What is Nano Banana Pro, and why is Google's new image model better at text?

Yesterday, Thursday, November 20, 2025, Google DeepMind introduced Nano Banana Pro, whose official name is Gemini 3 Pro Image. It follows the original Nano Banana, Gemini 2.5 Flash Image, which Google released a few months ago and which people used for restoring old photos and turning themselves into figurines. The new model is built on Gemini 3 Pro, the language model Google released just two days earlier.

The pitch is precision. Image generators have long been good at mood and style but poor at details that must be right, especially words. Nano Banana Pro aims at the jobs where accuracy counts: a poster with a readable slogan, a labeled diagram, a menu translated into another language, a product mockup that keeps a brand's look.

What Google released

Google now offers two image models side by side: the original Nano Banana for quick, fun edits, and Nano Banana Pro for complex, high-quality work. The Pro model is reaching many products at once:

For developers, TechCrunch reports the extra quality has a price. The original model cost $0.039 for a 1024-pixel image; the Pro model charges $0.139 per image at 1080p or 2K, and $0.24 at 4K, and it is slower.

How it works

Most image generators turn a short description into a picture by matching patterns between words and visual styles. Nano Banana Pro adds a thinking step on top, because it is built on Gemini 3 Pro, a model that reasons: it works out what the picture needs to contain before drawing it. Google lists several abilities that follow from this.

Knowledge in the picture. Asked for an infographic about a houseplant or a step-by-step recipe card, it can fill the image with correct, relevant facts, either from material you supply or from what it knows. It can also connect to Google Search for current information, such as the weather or sports results, a technique called grounding.

Readable text. Google calls it its best model for legible text in images, from a short tagline to a full paragraph, with a wider range of fonts, textures and calligraphy. Because Gemini handles many languages, it can also translate the text inside an existing image, for example turning English on drink cans into Korean while leaving the rest unchanged.

Consistency. It can combine up to 14 images in one composition and keep up to 5 people looking like themselves, useful for turning sketches into product shots or keeping a character the same across a storyboard.

Studio controls. You can edit just one part of an image, change the camera angle or focus, adjust color grading, or switch a scene from day to night. Output comes in several aspect ratios at 2K or 4K; the original Nano Banana was capped at 1024 by 1024 pixels.

A worked example

Suppose a small bakery wants a bilingual poster for its holiday hours. A prompt that uses the new strengths:

Design a vertical poster for "Maple Street Bakery" announcing holiday
hours. Include this exact text, spelled correctly, in English and French:
"Open Dec 24, 8 a.m. to 2 p.m. / Ouvert le 24 déc., de 8 h à 14 h."
Use a warm hand-lettered style, a photo-realistic loaf of bread at the
bottom, and leave space at the top for our logo. 4K, 2:3 ratio.

Then refine it in follow-up messages: "Make the French line the same size as the English," or "Change the lighting to early morning." Each edit keeps the rest of the design in place.

How to tell an image is AI-made

More realistic images raise the question of how to spot them. Every image made with Google's tools carries SynthID, an invisible watermark woven into the pixels. Starting now, you can upload an image to the Gemini app and ask whether Google AI generated it; the check starts with English prompts and images, with audio, video and more languages to follow.

Free and Google AI Pro users will also see a visible Gemini sparkle on their images. Google is removing that visible mark for Ultra subscribers and in AI Studio, for professional work, so the invisible watermark becomes the main way to check those.

Why it matters

Getting text and facts right has been one of the weakest points of AI images, and it is exactly what blocks everyday business use: a poster with a misspelled date is worse than no poster. By pairing image generation with a reasoning language model, Google is betting that an image tool should understand what it is drawing. Putting the model in Slides, Ads and Search also brings it to people who would never open a dedicated art app.

The trade-offs are cost and speed: the Pro model takes longer and charges more, which is why Google keeps the original model available for quick edits.

How do you write a good prompt for Nano Banana Pro?

Write it like a brief to a designer. Say what the image is for (a poster, a slide, an infographic), who will see it, and the format, including aspect ratio and resolution. Put any text that must appear inside quotation marks, exactly as it should read, and name the language.

Describe the look in concrete terms: lighting, camera angle, style, colors. If you have reference photos, upload them and say what to keep from each, such as "keep this person's face and this jacket." For facts, give the model your source material or ask it to use Search, then check the result.

Finally, edit in small steps. Ask for one change at a time, and say what must stay the same. That keeps the parts you like while you fix the rest.