Google released Nano Banana 2.1, the update to the Gemini image model. What is new: up to 14 reference images, 4K, more accurate text inside the picture and a character that stays the same character. How to write to it, and where to run it.
On October 6, 2026 Google released Nano Banana 2.1, the update to Nano Banana 2, the Gemini image model. It is not a new model from scratch but an upgrade to a model that was already among the best, and the upgrade touches exactly the things that get in the way day to day: text inside the picture, a character that changes from one edit to the next, and too few reference images. Here is what changed, why it matters, and how to write to it.
By Google's own measurements, Nano Banana 2.1 beats both Nano Banana 2 and Nano Banana Pro at making a picture from text. In practice, the difference you feel most is in working with references: a product, a character and a place that go into one picture together and stay true to the source.
The references are the heart of the model. Upload up to 10 images of objects, for example the product, the logo, a piece of furniture or a garment, and up to 4 images of characters, and describe the scene where they all meet. That lets you shoot a whole catalogue without a shoot: the same product, in the same packaging, in ten different places. Or put the same character in every slide of a carousel, without the face changing on the way.
Worth knowing: A good reference is a clear one: a visible, well-lit face, a product on a clean background, one image per thing. A blurry or crowded reference teaches the model things you did not mean.
One of the things that improved most is text inside the picture, Hebrew included. In our own test a Hebrew infographic of about 20 words came out with every word correct. The habit that works: write the description of the picture in English, and write the text that should appear in it in Hebrew, inside quotation marks, exactly as it should look. The shorter the text, the better the odds that every letter lands in place.
Like a photographer receiving a brief: what is in the frame, where it happens, how the light falls, from which angle and with which lens, and what is written in the picture if anything. One clear sentence for each of these beats a long list of magic words. Here is an example prompt with a product reference:
A product photo of the perfume bottle from the reference, standing on a wet black stone, soft morning light from the left, shallow depth of field, 85mm lens, the words "Happy Holidays" in clean white letters above the bottle
Nano Banana 2.1 works in every common ratio: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5 and 21:9, and there is also an automatic mode that takes the ratio from the first reference image. 1K suits drafts and testing a direction, 2K suits social, and 4K suits print, signs and anywhere people look up close.
Can I write to it in Hebrew? You can, but a description in English usually gives more precise control. Text that should appear in the picture is written in its own language, inside quotation marks.
What is the difference from Nano Banana 2? More references, 4K, more accurate text and a steadier character across edits. Anyone who worked with version 2 will feel it mostly when working with products and characters.
Do the images belong to me? Yes. What is made in your account belongs to you, for personal, business or commercial use.
On Cadabra AI the model went live the day Google released it, and it is the first tile in the Images tab: up to 14 references, every ratio, and 1K to 4K. It sits on the same balance as the video, music and voice engines, so the picture you made can become a video in the same place. The price is shown before every run, and a generation that fails is refunded automatically. And for anyone starting out: 10 free credits on sign-up, no credit card, at livinginmovie.com/en/start.
עמוד הבית · בלוג · וידאו AI · תמונות AI · קול AI · תמונה לסרטון · כתוביות בעברית · כמה זה עולה · אפקטים לעריכה · פריסטים · כל המנועים · קרדיטים ולא מנוי · AI בעברית · קורס מתחילים · מדריכים · נגישות