Nano Banana 2.1: What Changed, 4K, References and Prompts

Key facts at a glance

Developer

Google DeepMind

Model ID

gemini-nano-banana-2.1

Release date

October 6, 2026

Base model

Gemini 3.6 Flash

Resolution

1K, 2K, 4K (default 1K)

References

Up to 14 images; 4 characters, 10 objects

Open weights

No, closed model

Uncensored?

No, Google content policies and SynthID

What is Gemini 3.6 Flash Image?

Google DeepMind ships its Gemini image models under the consumer name Nano Banana, and version 2.1 is the latest Flash member of that series. Instead of retuning the earlier Gemini 3.1 Flash Image, Google moved the image model onto Gemini 3.6 Flash. The new model renders at 1K, 2K or 4K (1K by default), accepts up to 14 reference images, and adds a thinking setting with minimal, medium and high levels. Google lists it in the Gemini app, Search AI Mode, Flow, Google Ads and Stitch for everyday use, while Google AI Studio and the Gemini API are the developer routes. It is a closed model: no weights, no local build.

Why people look for an uncensored Gemini 3.6 Flash Image

The model inherits Google's filters. Prompts and uploaded photos are checked against Google's generative AI prohibited use policy, swimwear, lingerie or body edits are often refused even when tasteful, and every output carries a SynthID watermark with C2PA Content Credentials. That is why people search for an uncensored version or an NSFW mode. No such build exists, and we do not publish bypass tricks. If you want uncensored generation, you can do it on this site: Spicy AI's image editor does not host Google's model but runs Seedream, Krea 2 and Wan 2.6 for adults (18+) editing images they have the right to use.

Five things that changed in version 2.1

Version 2.1 is a new base model rather than a patch, so treat it as a separate tool with its own trade-offs. The main differences, all taken from Google DeepMind's model card and the Gemini API documentation:

• Base model: Gemini 3.1 Flash Image was replaced by Gemini 3.6 Flash, which also refreshes part of the knowledge to March 2026.
• Resolution: fixed 1K, 2K and 4K tiers with 1K as the default, instead of the 512px-to-4K range of version 2.
• Wide formats: stitching artifacts at 1:4, 4:1, 1:8 and 8:1 in 2K and 4K were fixed, so banners and panoramas hold together.
• References: up to 14 reference images per request, with consistency for up to 4 characters and 10 objects.
• Control: a thinking level (minimal, medium or high, medium by default) trades speed for layout planning.

Nano Banana 2.1 vs version 2 and the Pro tier

The three models overlap, so pick by job, not by version number. Version 2.1 is the efficient default for wide 4K layouts, reference-heavy edits and dense infographics where a thinking pass helps. Version 2 keeps the higher limit of 5 characters and 14 objects, which matters for crowded group scenes. The Pro tier, Gemini 3 Pro Image, is still the choice Google recommends when factual precision matters most. Google also lowered the per-image output price compared with version 2, although input tokens cost more, so a request with many reference photos may not end up cheaper.

Release dates for 2 and 2.1 come from Google DeepMind model cards; the Pro date follows press reports. Cells marked as not compared were not checked against a primary source.

14 reference images, 4 characters and 10 objects

The most common mistake in tutorials is reading 14 reference images as 14 consistent subjects. The model will accept 14 images as context, yet Google only promises consistent identity for 4 characters and faithful detail for 10 objects. A practical setup is one clean, front-facing photo per character, one image per product or prop, and the remaining slots for style, palette or layout references. Repeat a short fixed description of each character in every prompt, and expect some drift across long multi-turn sessions. For final output, iterate at 1K and render only the approved layout at 4K, because large renders are slower and cost more.

How the thinking levels work

Thinking is new in version 2.1. At minimal, the model answers fastest and suits simple edits such as an outfit color change or a background swap. Medium, the default, adds planning for most posters and multi-image blends. High spends more time on composition and is worth it for dense infographics, multi-panel layouts or prompts that combine many references. Google's model card credits this extra reasoning for part of its gains in infographic and multi-character Elo scores, but those are Google's own ratings and have not been independently reproduced. In the Gemini app the level is chosen for you; developers set it per request.

Is the 2.1 model open source or available offline?

No. Google has not released weights, code or an offline build of the model, and any GGUF file, ComfyUI checkpoint or Hugging Face upload that claims to be it is another model or a scam. Google offers the model as a whole through its own products; the API is only one way in. That also means you cannot fine-tune it, add LoRAs or remove its safety layer yourself.

What people make with the 2.1 model

Early galleries lean on what the new base model does better: wide formats, text and many references at once. These are the creations that show up most in Google's examples and shared community posts:

• Panoramic banners and website headers at 4:1 or 8:1 that no longer show seams in the middle.
• Dense infographics and explainer cards rendered on the high thinking level.
• Lookbooks that combine a model photo, several garments and a location reference in one request.
• Localized ad sets where only the language of the text changes between versions.
• Comic strips and storyboards with up to four recurring characters.

Prompt guide with copyable templates

The model follows prompts best when they are clear sentences: subject, change, setting, lighting, format, and what must stay fixed. Name each reference image by role, for example "image 1 is the model, image 2 is the jacket". Quote any text you want rendered. State ratio and resolution, and raise the thinking level only when the layout is complex. Copyable templates:

• Multi-reference: "Dress the person in image 1 in the jacket from image 2, place them in the street from image 3, keep the face unchanged."
• Panorama: "An 8:1 banner of [scene] at golden hour, continuous horizon, no repeated objects, 4K."
• Infographic: "A 4:5 infographic titled \"[title]\" with [n] labeled steps: [data]. Flat icons, large readable labels."
• Edit: "Change [element] to [new element]; keep pose, lighting and background identical."
• Storyboard: "Four panels of [two adult characters] doing [actions], same faces and outfits in every panel."

Known limits to plan for

Google's documentation lists weaknesses that version 2.1 did not remove. Small text is still often blurry at 1K, and long paragraphs or full-page documents remain weak. Character consistency is not always perfect. Mask or scribble edits may be followed only partly or leave marks. Left and right can be mixed up. The context figure is also inconsistent: the model card mentions up to 1M tokens, while the Gemini API model page lists 131,072 input and 32,768 output tokens, so plan integrations around the API page.

Is there an uncensored Nano Banana 2.1?

No. Google has not released an uncensored, NSFW or unfiltered build, and none of its channels has a switch that turns the content policy off. The Gemini API lets developers adjust some safety thresholds, but Google's core protections and usage policy still apply. Some third-party API resellers advertise looser filters, yet they still call Google's model under Google's terms, so that is not a true uncensored version. If you need adult (18+) image generation, the honest route is a substitute, described next.

Uncensored alternative on Spicy AI

Spicy AI's image editor is a separate, adults-only tool where you can generate without Google's content filters. It does not run any Google model; it runs Seedream, Krea 2 and Wan 2.6, so results will look different. Open the editor, choose text to image or upload a reference, paste one of the templates above, state what must stay unchanged, and generate. Only use photos you own or have permission to edit, and never depict minors or real people without consent.

Prompt tips for Gemini 3.6 Flash Image-style results

1. Label every reference image by its role in the prompt.
2. Draft at 1K and render only the final layout at 4K.
3. Use the high thinking level only for dense layouts; keep simple edits fast.
4. State that everyone shown is an adult (18+), and edit only photos you have the right to use.

Sources

• Version 2.1 model card (Google DeepMind)
• gemini-nano-banana-2.1 model page (Gemini API docs)
• Gemini API pricing (Google AI for Developers)
• Gemini 3.1 Flash Image model card (Google DeepMind)

More uncensored AI Image models

• uncensored nano banana 2
• Uncensored Midjourney
• Uncensored DALL·E
• Uncensored Stable Diffusion
• Uncensored Flux
• Uncensored Nano Banana
• Uncensored Nano Banana 2
• Uncensored Ideogram
• Uncensored Leonardo AI
• Uncensored Seedream
• Uncensored Seedream 5.0
• Uncensored Krea
• Uncensored Kroma
• Uncensored Qwen-Image
• Uncensored Adobe Firefly
• Uncensored GPT Image 2.5
• Uncensored Text to Image
• Uncensored Z-Image Turbo
• Uncensored Z-Image
• Uncensored FLUX.2
• Uncensored FLUX.2 [dev]
• Uncensored FLUX.2 [pro]
• Uncensored GPT Image 1.5