Nano Banana in Batch Banana: best uses and how to prompt it

Painted picture-book scene of Fern, a brown hare in an olive coat and leather satchel, setting a paper boat on a pond among giant backlit poppies while four ducklings paddle over

Best for

Nano Banana is Google's family of image models. Nano Banana 2 is the model new accounts start on. It is at its best when you hand it your own images: it reads up to 14 references, keeps a character or product recognizable in a new scene, and follows plain-language edits. Asked to make a chocolate wrapper green, it changed the paper and left the glove, the bench and the sea alone. Nano Banana Pro is Google's highest-quality version.

For long exact copy and dense layouts, try GPT Image. For posters and logos, try Ideogram. For files you can scale and edit, use Arrow. Which model for what shows all five families on the same prompts.

A gloved hand holding a snapped-off square of chocolate above a half-unwrapped cobalt Halvard bar on a ferry bench, a snowy coastline across grey water behind

Flat painted illustration of Fern holding up a glowing paper lantern in a dark jungle of huge leaves, a pair of yellow eyes watching from the shadows

Which version

Model	Best for	Credits per image	Reference images
Nano Banana 2	Pro quality at Flash speed, up to 4K	16	Up to 14
Nano Banana 2 Lite	Google's fastest and cheapest, built for iterating	8	Up to 14
Nano Banana Pro	Highest quality from Google, up to 4K	32	Up to 14
Nano Banana	Fast and lightweight, good for iterating	10	Up to 8

Start on Nano Banana 2. Explore on Lite, at 8 credits and about 5 seconds an image, then run the prompts you like again on Nano Banana 2. Switch to Pro for its finer detail and text, as in an infographic or packaging. Nano Banana is the older model; the others replace it.

The same prompt on each version:

Gouache hare in an olive coat with a bundle on a stick walking a hillside path in the lower left, green hills and a line of birds at dawn

Gouache hare in an olive coat and cap walking with a stick across rolling green hills, a line of birds against a pale dawn sky

Write the prompt

Describe the scene in sentences, not keywords. Google's guide says: "A simple list of keywords won't cut it; you need to describe the scene narratively." Their formula is subject, action, place, composition, style. A keyword list still gives you a picture, but sentences let you say where things go and how the light falls. Our keywords gave a hare in a plaid coat running through the middle of the frame. The sentences put a small hare in an olive coat in the lower left, as asked.

Watercolour hare in a plaid coat and scarf running along a hillside path at sunrise, crows overhead

Gouache hare in an olive coat walking a path in the lower left of a wide landscape of green hills under a pink dawn sky with distant crows

Say what the image is for. With only "A logo", the model invents the business. Name the brand, what it sells and the look you want. Here the bare prompt invented an eagle crest and the tagline "Heritage & Forge". The specific one followed the brief, but still added "Bergen, Chocolate maker" under the name. The next tip covers how to stop that.

Gold eagle-and-shield crest over the word HALVARD and the tagline HERITAGE & FORGE on navy

Cobalt square with cream wavy lines above HALVARD in a tall condensed serif, with BERGEN and CHOCOLATE MAKER beneath, on cream

Put exact words in quotes and describe the lettering. Google: "Enclose your desired words in quotes", then describe the type, such as "bold, white, sans-serif font". Write the final copy yourself. If the model writes the slogan, you'll likely spend credits rewriting it. If nothing else should appear, say so: "No other text."

Give each reference a job, and name your characters. Google's formula for references is the images, then how they relate, then the new scene. Point to each image by what is in it, so the model can't mix them up. Name characters too: DeepMind's guide says to "assign a distinct name to each character or object" so they stay consistent.

Fern is the brown hare in the olive coat. Image 2 is the Halvard chocolate bar. Put Fern on a ferry bench, breaking a square off the bar, in the same ink and watercolour style as Fern's picture.

More tips

When you edit, say what stays. Google's advice is to "be explicit about what to keep exactly the same". Name the change, then list what must not move: the face, the logo, the light. In our test a bare "Make the wrapper green." kept the rest of the scene anyway and picked its own green. The specific edit named the shade and the parts to keep, and got pine green. A list costs nothing and protects the parts you care about on a busier edit.

The ferry-bench photo with the Halvard wrapper turned green, everything else as before

Use camera words, not praise. "Low-angle shot with a shallow depth of field (f/1.8)", "wide-angle lens" and "macro lens" are Google's own examples. Words like "masterpiece" and "ultra detailed" don't describe anything the model can draw.

Rewrite the prompt, don't reply to it. Each Create starts fresh, so the model doesn't remember your last prompt. When a result is close, copy the prompt, change the part that was wrong, and run it again.

Settings

Resolution appears on Nano Banana 2 and Pro. Lite and Nano Banana have no size setting. Every version offers all ten aspect ratios, from 21:9 to 9:16.

Model	Credits per image	Other settings
Nano Banana 2	16	Resolution: 2K 24, 4K 40
Nano Banana 2 Lite	8	Same price
Nano Banana Pro	32	Resolution: 4K 64
Nano Banana	10	Same price

On Pro, 1K and 2K both cost 32 credits, so make it at 2K. On Nano Banana 2, draft at 1K for 16 credits and make the final at 2K for 24 or 4K for 40.

The aspect ratio you pick holds even when your reference images are a different shape. A 16:9 request with a square reference came back 16:9.

Limits

Text can come out wrong. In our comparison, Nano Banana 2 printed copy of its own on a wrapper, and one small line of it is not a word. Check every word before you use an image.
Nano Banana reads up to 8 references; the other three read up to 14.

Sources

Google's own guides: the Nano Banana prompting guide on Google Cloud and DeepMind's Gemini image prompt guide.

Images made in Batch Banana on October 2, 2026. Select Prompt on an image to read the prompt and model that made it. Where a run made more than one image, we show the one we liked best.