Nano Banana 2.1 AI Image Generator

Nano Banana 2.1 (model ID gemini-nano-banana-2.1) is the update to Nano Banana 2. The Gemini API treats it as the efficient default for image generation and conversational editing, and recommends it for new projects.

It is based on Gemini 3.6 Flash and became generally available on Google Cloud on October 6, 2026. It keeps Flash-level speed at 1K, 2K, and 4K, with better visual quality, text rendering, multi-turn consistency, and Google Search grounding. Text, images, and video can go in together. Generated images include a SynthID watermark.

Where Nano Banana 2.1 Fits

• Everyday generation and edits for new work: Flash-level speed and cost, with a step up from Nano Banana 2 on quality and multi-turn consistency.
• Infographics, menus, and short marketing titles: text rendering is stronger on short copy, and still weak on a full page.
• Ongoing edits of one character or product: continue in conversation instead of regenerating the whole image.
• A poster or thumbnail from video: the model reads visual themes and key events, then composes a new still.

Try Nano Banana 2.1

Nano Banana 2.1 Core Features

Conversational generation and edits

One model does text-to-image, and uses text to add, remove, or replace elements, or to change style and color. Google recommends iterating in conversation. A typical follow-up is to change only the language of an infographic and leave every other element alone.

1K, 2K, and 4K, with no 0.5K

The Gemini API lists 1K, 2K, and 4K for 2.1. It does not support the 512px (0.5K) size added on Nano Banana 2. Google Cloud's output tokens are about 1,120 at 1K (about 1 megapixel), 1,680 at 2K (about 4 megapixels), and 3,780 at 4K (about 16 megapixels). How many images come back is limited by the 32,768 output-token cap.

Up to 14 references, split between objects and characters

References total up to 14 images. The Gemini API splits that for 2.1 into up to 10 high-fidelity object images and up to 4 character images. Each input image costs about 1,120 tokens. Console or inline uploads are capped at 7MB per file. Cloud Storage files are capped at 30MB. Formats are PNG, JPEG, WebP, HEIC, and HEIF.

Thinking sketches first

The model creates temporary thought images to refine the composition. Google says those intermediate images are visible on the backend and are not charged separately. The file you receive is the final image. In DeepMind's October 2026 comparison, overall preference is 1050 (±14) with thinking and 1015 (±13) without it. Infographic factuality is 0.521 versus 0.328. Turn thinking off when speed matters more.

Web search and image search are separate

Web search checks live facts such as weather, prices, and recent events, then draws them. With web search alone, image results are not passed into the generator. Visual references from the web need image search turned on as its own search type. Nano Banana 2.1 and Nano Banana 2 both support it. Nano Banana 2 Lite does not.

A video can become one image

Video is input only. The model reads themes and key events, then combines them with the prompt into a thumbnail, poster, or summary infographic. You can pass a public YouTube URL or upload a local file. Google Cloud caps each prompt at 10 video files and 1 YouTube URL, about 25 minutes without audio. Audio is not a supported modality.

Nano Banana 2.1 Compared with Earlier Models

Positioning and the reference split come from the Gemini API image generation guide. Preference scores come from the DeepMind model card comparison in October 2026. Resolution tokens come from the Google Cloud docs.

Item	Nano Banana 2.1	Nano Banana 2	Nano Banana Pro
Official role	Efficient default for new projects	Previous efficient default	Premium tier for the hardest tasks
Base model	Gemini 3.6 Flash	Gemini 3.1 Flash Image	Gemini 3 Pro Image
Resolution	1K, 2K, 4K, no 0.5K	Includes 0.5K, plus 1K to 4K	Up to 4K
References	Up to 10 objects + 4 characters	Same split as 2.1	Up to 6 objects + 5 characters + 3 style images
Thinking, overall preference	1050 ± 14	990 ± 7	935 ± 8
Infographic factuality	0.521	0.179	0.265
Multi-character consistency	1106 ± 14	978 ± 10	1011 ± 10

Without thinking, 2.1 scores 1015 ± 13 on overall preference, 0.328 on infographic factuality, and 1068 ± 14 on multi-character consistency. With thinking, ink or mask edits score 1049 ± 15, and multi-reference edits score 1066 ± 22.

Use Nano Banana 2.1 for new projects. For 0.5K or an existing Gemini 3.1 Flash Image workflow, see Nano Banana 2. For stronger world knowledge, localization, and brand control, see Nano Banana Pro.

Nano Banana 2.1 Use Cases

• Infographics, menus, diagrams, and posters with a short title
• Multi-turn edits, including a language change that leaves the rest of the image alone
• Product shots with a new background or color, held to a reference
• Up to 4 character images to lock people, then a new setting
• A thumbnail or poster made from a video
• A chart drawn from a live forecast, an event, or a search result

How to Use Nano Banana 2.1

1. Open the AIGCVA image workspace and select Nano Banana 2.1.
2. Describe the subject, the scene, the style, and any short text that must appear. Use object references to lock a product and character references to lock a person. Stay within 14 images total.
3. Choose 1K, 2K, or 4K. Turn thinking on for infographics and facts you need to check.
4. If the result is close, describe only the change and say what must stay. Check small text, left and right, and whether the person still matches the reference.

Nano Banana 2.1 Best Practices

• In a follow-up, describe only the change. For example, "Change the words to English. Do not change layout, icons, colors, or numbers."
• References hold identity and product. The prompt names the one change.
• Keep on-image text short. The model card still lists small type at 1K, and page-length paragraphs, as weak spots.
• Use web search for live facts. Turn on image search when you need pictures from the web as visual context. Web search alone does not feed those pictures into the generator.
• For ink or mask edits, describe the marked region. Instructions are sometimes only partly followed, and ink does not always persist.
• The knowledge cutoff is about March 2026, and some domains still stop at January 2025. Send time-sensitive facts through search.
• The Google Cloud API rejects seed, temperature, topP, topK, and logprobs. Sending them returns an error.

Nano Banana 2.1 Prompt Examples

Infographic

Make a vertical infographic. The only title is "Three steps to list it." The steps are photograph, write the selling point, and publish. Each step gets one icon and one line of no more than 12 words. Do not add numbers that were not provided.

Change only the language

Change the words on this infographic to English. Keep the layout, icons, colors, and numbers exactly as they are.

A forecast chart

Draw the next five days of the weather forecast as a clear chart. Add a small icon for what to wear each day. Do not invent temperatures that were not retrieved.

A poster from video

Make a 16:9 poster from the key themes of this video. Keep the main locations from the video. Do not add people who are not in it.

Nano Banana 2.1 FAQ

What is Nano Banana 2.1?

Nano Banana 2.1 is the update to Nano Banana 2, and the Gemini API recommends it for new projects. The model ID is gemini-nano-banana-2.1. It is based on Gemini 3.6 Flash and became generally available on Google Cloud on October 6, 2026. It keeps Flash-level speed at 1K, 2K, and 4K, with better visual quality, text rendering, multi-turn consistency, and Google Search grounding. Inputs can be text, images, and video. Outputs are images and text. Generated images include a SynthID watermark.

How should I choose among Nano Banana 2.1, Nano Banana 2, and Nano Banana Pro?

Use Nano Banana 2.1 for new projects. It is the efficient model for conversational editing. Keep Nano Banana 2 when a workflow already runs on Gemini 3.1 Flash Image, which still has the 0.5K size that 2.1 does not. Nano Banana Pro remains the tier for the hardest world knowledge, localization, and brand control. In DeepMind's October 2026 comparison, 2.1 thinking scores 1050 on overall preference, Nano Banana 2 thinking scores 990, and Nano Banana Pro scores 935.

What are the limits on resolution, references, and aspect ratio?

It supports 1K, 2K, and 4K, and it does not support the 0.5K (512px) size on Nano Banana 2. 1K is about 1 megapixel and about 1,120 output tokens, 2K is about 4 megapixels and about 1,680 tokens, and 4K is about 16 megapixels and about 3,780 tokens. References total up to 14 images, with up to 10 high-fidelity object images and up to 4 character images. Each input image costs about 1,120 tokens. A console upload is capped at 7MB per file, and Cloud Storage at 30MB. Aspect ratios include 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9, 9:21, 1:4, 4:1, 1:8, and 8:1.

Which context window should I trust?

The model card lists an input context of up to 1 million tokens and a 64K-token text output. The Google Cloud API lists a 131,072-token context window and a 32,768-token output cap. The two official documents disagree, so use the limit of the endpoint you actually call.

What changes when thinking is off?

Thinking first creates temporary thought images to refine the composition. Google says those intermediate images are visible on the backend and are not charged separately. You receive the final image. The model card scores Thinking and No Thinking separately. Infographic factuality is 0.521 with thinking and 0.328 without it. Multi-character consistency is 1106 with thinking and 1068 without it. Turn thinking off when you need a faster image.

How reliable is small text?

Text rendering is improved over Nano Banana 2, and it fits infographics, menus, diagrams, and short marketing titles. The model card still says small text is often blurry at 1K, and long paragraphs or page-length text are still unstable. Left and right, character consistency, and a pose carried over from the input can still miss, so check the image before you publish it.

Can I use it commercially?

Yes, for posters, infographics, product shots, and multi-turn edits. Rights follow the terms of the service you use. Generated images include a SynthID watermark. Google Cloud lists Content Credentials (C2PA) as supported. You need the rights to any reference image you upload.

Nano Banana 2.1 Technical Advantages

• The Nano Banana 2 update, recommended for new projects, model ID gemini-nano-banana-2.1.
• 1K, 2K, and 4K, with no 0.5K. Up to 10 object references plus 4 character references.
• Thinking creates temporary draft images. Google says those drafts are not charged separately. Scores are higher than earlier models on infographic factuality and multi-character consistency.
• Web search handles live facts. Image search handles visual references. They are separate switches.
• Multi-turn edits, image-from-video, virtual try-on, and portraits. Outputs carry SynthID, and Cloud lists C2PA.
• Known gaps include small text at 1K, long copy, left and right, and a knowledge cutoff around March 2026.

Try Nano Banana 2.1

Related Resources

• Gemini API: image generation (Nano Banana 2.1)
• DeepMind model card: Nano Banana 2.1
• Google Cloud: gemini-nano-banana-2.1

Related Articles

• Nano Banana 2 AI Image Generator
• Nano Banana Pro AI Image Generator
• GPT Image 2.5 AI Image Generator
• AI Image Generator collection