Nano Banana 2 Lite image generator and editor

What Nano Banana 2 Lite is

Nano Banana 2 Lite is Google's speed-first image model, published on 30 June 2026 under the model ID gemini-3.1-flash-lite-image and the architecture name Gemini 3.1 Flash Lite Image. Google's own summary is blunt about the trade: it is the fastest, most cost-efficient image model in the Nano Banana family yet, built for high throughput, speed and scale.

It is not a replacement for Nano Banana 2 or Nano Banana Pro. Google positions it for near-real-time, high-volume work where ultra-low latency is critical, and names it as the recommended upgrade path for anyone still calling the first-generation gemini-2.5-flash-image model.

The honest way to read this tier is as a throughput decision. Google publishes a one-minute generation test in which Lite returned 21 images against 3 for the premier model on the same prompt, alongside a family chart that rates Lite one step below the tiers above it on both visual quality and reasoning. Both artefacts are reproduced further down this page.

Four figures Google puts on this model

These are the numbers the launch post states in its own words. Everything else on this page either repeats them, quotes a third party explicitly, or describes this workspace.

The one-minute test Google published

The launch post carries a side-by-side video titled How many animals can you generate in 1 minute? Two panels run the same prompt for sixty seconds, a watercolour painting of a number with that many animals around it, and each panel reports its own counter. It is the clearest throughput statement Google makes about this model.

The two totals are not a contradiction. The Lite panel spent more because it finished seven times as many images; per image it cost almost exactly half, and $0.0336 rounds to the $0.034 rate Google quotes. Seven times the throughput at half the unit price is the whole argument for this tier, and it is Google's own demonstration rather than a reseller's claim.

One caveat on those two figures. They are the counters inside Google's promotional video, not an independent benchmark, and the prompt is a single narrow task: a flat watercolour subject in a square frame. Latency on a denser prompt, a longer instruction or an edit with several references will not necessarily land at 2.8 seconds.

Where Nano Banana 2 Lite sits in the family

Google publishes a four-axis family chart with the launch. It is the most useful single artefact of the release, because it states plainly where this tier gives something up instead of only where it wins.

Read across the Lite row and the case for the model is clear: it is the only tier rated Low on both latency and cost. Read down the visual quality column and the price of that is equally clear, Medium against High for the two tiers above it. Reasoning drops a further step, to Low.

Google's model card headline for the same model reads No compromise on quality, and its family page calls Lite the most efficient model to date with the best balance of quality and speed. Both statements are fair; the four-axis chart is simply the more precise document, and it is the one worth reading before you commit a high-volume pipeline to this tier.

Prompts you can copy

Each prompt below is published by the source named beside it and is reproduced word for word. Four come from Google, and one comes from a third-party playground that publishes its own example inputs. Opening a prompt loads it into the workspace.

A charming watercolor painting of the number n, with n animals around it, on a clean white background, formatted as a perfect 1:1 square. Do not include any text, letters, words, or writing.

A lone astronaut stands poised at the threshold of a futuristic corridor, their back to the viewer, gazing out towards a distant, mountainous landscape visible at the far end. The composition is perfectly symmetrical, using the clean, repeating lines of the brightly lit hallway to create a powerful vanishing point. The corridor itself is a marvel of sleek, sterile design, with glowing white panels on the walls and ceiling, and a highly reflective, cyan-tinted floor that mirrors the entire scene, doubling its visual impact. The astronaut, clad in a white and gold-trimmed spacesuit, is the central, human element in this otherwise impersonal, high-tech environment. The lighting is bright and even, giving the scene an ethereal, almost heavenly quality. The color palette is a minimalist blend of stark white, cool cyan, and the earthy tones of the distant planet. The artistic style is one of clean, utopian science fiction, evoking a sense of calm, wonder, and the beginning of a grand journey into the unknown.

Create a beautiful travel lifestyle image of a woman in a light linen dress standing on a sunlit Mediterranean terrace. The background includes blue sea, white coastal houses, flowering bougainvillea, terracotta steps, and warm afternoon light. Use a soft travel-magazine aesthetic, natural colors, realistic facial and clothing details, and a balanced composition that feels bright, graceful, and shareable.

Keep the same woman's face, hairstyle, expression, pose, and clothing, but replace the plain studio background with a beautiful Mediterranean seaside terrace. Add blue ocean water, white coastal buildings, soft flowers, warm sunlight, and a refined travel-lifestyle atmosphere. Preserve the subject's identity clearly and make the final image feel elegant, bright, and natural.

Create a charming editorial illustration of an outdoor breakfast table under a lemon tree. Include croissants, honey, tea, fresh fruit, wildflowers, ceramic plates, linen napkins, and soft morning sunlight. Use warm hand-painted textures, tasteful color harmony, and a polished storybook-meets-lifestyle-magazine style. The composition should feel cozy, beautiful, and visually refined.

Jobs this model is suited to

Google's own framing is rapid ideation and high-velocity pipelines. The list below stays inside what either Google demonstrates or a published example shows, rather than listing everything an image model can theoretically do.

• Drafting at volume. Google's one-minute test is the whole pitch: 21 cheap 1K frames in the time a heavier model finishes three.
• Filling an interface as a person uses it. Google demonstrates the model inside Gridscape, an infinite canvas that generates imagery as a topic is explored, and inside Space Lift, which turns one photo of a room into several interior concepts.
• Rapid prompt exploration before you commit. Test a direction at 1K, then move the winning prompt to a tier with more resolution and reasoning.
• Personalised composites from a single photo. Google's Anywhere demo attaches one image and generates postcards of that subject at landmarks.
• Scene and item art where the current state changes faster than a design cycle, which is the reason Google cites the Latitude demo for using the model.
• Product and lifestyle scene drafts, using the reference-image workflow to keep the product unchanged and replace what surrounds it.
• Background and setting replacements on an existing photograph, which is the edit case measured further down this page.
• Localisation passes where the layout matters more than the language. Google's own limitations list flags grammar and cultural nuance as weaker than the visual result, so treat translated text as a draft.

What this workspace exposes

The picker is the authority on what you can submit here. Another provider's page describes that provider's account, and the price is resolved by the service rather than by this page.

The seven ratios above are the exact list this workspace accepts for Nano Banana 2 Lite, and it is not the same list as Nano Banana 2, which offers fifteen values including auto and the extreme 1:4 and 8:1 frames. A narrower picker is a limitation of this site, not of the model.

The five-image reference ceiling deserves the same honesty. Google publishes no reference-image limit for this model at all, and third-party schemas allow ten. Rather than advertise a number the service has not confirmed, this workspace keeps the conservative ceiling of the earlier Nano Banana tier. If your edit needs more references than the picker allows, that is a reason to use a different model here.

Nano Banana 2 Lite, Nano Banana 2 and Pro in this workspace

Only the first group of rows below is Google's own material. The control rows are this workspace's picker, which is why they can differ from what another site offers for the same three models.

Because Lite is the only one of the three capped at 1K here, the choice between it and Nano Banana 2 is rarely about resolution. It is about whether you would rather have five times the frames from the same budget, or fewer frames that Google rates a step higher on quality and reasoning.

Why other sites quote different numbers

Several of the pages ranking for this model quote settings that this workspace does not offer. None of them is necessarily wrong; they describe the model's own published schema against their own account. The gap is worth understanding before you compare running costs.

The nanobanana.im page lists 15 aspect ratios, up to 10 reference images and a cost of 2 credits per generation, the same as its earlier Nano Banana tier. kie.ai's published request schema for the same model accepts fifteen aspect-ratio values and a file limit of ten, which is where the fifteen and the ten come from. The model will take more than this workspace exposes; the picker here simply passes a narrower set.

On price, kie.ai's own page carries two different statements about its own rate for the same 1K tier, one saying four credits and one saying three, in adjacent strings. Treat both as that provider's marketing copy rather than a rate you can hold them to. The only figure that binds here is the quote the workspace prints before you submit.

• Fifteen aspect ratios on two competitor pages against seven in this picker: the model accepts more than this site passes.
• Ten reference images in a third-party schema against five here, and Google publishes no ceiling of its own.
• Credit prices are provider-specific and change; the workspace quote is the number that applies to your task.

Editing with references instead of masks

An edit starts with images you are allowed to use and an instruction that says what should change and what must survive it. There is no mask to draw: the model reasons about which parts the instruction refers to and leaves the rest alone. Google's own limitations note is less confident about the hard cases, warning that masked editing, day-to-night lighting changes and multi-image blending can still produce unnatural results, visual artifacts or disjointed scenes.

The comparison further down this page is a measured before-and-after rather than a decorated pair. Across a nine-tile grid, the two frames correlate at 0.88 to 0.96 in the centre column and near zero or below on both flanks. That is the signature of a subject that was kept while everything around it was replaced, and it is our check rather than Google's claim. The prompt and both frames come from the same published example, so the pairing is documented as well as measured.

Write the instruction in two halves. First the subject and the change. Then the parts that must not move, such as the silhouette, the camera angle or the colour of everything else.

• Check faces and hands first. Google names small faces as a weak point, and resemblance drift is the most common unwanted change in a portrait edit.
• Re-read small text at full size. Accurate spelling is the other weak point Google names, and a five-word sign is enough to expose it.
• Compare reflections, shadows and contact points where the background was swapped, since those are where a replacement shows seams.
• Change one instruction at a time so you can attribute a difference to the instruction rather than to the model's own variance.

How to generate or edit here

Four steps, in the order the workspace asks for them.

1. Paste your prompt, or open one from the list above with the copy button.
2. Leave the resolution at 1K and pick one of the seven aspect ratios, or start from 1:1.
3. Attach up to five reference images when the job is an edit rather than a new image.
4. Read the quote, sign in when the workspace asks, then submit. If a task fails, read its status before retrying so you do not submit a duplicate.

What this page does not promise

A model page should say where its own claims stop. These are the boundaries for this one.

• No free allowance is promised. Whatever credits or trial grant exist are shown in your account, not here.
• The 4-second and $0.034 figures are Google's published numbers for its own API. They are not a measurement of this workspace, and real latency depends on the load the service is under when you submit.
• Google publishes no reference-image limit for this model, so the five-image ceiling here is this workspace's conservative choice rather than a documented model limit.
• One-minute test results come from a promotional video on a single narrow prompt. Do not read 2.8 seconds as a guarantee for a long instruction or a multi-reference edit.
• The family chart ratings were read from the published chart image, not from a text specification, because Google publishes no table of those four axes.
• Elo scores and latency comparisons trace to arena.ai and artificialanalysis.ai as credited by Google, and to third-party readings of those charts. No benchmark on this page was run by this site.
• Outputs carry SynthID watermarking per Google. This site cannot remove it, and Google's own verification route is to upload an image to the Gemini app and ask whether Google AI generated it.
• Text inside an image, small faces, factual accuracy in infographics and character consistency are all named as improvable in Google's own limitations list. Proofread anything critical at full size.
• The gallery mixes official Google material with third-party playground examples. None of it was generated on this site, and a published prompt is not a guarantee that a given result will reproduce.
• The model ID is published by Google as a current identifier, and availability or behaviour can change without notice.

Where the samples and numbers come from

Every figure and every image on this page traces back to one of the sources below. Where a value comes from a third party rather than from Google, the row says so.