After making a strong impression last year, Google's image modelNano Bananareturned with another major release early this morning.
Google officially introduced its latest generation, Nano Banana 2.Nano Banana 2（Gemini 3.1 Flash Image）。
The new model advances performance, usability, detailed control, and accessibility at the same time.
Google CEO Sundar Pichai described it in a post as
'our best image model yet.'
A stronger understanding of the world
Nano Banana 2 combines Gemini'sdeep world knowledge with real-time web search.
Pichai's 'Window Seat' example shows what that means in practice.
When a user specifies a location and asks for the view through a window, the model can
- understand the geography,
- retrieve current local weather,
- generate a scene consistent with current real-world conditions,
- and produce output at 2K or 4K resolution.
Image generation is no longer limited to stylized imagination. It can now construct a visual grounded in a particular time and place.
A substantially better cost profile
Banana Pro previously charged$0.134 for a 1K image, close to RMB 1 per generation.That cost created real pressure in high-volume workflows.
Nano Banana 2 reduces the price of the same 1K outputto $0.067, or roughly RMB 0.50,cutting the cost in half.
The price of 4K output has also fallenfrom $0.24 to $0.151,meaningfully improving the economics across resolutions.
Generation is also noticeably faster in everyday use.
Nano Banana 2 additionally offersa broader range of aspect ratios.
New 4:1, 1:4, 8:1, and 1:8 formats extend the model to more production and publishing scenarios.
Subject consistency
Google says a single workflow can preserve consistency across up to five characters and maintain fidelity for as many as 14 objects.
One of image generation's persistent problems has been that a great first image rarely guarantees a matching second one.
Stronger consistency makes sequential storytelling and multi-image production considerably more practical.
We tested the model with GeekOnUp's own characters and saw good results overall.
One image still showed a small inconsistency between our Youyou and Bridge characters.
More precise instruction following
Consider this example.
Prompt:
show me a where’s waldo set in ancient Venice, but instead of waldo it is an otter wearing a blue striped pilots outfit.
'Create a Where's Waldo-style scene set in ancient Venice, where Waldo is an otter wearing a blue-striped aviator uniform.'
The instruction combines several difficult requirements:
- a highly detailed, densely populated composition in the style of Where's Waldo,
- an ancient Venetian setting,
- a main character replaced by an otter in a blue-striped aviator uniform,
- and exactly one matching subject among a large crowd.
This is a broad test of model capability. Despite the complexity of the scene, the result contained exactly one otter.
Text generation
Text rendering is another focus of this generation.
Google explicitly positions Nano Banana 2 as capable ofgenerating clear, readable text suitable for commercial imagery.
Across posters, greeting cards, and brand messages, lettering is less likely to appear blurred, distorted, or replaced with random characters.
Higher visual quality
The model creates more natural light and shadow, with tonal transitions, contrast, and depth that more closely resemble real-world photography. Fine-detail clarity has also improved substantially.
For advertising, product displays, and brand assets that require high-fidelity output, this is a practical improvement.
Nano Banana 2 still has limitations.Longer passages and small type can occasionally lose clarity, and Chinese text can still produce malformed characters.
After many generation rounds, or when reference images contain highly complex elements, individual details may still drift.
In the AI era, curiosity, creativity, and an appetite for the unknown remain essential.
These new capabilities give more people a practical way to turn an idea into something visible and real.
Share the most useful or unexpected workflows you discover.
