Nano Banana 2 taucht plötzlich auf. Es kann Formeln zeichnen und mathematische Probleme lösen. Selbst Überwachungsaufnahmen kann es fälschen.

The second - generation Nano Banana amazed everyone as soon as it made its debut.

The following image is completely AI - generated. Not only the content on the webpage, but even the entire browser interface and desktop are generated, and there is no reference image.

This time, Nano Banana 2 appeared in preview form on the third - party website Media.io and was subsequently removed. Only a few netizens with quick hands participated in the test.

The capabilities demonstrated by the preview version far exceed those of the previous generation. It performs outstandingly in handling complex prompts, including precise text rendering, ultra - realistic scenes, and fine simulation similar to a complete desktop interface.

However, Google DeepMind has not made any statements regarding the official release time.

Comprehensive Upgrade of Nano Banana 2

According to the test page instructions, this Nano Banana 2, also known as GemPix2, has mainly been improved in terms of authenticity, generation speed, and natural interaction control.

Judging from the existing test results, Nano Banana 2 can generate extremely complex UIs, and there are no flaws in text rendering. Many netizens would think it is a real screenshot.

It also performs well in terms of physical common sense and details in following prompts.

It can accurately draw a clock pointing to a specific time and a glass of full - poured red wine at the same time.

It can even forge realistic surveillance video footage. However, it is speculated that this ability should be weakened when the official version is released.

In addition, Nano Banana 2 has also mastered certain world knowledge and logical reasoning abilities.

In the comparative test of solving mathematical problems, the problem - solving idea of the first generation seems to be generally correct, but the formulas finally rendered are basically incomprehensible.

The second generation has some minor errors, but the results are impressive.

Nano Banana Team: The Quality of Image Generation is Approaching the Upper Limit

Nano Banana first anonymously appeared on the AI model evaluation platform LMArena in mid - August 2025. With its excellent image editing ability, it quickly topped the leaderboard, triggering heated discussions in the community.

At the end of August, Google officially revealed its identity as Gemini 2.5 Flash Image. The code name Nano Banana originated from the good results when Google internally tested the generation of "nanobots the size of a banana".

Within less than 10 days of its launch, users edited more than 200 million images, bringing 10 million new users to the Gemini app and once helping Gemini surpass ChatGPT to become the top - ranked free app on Apple.

The core advantage of the first - generation Nano Banana lies in its powerful image editing and understanding abilities.

This includes natural language editing and character consistency. Users can use daily language for multi - round and iterative image editing, and at the same time, it solves the common "identity drift" problem in AI image editing, keeping the character features highly consistent after multiple modifications.

Compared with previous models, it has added multi - image fusion and style transfer functions, supporting the seamless fusion of multiple different pictures into a coherent image, or applying the style of one picture to the objects in another picture, improving the creative efficiency of industries such as e - commerce and advertising.

It also has low cost and high speed. Optimized based on Google's TPU v5 architecture, the average response time of Nano Banana is only 1.3 seconds, and the cost of generating a single picture is about $0.039, only one - tenth of that of DALL - E 3.

Previously, the core team of Nano Banana revealed that the quality of image generation is approaching the upper limit, and the key in the future lies in improving the model's ability to understand the user's "intention".

Google itself is also accelerating the integration of Nano Banana into its core product ecosystem. In addition to providing services in the Gemini app and AI Studio, Google has started testing its integration into Google Photos, Search, Lens, and Circle to Search, aiming to create a seamless AI - driven visual experience closed - loop.

Reference Links:

[1]https://x.com/synthwavedd/status/1987259262322749784?s=20

[2]https://www.reddit.com/r/singularity/comments/1osolhn/nano_banana_2_vs_nano_banana_comparison_output

This article is from the WeChat public account "QbitAI". Author: Meng Chen. Republished by 36Kr with permission.