Nano Banana 2.1 image generator and editor

Nano Banana 2.1 is an image generation and editing model from Google. It was released on October 6, 2026, and in the Gemini API it is called gemini-nano-banana-2.1. Google describes it as an update to Nano Banana 2, the model also known as Gemini 3.1 Flash Image, and has named it the replacement for that model, which is now deprecated.

The idea behind the release is simple: keep the speed and low cost of a Flash-class model, but make the pictures better. Google lists improvements in visual quality, prompt adherence, character consistency across multiple turns, text rendering, and wide and panoramic aspect ratios. The model can create images from text alone, edit a photo you provide, or combine several reference images into a new scene.

On PixArmory you do not need a Google Cloud account, an API key or any code. You write a prompt, add images if you want to, and the result is saved to your history like any other PixArmory image.

If you already used Nano Banana 2, the controls will feel familiar. The differences show up in the results, and in the price.

Text is the most visible change. Menus, posters, labels and infographics come out cleaner and are misspelled less often, especially when you put the exact words in quotation marks. Character consistency is the second big one: when you edit the same person, pet or mascot several times, the new model is better at keeping their face, colors and proportions. Google also improved wide and panoramic formats, and lists 1:4, 4:1, 1:8 and 8:1 among the ratios that now work much better at 1K, 2K and 4K.

There is one thing it no longer offers: the small 512-pixel output that Nano Banana 2 supported. The new version starts at 1K.

Finally, the cost. Google's official API price for image output dropped from $60 to $30 per million tokens, which makes a 1K image about $0.034 instead of about $0.067. PixArmory passes that saving on, so the new model costs fewer credits here than Nano Banana 2 at every resolution.

All three models are available on PixArmory. For most everyday work, Nano Banana 2.1 is the best starting point. Nano Banana Pro, built on Gemini 3 Pro, is still the premium option when a picture depends on deep world knowledge or very fine creative control. Nano Banana 2 stays available for now, so existing workflows keep working while you compare results.

1. Write a prompt. Describe the subject, the setting, the style and any text that should appear in the image.
2. Add images if you are editing. Upload up to 14 JPG, PNG or WebP files, each up to 15 MB. Leave this empty to create a brand-new image.
3. Choose a resolution and an aspect ratio. Auto lets the model pick a frame that suits your prompt.
4. Turn on Google Search only when the image depends on current, real-world information.
5. Press the button. The exact credit cost is shown on it before you click, and the result appears on the right.

You can download the result, or use it as the starting point for the next edit. A good habit is to change one thing per round. Small, specific steps are easier for the model to follow than one long list of changes.

1K is ideal for drafts, social posts and quick tests. 2K gives you more detail for websites, presentations and product pages. 4K is worth the extra credits when the image will be printed, cropped heavily or shown on a large screen. A practical approach is to explore ideas at 1K and repeat your favorite prompt at 2K or 4K.

Match the aspect ratio to where the image will end up: 1:1 or 4:5 for social feeds, 9:16 for stories and phone screens, 16:9 for videos and slides, and 3:4 or 2:3 for posters. The panoramic ratios 4:1 and 8:1 suit website banners, and 1:4 or 1:8 suit tall formats such as a skyscraper ad or a long infographic. If you choose an extreme ratio, describe what should sit at each end of the frame, so the extra space is used well.

According to Google, Nano Banana 2.1 accepts up to 10 object images for high-fidelity details and up to 4 character images for consistency, 14 in total. In practice, that means you can upload a product from several angles, a logo, a texture and a couple of people, and ask the model to put them together.

References work best when your prompt says what each image is for. Write “the first image is the candle, keep its label unchanged” rather than “use these images”. Name the things that must not change, such as the label text, a face or a color, and then describe what should be new. For a recurring character, keep one clean reference image and reuse it every time, instead of feeding back an edited copy of an edited copy.

1. Put any text that must appear in quotation marks, and keep it short. One headline and a few lines work better than a full paragraph.
2. Describe the scene like a photographer or designer would: subject, setting, light, camera angle and mood.
3. State the purpose, for example “a website banner with space for a headline on the left”. The model uses this to plan the layout.
4. Say what to leave out. “No people, no extra text, no watermark” prevents unwanted details.
5. For edits, say what stays the same before you say what changes.

The new model is a clear step forward, but it is not perfect. Very small text, long paragraphs and unusual fonts can still contain mistakes, so always proofread before you print. Character consistency is much better, but a face can still drift after many edits. Complex spatial details, such as hands holding several objects or precise technical diagrams, sometimes need another attempt.

Google adds an invisible SynthID watermark to every image the model creates, so the image can be identified as AI-generated. Requests that break Google's or PixArmory's content policies are refused. Facts in an image are only as current as the model's training data, unless you turn on Google Search grounding.

A Nano Banana 2.1 image costs 10 credits at 1K, 13 credits at 2K and 26 credits at 4K. Google Search grounding adds 3 credits. The number of reference images does not change the price, and the total is always shown on the generate button before you confirm. If a generation fails, the credits are returned automatically. New accounts receive free credits, so you can try the model before choosing a plan.