Nano Banana 2 Prompting: The Practical Guide

Nano Banana 2 — Google's Gemini 3.1 Flash Image model — earned its reputation on two things: it renders text that stays legible, and it follows references instead of merely tasting them. Google's own prompting guidance describes the first and quietly skips the second. This guide is the version you can actually work from.

Every example below runs as-is in Imgo's image generator with Nano Banana 2 selected, and the method carries to the prompt guides in the library.

The anatomy of a Nano Banana 2 prompt

A prompt that lands first-try has five blocks, and the order matters more than the wording:

Block	What it does	Example
Subject	Who or what, with the two or three details that identify it	a matte-black ceramic pour-over kettle, 1 L, wooden handle
Scene	Where it sits and what the light is doing	on a linen cloth beside a north-facing window, soft morning light from the left
Style	The rendering register	photorealistic product photography, shallow depth of field
Technical	Framing and output discipline	3:4 vertical, product fills 80% of frame, no props
Variables	The bracketed parts you swap between runs	[product] [surface] [time of day]

Most failed prompts are missing the technical block — the model chooses a framing for you, and it will not be the framing you needed.

Reference images: the real superpower

Attach a reference and Nano Banana 2 treats it as evidence, not vibes. Three rules make references reliable:

1. One reference per job. A product shot attached to a style shot confuses which is which. Two references of the same subject from different angles are fine; two references of different subjects are not.
2. Say what the reference is for. "Using the attached kettle as the exact product, place it…" beats an unlabeled attachment on every model, this one included.
3. Keep the reference stable across a set. The same photo plus the same prompt with different [scene] variables is how a whole range ships consistent.

Text in images: where it wins, where it still fails

Short text is the model's documented strength: a three-word headline, a product label, a price tag render cleanly. Long copy is still a losing game — a paragraph of body text will come back with plausible-looking nonsense letters. The working split: generate the headline and the label in the image; set the paragraphs and the fine print in your design tool afterwards, where they are editable and crisp.

Editing: the prompt is an instruction, not a wish

In the editor, Nano Banana 2 takes edit instructions literally. "Make the background white" recuts everything; "replace the background with pure white, keep the product and its shadow untouched" keeps what you came for. Name what stays before you name what changes — that single habit prevents most re-runs.

Three prompts to steal

1. The white-background product shot

Using the attached photo as the exact product: [product name] centered on a pure white background (RGB 255,255,255), product fills 85% of the frame, soft even studio light, no props, no text, 1:1.

2. The character portrait with a fixed identity

Same character as the attached reference: [name], [one identifying detail], now [action] in [setting], consistent face and outfit, [light], photorealistic, 4:5.

3. The poster with real headline text

Vertical music festival poster for "[festival name]", headline "[three-word headline]" in bold condensed type at the top, one headline act name below it, textured paper background, risograph palette, empty lower third for lineup type, 2:3.

What to check before accepting a result

Labels and logos first — zoom to 200% and read every character. Then silhouette (does the product still look like your product), then edges of the edit (did the change leak outside the requested area). A result that passes those three survives contact with a real buyer.

The method above is the short version. The full prompt guides break it down per task — products, editing, references, characters, posters — with the inputs and checks for each, and the model page covers where Nano Banana 2 sits against Nano Banana Pro and GPT Image 2 in the three-way comparison.