Nano Banana 2.1
Nano Banana 2.1 is Google's newest Nano Banana image model, published as gemini-nano-banana-2.1 and generally available since 6 October 2026. This page keeps three kinds of statement apart: what Google documents, what this site's catalogue returns for the product, which is what the workspace actually obeys, and what neither source establishes.
Draft content — not yet reviewed by an editor.
Nano Banana 2.1
Choose a mode, write the prompt, then read the quoted cost before you generate.
Results
Generated content will appear here
What Nano Banana 2.1 is, in the publisher's own words
Google's model documentation introduces Nano Banana 2.1 with one sentence: it is optimised for multimodal image generation and editing, with good cost-effectiveness. The model identifier is gemini-nano-banana-2.1, the release stage is generally available, and the published release date is 6 October 2026.
That sentence is positioning, not a measurement, and this page repeats it as such. What the documentation establishes in concrete terms is a capability table rather than adjectives: image generation and image editing including multi-turn edits are supported, interleaved text and image output is supported, generation from video input is supported, and Google Search and image search grounding are supported. Content Credentials (C2PA) is listed as supported. Person portrait generation is listed as not supported, which is the one entry in the capability table that rules a whole task category out.
The documentation also fixes what the model does not accept. Nano Banana 2.1 does not support the seed, topK, logprobs, temperature and topP parameters, and setting any of them returns an API error. For a workspace that means there is no sampling control to expose and nothing to hide: the catalogue form for this product has no field for any of them.
The context window is documented at 131,072 tokens with an output cap of 32,768. Each input image consumes 1,120 tokens, and an output image consumes a token amount that scales with its resolution tier. Those figures describe the publisher's API accounting and are not the same thing as the points this site charges, which come from this site's own catalogue and are listed further down.
This page is not an official Google product description and is not endorsed by Google. It records what the publisher documented as of 7 October 2026 and what this site's catalogue returns.
What this catalogue declares for the product
The catalogue returns two callable, priced modes for nano-banana-2.1. They are what the workspace offers, and the task model sent with a request changes with the mode.
The form fields the catalogue publishes are exactly four: prompt, image_urls, aspect_ratio and resolution. The aspect-ratio field carries fourteen values and defaults to 1:1; the resolution field carries 1K, 2K and 4K and defaults to 1K. There is no quality or detail selector, no seed, no step count and no negative prompt, and the publisher's documentation independently confirms that the sampling parameters such a control would map to are unsupported by the model.
One number deserves its own paragraph. Google documents a ceiling of 14 reference images per prompt for this model. The catalogue entry publishes no reference-image limit at all, so the workspace applies its own setting of 14, which happens to match the publisher's documented figure. The agreement is coincidence of documentation rather than something this site copied from either source, and the workspace limit is the one a task actually runs into.
The charges above are this catalogue's figures, read from the live endpoint on 7 October 2026. They are not Google's API prices, which Google publishes separately on its own pricing page, and they are the numbers the quote shows before a task is accepted.
| Mode | Task model | Published charge | Status in the catalogue |
|---|---|---|---|
| text-to-image | nano-banana-2.1 | 3 points per result at 1K, 4 at 2K, 6 at 4K | callable and priced |
| image-to-image | nano-banana-2.1-i2i | 3 points per result at 1K, 4 at 2K, 6 at 4K | callable and priced |
Two aspect-ratio lists, and the ratio only one of them has
Two lists exist for this product and they are not identical. The catalogue this workspace talks to declares fourteen ratios. Google's documentation lists fifteen. The difference is a single value, and the catalogue is the list the selector obeys.
The practical consequence is one-directional. A ratio the catalogue lists is accepted by the service, because the service published the list. A ratio only the documentation lists, 9:21, cannot be selected here: the form sends aspect_ratio values from the catalogue's list, and no selector value would carry 9:21. If the catalogue adds the missing value later, the selector follows it without a change on this site's side, because the picker is built from the catalogue response.
Beyond that one value the lists agree, including the four extreme shapes, 4:1, 1:4, 8:1 and 1:8, that this generation of models accepts as generation targets. A ratio that extreme is a compositional decision rather than a crop, and it is worth choosing deliberately instead of leaving the default in place.
The comparison is between two sources this project reads directly: the live catalogue endpoint and Google's published model documentation, both accessed on 7 October 2026.
| Source | What it lists | Default |
|---|---|---|
| Catalogue entry (read 7 October 2026) | 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 9:21, no auto. | 1:1 |
| Google model documentation | 15 ratios: the same 14 plus 9:21. | not stated |
Input limits, as documented and as applied here
The publisher documents input limits precisely, and they are worth reading before a large edit task, because two of them bind earlier than a reader might expect.
- Reference images: up to 14 per prompt. Each input image consumes 1,120 tokens against the 131,072-token context window, so the ratio budget is not the binding constraint at fourteen images; the count itself is.
- File size: 7 MB per inline file, 30 MB per file through Google Cloud Storage, and a 500 MB total input size limit. This workspace uploads references to its own CDN and sends URLs, so the practical ceiling for an uploaded reference is the CDN path, not the console path.
- Formats: PNG, JPEG, WebP, HEIC and HEIF are accepted image input types. Video is a documented input modality for this model, and generation from video input is a supported capability, but the catalogue form for this product declares no video field, so a task placed here carries images and text only.
- Output: up to 32,768 output tokens per prompt, which the documentation notes bounds the number of output images. The workspace form has no image-count field, so each task requests one image, which is also what the per-result charging assumes.
How a task runs here
Arriving from this page pins the model; nothing else has to be chosen by hand. Everything below happens before anything is charged.
- Open this page in the workspace. The model is fixed to Nano Banana 2.1 and the form is built from the catalogue entry for it.
- Write the prompt. Both modes take the same kind of instruction: a plain description of the subject, the setting, the lighting and any text that must appear verbatim. There is no exclusion field, so anything to keep 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 then becomes image-to-image and the task model sent changes with it. Up to 14 references are accepted.
- Choose an aspect ratio and a resolution tier. The list and the tiers are the catalogue's own; 1:1 at 1K is where both start.
- Read the quote, then submit. The quote is computed from the catalogue's published per-result charges and is shown before the task is accepted. A failed task is reported rather than silently retried, and it appears in the task history with the settings it actually used.
Writing prompts for this model
The publisher's own examples across the Nano Banana family are one or two sentences of plain description, and nothing about this model's documented feature set asks for a different style. The suggestions below are working habits for this form specifically, not claims about the model.
The first habit is to name the text that must appear, in quotes, and to keep that text short. In-image text is a documented strength across this generation, but long strings give a task more chances to be wrong, and a check is cheaper than a retry.
The second habit applies to edit tasks: say what must not change. An edit instruction that names only the change leaves everything else to the model, while one that pins the label, the logo or the face gives the output something specific to preserve.
The third is to treat sampling as fixed. There is no seed or temperature here, so the way to get a different result from the same prompt is to change the prompt, the ratio or the tier, not to look for a randomness control this form does not have.
A flat-lay infographic of a coffee brewing guide on a light grey background, with four numbered steps connected by hand-drawn arrows, each step labelled in short handwritten English: grind, bloom, pour, serve.
Take the product photo and place the same bottle on a wet slate surface with condensation drops, keeping the label text and logo exactly as they are, lit from the upper left with a soft shadow to the right.
A single continuous skyline of a harbour city at dusk, composed for an extremely wide 21:9 frame, with ferries leaving light trails on the water.
Where it fits, and where it does not
The capability table and the catalogue together sort tasks into three groups: documented and offered, offered but worth checking, and not offered. The middle group is where expectations do the most damage.
- Fits: posters, social visuals and infographics with short in-image text; product and scene imagery; reference-guided edits where part of the frame must survive unchanged; wide or tall compositions for the extreme ratios the catalogue lists.
- Check before relying on it: person consistency across multiple references. Multi-reference input and multi-turn editing are documented capabilities, but person portrait generation is listed as not supported in the same table, and the boundary between an edited photo of a person and portrait generation is the publisher's to draw, not this page's. Users discussing the family in public forums have reported mixed results on identity retention; treat any such report, including that one, as an individual experience rather than a verdict, and judge your own output against the checklist below.
- Does not fit: tasks that need a seed, a temperature or any sampling control, which the model rejects at the API level; tasks that need video input, which the capability table supports but this catalogue form does not carry; and tasks that need portrait generation, which the publisher lists as unsupported.
Common failures, and what to do about them
The failure modes below are the ones this form's shape makes possible, with the recovery that costs the fewest points.
- A parameter error mentioning seed, temperature, topP, topK or logprobs means a request carried a field the model rejects. This form never sends those fields; if you see such an error, it came from a stale page or a modified request, and a reload fixes it.
- A reference image rejected on upload is usually a format or size problem: anything outside PNG, JPEG, WebP, HEIC or HEIF, or an individual file past the size limits, fails before a task is created and nothing is charged.
- Text in the image rendered wrong is an output problem, not a submission problem: shorten the string, put it in quotes, and regenerate. There is no edit control to fix a single glyph.
- An edit that changed too much is the mirror image: re-run with an instruction that names what must stay, not only what should change.
- A task that fails after acceptance is reported in the task history with its settings. This site's pricing rule for failures is that verified failed tasks are credited back; the history entry is what the credit check reads, so keep it rather than resubmitting blindly.
An output checklist before you use the result
Thirty seconds against this list catches most of what a second generation would not fix.
- Every in-image string spelled correctly, including the small ones at the edges.
- For edits: the parts you pinned are unchanged, at full resolution, not just at thumbnail size.
- The aspect ratio is the one you selected, checked against the file's real dimensions.
- Faces and logos read correctly at 100 percent zoom if the task involves either.
- The output tier matches what you paid for: a 2K or 4K task should return visibly more pixels than a 1K draft, and the file's dimensions say so directly.
Choosing between 2.1, 2 and 2 Lite
Three Nano Banana generation lines are on this site's shelf, and the publisher's own positioning plus this catalogue's numbers separate them more cleanly than any benchmark quote could.
Read the middle column as positioning and the third as this site's published charges, which are the only numbers a task here is actually billed in. On those numbers, 2.1 sits below 2 at every tier above 1K and above Lite at its single tier, which is the cost side of the publisher's cost-effectiveness sentence applied to this site's price list rather than an independent quality judgement.
The practical picker: drafts and bulk variants where the 1K tier and the lowest price matter go to Lite; 2K and 4K work where the ratio list is wide and the reference count is high goes to 2.1; and tasks built around 2's specific documented capabilities, such as Google Search grounding with image search documented for the Flash Image generation, stay a reasonable choice on their own page, which describes what that model carries in detail.
| Product | Publisher's positioning | This catalogue's charges | Reference images here |
|---|---|---|---|
| Nano Banana 2.1 | Optimised for multimodal generation and editing, good cost-effectiveness. GA 6 October 2026. | 3 / 4 / 6 points per result at 1K / 2K / 4K | up to 14 (site setting; publisher documents 14) |
| Nano Banana 2 | The generalist workhorse; balances speed against 4K generation, world knowledge and text rendering. | 3 / 6 / 9 points per result at 1K / 2K / 4K | up to 14 (site setting) |
| Nano Banana 2 Lite | Fastest and cheapest; not optimised for multiple reference inputs or multi-turn sequential edits. | 1 point per result, single 1K tier | up to 5 (site setting) |
Examples

Batch consistency with a product as the subject. The same bottle design, label text and palette survive across more than forty compositions, which is the capability a product-photoshoot task actually leans on. The text on the label stays short, which is part of why it holds.

Multi-reference composition: four inputs - a face, a flower arrangement, a sky and a watering can - blended into one scene. The thumbnail strip on the left is part of the published image, so you can check each reference against what survived. The face and the metal can are the two elements worth comparing closely.

Scene swap around fixed subjects. The sofa, the person and the small bears keep their design while the entire environment changes from a flower meadow to a plain room, which is the edit shape that turns a usable photo into a reusable one.

One persona across five scenes - the social-feed shape. Outfit palette, hair and makeup read as the same person throughout, at the candid image quality a feed post needs. Person consistency is also the capability the publisher notes has limits, so treat a set like this as something to check panel by panel, not to assume.

A product line with legible in-image text. Seven containers, each carrying a correctly spelled brand and product name at different sizes, on one consistent backdrop. This is the catalogue-shot task, and the text density here is what tends to break on a second attempt - count the strings that must survive before you raise it.

Style transfer with the reference kept honest. The inset photograph labelled Original is part of the image, so the illustration can be judged against its source: pose, headwear and flower placement carry over while the rendering changes completely. Publishing the reference inside the frame is also what makes this one checkable rather than a claim.
Questions
What is Nano Banana 2.1's model identifier?
Google publishes it as gemini-nano-banana-2.1, generally available since 6 October 2026. This site's catalogue lists the product as nano-banana-2.1, which is what the URL and the workspace model selector use.
Which modes are open on this site?
Two: text-to-image and image-to-image, both callable and priced in the catalogue. The image-to-image mode routes to its own task model, nano-banana-2.1-i2i, and is selected automatically when a reference image is attached.
How much does a task cost?
This catalogue publishes 3 points per result at 1K, 4 at 2K and 6 at 4K, in both modes. The quote shown before you accept a task is computed from these figures. They are this site's points, not Google's API prices.
Which aspect ratios can I use?
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, defaulting to 1:1. Google's documentation lists one more, 9:21, which the catalogue does not carry, so it cannot be selected here. If the catalogue adds it, the selector picks it up on its own.
How many reference images can I attach?
The workspace accepts up to 14, which matches the per-prompt ceiling Google documents for this model. Each input image consumes 1,120 tokens of the model's context window on the publisher's side, so very large reference sets also grow the request, but the count limit is what binds first.
Can I control seed or temperature?
No. Google's documentation states that Nano Banana 2.1 does not support seed, topK, logprobs, temperature or topP, and that setting any of them returns an API error. The form offers no such control, and two runs of the same prompt can differ.
Does it generate portraits of people?
The publisher's capability table lists person portrait generation as not supported for this model. Editing photos that contain people is a documented editing capability, but the boundary between the two is the publisher's to draw, so a person-focused task is one to check against the output checklist rather than assume.
Do outputs carry provenance metadata?
Google lists Content Credentials (C2PA) as supported for this model, and SynthID provenance is standard across the Nano Banana family per the publisher. Both are metadata-level mechanisms that need a verification tool to detect. This site adds no watermark and does not strip provenance metadata.
Explore next
AI image generator
Describe an image or upload a reference, choose a model, and read the quoted credit cost before generating.
Nano Banana 2
Nano Banana 2 is Gemini 3.1 Flash Image. What Google published about it, what this catalogue declares, and the controls the publisher ships that this workspace form does not carry.
Nano Banana 2 Lite
Nano Banana 2 Lite is Gemini 3.1 Flash-Lite Image. What the publisher published about it, what this catalogue returns for it today, and why the workspace form cannot send a task to it yet.
Image to prompt
Upload a picture, get a written description you can reuse as a prompt. Five target models, three detail levels, optional negative prompt, four output languages.
From prompt to finished image
A worked walkthrough of one image task: choosing a model, writing the prompt, reading the credit quote and judging the result.