Fewer Do-Overs. Nano Banana 2.1 Keeps Faces, Text and Small Edits in Place

The Image Is Almost Done, You Just Need One Fix

You ask an AI to remove an extra glass from the table. It removes the glass, but it also changes the lighting, moves the mug and gives your character a stranger's face. You start over. Ten attempts later, it feels like grabbing your own camera would have been faster.

That is exactly the problem Google is going after with this update. On October 6, 2026, it released Nano Banana 2.1, a new version of its image generation and editing model. The headline promises are about targeted edits, consistent characters and text inside images, which is everything that usually sends you back to the "generate again" button.

We went through Google's model card and developer docs and figured out how to work with the new version in Unitool.

What Nano Banana 2.1 Is and Where It Fits

Nano Banana is Google's name for its image models in the Gemini family. There are now several models with similar names, so here is the short version.

• Nano Banana is the first version, built on Gemini 2.5 Flash Image.

• Nano Banana 2, also called Gemini 3.1 Flash Image, is the fast everyday model.

• Nano Banana Pro, or Gemini 3 Pro Image, is the top model for the hardest jobs.

• Nano Banana 2.1 is an update to Nano Banana 2, built on Gemini 3.6 Flash.

Google describes 2.1 as the more efficient counterpart to Nano Banana Pro. It stays in the fast Flash class but moves closer to the senior model in quality. Pro is still around as a separate model.

What Changed Compared With Nano Banana 2

According to the Gemini API documentation, version 2.1 brings six notable changes.

• Better image quality and realism at 1K, 2K and 4K. The default is 1K.

• Fixed tiling artifacts on very wide and very tall formats such as 1:4, 4:1, 1:8 and 8:1 at 2K and 4K.

• More accurate text inside images and better infographic layouts.

• Up to 14 reference images. Up to 4 of them can be characters whose look the model tries to keep, and up to 10 can be objects.

• Optional grounding with Google web and image search, so the model can check real-world details before it draws.

• Three thinking levels, minimal, medium and high. Thinking is the stage where the model plans the composition before it renders.

For an everyday user this means fewer surprises. The model better understands which part of the image you want changed, and it forgets your character's face less often between edits. When accuracy matters, you can ask it to check search first.

How Google Tested the Model and What the Numbers Mean

Most of the tests are blind comparisons. People see two images made from the same prompt without knowing which model made which, and pick the better one. Thousands of those choices turn into an Elo rating, the same system used in chess. The higher the number, the more often people preferred that model's image.

For overall preference in text-to-image, Nano Banana 2.1 with thinking scored 1050, without thinking 1015, Nano Banana 2 scored 990 and Nano Banana Pro 935. Infographic design shows a similar gap, 1048 for the new version against 961 for the previous one and 912 for Pro.

The editing results are even more interesting, and they are the main reason to upgrade. The biggest jump came in scenes with several characters, 1106 points against 978 for Nano Banana 2 and 1011 for Pro. In mask-based editing, where only a marked area changes, the score is 1049 against 965 and 927. With multiple references it is 1066 against 988 and 989.

Google also measured the factual accuracy of infographics, this time scored by an automated rater. Nano Banana 2.1 with thinking got 0.521, without thinking 0.328, Nano Banana 2 got 0.179 and Pro 0.265. In other words, thinking mode helps a lot when the image has to get the facts right.

One important caveat. These are Google's own tests, and each rating has a margin of error between 7 and 22 points. A model rated 1050 does not draw "6% better" than one rated 990. It simply wins pairwise comparisons more often on Google's set of tasks. On your tasks the result may differ.

Price and an Honest List of Weak Spots

In the Gemini API, which is how services plug the model in, a 1K image costs $0.0336, compared with $0.067 for Nano Banana 2. A 2K image costs $0.0504 versus $0.101. That is roughly half the price with better scores.

Google lists the weak spots itself.

• Small text, long paragraphs and full pages of text come out poorly. At 1K, small lettering is often blurry.

• A character in the result does not always match the input photo exactly.

• Mask-based edits are sometimes only partly followed.

• The model mixes up left and right.

• World knowledge, 3D reasoning and factual accuracy still fall short on complex tasks.

Two practical rules follow from that list. Use 2K or higher for anything with text, and double-check results when your prompt depends on where things are placed. "Left" in the prompt and "left" in the picture do not always match yet.

How to Turn On Nano Banana 2.1 in Unitool

The new version is already in Unitool, and you do not need to hunt for it. Open the Nano Banana 2 chat, and you will see a settings panel on the right.

1. Model version. Choose Nano Banana 2.1. Nano Banana 2 and Nano Banana 1 are still in the same list.

2. Image Size. Auto, Square (1:1), Landscape 16:9, 4:3 and 3:2, Portrait 9:16 and 3:4.

3. Resolution. 1K, 2K or 4K.

4. Number of Images. From 1 to 4 per request.

5. Thinking level. Minimal, Medium or High. Unitool sets Minimal by default.

6. Google Search. Off, Web, Images, or Web and images.

A note under the panel says the model takes up to 14 reference images and that search may add to the price. You upload photos for editing with the image button on the left of the input bar, next to a 0/14 counter. Unitool shows the request price under the input bar before you run it. In our test, one Nano Banana 2.1 image at 1K with minimal thinking and no search cost 12.5 tokens, and at 2K with high thinking it cost 18.75 tokens.

Pay attention to the thinking setting. Google's default is medium, while Unitool starts at minimal. That works well for quick drafts. For images with text, infographics and scenes with several characters, switch to Medium or High. The price difference is small, and it costs less than three extra tries in Minimal mode.

Scenario. A Product Listing in One Conversation

Say you sell handmade ceramic mugs on Etsy or Amazon and need listing photos. You have no studio and no photographer, just a phone shot of the mug.

Step one. In the Nano Banana 2 chat, choose version 2.1, Square (1:1), 2K resolution, Medium thinking and 2 images. Upload the mug photo and type this.

"Place this mug on a light wooden table by a window, soft morning light, a croissant on a plate and a linen napkin next to it. Keep the mug's shape, color and glaze exactly as in the photo."

Step two. Pick the best version and change one detail without touching the rest.

"Replace the croissant on the plate with a glass of water with lemon. Keep the lighting, the mug and the napkin unchanged."

If the model mixes up sides, name the object instead of its position.

"The plate with the croissant" works better than "the item on the left."

Step three. Make an infographic for the second listing image. Raise thinking to High.

"Turn this image into a product listing card. On the right, add three short labels, 'Handmade', '12 oz', 'Dishwasher safe'. Simple, large font, no small print."

Check every letter. Text in 2.1 is more accurate, but a typo on a listing is something your customers will notice, not Google.

Step four, if you need a real background. Set Google Search to Images and ask for the mug on a windowsill overlooking a specific street or landmark. Search helps the model avoid inventing what a place looks like, but it makes the request more expensive.

Four requests give you a small set of photos and an infographic in one style. No studio, no photo editor and no shuffling files between services.

In Practice. Who Gets the Most Out of 2.1

Online sellers can build a series of listing images with the same product in different scenes. Upload up to 10 object photos, and the model keeps their look from frame to frame.

Social media managers can run posts with the same character or brand mascot. The model holds up to 4 characters, and multi-character scenes are where it improved most in testing. That does not mean a perfect match every time, but you will redo fewer images.

Teachers and trainers can make diagrams and infographics for lessons. Turn on high thinking and web search, then always check the facts. The accuracy score went up, but 0.521 is still a long way from 1.

Designers can sketch concept drafts in minimal mode with 4 options per request and render the final version in 2K or 4K.

The logic is the same in every case. Minimal thinking to find the idea, medium or high for the final, and 2K or higher for anything with text.

If you work in a Unitool text chat, for example discussing a concept with Claude or ChatGPT, you can order images right there through Unitool MCP. It is a feature that gives a text model access to Unitool's image, video and audio generators, Nano Banana included. Turn on Unitool MCP in the text chat's menu and just ask for the image in your own words. Before each generation you see a card with the model, the settings and the price, and nothing runs until you press the button.

Why 2.1 When There Is Nano Banana Pro?

Fair question. Pro is Google's senior model, and it sounds more serious.

But in Google's own tests, the new 2.1 beat Pro on every measure the company published in the model card. The price gap in the Gemini API is noticeable too, a 1K image costs $0.134 with Pro and $0.0336 with 2.1. Google openly positions 2.1 for everyday work and keeps Pro for the hardest images.

In Unitool the choice is simpler. If Nano Banana 2.1 cannot handle a task, you do not need another subscription. The same plan includes GPT-Image 2.5 for posters and images with text, Flux 2, which is good with lettering, Midjourney v7 and Seedream 4.5. You can send the same prompt to another model and compare.

We do not promise a perfect image on the first try with any model. But testing several options in one place is cheaper and faster than paying for three separate subscriptions.

Start With One Image

Take a task you already have this week. A product photo, a social post cover, a diagram for a presentation. Open Nano Banana 2.1, set thinking to Medium, make a first version, then try changing one detail in it.

In Unitool one subscription gives you every model in the service, and tokens come out of a shared balance. If you need images regularly, look at the Optimal plan for $7 a month, the one most users pick. The price of every generation shows up before you run it, so there are no surprises on your balance.

Try Nano Banana 2.1 in Unitool

The best AI image tool is not the one that draws prettiest. It is the one that makes you stop typing "try again."