Nano Banana 2
Nano Banana 2 is Google's general-purpose image model, published as Gemini 3.1 Flash Image. This page keeps three kinds of statement apart. The first is what Google 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.
What Google says Nano Banana 2 is
Google announced Nano Banana 2 on 26 February 2026 and gave it two names: the product name Nano Banana 2, and the model identifier Gemini 3.1 Flash Image. The announcement introduces it as combining the capabilities of Nano Banana Pro with Flash speed, and says it replaces Nano Banana Pro across the Fast, Thinking and Pro options in the Gemini app.
The developer documentation is where the two names are tied together, and it splits the family four ways. Gemini 3.1 Flash Image is published as the most versatile model, a generalist workhorse for all tasks, balancing speed against 4K generation, world knowledge and reliable text rendering, and excelling at multiple reference images and consistency. That sentence is the publisher's positioning, not a measurement.
The capability claims the announcement makes in its own words are narrow enough to repeat. Up to five characters and up to fourteen objects can keep their identity in a single workflow. Text renders legibly enough for infographics and marketing mockups, and text inside an image can be translated and localised. Aspect ratios and resolutions are described only as covering a range from 512px to 4K, with no ratio named. The announcement also promises stricter instruction following and a visual fidelity upgrade over the first Nano Banana.
Two of those claims are carried by the documentation rather than the announcement, and they are the ones worth holding separately. The documentation states that Gemini 3.1 Flash Image adds a 512px tier and that it adds Google Image Search grounding alongside web search. Neither claim appears in the announcement. Grounding with Google Search is documented as unsupported by the Lite model.
The documentation also describes a thinking stage, in which the model generates interim images that are visible in the backend and not charged, and it exposes a thinking level control whose default is minimal and whose only other supported value is high. That control is an API parameter. Nothing in the catalogue this site reads carries it.
This page is not an official product description and is not endorsed by Google. It records what the publisher published and what this site’s catalogue returns.
The Nano Banana family, split four ways
The Gemini API documentation lists four image models under the Nano Banana name, and the split matters because the product names do not sort into an obvious order. Two of the four sit on the same Gemini 3.1 Flash generation, and the one called Lite is the newer of the two.
Read that table as the publisher's own positioning rather than as a benchmark, because it is the former. The one ordering it does settle is that the Lite model is not simply a smaller version of this one. Google says Lite is not optimised for multiple reference inputs or multi-turn sequential edits, which is a scope statement rather than a quality statement, and this site exposes a reference-image control for Lite anyway. That discrepancy has its own page and is registered as an open question rather than settled here.
The price ladder is visible in this site’s catalogue instead, where the product is charged 3 points per image at 1K, 6 at 2K and 9 at 4K, with roughly 1 point added per reference image. That is a catalogue figure, not a Google figure, and Google publishes its own per-image API pricing separately.
| Product name | Publisher's model identifier | What the publisher says it is for |
| Nano Banana 2 Lite | gemini-3.1-flash-lite-image | Fastest and cheapest image model. Built for speed and scale. Not optimised for multiple reference inputs or multi-turn sequential edits. |
| Nano Banana 2 | gemini-3.1-flash-image | The generalist workhorse. Balances speed against 4K generation, world knowledge and text rendering. Handles multiple reference images. |
| Nano Banana Pro | gemini-3-pro-image | For complex visual tasks. Highest world knowledge, advanced localisation, brand consistency and precise creative control. |
| Nano Banana | gemini-2.5-flash-image | The legacy first-generation model. Google recommends moving to Nano Banana 2 Lite instead. |
The aspect-ratio list, and the value only this site’s fallback carries
Two lists exist for this product and they are not the same list. The service that this workspace actually talks to declares fourteen ratios. The local table kept in this site’s own source, used when that service publishes no ratio field, carries the same fourteen plus auto. Because the service does publish the field, the fourteen-win rule applies and auto never reaches the selector.
The four extreme shapes are the interesting part of the list, and they are specific to this generation on this site. Ratios of 4:1, 1:4, 8:1 and 1:8 appear in the catalogue for this product and in no other image model this site offers. Eight to one is a shape the service accepts as a generation target; whether a crop would have produced the same framing is not something this page can establish, and neither official source names these ratios.
The fallback table is worth knowing about precisely because it is invisible. It is what the workspace would build its selector from if the catalogue request failed, and it disagrees with the catalogue on both counts that matter, the list and the default. A reader who wants to know what happens on a working day should read the first row of that table. The second row is a contingency, and its auto value is not something the service declares.
This is an observation about two lists this project controls, not a defect claim about Google’s service. The catalogue entry was read from the live endpoint on 19 September 2026.
| Source | What it lists | Default |
| Catalogue entry for this product | 14 ratios: 1:1, 1:4, 1:8, 2:3, 3:2, 3:4, 4:1, 4:3, 4:5, 5:4, 8:1, 9:16, 16:9, 21:9. No auto. | 1:1 |
| This site’s local fallback table | The same 14 plus auto, listed first. Used only if the catalogue stops publishing the field. | auto |
Three controls the publisher ships that this form does not carry
Google documents three controls on this model that have no counterpart in the workspace form. None of the three is hidden by the interface; each is absent because the catalogue entry does not publish it. That distinction matters, because it means the gap is in the service this site reads rather than in the page built from it.
- The 512-pixel tier. The documentation states that Gemini 3.1 Flash Image adds the smaller 512px (0.5K) resolution on top of the built-in 1K, 2K and 4K, and that the Lite model is limited to 1K. This catalogue declares three tiers for the product: 1K, 2K and 4K. No selector value would send 512, so the cheapest tier the publisher ships is unreachable from this form.
- The thinking level. The documentation exposes a control whose default is minimal and whose other supported level is high, and it describes interim images that the model generates to refine composition, which are visible in the backend and not charged. The catalogue form for this product carries a prompt, an optional reference-image field, an aspect ratio and a resolution tier. There is no reasoning field, so every task runs at whatever the service applies when none is sent.
- Google Search grounding. The documentation describes the model using Google Search to verify facts and generate imagery from real-time data, and says this generation adds image search grounding alongside web search. The catalogue form has no field for it, and neither does this workspace. A task here therefore runs without the grounding the publisher documents, and no control on the page can turn it on.
- The reference-image count, where the number is real but not the one this site chose. Google documents a ceiling of up to 14 reference images for the Gemini 3 image models, and publishes a per-model breakdown of how those slots may be used. The catalogue entry for this product publishes no ceiling at all, so the workspace applies its own setting of 14. The two numbers agree by coincidence of this schedule rather than because the site copied one from the other.
What this catalogue declares for the product
These are the fields the service returns for this product, and they are what the workspace obeys when it builds the form. A model mnemonic and a marketing name are not evidence of anything, so the useful question is always what a task can actually carry.
Two fields a reader might expect are simply not there. There is no quality or detail level published for this product, so the form does not offer one. There is no seed or step control either, which means a repeat of the same prompt at the same settings is not reproducible by any parameter this workspace can set.
The declaration is also the entire surface. Anything a page claims about this model that cannot be traced to one of those fields, to a dated Google announcement, or to the developer documentation is a claim this site has no basis for, and this page does not make any.
| Field | Value declared | Where the workspace uses it |
| Task types | text-to-image, image-to-image | The mode tabs, and the task model sent with the request |
| Resolution tiers | 1K, 2K, 4K | The resolution selector, defaulting to 1K |
| Aspect ratios | 14 values, listed above | The aspect-ratio selector |
| Form fields | prompt, image_urls, aspect_ratio, resolution | Every control the form shows, and nothing else |
| Charges published for the product | 3 points per image at 1K, 6 at 2K, 9 at 4K, with roughly 1 point added per reference image | The quote shown before a task is accepted |
The consistency figure, and the source that disagrees with it
Two Google sources publish different limits for the same model. The announcement states that character resemblance holds for up to five characters and the fidelity of up to fourteen objects in a single workflow. The developer documentation publishes a table in which Gemini 3.1 Flash Image allows up to ten objects and up to four characters among its fourteen reference slots. Both are quoted above as published, and this page does not pick a winner.
The two numbers are not measuring the same thing, which is the most likely explanation and also the one that cannot be confirmed from outside. A workflow ceiling is the count a task may be built around; a reference-slot table is the count of attached inputs of a given kind. Fourteen objects with identity held, against ten object slots and four character slots, is a pair of statements about the shape of the claim rather than a contradiction anyone outside the publisher can resolve.
The announcement publishes an example that sits at the same figure, and it is worth reading as published rather than as evidence. Its farm prompt asks for fourteen characters and items in one scene, which is the announcement’s own ceiling used as a demonstration. The prompt was written by the same team that wrote the figure.
The treehouse example is the harder test in the set and the more useful one. Its published prompt asks for a six-part story in which three characters keep their identity while their expressions and angles change, and it asks for the images one at a time. Holding identity across separate generations is a different claim from holding it inside one frame, and it is the one more likely to break on a real task of yours.
What none of this establishes is how consistency holds up on a task of yours. Every example was chosen by the publisher from outputs that worked, and no failure rate accompanies any of them. Treat the figures as the shape of the claim and the images as illustrations of it.
Provenance: SynthID plus Content Credentials
Google states that all generated images carry a SynthID watermark, and that it couples that technology with interoperable C2PA Content Credentials. Both are metadata-level provenance mechanisms, which has a practical consequence for anyone reading a gallery of AI images.
- SynthID is not a visible mark. It survives inside the file rather than across the picture, so an image carrying it looks untouched. Inspection tools are what detect it, not your eye.
- Google says its SynthID verification feature in the Gemini app had been used more than 20 million times across various languages since a November launch, and that C2PA verification was coming to the app. Those are the publisher's own usage figures and are repeated here as claims.
- Every image in the gallery on this page came from Google's own announcement, so the provenance mechanism attached to them is Google's, and any claim this page makes is limited to the absence of a watermark added by some other site.
- This site does not strip, re-encode away or claim to detect provenance metadata. The images here are resized for delivery, which is a separate operation from the watermarking the publisher describes.
How a task runs here
The workspace pins the model when you arrive from this page, so nothing else has to be chosen by hand. Everything below happens before anything is charged, and the quote is the last thing you see rather than the first.
- Open this page in the workspace, which fixes the model and reads the current catalogue entry for it.
- Write the instruction. Google's own examples are one or two sentences of plain description naming the subject, the setting, the lighting and any text that must appear verbatim. Because there is no exclusion field, anything you want kept out has to be written into the same sentence as the thing it applies to.
- Attach reference images if the task needs them. The mode becomes image to image and the task model sent changes with it. The workspace allows up to 14 references, which is its own setting; the catalogue publishes no ceiling, although Google documents 14 as the model’s limit.
- Choose an aspect ratio deliberately. The list the service declares is what the selector shows, and extreme shapes such as eight to one are available here and unusual elsewhere, so a specific shape is worth choosing rather than leaving the default in place.
- Choose a resolution tier. The 512-pixel tier the publisher describes is not offered here, so the fastest and cheapest option Google ships is not reachable from this form.
- Read the quote, then submit. If the task fails, the failure is reported rather than silently retried, and the task appears in your history with the settings it actually used.
What the artwork shows, and what this page does not claim
The gallery and the case list on this page come from Google's own announcements, which is an unusually good provenance for this site. It comes with limits that are easier to state than to work around.
- The images are presented by Google as artwork produced with this model, and the announcement does not label each image with the model that produced it. This page therefore attributes the artwork to the announcement as a set and not to a specific generation. Every image in the set carries a published prompt, so the prompt is not the weak link; the model attribution is.
- Each prompt is reproduced from Google’s own numbered footnote for that image, with no rewriting and no summarising, because a prompt is a model input. The description shown alongside each image is the publisher’s own alternative text for it, translated for the language of the page.
- The images demonstrate legible in-image text, in-image localisation across languages, subject consistency, and one portrait the publisher presents across several shapes. They do not demonstrate a latency figure, a cost figure, or a comparison against any other model, and this page uses them for nothing beyond the four things they show.
- One downloaded candidate was rejected during review because it is a screenshot of the publisher's own application rather than a generated image. It is excluded from the gallery, and the rejection is recorded in the project evidence rather than quietly dropped.
- This page carries no video, and the announcement provides none for this model: the media carousel behind it exposes empty video slots and no file addresses. There was therefore no video to embed and no frame to use as a poster, rather than a retrieval failure on this project’s side.
- Nothing here is a statement of quality. Google's own positioning language, such as calling this the generalist workhorse, is quoted as positioning, and no Elo score, ranking or benchmark number from the publisher is restated as an independent result.
Questions
Is Nano Banana 2 the same thing as Gemini 3.1 Flash Image?
Yes. Google announced the product under the name Nano Banana 2 and gives its identifier as gemini-3.1-flash-image. This site's catalogue calls it nano-banana-2, which is what the URL and the workspace model selector use.
Which aspect ratios does this model support?
The catalogue declares fourteen: 1:1, 1:4, 1:8, 2:3, 3:2, 3:4, 4:1, 4:3, 4:5, 5:4, 8:1, 9:16, 16:9 and 21:9. The service publishes that field itself, so those fourteen are what the selector shows and 1:1 is what it starts on. The site’s own source file carries the same fourteen plus auto, but that table is only a fallback for when the catalogue stops publishing the field, which is not the case for this product.
Can I ask for a 512-pixel image here?
No. The developer documentation states that Gemini 3.1 Flash Image adds the smaller 512px (0.5K) resolution alongside 1K, 2K and 4K, but the catalogue declares only 1K, 2K and 4K for the product, so no selector value would send it and no task placed from this page can request it.
Can I turn the model's thinking level up?
Not from this form. Google documents a thinking level control whose default is minimal and whose only other supported value is high, and describes interim images generated to refine composition, which are visible in the backend and not charged. The catalogue form for this product has no reasoning field, so every task runs at whatever the service applies when none is sent.
How many reference images can I attach?
The workspace allows up to 14, which is this site’s own setting. The catalogue returns no reference-image ceiling for this product, so there was nothing in the entry to copy. Google does document a real limit of up to 14 reference images for the Gemini 3 image models, and publishes a per-model breakdown of how those slots may be used, with ten object slots and four character slots for this model. The two numbers agree, but the site’s setting was not derived from the publisher’s.
How many characters can it keep consistent across images?
Google publishes two different figures. The announcement says five characters and fourteen objects can keep their identity in one workflow; the developer documentation allows ten object slots and four character slots among this model’s fourteen reference images. Both are quoted as published, and the difference looks like a workflow ceiling against an input-slot count rather than a correction of one by the other.
Is the image quality adjustable?
No. The catalogue publishes no quality or detail level for this product, and the form offers none. There is no seed or step control either, so the same prompt at the same settings is not reproducible by any parameter this workspace can set.
Do images from this model carry a watermark?
Google states that all generated images carry a SynthID watermark and that it couples SynthID with C2PA Content Credentials. Both are metadata-level mechanisms, so the image looks untouched and needs a verification tool to detect. This site adds no third-party watermark and does not claim to detect or remove provenance metadata.
Why is there no video on this page?
Because the announcement does not publish one for this model. The media carousel behind the announcement exposes empty video slots and no file addresses, so there was no video to embed and no frame that could honestly be presented as a poster. The page shows nothing rather than a substitute.
Is this an official Google page?
No. It records the fields this site's service catalogue returns for the product and quotes Google's published statements with attribution. It is not a statement by Google and is not endorsed by them.