Nano Banana 2 Leaks for Just One Hour — Nails AI's Toughest Image Test
Remember Nano Banana, the mysterious image model that went viral earlier this year? Well, it’s back — and the Nano Banana 2.0 update looks like a serious leap forward.
As a team dedicated to the future of design, we at Mew Design have been following this development closely.
Nano Banana Is Available in Mew Design!
Powered by Seedream 4.0 and Nano Banana, Mew Design is an AI design agent for creating artistic images and graphic designs without text illusions. It helps you produce everything from posters and social posts to brand assets and slides, giving you precise control over text and layout.
Try Mew Design for Free
To see how powerful Nano Banana 2 really is, just take a look at this now-famous photo of a clock showing 11:15 next to a wine glass filled to the brim.
Image source：x@synthwavedd
Why is this so impressive? In the AI image world, there’s a notorious stress test that no AI has consistently mastered: the clock and full wine glass challenge.
Why is this image such a big deal?
In the world of AI image generation, there’s a notorious benchmark that no model has ever passed reliably — the “Clock and Full Wine Glass Test.”
The challenge is simple: prompt the AI with:
“11:15 on the clock and a wine glass filled to the top.”
and see if it can render both correctly. Most models — from Arena, Seedream, to Nano Banana 1.0 — completely fail this test.
But Nano Banana 2 aced it.
Nano Banana 2.0 vs 1.0: A Massive Step Up in Text, Logic, and Visual Fidelity
Compared to the first version, Google Nano Banana 2 shows huge improvements.
- It can now render readable text directly in images, across multiple styles and languages.
- Its world knowledge has expanded — generating infographics and screenshots with precise details.
- And it follows instructions far more accurately than before.
Unlike the original model that appeared on AI Arena as a mystery “card draw,” Nano Banana 2 surfaced briefly on Media.io, under the Media AI platform — home to other models like Seedream.
However, the model vanished just an hour after launch. Reddit users confirmed that Media AI accidentally (or intentionally) made the model public for about one hour, and that was enough to cause a storm across social media.
Users quickly began sharing results that clearly looked like Nano Banana 2.0 — smoother colors, better lighting, and none of that plasticky “AI look” we’ve come to expect.
Take a look at the two images below — both featuring a car model posing along a seaside highway.
Can you tell which one was created by Nano Banana 2?
Image source：Image source: x@Azinha810
At first glance, they might look similar. But notice the color tone and overall vibe — the first image feels more natural, cinematic, and tonally accurate, while the second carries that slightly “plastic” look typical of earlier AI-generated photos.
Image source: x@Azinha810
Now compare these two sci-fi–style renders — the difference is even more obvious.
The second image has a richer cinematic atmosphere and far more refined visual details than the first.
Look closely at the control panel and dashboard elements in both images.
In the first, the instruments appear stylized and synthetic — very much that old “AI art” look.
In the second, Nano Banana 2 delivers a much more realistic, tactile, and technically accurate dashboard, proving how far its rendering engine has evolved.
Built on Gemini 2.5 Flash — Nano Banana 2 Still Ahead of the Game
According to x@testingcatalog, Nano Banana 2 still runs on Gemini 2.5 Flash rather than Gemini 3.0 Pro, suggesting an early release ahead of Google’s next-gen model.
Leaked info points to several standout upgrades:
- Handles complex visual logic — accurate text, advanced lighting, and perspective control that 1.0 couldn’t manage.
Example: Prompted to “turn the water pink,” Nano Banana 2.0 rendered subtle, realistic changes, far surpassing Nano Banana 1.0. Image source: x@Angaisb_
- A new multi-step workflow that feels more like a designer than a generator:
- Plan the output
- Generate a draft
- Review internally
- Fix detected issues
- Iterate before delivering the final image
- Higher resolution and flexible aspect ratios. It now supports a much wider range of aspect ratios — including 1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, and even ultra-wide 21:9 formats — giving creators more flexibility for everything from social media posts to cinematic renders. According to leaks, Nano Banana 2 also offers multiple output resolutions, including 1K, 2K, and full 4K modes.
Among all the upgrades, the new multi-step workflow might be Nano Banana 2’s biggest game-changer.
We’ve seen similar concepts before in tools like ComfyUI, which allow users to build image or video generation pipelines.
But this time, Nano Banana integrates that process directly inside the model itself — no extra setup, no node editing required.
For high-precision image generation, that’s a massive leap forward.
And of course, the real proof is in the pictures — so let’s take a look at some of the Nano Banana 2 test results that have been circulating online.
Nano Banana 2 Highlights & Test Results (With Example Prompts)
1. Stronger Text Rendering
Nano Banana 2’s text rendering is where things get truly wild.
In one example, it generated a Windows 11 desktop screenshot showing YouTube open in Chrome, complete with a MrBeast thumbnail — pixel-perfect and fully legible.
Simply upload a Windows desktop screenshot. Prompt: Generate a screenshot of a Windows 11 desktop, with Chrome open, showing a YouTube thumbnail of MrBeast on YouTube.com. | Image source: x@synthwavedd
Taking things a step further, Nano Banana 2 can even generate a fully detailed screenshot of a Google DeepMind webpage — and impressively, all the dense text renders without any gibberish.
Image source: x@synthwavedd
That said, attentive users have noticed that Nano Banana 2 isn’t perfect. In this example, there are minor typos like “Gemini 31” and “Google DeepMind – Nevdscl.”
At first glance, if you only look at the main headlines and subheadings, these mistakes are barely noticeable. However, when the text content is smaller or less dense, the likelihood of errors occurring is significantly reduced.
In this live TV news broadcast screenshot, all the text on the screen is accurate and consistently formatted. Impressively, even the reflections of the text on the floor are rendered correctly.
Image source: x@synthwavedd
Similarly, another TV broadcast screenshot shows the anchor positioned on the right side of the screen, with the classic news ticker running along the bottom — and once again, every piece of text is rendered flawlessly.
Prompt: 50mm big budget live broadcast 8k photograph close-up an attractive italian age 30 wry female news anchor. She is at a news desk and on a screen behind her is a bio engineered Banana with text “Nano Banana 2 live on Media .IO?” Live news shot. There is a red and blueckhyron on the bottom of the image that says “Breaking News: Nano Banana 2 on Media.IO?”
Image Source: x@BrentLynch
We ran the same prompt through the Nano Banana 1.0 as well.
This version has a slightly less “AI-generated” feel. However, the background screen is subtly blurred, making it look less like an actual news set. The reflections of the text also appear somewhat artificial, though the focus on the anchor herself remains strong and visually clear.
2. Better World Knowledge Means It Actually Gets Us
That infamous clock-and-wine test wasn’t a fluke — Nano Banana 2 consistently understands prompts that stump other AIs. Reddit users flooded the comments with images generated using other models like Grok and ChatGPT — but none came close to the accuracy of Nano Banana 2.
Some users even challenged Nano Banana 2 to turn a hamburger into glass. Remarkably, only Nano Banana 2 managed to transform the entire burger into glass, while other models like Bytedance’s Seedream and Riverflow only rendered the buns as transparent.
Prompt: Make the hamburger made of glass.
The first image was generated by Nano Banana 2. | Image Source: x@Angaisb_
Its improved world knowledge allows Nano Banana 2 to better understand prompts, strictly follow instructions, and even let users simplify their prompts without losing precision.
It also nails pop culture context. For example, one user asked Nano Banana 2 to generate a GTA 6 trailer — it generated a YouTube page complete with the actual GTA 6 logo.
Image Source: x@jewgibor
Another test involved the prompt “One Piece live action Netflix”. Nano Banana not only recognized the Netflix logo but also correctly identified the live-action cast.
Prompt: One Piece live action Netflix.
Image Source: x@jewgibor
Combining world knowledge with precise text rendering, Nano Banana 2 can even solve math problems on a chalkboard — a much more satisfying experience than scrolling through long math explanations in a chat window.
Upload the screenshot of the math problem. Prompt: Solve this math problem and write the full solution on a whiteboard.
Image Source: x@MagusWazir
Here’s a side-by-side comparison of Nano Banana 1.0 and Nano Banana 2 for this case:
Image source: https://www.reddit.com/r/singularity/comments/1osolhn/nano\_banana\_2\_vs\_nano\_banana\_comparison\_output/
3. It’s Not Just Fixes: Core Generation Got a Boost, Too
Visual generation has taken a big leap too.
In tests shared by x@legit_api, Nano Banana 2 recreated an emotional Tokyo Ghoul scene — Ken Kaneki carrying his friend through the snow — with cinematic realism.
Prompt: Ken Kaneki carrying his friend in his arms in the snow, Tokyo Ghoul.
Image source: x@legit_api
It also excelled at cyberpunk, portrait, and film-style shots.
Prompt: Cyberpunk hacker robot working in front of many monitors.
Image source: x@testingcatalog
Even “meme-like” portraits and stylized group photos showed remarkable cohesion and mood.
Image source: x@MicahBerkley
But Is It Really Google’s Model?
Some users speculate that Media.io AI, the Chinese company hosting Nano Banana 2, might not be officially affiliated with Google — similar to when DeepSeek caused confusion at launch.
Still, regardless of origin, the output speaks for itself.
Whether or not it’s truly Google-backed, Nano Banana 2 undeniably represents a new frontier in image generation.
And if this is just version 2.0 — imagine what Nano Banana 3 could do.
