You can’t talk about AI image generation without including Google’s nano banana models, for good reason. The two versions, the original (Gemini 2.5 Flash Image) and the new pro (Gemini 3 Pro Image), have only been around for a couple of months, but they’ve quickly redefined what’s possible with AI image generation and editing. 

The pro version uses Gemini 3’s reasoning model to power results. That means it takes a little bit longer to generate, but the images are more detailed. You can also add swaths of legible text to your images, an industry first. The pro version is the best AI image generation tool CNET has ever tested, hands down. But that isn’t necessarily a good thing. 

Our testing revealed how easy it is to create ultrarealistic images and infographics with wrong information. The Gemini-made images lacked many of the most common errors in AI-generated images, which is good for creators. Yet it also means that anyone who comes across the Gemini-created images online will have a harder time deciphering if they’re real or AI. Google removes the sparkle-shaped watermark from some images, which doesn’t help. Although Google’s SynthID or invisible watermark is embedded in all AI-generated images, the detector technology in Gemini is too limited to make a difference. These aren’t new problems, but nano banana pro’s capabilities make them even more prevalent and difficult to solve.

Whether you use nano banana through Gemini or take advantage of more settings in Google’s AI Studio or Flow, it’s easy to use the model to create. Here’s everything you need to know about nano banana pro based on our testing. For more, check out the best AI image generators and our AI essentials.

How good are the images, and how well do they match prompts?

Nano banana pro makes some of the most impressive AI images I’ve ever seen. Even after the shine wore off following my initial hands-on experience, I was continually awed by the nano banana pro throughout my in-depth testing. For many of the images I generated, I wouldn’t be able to immediately tell if they were real or created with AI — a stunningly horrifying thought. While bouncing between being impressed and unnerved, one truth remained: Nano banana pro obliterates the line between reality and AI. And this is the worst this model will ever be.

Nano banana pro images made in Gemini automatically generate as 16:9 landscape images; you’ll need to use AI Studio or Flow to change the aspect ratio or any other customization settings. Gemini is very good at adhering to your prompts, meaning it will give you what you ask for. You can check out our guide to writing the best AI image prompt, or if you’re pressed for time, you can ask Gemini to improve your prompt.

Nano banana pro excels at realism. I asked for an image of a score showing UNC-Chapel Hill winning over Duke University. Gemini used the ESPN graphic layout to represent the final score and added logos from ESPN, ACC and the universities. The text is intelligible and the iconic shades of blue are spot-on, and it even added a reflection in the TV screen to show a lamp lighting the scene. It’s these little details that make nano banana pro images truly extraordinary.

There are some limits to what Gemini can generate. For example, Gemini had no problem creating an image based on the song Mr. Brightside by The Killers, but when I asked for a similar image inspired by a Taylor Swift song, it refused to generate, stating it didn’t have access to that content. Never mind that Google Search certainly has oodles of content on Swift; Google is likely aiming to avoid any copyright infringement allegations.

The biggest leap forward is nano banana pro’s ability to create legible text. AI creative models have long struggled with this — we used to be able to point to mangled text as obvious signs that a photo or video was AI-generated. But nano banana pro is one of the most competent models I’ve ever tested, and it easily handles integrating clear text into imagery. It’s scarily good.

While the majority of the graphics I made with Gemini were virtually flawless, I did notice a few errors. Some were so small you wouldn’t immediately notice, like with this Fourth Wing design. If you zoom in very closely, there’s a nonsense word above “Fourth.” Hallucinations like these are annoying, but there’s a bigger problem when the model starts to make up information.

I asked Gemini to create a series of infographics about the new iPhone 17, and CNET’s mobile experts spotted issues immediately. Gemini couldn’t distinguish between the features of the base and pro models, often mixing them up. The AI couldn’t correctly render what the back of any iPhone 17 looked like; CNET’s Patrick Holland said one image actually looked like an earlier mock-up of what people suspected the new iPhone could look like based on the Google Pixel, which is why the AI version looks like a Pixel painted orange with Apple’s logo added. I asked Gemini to limit its information to CNET’s own specs comparison page, and it still made an image full of false information.

Beyond the immediate issues, the more pressing issue is how convincing the image looks. If you don’t have in-depth knowledge of the topic you’re asking Gemini to visualize, you might not notice when the AI starts making stuff up. This is one of the most common and concerning issues with chatbots, and it’s likely to persist as AI image generation relies on chatbots’ knowledge bases. Future model updates might help with this, but for now, you need to keep a close eye on the information Gemini is including in your images.

Editing photos with nano banana pro

My biggest complaint about the original nano banana model was its image editing, which didn’t work in the way it was supposed to. I’m pleased, and somewhat unnerved, to report that the pro model fixes most of these errors.

Nano banana pro is great at changing the background scenery of photos. I uploaded a picture of my family we took before attending a snowy football game at my sister’s college last year. It was so cold we didn’t get another family shot that day, but Gemini was able to fix that.

The pro result was stunning — the stadium is an exact replica, the signs are crystal clear and the snow looks very natural. I tricked a number of family members with that image on Thanksgiving. I gave the same prompt to the original model, and there’s no doubt the pro’s image was better. 

I did try to achieve a similar result by uploading multiple reference images of my family and asking Gemini to combine them, but the results weren’t as good — facial expressions took on a plastic-y AI look, the signs in the background weren’t as clear, etc. To get the best results with multiple reference images, it helps to use high-quality images with plain backgrounds, which I didn’t have.

I also asked Gemini to relight my headshot. I asked for a hunter green background and a spotlight on my face. You can see how the original model distorted the shadowed part of my face and how the spotlight completely missed most of my face. The pro model used much more natural lighting and correctly lit up my face.

The one edit that drove me slightly bananas (sorry, couldn’t resist) with the original model was trying to remove reflections. Granted, removing reflections is a hard photo edit. The pro model beat the original in that it actually removed the reflection, but it still compromised the fine details and distorted the faces. Apologies, Lindsay Lohan and Jamie Lee Curtis. Maybe the next nano banana will finally pass this test?

Editing photos with Gemini’s nano banana pro is certainly better than with the original model. The conversational flow is great for people who like chatting to edit. But a complete lack of editing tools means it’s not great for people who want any level of hands-on control. Even the pro model can only handle bigger edits, and too many iterative edits dilute the quality of the image. If you want more granular control, it’s worth heading to Photoshop or Canva and using their AI and tools. This is one area where it’s very clear Google is not a creative software company.

Google’s original model vs. pro

Nano banana pro is the best model, but there are times when you may not want to use it. The original model is much quicker than the pro, and it can handle any kind of wholesale or bespoke AI image generation well. The pro model should be the go-to choice for users who need more in-depth image edits and those who need to rely on Gemini’s world knowledge for infographics. 

Here’s a fuller breakdown of the original nano banana model versus the new pro model.

How fast do images arrive?

Nano banana pro uses a reasoning model, which tends to take a little bit longer to run. I found that my images were generated in anywhere from 50 to 120 seconds. Two minutes would be considered long for other image models, but it’s not unreasonable. If you want quicker results, switch to the fast model to use the original nano banana model.

Final thoughts

Nano banana pro is awe-inspiring. Some of the results are awesome — they’re grounded in reality in a way that competitors like Midjourney and OpenAI clearly aren’t. That’s why I said that Nano banana pro is one of the best AI image generators CNET has ever tested. It’s also a deeply scary tool: The results are scarily good, and it’s unnervingly capable of creating almost anything. And scariest of all, just because it can create clear text in images doesn’t mean it can produce reliable sources of information. It’s all the best parts of image generators and the worst parts of chatbots.

The pro model is a powerful tool, so much so that I anticipate it will quickly become a go-to for anyone interested in creating media with AI. Creators of all skill levels will likely be able to find uses for it. But I worry that it will also become a go-to tool for bad actors who can take advantage of its quick and easy process to create harmful results, ones that we won’t be able to immediately identify as AI-generated. We’ve already seen how AI companies’ guardrails can fail. 

When images like the ones created with nano banana pro look convincing, we’re not going to be automatically suspicious of their origins. That’s a problem, especially with information-heavy designs like infographics. It doesn’t matter that Gemini can create legible text if the information is wrong. And when we’re idly scrolling through our social media feeds, we don’t always scrutinize the posts we see or read the fine print to see if the headlines check out. It sets the stage for a lot of confusion, chaos and misinformation.

In short, nano banana pro is the future of generative media. But that future could be treacherous if we aren’t careful.

How CNET tests AI image generators

CNET takes a practical approach to reviewing AI image generators. Our goal is to determine how good it is relative to the competition and which purposes it serves best. To do that, we give the AI prompts based on real-world use cases, such as rendering in a particular style, combining elements into a single image and handling lengthier descriptions. We score the image generators on a 10-point scale that considers factors such as how well images match prompts, creativity of results and response speed. See how we test AI for more.

Nano banana pro pricing, availability and privacy

You can use nano banana pro for free in the Gemini app and web browser, but you’ll want to upgrade to one of Google’s AI subscriptions for higher usage limits. Those plans start at $20 per month and come with a bunch of other services, like more Google storage. If you’re a paying Gemini subscriber, I highly recommend using nano banana pro in Google’s AI Studio or Flow so you can take advantage of the hands-on controls to customize your images.

Google’s general Gemini privacy policy says it can use any information you give it, like reference images, for improving its AI products. That’s why the company recommends you avoid uploading sensitive or private information. Like all major AI companies, Google has a prohibited use policy aimed at preventing the creation of illegal and abusive material.

For more, check out our expert-backed ways to fix AI image hallucinations and OpenAI’s Sora vs. Google’s Veo 3 AI video generators.