Nano Banana 2 Complete Guide

Updated: 2026-08-21. Model IDs, resolution ceilings, and billing follow the Gemini image generation docs for that day.

Introduction

Nano Banana 2 is what product copy often calls Gemini 3.1 Flash Image, commonly tagged in the API as gemini-3.1-flash-image (confirm in docs that day). It is the family’s balanced workhorse: stronger multi-reference and multi-turn than Lite, better daily throughput than jumping straight to Pro; docs permitting, roughly 4K. This guide is about shipping images reliably.

What this guide solves

• Where Nano Banana 2 sits—and when to switch to Lite / Pro
• A reproducible workflow: composition first → local edits → then resolution
• Aspect and resolution settings that raise on-image text success
• A delivery checklist that catches common failure modes

Where Nano Banana 2 fits

Treat it as the default image tier: article heroes, marketing key visuals, multi-turn character/product polish, jobs that need reference-driven consistency. Quality and speed land in the daily delivery sweet spot; multi-ref and multi-turn are steadier than Lite; resolution can reach HD / ~4K when docs allow; cost usually sits above Lite (pricing that day).

Product name, UI label, and API model ID may not match character-for-character. In specs, record: entry URL + on-screen name + doc model ID + date. Identity checks: What is Nano Banana?.

Choosing among Lite, 2, and Pro

Dimension	Nano Banana 2 Lite	Nano Banana 2 (this guide)	Pro
Role	Faster, cheaper draft engine	Balanced workhorse	High-precision hard jobs
Common ID hint	gemini-3.1-flash-lite-image	gemini-3.1-flash-image	Pro image ID per docs that day
Resolution bias	~1K oriented	Can push HD / ~4K	Prefer for stacked detail & constraints
Multi-ref / complex multi-turn	Not emphasized	Stronger	Best when hard constraints pile up
Recommended use	Batch layout probes, A/B sketches	Daily final driver	Final polish, hard-acceptance frames

Fast decisions:

1. Need layout direction across many versions → Lite first: Lite guide.
2. Need finals, multi-ref consistency, higher resolution → use 2.
3. Ultra-complex layout, tiny product structure, many hard constraints at once → step up to Pro, or “2 locks composition → Pro final polish.”

Generate + edit workflow (reproducible)

The goal is not a novel-length prompt—it is one verification per round.

Step 0: Write the acceptance card (before prompting)

In notes, write four lines, then open chat:

1. Use: hero / poster KV / ecommerce main / slide art
2. Canvas: aspect (e.g. 16:9), title safe area or not
3. Subject & count: how many objects, facing, faces shown
4. Bans: no watermark, no garbled text, no extra limbs, no unauthorized brands

Round 1: Composition only

Prompt short in this order: use → subject → framing/lens → light & color → material/style → on-image text (optional, often skip) → bans. Don’t dump every style adjective at once.

Use: 16:9 blog hero, right side reserved for a title.
Subject: a matte ceramic desktop speaker, centered slightly left.
Camera: eye-level, ~50mm, soft background bokeh.
Light: window soft light; cool gray + warm wood.
Style: clean product photography—not illustration.
Ban: watermarks, borders, extra cables, warped buttons, any on-image text.

Accept only: correct subject, correct aspect, enough margin, no broken structure. Wrong composition → full redraw, not local patches on a bad base.

Round 2: Change one variable

State what stays fixed, then the change:

Keep the same camera, subject pose, palette, and background.
Do only one thing: change the speaker finish from matte ceramic to brushed aluminum.
Do not change composition; do not add text or logos.

On multi-turn edits, Nano Banana 2’s edge is steadier constraint following. Changing light, material, text, and background in one go makes failures unattributable.

Round 3: Add references (when consistency matters)

Upload 1–several licensed refs and say what each controls:

Image 1: product geometry and button layout (must match).
Image 2: brand primary colors (palette only—do not copy packaging copy).
Generate: same product on a dark desk, 16:9, still no on-image text.

Don’t ask to imitate a living artist’s unique style, and don’t upload unlicensed assets. Product shots must be checked against source files for logo and structure.

Round 4: Raise resolution and final check

Only after composition and details pass, raise to the target resolution options the docs support (including ~4K if available that day). Higher resolution does not fix wrong composition or wrong glyphs—it magnifies them.

Zero-to-one flow: Getting started. Prompt craft: Prompt playbook.

Aspect ratio and resolution

Common aspects: article/video covers 16:9 (title safe zone), social square 1:1, vertical 9:16 or 3:4, slides 16:9 or 4:3.

• Lock aspect before describing the subject; put aspect early in the prompt.
• Draft at lower resolution or on Lite; finalize on 2 at higher res—price and quota per docs that day.
• Before delivery, inspect thumbnail and full size. Exact pixel tiers follow UI and the image generation docs—no frozen pixel table here.

Raising on-image text success

Nano Banana 2 can draw letters; it is not a print typesetting engine.

1. Keep copy short: title-length phrases; don’t expect legal microcopy, QR codes, or dense tables to nail in one pass.
2. Quote exact text and state language/case: White sans title top-right: "Nano Banana 2".
3. Generate without text, then type in design tools when spelling must be perfect, languages mix, or brand VI is strict.
4. Edit text in its own round: Keep composition and subject; change title only to: "…". Don’t casually rewrite the background.
5. Final check: character-by-character against source copy—mixed scripts, numbers, and trademark symbols fail most often.

Delivery checklist

• Entry and model identity logged (UI name / doc ID / date); you confirmed Nano Banana 2, not Lite
• Aspect matches use; title/crop safe area is enough
• Subject count, hands/edges, reflections and shadows look intact
• Each multi-turn round changed one variable and stated what stays fixed
• Reference roles explained; usage rights recorded
• On-image text proofed character-by-character, or switched to “blank plate + post layout”
• No surprise logos, watermarks, signatures, unauthorized brands
• Final pass at target resolution; thumbnail and full size both reviewed
• Prices, quotas, and model IDs are not hard-coded into external promises (docs that day)

Quick access

• China-friendly third-party: Nano Banana 2
• Official Gemini: gemini.google.com
• Image docs: ai.google.dev image generation

Third-party ≠ Google’s account system; read that entry’s privacy/data notes before uploading commercial assets. China details: China access.

FAQ

Are Nano Banana 2 and Gemini 3.1 Flash Image the same?

Product narratives often map Nano Banana 2 to Gemini 3.1 Flash Image / gemini-3.1-flash-image. Marketing labels and API IDs can diverge slightly—trust your entry’s model picker and official docs.

When must I move to Pro instead of staying on 2?

When many hard constraints stack (precise product structure + multi-ref consistency + complex typography + fine materials) and 2 still misses the acceptance line after two solid rounds, Pro—or “2 locks composition → Pro polish”—usually saves total time.

Can I demand “4K + perfect text + all refs followed” in one round?

Not recommended. Split composition → reference constraints → text → resolution so failures are local and quotas stretch further. If multi-turn drifts, return to a correct composition branch or start a new chat with the good frame as reference.

What are the price and rate limits?

No frozen unit prices or quotas here. See the image generation docs and your account pages for that day.

Official resources

• Gemini
• Gemini API image generation
• Google AI for Developers

Further reading

• Nano Banana guides hub
• What is Nano Banana?
• Nano Banana 2 Lite
• Getting started
• Prompt playbook

Summary

Nano Banana 2 is the default daily driver: steadier multi-ref and multi-turn than Lite, with a higher resolution ceiling when docs allow. Drive prompts from an acceptance card, lock composition before single-point edits, postpone text when you can, and run the checklist before delivery.