Nano Banana 2.1 Prompt Guide: Generate, Edit and Combine Images

All guides · Gemini Image in Fuser

Quick answer: To prompt Nano Banana 2.1 in Fuser's Gemini Image node, where it is the default model, describe the scene in full sentences, not a list of keywords. Name the subject, the setting, the light, the camera angle and what the image is for. To edit, pass the image back in and say what to change and what must stay the same. To combine images, refer to each one by its order ("the logo from the second image"). For text, put the exact words in quotation marks and describe the lettering. Google's own advice is to "describe a scene in rich detail. The more specific you are, the more control you have over the results" (Gemini API docs).

Which Nano Banana is which

"Nano Banana" is Google's name for Gemini's native image generation, and it now covers five models (Gemini API docs). In Fuser they sit in one Gemini Image node, chosen from the Model dropdown:

• Nano Banana 2.1 (gemini-nano-banana-2.1) is the default. Google calls it the primary high-efficiency workhorse for generation and conversational editing: an update to Nano Banana 2 with improved visual quality, text rendering, multi-turn consistency and Google Search grounding across 1K, 2K and 4K (Gemini API docs).

• Nano Banana Pro (Gemini 3 Pro Image) is Google's premium choice for the most complex work, with the most world knowledge, advanced localisation, accurate brand consistency and precise creative control (Gemini API docs).

• Nano Banana 2 (Gemini 3.1 Flash Image) is the previous generation. Google deprecated it on 6 October 2026, when Nano Banana 2.1 launched as its replacement, and as of 23 October 2026 has announced no shutdown date (Gemini API deprecations). Switch saved nodes to 2.1; what changed in Nano Banana 2.1.

• Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) is the fastest and cheapest. It renders at 1K only, and Google notes it is not optimised for multiple reference images or chained edits.

• Nano Banana (Gemini 2.5 Flash Image) is the legacy model. Google recommends moving to Nano Banana 2 Lite.

Everything below was tested on Nano Banana 2.1 on 6 October 2026 unless we say otherwise. Fuser calls Google's API directly for this node; our test images were generated with the same model versions (gemini-nano-banana-2.1, and gemini-3-pro-image for the text comparison) at 1K.

Write the scene, not a tag list

We ran one subject two ways. The scene prompt:

"A high-resolution, studio-lit product photograph of a speckled cream ceramic pour-over coffee dripper standing on a pale travertine block. Soft window light from the left casts a long, gentle shadow to the right; the background is a warm sand-coloured plaster wall. The camera angle is a slightly elevated 30-degree shot that shows the spiral ridges inside the cone. Ultra-realistic, with sharp focus on the glaze speckles."

Run this prompt in Gemini Image →

Every detail came through: the speckles, the travertine, the plaster, the light direction and the ridges inside the cone. The keyword version still produced a clean photo, but the model filled the gaps with its own choices: a matching carafe under the dripper, a paper filter in the cone and a plain, unspeckled glaze. Keywords don't fail; they just hand the art direction to the model.

That prompt follows Google's product-photography template almost word for word (Gemini API docs):

1. Shot type and subject. "A high-resolution, studio-lit product photograph of [product]".

2. Surface and setting. "on a [background surface/description]".

3. Light, with a purpose. "[lighting setup] to [what it should do]".

4. Camera. "a [angle] shot to showcase [feature]". Google suggests photographic terms such as wide-angle, macro and low-angle (Google Cloud best practices).

5. Focus. "sharp focus on [key detail]".

Three more rules from Google's best practices are worth keeping in your head (Google Cloud best practices):

• Say what the image is for. "A logo for a high-end, minimalist skincare brand" beats "a logo".

• Describe what you want, not what you don't. Instead of "no cars", write "an empty, deserted street with no signs of traffic". The Gemini Image node has no negative-prompt field, so this is how you exclude things.

• Split crowded scenes into steps. "First, create a misty forest at dawn. Then, in the foreground, add a stone altar. Finally, place a glowing sword on the altar."

Google also advises phrasing the request as "create an image of" or "generate an image of" so the model doesn't answer in text (Google Cloud best practices). In Fuser you can skip that: the node always asks Gemini for an image.

Edit by changing one thing

Connect an image to the node's Image input and the prompt becomes an edit instruction. Google's pattern for a targeted edit is "change only the [element] to [new element]. Keep everything else in the image exactly the same" (Gemini API docs). Our edit:

"Using the provided image, change only the travertine block to a dark walnut wood block with visible grain. Keep the dripper, the window light, the shadow and the plaster wall exactly the same, preserving the original style, lighting and composition."

Run this prompt in Gemini Image →

The dripper, the speckles, the wall texture and the shadow direction held; only the block changed. Two habits make this reliable:

• Name what must stay. Listing the things to keep ("the dripper, the light, the shadow") does more than a general "keep everything else".

• One change per pass. Google recommends iterating in small steps ("make the lighting warmer") rather than rewriting the whole prompt (Google Cloud best practices). In the Gemini API that is a multi-turn chat. In Fuser, you wire the output into another Gemini Image node, so every version stays on the canvas and you can branch from any of them.

Combine several reference images

The same Image input takes up to 10 images in Fuser. Refer to them by order, the way Google's examples do: "the dress from the first image", "the woman from the second image" (Gemini API docs). For the third panel above we passed the dripper as the first image and our generated logo as the second:

"Take the speckled ceramic coffee dripper from the first image. Add the circular KILN & CO logo from the second image to the front of the dripper as a small stamp pressed into the clay, in dark brown underglaze, following the curve of the cone. Ensure the dripper's shape, glaze speckles, the travertine block, the lighting and the background remain completely unchanged."

Run this prompt in Gemini Image →

The logo landed on the front of the cone as a pressed stamp, lettering and flame intact, and the rest of the shot stayed the same. This is Google's "high-fidelity detail preservation" pattern: describe the thing that must not change in detail and say so explicitly.

How many references each model handles well, per Google: Nano Banana 2.1 keeps the likeness of up to 4 characters and the fidelity of up to 10 objects, from up to 14 reference images (model card); Nano Banana Pro handles 5 images with high fidelity; the legacy 2.5 model works best with up to 3 (Gemini API docs). For character work across many images, see consistent characters with AI.

Get text right

Google's text template is: create a [image type] for [brand] with the text "[text]" in a [font style], with a [style] and a [colour scheme] (Gemini API docs). What mattered in our runs:

• Quote every string exactly. Our logo prompt asked for "KILN & CO" and got it, ampersand included, on the first try.

• Describe the lettering, not a font name. "Hand-lettered white chalk capitals", "clean, bold, geometric sans-serif".

• Say where each piece of text goes. The menu prompt named a heading, four lines and prices aligned right, and both models followed that layout.

• Watch the background. Nano Banana Pro set our menu in a busy café with a second, unrequested menu board behind it, and drew a cup where we asked for a dripper. If stray text is a problem, describe a plain setting, such as "a plain white tiled wall with nothing else on it".

Put your own copy in the quotes and run the text template in Gemini Image →

Google positions Nano Banana Pro for professional asset production, while listing better text rendering and infographic layout as Nano Banana 2.1 gains (model card). On this five-string menu both models spelled everything correctly, so try Nano Banana 2.1 first and switch when the layout gets dense. For signs inside a photographic scene, try GPT Image: in our Nano Banana 2.1 vs GPT Image 2.5 test both spelled a poster correctly, but only GPT Image 2.5 Flare rendered readable shop signs in a night street scene. Google also notes the model does best when you settle the wording before asking for the image (Gemini API docs), so write your copy first and paste it in. Our best AI models for text in images compares other models on the same job.

Settings in Fuser

The Gemini Image node exposes:

• Model: Nano Banana 2.1 (default), Nano Banana 2, Nano Banana 2 Lite, Nano Banana Pro, or the legacy Nano Banana.

• Image: up to 10 reference images for edits and composites.

• System Prompt: standing instructions for style or behaviour, kept separate from the per-image prompt.

• Aspect ratio: Auto, 1:1, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5, 16:9, 9:16 or 21:9. Google says the output follows the input image's size by default, or is square when there is no input (Gemini API docs), so set a ratio when you want a different shape.

• Resolution: 1K, 2K (default) or 4K. Nano Banana 2 Lite always renders at 1K, and Fuser warns you if you pick more.

• Search Grounding: lets Nano Banana 2.1, Nano Banana 2 and Nano Banana Pro use Google Search for current information such as weather, scores or recent events. It is not available on Lite or the legacy model.

• Output format: JPEG (default) or PNG.

Google adds a SynthID watermark to every generated image, and lists the languages that perform best, including English, Spanish, French, German, Japanese, Korean, Hindi, Portuguese and Chinese (Gemini API docs).

Build it as a workflow

The hero image above is the workflow this guide used: one prompt node per idea, a Gemini Image node per step, and every edit wired from the output it builds on. From there you can send the finished product shot to an image-to-video model such as Veo or Kling 3.0; product photo to video ad walks through that chain. To see how Nano Banana compares with other models on the same prompts, read Nano Banana vs GPT Image, Flux vs Nano Banana and Seedream vs Nano Banana, or the wider best AI image editing models.

Copy-paste Nano Banana prompts

Six Nano Banana prompt examples for common jobs, each run once on Nano Banana 2.1 on 23 October 2026. Paste one into a Text node, wire it into Gemini Image and swap the subject for yours. Prompts 3 and 6 are edits, so connect the earlier outputs to the node's Image input first.

1. Product shot. "A studio product photograph of a compact tabletop radio with a pale warm-grey plaster-textured body, a perforated aluminium speaker grille and a single cobalt-blue tuning dial, standing on seamless off-white paper. Soft north-window light from the left gives a gentle falloff and a short, soft shadow to the right. Shot straight-on at eye level on medium-format film, with the radio in the centre of a square frame and generous space around it. Sharp focus on the dial's knurled edge. No logos or text anywhere." Result: Plaster body, grille and cobalt dial as written, with the shadow falling to the right.

2. Logo with text. "A logo for a small, high-end audio company called LOW TIDE AUDIO. A simple circular mark of three horizontal wave lines inside a thin circle, above the words "LOW TIDE" in a clean, bold, geometric sans-serif and "AUDIO" in small, widely spaced capitals underneath. Cobalt blue on a flat off-white background, centred, with nothing else in the image." Result: Both lines spelled correctly on the first run.

3. Two-image composite. Connect the radio first and the logo second. "Take the radio from the first image. Print the LOW TIDE AUDIO logo from the second image onto the radio's front panel, below the cobalt dial, as a small, crisp cobalt-blue screen print that sits flat on the plaster texture. Keep the radio's shape, the cobalt dial, the speaker grille, the paper backdrop, the window light and the shadow exactly the same." Result: The logo landed under the dial; the radio, backdrop and shadow stayed put.

4. Poster with type. "A square two-colour risograph poster on cream paper for a fictional architecture lecture series. A grainy black-and-white photograph of a concrete stairwell lit by a single skylight fills the top two thirds; a flat terracotta rectangle overlaps its lower-left corner. Below, on a strict Swiss grid and aligned left, the title "ROOMS WITHOUT WALLS" in a heavy grotesque sans-serif, then "Four evening lectures on light and mass" in a lighter weight, then "Thursdays in March · 7 PM · Hall B" in small type. Visible risograph grain and slight misregistration. No other text." Result: All three lines spelled exactly, middle dots included, with the terracotta block where we asked.

5. Editorial portrait. "An editorial portrait of a fictional ceramicist in her sixties with short silver hair and clay-dusted forearms, wearing a faded indigo work jacket, seated beside a pottery wheel in her studio. Soft north-window light from the right, pale plaster walls and shelves of unglazed bowls behind her, slightly out of focus. She looks just past the camera with a calm, direct expression. Shot on 35mm documentary film with natural skin texture, square frame, chest-up." Result: Everything as described, though it framed her waist-up rather than chest-up.

6. Same person, new setting. Connect the portrait to the Image input. "Using the woman from the provided image, show her standing on a windswept shingle beach on an overcast morning, holding a large unglazed bowl against her hip and wearing the same faded indigo work jacket. Keep her face, silver hair, age and expression exactly the same. Soft, even grey daylight, a pale sea and sky behind her, 35mm documentary film look, square frame, waist-up." Result: Her face, hair and jacket carried over; she holds the bowl in front of her rather than against her hip.

Run these prompts in Gemini Image →