Skip to content
Gemini Link
English简体中文繁體中文Čeština한국어Türkçe

Nano Banana 2

Nano Banana 2 — Google’s Gemini 3.1 Flash Image — combines Flash-speed generation with world knowledge, legible text rendering and consistent subjects across image creation and editing.

AI Image

Generated content will appear here

Examples

Photorealistic image of a corgi wearing a tiny cowboy hat, sitting on a sunny lawn

From Google's Gemini image-generation gallery: a photorealistic corgi in a cowboy hat, demonstrating natural light and fur texture at Flash speed.

Google Gemini official image-generation gallery — vendor demonstration, not generated by this site · Official vendor demo artwork, reproduced for factual reference; no third-party result is presented as this site's own test.

Concept car built from ramen noodles, springy noodles forming the body, a soft-boiled egg as the headlight and chashu pork as the spoiler, next to the original line-sketch reference

Reference-driven material transfer: the same line sketch of a car re-imagined with ramen as the source material, from Google's style gallery.

Google Gemini official image-generation gallery — vendor demonstration, not generated by this site · Official vendor demo artwork, reproduced for factual reference; not generated by this site.

The same concept-car sketch re-imagined as a translucent pink jellyfish vehicle, next to the sketch and a jellyfish photo reference

The control case: identical sketch, different reference material — a jellyfish photo instead of ramen. Comparing the two shows how tightly the model follows the reference while keeping the composition fixed.

Google Gemini official image-generation gallery — vendor demonstration, not generated by this site · Official vendor demo artwork, reproduced for factual reference; not generated by this site.

The word HELLO rendered as glossy 3D letters with an iridescent pink and blue gradient on a black background
Example prompt
The word "HELLO" as glossy inflated 3D letters, iridescent pink-to-blue gradient chrome, studio reflections, on a pure black background.

Text rendering showcase: HELLO as glossy 3D lettering, the model's signature legible-typography ability.

Google Gemini official image-generation gallery — vendor demonstration, not generated by this site · Official vendor demo artwork, reproduced for factual reference; not generated by this site.

What Nano Banana 2 is

Nano Banana 2 is Google's Gemini 3.1 Flash Image model (API ID gemini-3.1-flash-image) and the general-purpose member of the four-model Nano Banana family. It previewed on February 26, 2026 and reached general availability on May 28, 2026 alongside Nano Banana Pro. Its pitch is the best of both lineages: Pro-level world knowledge, legible text rendering and multi-character consistency, delivered at Flash speed, with 1K and 2K output generally available, 4K in preview and a new 512 px tier for rapid iteration. Google serves it through the Gemini API and AI Studio (paid key required), Vertex AI for enterprise, Antigravity and Firebase, and in the Gemini app and Search. In this workspace it is served from this site's own runtime catalog, which currently lists image-to-image alongside text-to-image modes.

Model facts

  • Gemini 3.1 Flash Image official model identity API ID gemini-3.1-flash-image
  • 512 px – 4K output resolution tiers 1K and 2K generally available; 4K still in preview
  • 14 reference images per prompt png, jpeg, webp, heic, heif
  • 131,072 / 32,768 max input / output tokens
  • January 2025 knowledge cutoff later facts come from web-search grounding, not training data

Capabilities at a glance

The documented strengths of the model map onto the editing workflows this workspace exposes. Which modes are callable at any moment is decided by the live catalog.

World knowledge

Draws on Gemini knowledge plus real-time search grounding to render specific subjects, infographics and data visualizations accurately.

Text rendering and translation

Produces legible text for posters, cards and mockups, and can translate and localize text inside an image.

Subject consistency

Keeps up to five characters and ten objects faithful within one workflow, per Google, which is what makes storyboards viable.

Instruction following

Adheres strictly to complex, multi-layered prompts; a configurable thinking level lets it reason first on hard briefs.

Production specs

Native aspect ratios including 4:1, 1:4, 8:1 and 1:8 — fourteen ratios from 1:1 through 21:9 — and resolutions from 512 px to 4K.

Content credentials

Google states every generated image carries C2PA Content Credentials and an invisible SynthID watermark.

Which Nano Banana model to use

Google's own documentation splits the family four ways, and that split is the most useful thing to know before spending a credit. The table quotes what each model is for and what it is not for.

ModelWhat it isChoose it whenDocumented limits
Nano Banana 2 Lite (gemini-3.1-flash-lite-image)Fastest and cheapest family memberDrafting, volume and A/B variants1K output; not optimized for multiple reference inputs or multi-turn sequential editing; no search grounding
Nano Banana 2 (gemini-3.1-flash-image)General-purpose workhorseEveryday generation and editing, 2K–4K, multiple references4K is still in preview
Nano Banana Pro (gemini-3-pro-image)Premium tierHighest world knowledge, advanced localization, brand consistency, precision controlHighest cost; 65,536-token input ceiling
Nano Banana (gemini-2.5-flash-image)Legacy pioneerNothing new — Google recommends migratingSuperseded by Nano Banana 2 Lite

Production specifications

Google's model documentation publishes the technical envelope. Treat it as the model's capability; the workspace may expose a subset per session, and the catalog remains authoritative.

SpecificationPublished value
Official model IDgemini-3.1-flash-image
Launch stagePreview February 26, 2026; generally available May 28, 2026
InputsText, images, PDFs, and video (preview)
OutputsText and images
Max input tokens131,072
Max output tokens32,768
Reference images per promptUp to 14
Reference image formatspng, jpeg, webp, heic, heif
Document inputsUp to 50 MB per file via API or Cloud Storage; 7 MB via console upload
Aspect ratios1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9, plus 1:4, 4:1, 1:8, 8:1
Resolutions512 px, 1K, 2K; 4K in preview
Search groundingSupported
Knowledge cutoffJanuary 2025
Content credentialsC2PA Content Credentials and SynthID watermark

Configurable thinking levels

Nano Banana 2 exposes a reasoning control earlier image models did not have: Minimal, the default, renders immediately, while High and Dynamic let the model reason through a prompt before drawing. Google positions this as the main lever between latency and prompt adherence, and it is where long, multi-constraint briefs change the most.

The practical rule: stay on Minimal for volume work, and raise the level when one brief carries several simultaneous constraints — exact text, a fixed layout, and a subject that must stay identical.

How to use it in this workspace

Four steps keep image work predictable, whether you start from text or from a reference you are allowed to process.

  1. Pick Nano Banana 2 in the workspace and choose the mode that matches your input: text for new artwork, image mode when the subject or composition is already fixed.
  2. Start at 512 px or 1K to settle the direction cheaply, then re-run the winner at 2K or 4K once the composition is right.
  3. Review the server quote for your selected settings, then submit and inspect the returned image for text, identity and composition errors.
  4. Iterate with edit instructions instead of re-rolling from scratch; the ledger records the final charge for the completed task.

Prompt patterns that work well

Two patterns come from Google's own prompting guide for this family. Copy them into the workspace and adapt the bracketed parts.

Five-part generation formula
[subject] [action], in [location or context], [composition: framing, angle, distance], [style: medium, lighting, mood].
Reference-driven relationship instruction
Using the attached [sketch] as the structure and the attached [material] as the texture, transform this into a high-fidelity [outcome]. Place it in [new scene].

Three rules from that guide apply to every brief here: describe what you want rather than what you want to avoid — an empty street, not no cars; control the camera with photographic terms such as low angle or aerial view; and iterate in conversation instead of restarting from scratch.

How this image model relates to Gemini 4 Argon

Nano Banana 2 is an image model served from this site's own runtime catalog. It is separate from Google's Gemini 4 Argon text-generation family announced on October 13, 2026: Gemini 4 Argon is a separate chat model and is callable here when the live catalog returns it as available and priced. What you can call today is always what the live catalog returns.

Questions

Can I start from an existing image?

Yes. The current site catalog lists an image-to-image mode alongside text-to-image. The live workspace remains authoritative if availability changes.

How is the charge determined?

The server quote reflects the selected runtime settings, and the completed-task ledger records the final deduction.

Is this model part of Google's Gemini 4 Argon?

No. Nano Banana 2 is the Gemini 3.1 Flash Image model, part of Google's Gemini 3.1 image line. Gemini 4 Argon is Google's newly announced text-generation family, first model Argon; it is callable here through the live catalog.

What can it do with text inside images?

Google documents accurate, legible text rendering for mockups and cards, plus in-image translation and localization — you can regenerate a sign or poster with its text carried into another language.

How many characters or objects can stay consistent?

Google's announcement cites up to five characters and fourteen objects held faithful within one workflow, which is what makes storyboarding viable.

Which sizes and aspect ratios are supported?

The model documentation lists resolutions of 512, 1K, 2K and 4K (preview) and fifteen aspect ratios from 1:1 through 21:9, including vertical 9:16 and extreme 1:8. The workspace may expose a subset; the catalog decides.

How many reference images can one prompt take?

Google Cloud's documentation allows up to 14 reference images per prompt within a 131,072-token input window.

Does it know current events or specific people?

The model combines Gemini world knowledge with real-time Search grounding, which is why it can render specific subjects and up-to-date visual information more accurately than earlier image models.

What is the knowledge cutoff?

January 2025 for Nano Banana 2 and Nano Banana Pro alike, per Google. Anything newer that appears in an image comes from web-search grounding rather than from training data.

Do generated images carry content credentials?

Yes. Google states that every generated image includes C2PA Content Credentials and a SynthID watermark. Both travel with the file, so treat them as part of the output when you reuse it.

Which Nano Banana model should I pick?

Follow Google's own split: Lite for speed and volume, Nano Banana 2 as the general-purpose default with 2K–4K output and multiple references, Pro when the highest world knowledge, localization and brand consistency matter. Legacy Nano Banana users are advised to migrate to Lite.

Explore next