Nano Banana Prompts: A Copy-Paste Library That Actually Works
The fastest way to get a usable image out of Nano Banana is to stop writing prompts and start writing edits. Nano Banana rewards plain-language editing instructions — name what to change, name what to protect, describe the light — not keyword piles or from-scratch scene descriptions. It is Google’s Gemini image model, and its whole reputation is conversational photo editing that keeps a face, a product, or a character steady across changes. Below is the mental model, the shape of a prompt that lands, and a copy-paste library of fifteen prompts grouped by job. Want to run them inside a directed studio instead of retyping the look every time? The Illustration Canvas agent carries the craft for you. For the model itself, see Nano Banana explained.
What Nano Banana is built to do
The single most useful thing to know before you type a word: Nano Banana is an editor first, a generator second. Give it a photo and a sentence and it re-renders the scene while holding the subject’s identity — same face, same hands, same product geometry. That is the sweet spot, and prompts that lean into it land far more often than prompts that ask it to paint a complicated scene from nothing.
- Conversational edits. Each instruction builds on the last render, so a session feels like directing a retoucher: relight, swap a background, change a garment, remove a bystander — one note at a time.
- Character and product consistency. It carries a likeness across poses and lighting, and on the Pro tier blends a stack of reference images without inventing a lookalike. This is the trait that made it break out.
- Blends and world knowledge. It composites multiple inputs into one believable photograph and, because it draws on Gemini, renders real places, brands and objects more accurately than a model working from pixels alone.
As of August 2026 the everyday default is the standard Nano Banana tier; reach for Nano Banana Pro when you need higher-resolution output or the hardest in-image text. The prompting approach below is the same across the whole family — the tier changes the ceiling, not the grammar.
What a prompt that works looks like
Whatever the job, a good Nano Banana prompt has the same three moving parts, roughly in this order: the change, the protection, and the light.
- The change — one edit, stated plainly. "Make the jacket navy," not "transform this into a stunning editorial masterpiece." One scene change per turn is where identity holds.
- The protection — say what not to touch. "Keep her face and pose exactly" is the single fastest way to stop face drift. The model treats the rest of the frame as fixed and works around it.
- The light — describe it, don’t just name a place. "Soft window light from the left, warm" gives the model something to solve for, so shadows and reflections re-solve instead of the edit reading as a paste-in.
Notice what’s absent: no "8k, cinematic, hyper-detailed, masterpiece." Those are ranking tags borrowed from search, and an editor has nothing to rank. Describe the photograph you want, protect what matters, and let the render carry the quality.
You are not writing a wish. You are handing a retoucher a note: change this, keep that, light it this way.
A copy-paste prompt library
Here are fifteen prompts that work, grouped by the job you’re doing. Each is a full instruction you can paste as-is and then adjust. Read the right-hand note — the reason a prompt works travels further than the prompt itself.
Portrait and headshot edits
| Prompt | Why it works |
|---|---|
| Keep her face, hair and expression exactly the same. Relight with soft window light from the left, warm, late afternoon. Gently blur the background. | Names what to protect first, then gives a single light direction — the edit re-solves the whole frame instead of just pasting a face onto new light. |
| Change the jacket to charcoal wool, same fit and pose. Leave the face, hair and hands untouched. | Scopes the edit to one garment and explicitly fences off the identity anchors, so nothing drifts while the coat changes. |
| Remove the sunglasses and show the eyes, matching the skin tone and lighting already in the photo. | A removal grounded to the existing light — "matching what’s already there" is what stops the eyes reading as a sticker. |
Product shots
| Prompt | Why it works |
|---|---|
| Place this matte-white ceramic mug on a walnut table by a window, soft morning light from the right, faint steam rising. Keep the mug’s shape, logo and proportions identical. | Real scene plus a hard protect-clause on the product geometry — the model can build a room around the SKU without redrawing the SKU. |
| Put the sneaker on a seamless light-grey studio backdrop, soft top key light, subtle contact shadow under the sole. Do not change the colourway or the laces. | Clean PDP framing and a named shadow — the contact shadow is what keeps the shoe from floating; the lock stops the model "improving" the design. |
| Show the same perfume bottle from a low three-quarter angle on wet slate, one hard rim light from behind. Keep the label text sharp and legible. | Leans on Nano Banana’s text-survival strength and asks for one clear light, so the label stays readable instead of turning to mush. |
Background swaps
| Prompt | Why it works |
|---|---|
| Keep the subject exactly as-is. Replace the background with a blurred rooftop at golden hour, and match the warm rim light on her shoulders to the new sunset. | The "match the rim light to the new background" clause is the whole trick — it’s what makes a swap read as one photograph rather than a cut-out. |
| Cut the person out and drop them into a snowy street at dusk. Add cool ambient light on the skin and a faint breath fog. | Environmental light spilling onto the subject sells the composite; the breath fog is a small realism cue the model handles well. |
| Move this product from the white background to a marble kitchen counter, and keep the original soft shadow direction. | Preserving the existing shadow direction avoids the tell-tale floating-object look you get when the light suddenly comes from nowhere. |
Style changes
| Prompt | Why it works |
|---|---|
| Redraw this photo as a clean flat-vector illustration, bold outlines, a limited five-colour palette. Keep the pose and composition. | Names a concrete style and preserves the layout, so the result is recognisably the same subject — not a random illustration that happens to rhyme. |
| Turn this portrait into a soft watercolour with visible paper texture. Keep the likeness readable. | Asks for the medium and its texture, then guards the face so the person survives the stylisation. |
| Restyle as 1970s film photography: warm grain, slight halation, faded blacks, same framing. | Era-specific grade cues ("halation, faded blacks") beat a vague "vintage," which the model would average into nothing in particular. |
Character and product consistency
| Prompt | Why it works |
|---|---|
| Using these three reference photos of the same woman, place her in a cafe reading a book. Keep her face, hair and freckles consistent. | Multi-reference anchoring is the model’s core strength; naming specific features (freckles) gives it something concrete to hold onto. |
| Same character as the last image, now standing in the rain holding an umbrella, same outfit and hairstyle. | Conversational continuation leans on the running session context, so you extend a character instead of re-describing them from zero. |
| Keep this mascot’s proportions, colours and logo exactly. Show it waving on a plain background, framed for a die-cut sticker. | Brand-asset consistency: protect the defining traits by name and the model reposes the character without redrawing its identity. |
Save the prompts that land — with their why — in a notes file. After a few projects you’ll have a personal library that beats any generic prompt pack, because it’s tuned to the edits you actually make.
The turn-by-turn editing workflow
The library gets you a strong first render. The quality comes from what you do next. Treat every image as a starting frame and direct it in small moves rather than firing a fresh prompt each time.
- Start with one clean instruction. Name the change, the protection, and the light. Resist stacking five edits into the opening prompt — you won’t know which one caused what.
- Read the render, then change one variable. "Push the light warmer" or "lower the camera" — one note at a time, so you learn what each move does to the image.
- Repeat the protection on every follow-up. Keep saying "hold the face and pose" so an unrelated edit doesn’t quietly drift the subject two turns later.
- Correct regionally when you can. If only the background is wrong, ask for the background. Re-rolling the whole image gambles the parts that were already right.
- Anchor identity with a clean reference. Feed a sharp, well-lit reference of the person or product so the model matches it instead of inventing a convincing lookalike.
- Keep the render you like before a bold move. Save the good frame first — an ambitious edit can lose it, and you want a fallback to return to.
Where the prompts stop working
Consistency is reliable, not guaranteed. A few situations will fight you no matter how clean the prompt, and it’s better to know them going in than to burn ten tries:
- Big lighting swings. Day to night, or hard flash to soft daylight, can shift skin and colour enough that the face reads slightly different. Move in stages instead of one jump.
- Heavy multi-image blends. Stacking many references or several people raises the odds of a merged feature or a lost detail. Fewer, cleaner inputs beat a big pile.
- From-scratch art. Ask for a complex illustrated scene with no input photo and you’re outside the sweet spot — a from-scratch aesthetic engine like Midjourney v7 will usually beat it there.
- Tiny text and exact logos. In-image text is a genuine strength, but very small type or a precise brand logo can still garble. Check it at full size and budget a retry.
Let a studio hold the craft
Raw Nano Banana gives you the engine, but you still supply the look on every single prompt. That’s the part an image agent removes. ReelWand’s Illustration Canvas carries a permanent style DNA on the server — the medium, the line and colour language, the finish — assembled into every request before it reaches the model. So your prompt stays short (name the change, protect the subject) and the render lands in your look without re-typing the vocabulary. Session memory means each instruction iterates on the previous frame inside a 2-hour window — directing, not re-rolling — and because the style lives server-side, the recipe never leaks through your prompt. Stills sit below video on the credit system, so trying ten versions of an edit costs little.
Name the change, protect the subject, and let the agent carry the style.
Run your prompts in the Illustration CanvasFrequently asked questions
What are the best Nano Banana prompts?
The best prompts are plain-language edits, not keyword piles: name one change, say what to protect ("keep her face and pose"), and describe the light ("soft window light from the left"). That structure works across portraits, product shots, background swaps and style changes. Paste one from the library above and adjust a single variable at a time.
How do you write a good Nano Banana prompt?
Write it as an instruction to a retoucher, with three parts: the change, the protection, and the light. Make one scene change per turn, explicitly fence off the face, hands or product you want held steady, and describe the light direction and quality rather than just naming a location. Skip "8k, cinematic, masterpiece" tags — the model composes, it doesn’t rank.
Why does Nano Banana change the face when I edit a photo?
Usually because the prompt didn’t protect it and tried to do too much at once. Add an explicit protect-clause — "keep her face and pose exactly" — and make one change per instruction. Big lighting swings (day to night) and heavy multi-image blends are the situations most likely to drift a face, so move in small stages and anchor with a clean reference photo.
Can Nano Banana keep the same character across multiple images?
Yes — character consistency is its defining strength. Feed clean reference photos of the same subject and name specific features to hold (hair, freckles, an outfit), and it carries the likeness across poses and lighting. On the Pro tier it can blend a larger stack of references and keep several people consistent, though fewer, sharper inputs tend to hold better than a big pile.
Does Nano Banana work without an input photo?
It can generate from a text prompt, but that’s not where it shines. Nano Banana is an editor first — its edge is re-rendering a real photo while keeping the subject consistent. For a complex illustrated scene from nothing, a from-scratch aesthetic engine like Midjourney usually wins; reach for Nano Banana when you have a photo to change.
Put it into practice
Specialized image agents carry the craft this guide describes. Pick an available agent and start creating.
Part of ReelWand's AI Art & Illustration Tools tools.