Gemini's Nano Banana: How Google's AI creates lifelike photo edits

Google's Nano Banana, powered by Imagen 4 in the Gemini app, uses LLMs, diffusion models, and watermarking to produce more realistic and consistent AI images

Google's new image generation and editing tool, informally known as Nano Banana, has been rolled out in the Gemini app. Trends have quickly emerged with people using the tool to create 3D figurines, 90s-style saree pictures to polaroid photos with celebrities. While image generation is not new, Nano Banana has captured public attention with its attention to detail and ability to recreate images to the likeness of the people sharing their photos with the platform.

 

But what's behind the tech? Where does Gemini hold the edge over its rivals such as OpenAI and xAI?

 

What is Nano Banana?

 

Nano Banana is the name for Google DeepMind’s latest AI-powered image generation and editing feature inside the Gemini app. It lets users create images from text, edit photos, combine multiple pictures, and maintain consistency of subjects such as people or pets across different edits. 

 

 

What technology powers it?

At the core of Nano Banana is Imagen 4, Google DeepMind’s latest text-to-image model released in May 2025. Google Cloud page states, "With Gemini Nano Banana and AI image generator, you can create high-quality images in seconds with Imagen 4 - plus image editing!"

 

Imagen started at Google Brain as part of experimental text-to-image models. In April 2023, Google merged Brain with London-based DeepMind. Development of Imagen then shifted under the combined unit, now known as Google DeepMind.

 

This merger allowed Google to pool research talent and computing resources, speeding up advances in areas like text-to-image generation. The result is Imagen 4.

 

How does Imagen 4 work?

Imagen builds images from scratch using two key techniques:

1. Large language models (LLMs), such as Google’s T5, to understand and encode natural-language prompts.
2. Diffusion models, which start with random noise and refine it step by step into a full image.

The system generates an initial low-resolution image and then “upsamples” it in stages until it reaches 1024x1024 pixels, or higher when integrated into Gemini.

 

This allows the system to generate sharper details, better lighting, and improved text rendering compared with earlier versions. 

 

Why Nano Banana images looks more realistic

Unlike earlier systems, Nano Banana places emphasis on editing without losing core details. It maintains consistency in perspective, lighting, and subject identity. For example, when two photos are combined, the final image retains recognisable features of both.

 

This approach is designed to address a common criticism of AI imagery — that edits often distort or erase key details.

 

It is also better at generating text inside images such as shop signs, labels, packaging. This is an area where earlier AI models often failed.

 

What can users do with it?

Key features of Imagen 4 include:

• Generating images from written descriptions.
• Editing images by adding or removing objects.
• Blending multiple photos into one.
• Preserving subject consistency across different images.

 

It supports resolutions up to 2048 x 2048 pixels and provides up to four variations in a single request. 

 

SynthID: How Google ensures authenticity

Every AI image created in Gemini carries an invisible SynthID watermark, embedded at the pixel level. This hidden signal can usually survive everyday edits like cropping or resizing, allowing verification later.

 

In addition, Google supports Content Credentials (C2PA), an industry standard that adds tamper-proof metadata to record when and how an image was created. These measures are meant to build trust as AI imagery becomes more widespread.

 

Users will soon be able to check images using a SynthID Detector, currently in testing.   

 

Bottom line

Generative AI tools are becoming more common, but they also raise questions of authenticity and misuse. While companies like Google, OpenAI, and Elon Musk’s xAI claim uploaded images are not permanently stored, privacy advocates urge caution. Verification remains a challenge since SynthID detection tools are still not public, and critics warn that watermarks can be faked, removed, or simply ignored. Therefore, users need to remain vigilant and critically evaluate the content they encounter and share online.   

The technology is available in the Gemini app, via the Gemini API, and on Vertex AI for enterprise users.