Skip to content
Grok image
English
English简体中文繁體中文DeutschEspañolPortuguês

Nano Banana 2 Lite

Nano Banana 2 Lite is the efficiency model in Google’s image family, published as Gemini 3.1 Flash-Lite Image. This page keeps three kinds of statement apart. The first is what the publisher published, quoted and attributed. The second is what this site’s catalogue returns for the product, which is what the workspace actually obeys. The third is what the published artwork can and cannot support. A fourth had to be added for this page alone, because the catalogue returns the product as one the workspace cannot run.

Updated:

Nano Banana 2 Lite

Choose a mode, write the prompt, then read the quoted cost before you generate.

AI Image0

Generated content will appear here

Publisher’s own comparison

The five published comparisons, with their published prompts

Five slides from the publisher’s comparison carousel. Each is a single file that already holds both models side by side, with Nano Banana 2 on the left and Nano Banana 2 Lite on the right. Every slide came with a prompt published on the model page and each prompt is reproduced here from that page. The site did not write any of them.

All five images are the publisher’s own artwork, retrieved on 19 September 2026. Rights remain with the publisher; they are shown with source attribution and no claim of ownership. Every file arrived pre-composed, so the two models in each frame sit where the publisher placed them and not where this site placed them. Every image on this page came from the publisher’s own pages, so any provenance marking it carries is theirs and this site added none. Prompts stayed in English in every language version of this page, because a prompt is the model input and translating it would change the output.

A composite holding two generated images side by side. Both show an astronaut seen from behind inside a symmetrical white corridor whose reflective cyan floor doubles the scene, and in both the corridor opens onto a distant mountain landscape. Details and published prompt ↗

01 / Published comparison with its published prompt

An astronaut at a corridor threshold

Google DeepMind ↗

Read the published prompt

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.

Send this prompt to the workspace ↗

Copied from the model page, where the prompt is published above this image. The only edit is the removal of the leading Prompt label, which was stripped when the prompts were archived; the wording is otherwise unchanged, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page cannot currently send it, because the catalogue returns this product as not callable.

A composite holding two generated images side by side. Both are steep top-down views of a coastline split on a diagonal, with dark navy water, white surf, a pale beach, turquoise shallows and rocky ochre land. Details and published prompt ↗

02 / Published comparison with its published prompt

An aerial coastline, read as an abstraction

Google DeepMind ↗

Read the published prompt

From a breathtakingly high, top-down perspective, the image captures a dramatic, abstract view of a coastline. The composition is dynamically split by a diagonal slash of land and sea, running from the upper right to the lower left. On the left, a deep, inky navy-blue ocean churns into a brilliant white surf that resembles delicate, intricate lace or the marbled patterns in polished stone. The foam crashes against a narrow strip of pale, sandy beach, which gives way to shallow waters of stunning aquamarine and turquoise. The land itself, occupying the upper right corner, is a textured, rugged terrain of ochre, beige, and dark, earthy tones, appearing almost like a geological cross-section. The overall effect is less of a literal landscape and more of a powerful, painterly abstraction. The camera angle removes all sense of a horizon, transforming the familiar elements of water, sand, and rock into a sophisticated composition of color, texture, and dynamic, diagonal energy. The scene feels both immense in scale and microscopic in its fine detail, capturing a moment of powerful, natural artistry.

Send this prompt to the workspace ↗

Copied from the model page, where the prompt is published above this image. The only edit is the removal of the leading Prompt label, which was stripped when the prompts were archived; the wording is otherwise unchanged, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page cannot currently send it, because the catalogue returns this product as not callable.

A composite holding two monochrome generated images side by side, each showing a large dark bird in flight above a valley filled with white mist, with the silhouettes of pine trees in the foreground and ridges behind. Details and published prompt ↗

03 / Published comparison with its published prompt

A bird held above a valley of mist

Google DeepMind ↗

Read the published prompt

This cinematic, high-angle shot plunges the viewer directly into the action of soaring flight. A lone, dark bird, wings powerfully outstretched, glides with purpose through the center of the frame. The camera seems to be flying alongside it, capturing the majestic sweep of its wings against an epic backdrop. Below, a deep valley is completely choked with a sea of roiling white mist, swirling and shifting as if it were a living, breathing entity. The action is layered: the bird’s determined forward motion contrasts with the slow, churning chaos of the fog beneath it. Emerging from the mist, the dark, skeletal tips of a vast pine forest reach up like grasping hands. The entire scene is framed by the dark silhouettes of trees in the immediate foreground, creating a sense of vertigo and placing the viewer precariously on a high precipice, witnessing this primal scene unfold. The dramatic monochrome palette and sophisticated composition transform a simple act of flight into a thrilling, atmospheric event—a journey through a mysterious, untamed world where the only certainties are the driving beat of wings and the swirling abyss below.

Send this prompt to the workspace ↗

Copied from the model page, where the prompt is published above this image. The only edit is the removal of the leading Prompt label, which was stripped when the prompts were archived; the wording is otherwise unchanged, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page cannot currently send it, because the catalogue returns this product as not callable.

A composite holding two generated images side by side, each a low-angle macro of a clear glass sphere resting on dark reflective water, with a swirl of orange and gold light refracted inside it and out-of-focus city lights behind. Details and published prompt ↗

04 / Published comparison with its published prompt

A glass sphere on rippling water

Google DeepMind ↗

Read the published prompt

A perfect, clear glass sphere rests on a rippling water surface, captured in an artistic macro photograph. The composition is compellingly off-center, with the sphere placed to the left, its weight creating concentric circles that spread across the dark, reflective water. Inside the sphere, a miniature, dreamlike world is refracted and inverted: a swirling vortex of fiery orange and brilliant gold light, resembling a sunset galaxy or a city street blurred into a whirlwind of motion. The camera angle is extremely low, almost at water level, which exaggerates the sphere’s presence and the intricate patterns of the ripples. The lighting is dramatic and high-contrast; the background is dark, punctuated by soft, out-of-focus orbs of blue and orange light—a classic bokeh effect—that suggest a city at night. This gives the scene a cinematic and mysterious atmosphere. The sophistication lies in the masterful control of light, reflection, and refraction, turning a simple object into a vessel holding a tiny, dynamic universe, creating a profound sense of wonder and technical artistry.

Send this prompt to the workspace ↗

Copied from the model page, where the prompt is published above this image. The only edit is the removal of the leading Prompt label, which was stripped when the prompts were archived; the wording is otherwise unchanged, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page cannot currently send it, because the catalogue returns this product as not callable.

A composite holding two generated images side by side, each showing the profile of a woman against a textured golden-brown background while her dark hair erupts into a flock of black birds sweeping to the left. Details and published prompt ↗

05 / Published comparison with its published prompt

A woman whose hair becomes birds

Google DeepMind ↗

Read the published prompt

In a breathtaking display of surreal transformation, a cinematic moment of pure action unfolds against a textured, golden-brown canvas. From the right, a woman is seen in elegant profile, her lips a startling slash of red. Her quiet composure is violently contradicted by the explosive event taking place: her dark hair is erupting, bursting forth into a chaotic flock of black birds. These creatures are captured mid-flight, a dynamic swarm launching from her mind and sweeping across the frame to the left. The motion is palpable; you can almost hear the frantic beating of a hundred wings as they break free. Some birds are still tangled, emerging from the dark mass of her hair, while others are already soaring into the open space. This is not a gentle release, but a powerful, almost violent, act of creation or liberation. The camera captures this impossible metamorphosis head-on, focusing on the sheer force of the birds' exodus. This highly sophisticated image is a kinetic masterpiece, portraying an internal storm made external, a thought process so intense it literally takes flight in a beautiful, dark, and unstoppable flurry of action.

Send this prompt to the workspace ↗

Copied from the model page, where the prompt is published above this image. The only edit is the removal of the leading Prompt label, which was stripped when the prompts were archived; the wording is otherwise unchanged, because a prompt is a model input and rewriting one changes what it produces. The workspace on this page cannot currently send it, because the catalogue returns this product as not callable.

Three names, one model, and a positioning that reads smaller than it is

The publisher gives this model three labels. Gemini 3.1 Flash Lite Image is the model identifier used in the API documentation, Nano Banana 2 Lite is the product name used on the model page and printed into the comparison artwork, and nano-banana-2-lite is the identifier this site’s catalogue returns for it. All three point at one product, and the model code is gemini-3.1-flash-lite-image.

The model page places it as the efficiency specialist of the image family. The claims made for it there are sub-2 second end-to-end latency, a significant reduction in TPU compute cost against its siblings, and a target of high-volume interactive and real-time consumer use. That is a statement about cost and speed rather than about capability, and it reads best as exactly that.

The word Lite in the name invites the reading that this is a weaker model. The publisher’s own charts are the place to test that reading, and they do not support it cleanly. On image editing this model is scored above every model on the chart except its larger sibling. On image generation the same is true. On both latency and price it leads the chart outright. The name describes where the model sits on cost, not where it sits on quality.

Model page, https://deepmind.google/models/gemini-image/flash-lite/ , and the API reference for the model code, https://ai.google.dev/gemini-api/docs/models/gemini-3.1-flash-lite-image . Both retrieved 19 September 2026 and archived under docs/qa/official/.

Where Lite sits in the family, resolution by resolution

The developer documentation separates the image family by the resolution each model is allowed to produce. That one line does more to explain the difference between the models than any capability list, and the wording below is the publisher’s own.

The four statements are quoted from the image generation page of the Gemini API documentation, retrieved 19 September 2026. The model page separately records 2K and 4K as unsupported for this model, and four of the five Nano Banana family members are represented on this chart of the family.

ModelWhat the publisher says about its output size
Nano Banana Pro (Gemini 3 Pro Image)can generate images of up to 4K resolutions
Nano Banana (Gemini 2.5 Flash Image)generates images at 1024px resolution
Gemini 3.1 Flash Imageadds the smaller 512px (0.5K) resolution
Gemini 3.1 Flash Lite Imageonly supports 1K resolution

Read down that column and the family stops looking like a ladder. The top model reaches 4K. The first model sat at 1024px. The general-purpose Flash model added a smaller 0.5K step. The Lite model is confined to the middle step alone, so it gives up the 512px tier its larger sibling gained and never had a route to 2K or 4K. What it keeps is the 1K step, at the lowest price on the publisher’s own chart.

The four charts the publisher publishes, and who was measured with what switched on

The model page carries four bar charts comparing six models. Each chart is an embedded component whose data can be read rather than an image, so the figures below are the publisher’s own numbers and are not estimated from a screenshot. Two of the six models are the publisher’s own and four are competitors. Lower is better on the latency chart and on the price chart, and the publisher says so in the subtitle of each.

The two Elo charts run from zero to fifteen hundred on the vertical axis, which is worth holding in mind while reading a nineteen point gap between Lite and its larger sibling. Nineteen points on that scale is a little over one per cent of the drawn range. The publisher published the numbers and did not publish an interval around them, and the bars are drawn with the confidence band hidden.

The caption under each chart names the configuration each model was measured in, and the two models from this publisher were not measured the same way. Nano Banana 2 Lite is labelled as measured with thinking on. Nano Banana 2 is labelled as measured with thinking on and search on. Search grounding is a capability the Lite model does not have at all, so one of the two was measured with a feature the other cannot use. That configuration is the publisher’s own and it is reported here rather than corrected.

The four charts were decoded from the model page’s own embedded chart data and archived at docs/qa/official/nano-banana-2-lite-benchmark-charts-2026-09-19.json . Elo is given to the point, latency to one decimal place in seconds, and price per 1K resolution image in United States dollars. The two price bars for the publisher’s own models carry a plus sign after the figure in the chart itself.

ModelGeneration EloEditing EloSeconds per 1K imagePrice per 1K image
Nano Banana 2 Lite125113084$0.034+
Nano Banana 21270138720$0.067+
Nano Banana115112957$0.039
Grok Imagine Image117413296.4$0.02
Seedream v5 Lite1132129445.1$0.035
Flux 2 Klein 9B106912244.4$0.015

The 1K ceiling, the tier Lite does not get, and two ratio lists that disagree

The publisher model page and the catalogue the workspace obeys disagree about how many shapes this model takes. The model page says the model is optimised for 1K, written out as 1024 by 1024 pixels, and that 2K and 4K are unsupported. It lists fourteen supported aspect ratios. The catalogue publishes seven.

Fourteen ratios cannot all be 1024 pixels wide, because the list includes extremes. A 21 to 9 frame at 1024 pixels wide is 439 pixels tall. A 1 to 8 frame at 1024 pixels wide is 128 pixels tall. Whatever the model returns for those two, it is not a square 1024 pixel image, and the page does not say what it returns instead.

The catalogue sidesteps the question by not carrying the seven ratios it cannot account for. The seven it does publish are a strict subset of the publisher list and they are the ordinary square, landscape and portrait shapes. What is left is a 1K budget attached to a shape rather than to a pixel count.

The fourteen ratios are published as a discrete list on the model page and the count was taken from that list rather than inferred. The catalogue publishes 1:1, 3:2, 2:3, 16:9, 9:16, 4:3 and 3:4 for both of the product two modes, and all seven appear in the publisher list. Neither document states a pixel height for the wide or tall ratios.

StatementWhere it comes fromWhat it leaves open
Optimised for 1K, written as 1024 by 1024 pixelsModel pageThe pixel size for every ratio that is not square
2K and 4K unsupportedModel pageWhether a non-square 1K frame still counts as 1K
Fourteen aspect ratiosModel pageThe height that goes with each of the fourteen
Seven aspect ratiosCatalogue entryWhy the extreme ratios on the publisher list are absent
1K as the only resolutionCatalogue entryWhether that means width, area or longest edge

What this catalogue reports about the product today

This site does not ship the model. It reads a catalogue published by a separate service, and that catalogue is the only thing the workspace obeys. For this product the catalogue returns a state the page has to report rather than smooth over.

The catalogue returns nano-banana-2-lite as not callable, with an availability of temporarily unavailable and a reason recorded as a missing local price. Its cost is zero points and its billing mode is recorded as a fallback. A second read on 19 September 2026 returned the same values. This is not the ordinary state of a listed product.

Four products in the whole catalogue carry a missing local price, and three of them are variants of a single music model. This is the only image product in that state, which matters here because this site publishes both a landing page for the product and a workspace that lists it.

The visible consequence is in the form rather than in the prose. The workspace still renders the model controls, because the ratio list and the reference-image control come from fields the catalogue does publish. What the workspace does not do is enable them. Every mode resolves to disabled and the generate button is disabled with them, so a visitor can read this page and open the workspace and cannot send a task.

The catalogue value and the interface behaviour are two different facts and both are reported as they stand. Neither is presented as a decision this site has taken, because it is not one.

Read from this site’s own catalogue response for the product and archived at docs/qa/official/catalog-site19-2026-09-19.json . The entry is registered in the ledger for a human decision, because publishing a landing page for a product the workspace cannot run is a commercial choice rather than a content one.

Catalogue fieldValue returned
callablefalse
availabilitytemporarily_unavailable
unavailable_reasonlocal_price_missing
cost_points0
billing_modefallback
capabilities.resolutions1K, on both modes
capabilities.aspect_ratiosThe same seven ratios on both modes
reference_price.display_labelZero points per image, flagged as a fallback

What the workspace does with it, control by control

The form a visitor sees is assembled from three sources that do not always agree. Taking the controls one at a time is the only way to say which source each one came from, because the resulting form looks like a single design.

The rule that a published runtime list beats the bundled table is visible in the workspace source, where a catalogue ratio list is preferred whenever it is present. The fallback ceiling of five reference images is a value on this site, not a published limit of the model, and it can change with the next catalogue read.

  • Model selection comes from the catalogue. The product is listed, so it appears in the list. Whether a mode inside it can be started comes from the same response, and for this product that answer is no.
  • The aspect-ratio control comes from the catalogue whenever the catalogue publishes a ratio list for the mode, and from this site’s local fallback table only when it does not. Both hold the same seven ratios for this product, so the control reads the same either way, and the seven are fewer than the fourteen the publisher lists.
  • The reference-image control comes from this site’s local fallback table, because the catalogue publishes no ceiling for it. The fallback allows five images, and that number is this site’s own rather than a figure the publisher states.
  • A reasoning control does not exist in this workspace at all. The publisher documents thinking levels for the model and the catalogue carries no field for them, so there is nothing for the form to render and no control to disable.

Read together, the three sources produce a form whose ratio list and reference-image allowance both look more generous than the publisher’s own description of the model. Neither is a claim about what the model does well. Both are what the form will accept.

Reference images, one caveat in the documentation and one control in this site’s source

The publisher’s own description of this model warns against the exact thing this site’s form invites, and the gap is worth stating plainly rather than rounding off.

The documentation caveat is that the model is not optimised for multiple reference inputs. The control in this site’s source is a reference-image picker built from a local fallback table that allows five images, because the catalogue publishes no ceiling of its own for this model.

The same class of gap runs on a second axis. The publisher documents thinking levels and the catalogue carries no field for them, so no reasoning control can be offered. Where the catalogue is silent, this site either falls back to its own table or shows nothing at all. A reference ceiling falls back and a reasoning control shows nothing.

Three things a visitor therefore cannot tell from the form.

The warning is quoted from the model page. The five-image ceiling is read from this site’s fallback table. No published example on this page uses more than one reference image, so the page makes no claim either way about series consistency.

  • Whether the five-image allowance reflects the model or reflects this site, since the form presents it without saying where it came from.
  • Whether a supplied reference is being used as a style reference or as an edit target, because the form does not distinguish the two roles.
  • Whether the model handles several references better in practice than the publisher’s warning suggests, since no published example on this page uses more than one.

What the publisher says the model gets wrong

The model page collects the model limitations in one place and the list is short and specific enough to repeat in full rather than summarise.

The first three items are the publisher’s own limitation statements, quoted. The fourth restates a documented default rather than a limitation. The marking described there is a watermark and a metadata record rather than a visible stamp on the picture, and nothing on this page was re-marked by this site. The model page records the most recent update as June 2026.

  • The model is not optimised for multiple reference inputs, which is the caveat the previous section turns on.
  • Search grounding is not supported, so the model cannot look anything up while generating. That is also why the publisher’s benchmark comparison measured it against a version of its larger sibling that had search switched on.
  • Output is confined to 1K, with 2K and 4K recorded as unsupported.
  • SynthID watermarking is always on and Content Credentials are attached to every output, so nothing leaves the model unmarked.

What the artwork on this page is, and what it cannot show

Every image here belongs to the publisher and came from the model page for Gemini 3.1 Flash Lite Image. Twenty-three published assets were downloaded and screened one at a time; ten survived and produced the seventeen files in this site’s media manifest.

  • Thirteen published assets were rejected rather than used as filler. Six are corporate logos in the page Showcase carousel, belonging to Figma, Locket, Manus, Artlist, Weekend and Latitude. They are third-party trademarks on a plain background and not model output, so presenting one as a case would claim a customer relationship as this page’s evidence.
  • Three more are capability icon cards drawn as a stopwatch, a document and an HD badge, and four are poster frames for hands-on videos, which are captures of four applications rather than images the model produced.
  • The five comparison slides are single files that already contain both models side by side. This page’s gallery renders them as ordinary single images and every caption says the frame holds two models, because the gallery has no renderer that can split a composite that arrived pre-composed.
  • Each of the five slides carries a caption the publisher printed into the artwork itself, naming both models and giving a latency average. Nano Banana 2 is on the left in all five and Nano Banana 2 Lite is on the right, and the printed figures are nineteen seconds against three.
  • An official moving comparison exists and is not on this page. It runs eighty seconds at 1916 by 1080 with audio, thirty megabytes, and it repeats what the five stills already show. It is registered in the ledger for a decision rather than discarded.

The one thing this artwork cannot show is whether the model suits the task a particular visitor has in mind. These are the publisher’s selected examples, published alongside the model, and every prompt reproduced on this page came with an image rather than with a failure. A page built only from those cannot measure an error rate and this one does not try to.

Examples

A man in a dark coat stands on a tall stack of books in a vaulted hall while a wave of loose white paper sheets breaks across the space in front of him.

The first slide of the model page’s hero carousel, and the one used as this page’s open graph image. It is a physics problem dressed as a portrait, since the prompt has to hold a stack under a person, a vaulted interior and a mass of individual sheets in one frame without any of the three losing its own structure.

Google DeepMind, model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Published by the publisher as part of the model page’s hero carousel. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A macro photograph of a brightly coloured spider resting at the centre of a web whose threads are strung with dew drops.

A resolution case rather than a composition case. At 1K the interest is almost entirely in the dew and the leg detail, so this is the kind of image that would gain most from a higher tier and is the clearest illustration of what a 1K-only model gives up.

Google DeepMind, model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Published by the publisher as part of the model page’s hero carousel. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A swimmer with both arms out of the water in the middle of a butterfly stroke, seen against backlit water that is breaking into spray around them.

A single frozen instant with water moving in several directions at once. The arms, the shoulders and the spray all have to agree on one moment, which is the part that tends to come apart when a model treats a motion prompt as a pile of objects.

Google DeepMind, model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Published by the publisher as part of the model page’s hero carousel. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A spiny lizard with a patterned body basking on a surface of worn red bricks.

A pattern test on a shallow subject. The lizard scales and the brick surface are two regular patterns of different scale and colour sitting directly against each other, so a failure here shows up as either pattern bleeding into the other.

Google DeepMind, model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Published by the publisher as part of the model page’s hero carousel. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A car on a wet road at night under a teal street light, with palm trees standing behind it and the road surface reflecting the light.

The still the model page pairs with its prompt guidance, so it is the publisher’s own example of a plain photographic brief. It is used here as the one image in the set that looks like an ordinary request rather than a demonstration.

Google DeepMind, model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Published by the publisher as part of the model page’s hero carousel. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A composite holding two generated images side by side. Both show an astronaut seen from behind inside a symmetrical white corridor whose reflective cyan floor doubles the scene, and in both the corridor opens onto a distant mountain landscape.

The pair that shows most clearly what the two models were asked to do differently rather than badly. Both frames keep the symmetry, the mirrored floor and the vanishing point, and both place a small human figure at the centre of a large sterile space. The differences are in the far end of the corridor and in how much surface detail the floor carries. The publisher printed a caption into the artwork itself, naming both models and giving a latency average; this slide carries nineteen seconds for Nano Banana 2 and three seconds for Nano Banana 2 Lite.

Google DeepMind, comparison carousel on the model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Prompt published alongside the image and reproduced verbatim on this page. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A composite holding two generated images side by side. Both are steep top-down views of a coastline split on a diagonal, with dark navy water, white surf, a pale beach, turquoise shallows and rocky ochre land.

A test of texture rather than of objects. The prompt asks for surf that reads as lace or marbled stone and for the frame to lose its horizon, and both results obey: neither frame gives the viewer a sky, and in both the water does the compositional work. What separates them is how finely the foam breaks up at this display size. The publisher printed a caption into the artwork itself, naming both models and giving a latency average; this slide carries nineteen seconds for Nano Banana 2 and three seconds for Nano Banana 2 Lite.

Google DeepMind, comparison carousel on the model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Prompt published alongside the image and reproduced verbatim on this page. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A composite holding two monochrome generated images side by side, each showing a large dark bird in flight above a valley filled with white mist, with the silhouettes of pine trees in the foreground and ridges behind.

The most demanding frame in the set for placement, because the prompt fixes the camera alongside a moving bird and stacks a foreground silhouette against a middle distance of churning fog. Both results hold the layering. The bird sits at a different angle in each, which is an input-level difference rather than a quality difference. The publisher printed a caption into the artwork itself, naming both models and giving a latency average; this slide carries nineteen seconds for Nano Banana 2 and three seconds for Nano Banana 2 Lite.

Google DeepMind, comparison carousel on the model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Prompt published alongside the image and reproduced verbatim on this page. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A composite holding two generated images side by side, each a low-angle macro of a clear glass sphere resting on dark reflective water, with a swirl of orange and gold light refracted inside it and out-of-focus city lights behind.

A refraction test, and the hardest thing on the page to fake, because the sphere has to show an inverted world that is consistent with the lighting around it. Both frames produce a legible inverted vortex. The ripple rings are the visible difference, since the prompt makes them the measure of the sphere resting on the surface. The publisher printed a caption into the artwork itself, naming both models and giving a latency average; this slide carries nineteen seconds for Nano Banana 2 and three seconds for Nano Banana 2 Lite.

Google DeepMind, comparison carousel on the model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Prompt published alongside the image and reproduced verbatim on this page. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A composite holding two generated images side by side, each showing the profile of a woman against a textured golden-brown background while her dark hair erupts into a flock of black birds sweeping to the left.

The kinetic case, and the one where the two frames read most differently at a glance. The prompt asks for hair that is still partly tangled where the birds emerge, so the transition has to look continuous rather than assembled. In one frame the flock is tighter to the head, in the other it has already spread across the frame. The publisher printed a caption into the artwork itself, naming both models and giving a latency average; this slide carries nineteen seconds for Nano Banana 2 and three seconds for Nano Banana 2 Lite.

Google DeepMind, comparison carousel on the model page for Gemini 3.1 Flash Lite Image, retrieved 19 September 2026. Prompt published alongside the image and reproduced verbatim on this page. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Questions

Is Nano Banana 2 Lite the same thing as Gemini 3.1 Flash Lite Image?

Yes. The publisher released the product under the name Nano Banana 2 Lite and gives the model code gemini-3.1-flash-lite-image, which is how the API documentation names it. This site’s catalogue registers it as nano-banana-2-lite, which is the identifier that appears in the URL and in the workspace model list.

Can I send a task to it on this site right now?

No. The catalogue this site reads returns the product as not callable with an availability of temporarily unavailable and a reason recorded as a missing local price. The workspace still renders the model controls, but every mode and the generate button are disabled. That state comes from the catalogue rather than from this page.

How does it differ from Nano Banana 2?

Four documented differences. Output is confined to 1K where the larger model adds a 512px step and reaches 4K. Search grounding is not supported. The publisher documents thinking levels for it but the larger model was benchmarked with both thinking and search on. And it is measured at roughly half the price and a fifth of the latency, at four seconds against twenty seconds per 1K resolution image.

Why is it called Lite if it scores near the top of the publisher charts?

Because the name describes cost rather than quality. On the publisher’s own charts this model is above every model on the page except its larger sibling for both image generation and image editing, and it leads both the latency and the price charts. The publisher positions it as the efficiency specialist for high-volume and real-time use, which is a statement about cost and speed.

Does it support 2K or 4K output?

No. The model page records 2K and 4K as unsupported and describes the model as optimised for 1K, written as 1024 by 1024 pixels. The API documentation states the same limit more briefly, as supporting only 1K resolution. Both documents agree that the model does not reach the higher tiers.

How many reference images can I supply?

The form allows up to five, but that figure comes from this site’s own fallback table rather than from the publisher, because the catalogue publishes no ceiling for this model. The publisher documentation separately warns that the model is not optimised for multiple reference inputs. This page reports both facts rather than picking one.

Does it support search grounding?

No. The publisher lists search grounding as unsupported, so the model cannot look anything up while generating. This is also why the benchmark comparison on the model page measures it against a version of its larger sibling with search switched on, a configuration the Lite model cannot match.

Explore next