Nano Banana 2 Lite

Nano Banana 2 Lite Image Editing API on E2X

Nano Banana 2 Lite is Google's gemini-3.1-flash-lite-image, and its editing capability lives at google/nano-banana-2-lite/edit-image on our platform. We charge $0.0238 per edited image — about 43% below what the same model costs on fal.ai. It accepts up to fourteen reference images per call and returns at 1K resolution, which is the whole bargain in one sentence.

No source image to work from? Then you want Nano Banana 2 Lite text-to-image instead — same model, same price, prompt only.

This page is really for one group of readers: anyone whose edit pipeline still points at the original Nano Banana, gemini-2.5-flash-image. Google shuts that model down on 2 October 2026 and says so plainly — "We recommend upgrading to Nano Banana 2 Lite for better quality, faster speeds and lower costs." For editing, the case is stronger than the price alone. Legacy Nano Banana wants a maximum of three input images. Lite takes fourteen. That is a different class of composite in one request — and it costs less than you pay now ($0.0238 here versus $0.0312 on the legacy edit endpoint).

Our price when we last checked it on August 26, 2026:

Prices and product terms can change. Check each provider's current pricing before making a purchasing decision. Google Batch or Flex pricing is not directly equivalent to a standard on-demand API request because scheduling, availability, and processing conditions differ; it is therefore excluded from this comparison. This is a scoped comparison, not a claim that E2X is the world's cheapest option in every configuration.

Fourteen references at the cheap tier

The reference budget is the part of Lite we did not expect to survive the cut. It did — fourteen images per request, identical to full Nano Banana 2. A person, a product, three lighting references and a stack of angle shots all go into one call.

What that unlocks: catalogue work where a garment stays the same garment across eight scenes, character consistency across a comic panel set, product placement where you feed the object from several angles so the model stops inventing a back you never showed it. Exactly the jobs the legacy model choked on at three inputs.

The trade-offs are real and we would rather you hit them here than in production:

• 1K output, full stop. Google's docs say the model "only supports 1K resolution". There is no resolution parameter on this endpoint because there is nothing to set. Retouching a photo that has to survive a print crop? Use Nano Banana 2 edit-image, which goes to 2K and 4K.
• Editing is slower than generating. Google's four-second figure describes generation from a prompt. Pushing fourteen references through conditioning adds time, and more references means more of it. Still fast — just not four seconds.
• No search grounding. Lite cannot look anything up. "Make the logo match the current brand colours" gets you a guess.
• Crowded instructions lose to Nano Banana 2. Change one thing, or a few related things, and Lite matches its bigger sibling. Hand it six simultaneous edits with spatial constraints and the deeper model pulls ahead noticeably.

Text inside the image is the pleasant surprise. Lite renders legible, styled copy at the same level as Nano Banana 2 in head-to-head testing, so swapping a headline or fixing a label on a packshot is genuinely on the table at this price.

API request: sending references

Two fields are required: prompt and image_urls. Keep your key server-side.

curl -X POST https://api.e2x.ai/v1/jobs/submit \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/nano-banana-2-lite/edit-image",
    "input": {
      "prompt": "Put the sneaker from image 1 on the studio cyclorama from image 2. Match the grey backdrop and keep the sole tread visible.",
      "image_urls": [
        "https://example.com/sneaker-side.jpg",
        "https://example.com/studio-set.jpg"
      ],
      "aspect_ratio": "1:1"
    }
  }'

We return a job ID. Poll it until it settles, or send a webhookUrl in the submit body and skip the loop:

const headers = {
  Authorization: `Bearer ${process.env.E2X_API_KEY}`,
  "Content-Type": "application/json",
};

const submitted = await fetch("https://api.e2x.ai/v1/jobs/submit", {
  method: "POST",
  headers,
  body: JSON.stringify({
    model: "google/nano-banana-2-lite/edit-image",
    input: {
      prompt: "Keep the woman's face, hair and pose exactly as they are. Replace the denim jacket with a cream linen blazer.",
      image_urls: ["https://example.com/model.jpg"],
      aspect_ratio: "4:5",
    },
  }),
}).then((r) => r.json());

const jobId = submitted.data.jobId;

while (true) {
  const job = await fetch(`https://api.e2x.ai/v1/jobs/${jobId}`, { headers })
    .then((r) => r.json());

  if (job.data.status === "completed") {
    console.log(job.data.outputs[0].url);
    break;
  }
  if (job.data.status === "failed") {
    throw new Error(job.data.error?.message || "Edit failed");
  }
  await new Promise((r) => setTimeout(r, 1500));
}

Jobs run pending → processing → completed, or stop at failed or cancelled. Our aspect_ratio default on this endpoint is 1:1, so set it if your source is portrait or wide — the enum runs from 21:9 through the ultra-thin 8:1 and 1:8 strips, and auto lets the model decide. The machine-readable spec carries the live schema and price.

Choosing between our image editing models

The decision is narrower than the table looks. Nano Banana 2 costs 68% more than Lite and buys exactly three things: resolution above 1K, search grounding, and better behaviour under complicated instructions. If none of those is on your requirement list, you are paying for headroom you will not use. We would rather say that than sell it.

Where we push back on Lite: anything destined for print, and anything where one call has to do six things at once. Both are worth the upgrade. The full set lives in the image editing category and the model catalog.

Writing an edit prompt that sticks

An edit prompt is not a scene description. Say what changes, then say what must not.

Label your references when you send more than one — "image 1 is the subject, image 2 is the background plate, images 3–5 are colour reference only." Nothing else tells the model which attachment was a palette. Then anchor what has to survive: face, pose, label text, brand colour. Lite's shallower reasoning pass drifts on unstated details more than Nano Banana 2 does, so state them.

When you edit text, quote the old string and the new one. Paraphrasing is an invitation to reinterpret.

Before you ship

Every edited image comes back with a SynthID watermark baked in. Google applies it to all output, no provider exposes a toggle, and that includes us. Check your client contracts before you build on it.

Watch your input sizes. You can hand us a 4K source, but the result returns at 1K — one output tier, and it does not track your input. Migrating from the legacy endpoint? Drop the three-image chunking logic you wrote around the old limit. Dead code now.

Frequently asked questions

How many reference images can one edit request take?

Fourteen. Nano Banana 2 Lite matches full Nano Banana 2 on reference count — Google did not reduce it for the Lite tier. That is more than four times what the legacy Nano Banana recommended, which capped out at three input images.

How much does Nano Banana 2 Lite image editing cost on E2X?

We charge $0.0238 per edited image, flat — no resolution tiers, and no surcharge for sending fourteen references instead of one. That figure comes from our August 26, 2026 price check, when fal.ai listed $0.042 for the same model. Verify on the live model page before committing a budget.

Can I get 2K or 4K edits from this model?

No. Google's documentation is explicit that the model only supports 1K resolution, so there is no resolution parameter on the endpoint at all. Use Nano Banana 2 or Nano Banana Pro when you need larger output.

Is it worth migrating from the original Nano Banana?

Yes, and you have a deadline — Google retires gemini-2.5-flash-image on 2 October 2026. For editing you get fourteen references instead of three, working in-image text rendering, faster generation, and a lower price on our platform ($0.0238 versus $0.0312).

How long does an edit take?

Longer than a text-to-image generation. Google's "as little as four seconds" claim describes prompt-only generation; conditioning on references adds time, and it scales with how many you send. It is still the fastest editor in the Nano Banana family.

Should I use Lite or Nano Banana 2 for editing?

Use Lite unless you need output above 1K, search grounding, or reliable handling of many simultaneous instructions in one call. For background swaps, wardrobe changes, object removal and colour work at web resolution, the difference rarely shows and Lite costs 40% less.

Do edited images carry a watermark?

Yes. SynthID, Google's imperceptible marker for AI-generated content, is applied to every output. No provider offers a way to disable it.