A practical guide to Nano Banana prompts: write clearer scene descriptions, create natural-looking portraits, preserve likeness with references, and make precise image edits.
Nano Banana is the familiar name for Google's Gemini image-generation capabilities. The original examples in this guide focus on Gemini 2.5 Flash Image and Nano Banana Pro / Gemini 3 Pro Image. The practical lesson is simple: do not treat this as a Midjourney or Stable Diffusion command box.
Describe a scene as you would explain it to a photographer or retoucher—not as a pile of tags and settings copied from another model.
Gemini works with text and images together. Clear, connected prose helps explain what belongs in the scene and how its parts relate. Positive descriptions tell it what to create; conversational editing tells it what to change and what to preserve. Weight syntax such as “::” and flags such as “--no” are not a portable command language for Nano Banana.
If you remember only three rules, make them these:
The sections below turn those principles into a practical workflow for realistic images, reference-based portraits, style transfer, and editing.
Sign in to leave a comment.
| Habit carried over from other image workflows |
|---|
| A better starting point for Nano Banana |
|---|
| A long tag list with numeric weights | A connected description of the scene |
| A dedicated negative-prompt recipe | A positive description of the desired result |
| Regenerating everything to change one detail | A focused edit request explaining what must stay |
| Aspect-ratio flags embedded in the text | The generator's aspect-ratio control |
| Increasing a term's numeric weight | Clarifying its importance and its relationship to the scene |
Think in terms of visual relationships: who → doing what → where → how the light falls → what the image should feel like. A coherent paragraph makes those relationships explicit. A tag list leaves the model to guess how they fit together.
This is a prompting approach, not a promise that every instruction will be followed perfectly. Clear language improves the brief; you still need to inspect the result.
Google's image-prompting guidance puts it succinctly:
“Describe the scene, don't just list keywords.”
The point is to connect the subject, action, setting, lighting, and visual treatment into a scene.
A weak starting point:
woman, red dress, cafe, paris, cinematic, 8k, bokeh, golden hour, highly detailedA more useful Nano Banana prompt:
A young woman in a flowing red dress sits at a small marble table outside a Parisian cafe,
gently stirring her espresso. Warm golden-hour light rakes across the scene from the left,
throwing long soft shadows. Shot on an 85mm lens with a shallow depth of field, the background
dissolves into creamy bokeh of cafe awnings and cobblestones.The second version explains where the light comes from, what is in focus, what is blurred, and what the subject is doing. It gives the image a plan instead of a mood board of unrelated words.
Do not repeat the claim that narrative prompts produce “73% better results.” The original guide did not find a reliable source for that figure. The useful principle does not need a made-up percentage.
Build these elements into a connected paragraph rather than a checklist pasted verbatim into the prompt:
[Subject] + [action or expression] + [setting and surroundings] +
[composition and viewpoint] + [lighting] + [style or color grade]Clarity and specificity matter more than length. Adding text just to make a prompt longer can blur your priorities.
Photography terms can help communicate the kind of image you want. Use them as visual direction, not as a guarantee that the output reproduces a real camera's exact optical behavior.
Shot sizes and viewpoints:
Wide-angle shot; close-up; extreme close-up; macro; low-angle perspective for a sense of power; high angle for vulnerability; bird's-eye or overhead; Dutch angle for tension; over-the-shoulder; eye-level.
Useful lens and camera descriptions:
| Desired look | Example wording |
|---|---|
| Portrait with flattering perspective | 85mm portrait lens, f/1.8, shallow depth of field, soft bokeh |
| Natural-looking perspective | 50mm lens, eye-level, natural perspective |
| Street or documentary image | 35mm lens |
| Moderately wide view with broad focus | 24mm wide-angle lens, f/11, deep depth of field, slight barrel distortion |
| Strong wide-angle look without a circular frame | 14mm wide-angle lens, strong wide-angle barrel distortion |
| Immersive action-camera look | GoPro action camera, ultra-wide immersive perspective |
Be careful with “fisheye.” It can suggest a circular image with dark corners. If you want an extreme wide-angle image that still fills a rectangle, describe “14mm wide-angle lens” or an action-camera perspective, then specify “fills the whole rectangular frame, bright clean corners.”
Lighting deserves its own clear instruction. Useful descriptions include golden hour, softbox studio light, rim light, low-key lighting, high-key lighting, window light, dramatic chiaroscuro, and backlight. Say where the light comes from and what it does to the subject.
Camera and film references can shape the aesthetic: shot on Canon 5D, Fujifilm color science, Kodak Vision3 500T film, disposable camera with flash, or Polaroid. These communicate a visual character; they are not evidence that a real camera or film stock was used.
A common complaint is that the image looks convincing at first glance but still feels too smooth, glossy, and advertising-like. Think of the target as the middle of a dial:
POLISHED ADVERTISING ←—— NATURAL PHOTOGRAPH ——→ EXCESSIVE GRITTo move toward a more natural photograph, add positive documentary-style cues:
If you want a casual photograph rather than a glossy campaign, reconsider words such as cinematic, ultra-detailed, 8k, HDR, glossy, perfect, glamour, studio lighting, and hyperreal. They may pull the brief toward a different aesthetic.
Do not overcorrect. Repeated requests for imperfections, grit, sweat, and rough skin can push the image far beyond the natural look you intended.
For an ordinary, healthy complexion, define the desired balance:
Clear, healthy skin with natural texture and an even complexion,
natural matte skin, no heavy retouching, no acne or blemishes.The key is to describe healthy, natural-looking skin first, rather than relying only on “no plastic skin” or “no CGI.” The wording guides the model; it does not guarantee an identical result every time.
Reference-based portraits are a major creative use case: place the person from a supplied photo in a new location, outfit, or situation while keeping them recognizable.
There is no verified magic phrase that guarantees a one-to-one biometric copy. Phrases such as “this exact character” or “identical face” are instructions, not guarantees. Likeness is a workflow: start with a suitable reference, avoid conflicting descriptions, generate alternatives, and choose the strongest result.
1. Do not invent a competing description of the face.
If identity comes from the reference, avoid adding a different eye color, nose shape, age, or especially hair color and texture. Describe the scene, clothing, pose, and lighting. Let the reference supply the person's appearance unless a deliberate styling change is part of the brief.
2. Specify which identity details should remain unchanged.
Rather than relying on “make the face 1:1,” give a concrete preservation instruction:
Keep the exact same face, identity and natural hair color as the reference photo — same facial
features, eye shape, nose, mouth, jawline, skin tone and proportions. Do not beautify, stylize,
slim, age, recolor the hair or alter the face in any way. Photorealistic, natural skin texture.This names the details you care about and discourages common unwanted changes. It still does not turn generation into an identity-lock guarantee.
3. Return to the original reference for important variations.
Repeatedly editing the previous output can accumulate drift. For a fresh scene or an important portrait, start again from the original reference rather than extending a long chain of edits.
4. Use a clear reference.
A sharp, evenly lit, front-facing photo is a better starting point than a blurry or heavily filtered one. When the selected version and interface allow it, additional views such as front and three-quarter can help communicate the person's appearance.
5. Keep identifying features visible.
Closed eyes, dark glasses, a strong squint, or a diving mask can make likeness harder to judge. If the intended composition allows it, “eyes open, looking at the camera” gives a clearer portrait brief.
Using the person from the reference photo, a [shot type] photo of them [doing action] in
[location/setting]. They wear [outfit]. [Lighting]. Shot on [lens, e.g. 85mm portrait lens],
shallow depth of field. [Mood / style].
Keep the exact same face, identity and natural hair color as the reference photo — same facial
features, eye shape, nose, mouth, jawline, skin tone and proportions. Do not beautify, stylize,
slim, age, recolor the hair or alter the face. Photorealistic, natural skin texture.Attach the inspiration image, describe a new scene, and explain exactly which qualities to borrow:
[New scene]. Apply the color palette, lighting mood and overall visual style of the reference
image — its [warm muted tones / film grain] — but do not copy its objects, subject or composition.Separate palette, lighting, grade, and mood from objects and composition. Without that distinction, the model may reproduce the contents of the reference when you only wanted its visual treatment.
For a close recreation, give a direct brief:
Recreate this exact image as faithfully as possible: same composition, subject, pose,
lighting and colors. Do not add, remove or restyle anything.The realistic target is a close recreation, not a pixel-for-pixel copy. This is a generative process, not a duplicate-file operation.
Explain both what should change and what should stay:
Change only the [element] to [new]. Keep the background, lighting, composition and everything
else in the image exactly the same. Do not change any other elements.This acts as a semantic mask: a description of the intended edit boundary, not a guarantee that every untouched pixel will remain identical. Conversational editing is useful, but repeated edits can reduce quality or introduce identity drift. Return to the source image when preservation matters.
Do not trust important text without checking it. Short words may work well, but longer copy, unusual spellings, non-Latin scripts, dates, and numbers still need close inspection. Results vary by model version and by the complexity of the design.
For advertising, covers, invitations, and other finished assets, the reliable production workflow is:
If you need a particular logo, supply the actual visual reference rather than assuming that a brand-name instruction will reproduce every detail correctly. Check the output against the original artwork.
The point is not that the model can never render text. It is that important copy needs a verification step, and a separate text layer gives you predictable spelling and layout.
An otherwise ordinary fashion or lifestyle scene can be described in a way that makes its intent unclear. If the intended scene is non-explicit, make the clothing and context genuinely non-explicit as well. This is not a way to bypass safety rules, and rewording does not guarantee approval.
| Wording that can imply a different kind of scene | A clearer brief for a non-explicit image |
|---|---|
| Sheer, see-through, wet-clinging clothing | Opaque cotton or linen clothing, with the intended coverage clearly described |
| Sultry, seductive, sensual, or provocative mood | Serene, confident, joyful, or candid mood |
| A wet white top with emphasis on body shape | An opaque colored top, such as a fitted sage-green top, with attention on lighting and location |
| Lingerie when the actual goal is an ordinary lifestyle image | An elegant dress or activewear appropriate to that scene |
Keep the emphasis on light, location, visual style, and emotion. Choose clothing and a context that actually match the safe editorial, lifestyle, or wellness image you want to make. If the underlying request is prohibited, changing a few words does not make it acceptable.
| Shortcut or claim | A more useful approach |
|---|---|
| Negative-prompt flags such as “--no text” work like they do in other tools. | Describe the wanted image in plain language. For example, ask for an empty street rather than relying on a special “no cars” flag. |
| Numeric weights such as “red dress::1.5” provide portable control. | Explain the priority and the scene in prose instead of importing another tool's syntax. |
| “Fisheye” is a reliable way to get any wide-angle image. | If you need a rectangular frame, describe the wide-angle perspective and the desired corners explicitly. |
| One magic phrase guarantees the same face in every image. | Use the reference-based workflow, inspect likeness, and select the best result. |
| Multi-person consistency is guaranteed. | Treat it as a capability to test, not a promise of exact identity preservation. |
| Pro renders every piece of text perfectly. | Proofread all important text; add final copy in a design tool when accuracy matters. |
| “No CGI” and “no 3D” are enough to remove a plastic look. | Add positive photographic cues: natural skin texture, appropriate light, and a coherent camera or film aesthetic. |
| A longer prompt is automatically better. | Make the description coherent and specific. Remove repetition and conflicting instructions. |
| Narrative prompts are “73% better.” | Do not repeat an unsupported number. Use the practical principle without the statistic. |
Check the selected version, not just the Nano Banana name. This guide grew out of work with Gemini 2.5 Flash Image and Nano Banana Pro / Gemini 3 Pro Image. Availability, reference limits, output sizes, and prices can differ across versions and interfaces. The original article's retirement forecasts and “current flagship” labels are not useful as permanent instructions.
For version-specific details, consult Google's Gemini image-generation documentation.
A. Turn a vague brief into a prompt. Identify the subject, action, location, mood, and intended use. Turn those details into a connected paragraph, then add the lighting, viewpoint, and visual treatment.
B. Improve an existing prompt. Convert tags into prose, replace vague negatives with desired outcomes, remove imported weight syntax, and check for missing lighting or composition. If the result looks too synthetic, refine the photographic cues instead of simply adding “no CGI.”
C. Transfer a reference style. Read the palette, lighting, grade, and mood—not just the objects. Describe a new scene and state that you want the style rather than the original contents.
D. Create a reference-based photo shoot. Let the reference define the person's appearance, describe the new scene around them, specify what must remain unchanged, and generate from the original reference when likeness is important. Expect to compare alternatives rather than find a perfect identity-lock phrase.
For a complex composition, divide the work. If the brief combines many objects, precise typography, and several distinct panels, consider generating the components separately and assembling the finished layout in a design tool. This gives you more control over the final result.
Neylora's AI prompting assistant follows these principles to help you build a clearer brief. Try the Nano Banana generator.
Nano Banana rewards clear human language and realistic expectations. Describe a connected scene, use positive visual direction, and explain edits in terms of what should change and what should remain.
For portraits, treat likeness as a reference-based workflow. For realism, aim for a deliberate photographic look rather than a collection of “quality” tags. For finished designs, keep typography and final quality checks under your control.
The goal is not a longer prompt. It is a brief the model can understand—and a result you can evaluate against what you actually wanted.
Adapted for Neylora Academy from the original Nano Banana prompting guide. The examples and practical workflows are retained, with cautious wording around guarantees and version-dependent limits. Model behavior and available controls can change; inspect outputs and check current settings before relying on a workflow.