowing environment light to change naturally.
Too many edits at once
A new outfit, location, camera angle, season, action, poster title, and color grade create several failure points. Split them into passes. Each turn should make one main change and protect the details needed for the following step.
Exact-match expectations
A reference-aware model can maintain key features without reproducing every pixel. Text cannot recover hidden source files or guarantee exact identity, type, or product geometry. Use the Nano Banana prompt for directed variation and review high-risk details manually.
Reference-edit questions
Nano Banana prompt questions
What is a Nano Banana prompt?
A Nano Banana prompt is a natural-language instruction used with Gemini image generation or editing. It can define a subject, setting, action, composition, style, reference role, preservation rules, requested changes, and final format.
Can a Nano Banana prompt use an uploaded image?
Yes. Gemini image tools support uploaded references for generation and editing. State what the image contributes and what must stay fixed. With multiple images, label each source and assign its subject, style, background, or product role.
How do I keep a person consistent in a Nano Banana prompt?
Use a clear reference, give the subject a distinct name, and specify which facial, hair, age, skin, clothing, pose, and camera details must remain. Change one major variable at a time and inspect each result before continuing.
Can this recover the original Nano Banana prompt?
No. A finished image does not contain the exact Nano Banana prompt, Gemini conversation, model mode, references, masks, or edit history. The generator writes a new instruction from visible evidence and marks it for user review.
Should a Nano Banana prompt include aspect ratio and resolution?
Include the intended format when it affects composition, and use the current product controls for available resolutions or aspect ratios. Google guidance supports production-oriented formats, but the exact interface options can change.
Is the Nano Banana prompt generator free?
Visitors receive three daily credits. One analyzed image uses one credit and returns all six prompt versions, including the Nano Banana prompt. Accounts add editable History, and Pro subscribers can analyze several images in a batch.
Run one bounded edit
Give the next Nano Banana prompt one reference map and one clear edit
Upload the image, assign its role, verify the preserve list, and edit the generated Nano Banana prompt before copying it into Gemini. A Nano Banana prompt should make the reference map readable without guesswork. During each pass, keep the Nano Banana prompt focused on one primary edit. Save the final Nano Banana prompt with the source and accepted output so the decision trail remains clear.