Nano Banana 2 Lite is Google’s speed-first member of the Gemini image family. Released on June 30, 2026, it targets high-throughput generation and editing where a four-second draft and a low per-image price matter more than maximum resolution or complex reference control.
The official API model ID is gemini-3.1-flash-lite-image, and Google lists it as stable rather than preview. It is available through Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform, with consumer rollout across products including the Gemini app, NotebookLM, Google Photos, Flow, Search, Stitch, and Google Ads.
Google also recommends it as the replacement for the original Nano Banana model, gemini-2.5-flash-image. “Lite” does not mean a reduced trial surface; it identifies the efficiency specialist in a broader family.
The headline numbers
Google reports that Nano Banana 2 Lite produces text-to-image output in about four seconds and costs $0.034 for a 1K image at launch. Those figures make it suitable for interactive prototypes and pipelines that need many candidates.
The model documentation lists:
- text and image input;
- image and text output;
- generation and image editing;
- a 65,536-token input limit;
- a 4,096-token output limit;
- 1K image output only;
- multiple aspect ratios from square to 21:9;
- Batch API and function calling support;
- minimum and high thinking levels;
- always-on SynthID and C2PA provenance.
It does not support 2K or 4K output. Caching, file search, code execution, Google Search grounding, structured output, and the Live API are not supported on the documented model surface.
How it fits the Nano Banana family
Google now describes four native Gemini image models:
| Model | Best fit | Main tradeoff |
|---|---|---|
| Nano Banana 2 Lite | Fast, low-cost, high-volume iteration | 1K and lighter reference workflows |
| Nano Banana 2 | General-purpose production balance | More cost and latency than Lite |
| Nano Banana Pro | Complex professional control and grounding | Premium tier |
| Original Nano Banana | Legacy Gemini 2.5 model | Migration recommended |
Nano Banana 2 supports 4K generation, strong multiple-reference processing, and a broader balance of quality and efficiency. Nano Banana Pro targets world knowledge, localization, brand consistency, and precision control. Lite is the route for speed, not the automatic answer for every image.
What Lite retains
Google says the model retains reliable prompt adherence, character consistency, and legible in-image text despite its lower price and latency. It can generate new images or edit an uploaded image through a conversational text-and-image interface.
Strong use cases include:
- rapid mood boards;
- thumbnail candidates;
- ecommerce concept variations;
- ad-format drafts;
- UI and game-asset ideation;
- background and color experiments;
- high-volume personalization where 1K is sufficient.
The low price also makes it useful as a routing stage. An application can create several coarse candidates with Lite, have the user choose one, and send only the selected concept to a higher-tier model for final resolution and control.
Where Lite is intentionally weaker
Google’s image-generation guide says Nano Banana 2 Lite is not optimized for multiple reference inputs or multi-turn sequential editing. A project that must preserve the same person, product, logo, and room across many changes may be better served by Nano Banana 2 or Pro.
The 1K output ceiling also matters. A social thumbnail or storyboard frame may be ready at 1K, while a print layout, detailed product crop, or large display will need a higher-resolution source or separate upscaling workflow.
Do not interpret fast generation as guaranteed deterministic generation. Small prompt changes can alter composition, text, or identity. Store seeds or configuration where available, keep the approved source, and compare every revision visually.
Cost beyond the headline price
At $0.034 per image, one thousand generated drafts would cost $34 at the announced rate. But effective cost depends on acceptance rate. If a user keeps one image from ten, the useful-image cost is roughly ten times the single-generation price before storage and review.
Measure:
- generated images per accepted asset;
- latency at p50 and p95;
- policy rejection and technical failure rates;
- edit rounds per final;
- reviewer time;
- escalation rate to Nano Banana 2 or Pro;
- downstream restoration, upscaling, or manual design time.
Batch support can improve throughput, but interactive and batch workloads have different latency expectations. Do not promise the four-second launch figure as a service-level guarantee without your own production data.
High-volume generation also needs asset governance. Assign every candidate a request ID, model ID, prompt hash, aspect ratio, and approval state. Store rejected drafts only as long as policy requires, and prevent a downstream publishing job from selecting an unreviewed image simply because it is the newest file in a folder.
For large variant campaigns, define a diversity budget and a consistency budget. Some fields—background color or crop—may vary, while product shape, legal copy, and brand colors must remain fixed. Automated vision checks can flag obvious deviations, but a human should inspect a representative sample and every high-impact asset. Cheap generation makes overproduction easy; disciplined selection keeps review and storage from becoming the real bottleneck.
Prompting for a speed-first model
Clear constraints reduce waste. Specify subject, composition, viewpoint, lighting, palette, aspect ratio, and any exact text. Avoid packing several incompatible creative directions into one request.
For editing, state what must change and what must remain. “Replace the background with a studio gradient; preserve the person, clothing, pose, and crop” is safer than “make this more professional.”
When exact brand marks or legal text are required, generate the composition and add those elements deterministically later. Even models with strong text rendering can miss a character or reshape a logo.
Safety, rights, and provenance
All generated Nano Banana images include SynthID, and the model page also lists C2PA. These mechanisms help signal origin but do not establish that the prompt, reference image, or depicted identity was authorized.
An application should require users to confirm rights to uploaded photos and brand assets, preserve the edit history, and label generated imagery appropriately. Human review is essential for faces, sensitive contexts, factual diagrams, and ads that make product claims.
Choosing Lite with a routing policy
Use Lite first when the request is low-risk, 1K is enough, and variation is valuable. Escalate when:
- multiple references must remain consistent;
- exact typography or branding is critical;
- a 4K output is required;
- world grounding matters;
- repeated Lite edits drift from the approved concept;
- the asset will receive expensive downstream production.
The best routing policy compares cost per accepted output, not marketing tier names.
From a generated image to a controlled media workflow
Nano Banana 2 Lite can create a clean concept or prepare a visual reference, while a specialized service performs a separate transformation. Medux does not natively run Nano Banana, and Google’s model is not automatically connected to Medux.
For a fashion concept, a team might use Lite to draft a mood board, select an approved garment reference, and then ask Claude to call Medux for an outfit change on an authorized source portrait. The Claude MCP outfit-change tutorial demonstrates the signed-upload, task, polling, and download sequence.
For archives, Lite could generate a color-reference palette or help classify damaged photographs, while Medux performs the actual repair through the Codex MCP old-photo restoration workflow. Generated details should never be presented as historical evidence.
A reliable handoff:
- preserve the original source separately from generated drafts;
- select one approved reference and record its provenance;
- verify face, clothing, and restoration authorization;
- upload only the required assets;
- create one idempotent Medux task;
- compare the result with both source and approved concept;
- require human review before publication.
Lite makes ideation cheap and quick. The production value comes from knowing when a draft is good enough, when to escalate, and when to use an explicit transformation tool instead of asking the generator to approximate the final operation.
That distinction also keeps evaluation honest. Judge Lite on draft speed, useful variation, and cost per approved concept; judge the downstream transformation on fidelity to the authorized source and the requested operation. Combining those scores would hide which stage actually needs a better model, prompt, or review rule.