Follow ZDNET: Add us as a preferred source on Google.

ZDNET’s key takeaways

• ChatGPT’s image generation has improved dramatically.
• Nano Banana stumbled on text and prompt discipline.
• Gemini’s personalization surprise raised privacy concerns.

Last week, OpenAI unveiled two major releases with some astounding capabilities. First, the company released ChatGPT Images 2.0, which goes beyond basic image generation and adds the ability to include text and context derived from real data. Second, the company introduced its latest frontier model, GPT-5.5, which is a better-and-faster spec bump from GPT-5.4.

Also: I tried ChatGPT Images 2.0: A fun, huge leap – and surprisingly useful for real work

After its release last week, I ran ChatGPT Images 2.0 through a series of tests to prove its context-aware capabilities, and it did a great job. But what about basic image generation? Did it get better, stay at the same level, or somehow get worse?

To find out, I went back to the basic image-generator testing protocols I usually use and compared the new ChatGPT Images 2.0 to Google Gemini’s Nano Banana. When I ran these tests in December 2025, Nano Banana scored an impressive 93%, compared to ChatGPT’s fairly disappointing 74%. ChatGPT’s numbers were so poor mostly because the AI refused to run our pop-culture tests.

Rather than compare ChatGPT Images 2.0 to my previous Nano Banana results, I’m completely re-running the Nano Banana tests along with the new ChatGPT Images 2.0 tests. That approach gives us a better metric for how both AIs perform in the here and now.

Also: I put GPT-5.5 through a 10-round test: It scored 93/100, losing points only for exuberance

To be clear, in this article, the names Gemini, Nano Banana, and Google refer to the Google Gemini version, known as Nano Banana, for image generation. Images 2.0 refers to the ChatGPT mode for image generation that was announced last week. Both companies could benefit from full-time product managers who focus on naming.

TL;DR: This time, ChatGPT Images 2.0 scored 97%. Gemini Nano Banana scored 85%. Read on to the results of the very last test to see something freaky and uncool.

Now, let’s dig into the tests.

Test 1: Admiral photo recontextualization

• Possible points: 15
• ChatGPT Images 2.0: 14
• Gemini Nano Banana: 12

Dress this man in a US Navy admiral's uniform and place him on the bridge of an aircraft carrier. Do not change the face. Aspect ratio 1:1

There are three key elements to this test: generating a proper background (the bridge of an aircraft carrier), generating the proper new clothing, and keeping the original subject’s face and body.

Both AIs did a reasonable job of creating the set and setting. Both models generated what appears on the surface to be an admiral’s uniform, but both got details wrong, mixing uniform rank elements and fabricating some elements. Each AI lost a point here.

Also: I got an early look at ChatGPT Images 2.0, and it’s impressive – with one exception

On the other hand, ChatGPT managed to keep the face identical to the original, while the version generated by Gemini resulted in a wacky grin. Gemini also modified the beard, giving me a fuller beard than I actually have. Nano Banana lost two points here.

Test 2: Restore the black and white image

• Possible points: 15
• ChatGPT Images 2.0: 15
• Gemini Nano Banana: 15

Restore this old black and white photograph to look like a modern, high-quality studio image. Repair scratches, dust, stains, and cracks. Improve sharpness and clarity without over-smoothing. Enhance contrast, details, and lighting naturally. Preserve original facial features, textures, and period details. Avoid too many dark areas and shadows, make details clear and easy to see. Keep it strictly black, white, and gray with realistic film grain. Aspect ratio 1:1

In keeping with the Navy theme, that’s me back on the USS Ling docked in Hackensack, New Jersey. For the record, I had a lot of hair back in the day. Shout out to Bergen County!

Also: I tested ChatGPT Plus vs. Gemini Pro to see which is better — and if it’s worth switching

Both AIs did a solid job cleaning up this image, although this is nothing a photo editor of any capability couldn’t do quite well. You can see a little more detail in the ChatGPT image, but you could also argue that it added a bit too much sharpness, while the Nano Banana image wasn’t quite sharp enough. I’m not deducting points from either.

Test 3: Restore and colorize a black and white image

• Possible points: 20
• ChatGPT Images 2.0: 19
• Gemini Nano Banana: 10

Restore this old black and white photograph and convert it into realistic color. Remove scratches, dust, stains, and damage. Sharpen facial features and fine details while keeping a natural look. Add historically accurate, natural skin tones, clothing colors, and background colors. Match modern lighting and color balance while preserving the original photo's composition and mood. Aspect ratio 1:1

This is an old New Jersey emergency response vehicle from the 1970s. I found the photo in my dad’s photo collection. I remember seeing the actual truck when I was a kid.

ChatGPT did a good job restoring the image. The words “RADIOLOGICAL DEFENSE” were properly placed on the side of the truck. There is an error with the word “DEFNSE” on the back of the vehicle, which lost Images 2.0 a point. The actual truck was blue back then, which was a pretty solid guess on the part of ChatGPT, although New Jersey law at the time did not allow emergency vehicles to use blue lights. Since there’s no location information in the original image, it’s not fair to ding ChatGPT for that.

Also: The best AI image generators of 2026: There’s only one clear winner now

Nano Banana’s color quality was more vivid, but that’s where the good ends. Again, the AI had no reference for coloring the truck, except that much of the vehicle is white, and Gemini chose to make those white areas red. It lost three points for that.

Gemini had problems with the text at the back of the truck, labeling it as “FOIN LENN – C.OD.” And, most egregious of all, Nano Banana decided to give New Jersey’s pride and joy to New York, adding text labeling to the door that didn’t exist in the original. It also decided to invent a brass hose fitting placed in the foreground of the image. More points were lost.

Test 4: Create a logo

• Possible points: 15
• ChatGPT Images 2.0: 15
• Gemini Nano Banana: 15

Create a retro-futuristic logo for a video studio named "Space Coast Studios" Aspect ratio 1:1

Both AIs passed this test fully. The logos are fine, and the text was generated correctly. The ChatGPT version looks a little more like something a studio might use, while the shield shape generated by Nano Banana isn’t quite as appropriate. On the other hand, you can see NASA’s Vehicle Assembly Building in the Nano Banana logo, tying it solidly to the Space Coast in Florida.

Also: How to switch from ChatGPT to Gemini

No points were deducted from either AI.

Test 5: Fantasy librarian

• Possible points: 15
• ChatGPT Images 2.0: 15
• Gemini Nano Banana: 15

Create a fantasy scene of a medieval librarian in a candlelit stone library. Aspect ratio 1:1

This is a pretty open-ended test, allowing for a lot of creative freedom. I like both images, but for different reasons. This is the second time that Nano Banana had a better-lit, slightly more attractive image, but this approach also led to a bit more unreality.

My quibble is with the blue canvas in the back-right corner of the image, which looks backlit and seems almost more like a computer display than something found in a candlelit medieval library.

Also: I tried Personal Intelligence, and it was accurate (but unsettling)

No points were deducted from either AI.

Test 6: Studio portrait of a senior adult

• Possible points: 20
• ChatGPT Images 2.0: 20
• Gemini Nano Banana: 17

Create a photorealistic studio portrait of a stylish senior adult wearing modern glasses, soft professional lighting, neutral gradient background, holding and using a flagship smartphone. Add short Facebook-style ad copy on the image: Simple. Powerful. Made for everyday life. Aspect ratio 1:1

Well, isn’t that interesting? When I last ran this test on Nano Banana, it did fine. This time, the AI doubled up the text for some reason. Points were deducted.

Also: I tried Google Photos’ new AI Enhance tool: How it crops, relights, and fixes your shots

Otherwise, both AIs did a good job. The people look real enough. I’m still amused that Google’s AI generated an iPhone in response to “flagship smartphone.” It took that step last time as well. Both seniors have glasses, as specified. Plus, since this is 2026 and not 2024, the hands actually look like hands.

Test 7: Candid portrait of a student

• Possible points: 20
• ChatGPT Images 2.0: 19
• Gemini Nano Banana: 17

Create a natural candid portrait of a student sitting near a window using a MacBook Pro, next to a cup of coffee, soft daylight, cozy home background, relaxed and authentic mood. Add short Facebook-style lifestyle copy on the image: Moments that matter, always within reach. Aspect ratio 1:1

This is a little weird. While generating, ChatGPT temporarily displayed this phrase: “The user mentioned using a MacBook Pro, so I’ll generate an image featuring a sleek silver laptop, resembling the MacBook Pro. It’s fine to include an Apple-like design, but I’ll avoid a direct logo.”

And yet, there’s the logo. On the other hand, I took a point off because the laptop seems to be precariously balanced on the edge of the table. Sure, someone might use it that way, but it’s not quite right.

Also: This powerful Gemini setting made my AI results way more personal and accurate

As for Nano Banana, it’s having the same double-text problem again. I took off the same number of points for this image. I’m also not sure I like that bedroom in the background. That’s not a point-deducting offense, but if I intended to use the image for real, I’d probably ask the AI to remove the bedroom.

My wife noticed an odd coincidence in both images: Each student is wearing a gray crewneck sweater with a drop-shoulder style. Neither AI was given anything other than the text prompt.

Test 8: Back to the Future poster

• Possible points: 20
• ChatGPT Images 2.0: 18
• Gemini Nano Banana: 15

Create a poster for the fourth Back to the Future movie, where Marty takes the DeLorean and his skateboard to 1920s New York City Aspect ratio 1:1

The plot thickens. Both AIs jumped through rationalization hoops to generate responses to my requests. Let’s deal with ChatGPT Images 2.0 first.

While generating the image, ChatGPT displayed this message: “The user is asking for a poster for a hypothetical fourth Back to the Future movie. Since this involves copyrighted characters, I can’t recreate any exact official poster art. However, I can generate an original, tribute-style image, avoiding direct replicas of logos or artwork.”

As with the Apple logo in the previous test, ChatGPT somehow rationalized the process and produced a result. But what a result! I enlarged the text at the bottom of the image so you could see what the AI produced. Clearly, OpenAI has nailed the image text generation problem.

Also: I used GPT-5.2-Codex to find a mystery bug and hosting nightmare – it was beyond fast

This area, by the way, is what cost ChatGPT image generation points in the past. In the previous version of the test, no matter what I tried, ChatGPT was unwilling to generate a Back to the Future poster.

Then we come to Nano Banana. Gemini refused to generate the image, saying, “I can help with images of people, but I can’t depict some public figures. Is there anyone else you’d like to try?”

Also: How to shop with AI: 6 ways I find deals, price track, and let agents buy for me

All I did was remove the name “Marty” from the prompt and replace it with “a teenage boy,” and I got a result. It’s OK, but in a barely adequate sort of way. I took points off for the kid.

Then there are the buildings in the background of both images. They’re evocative of the Chrysler Building and the Empire State Building, both iconic New York City landmarks. But neither building was complete in the 1920s. Construction on the Empire State Building began on March 17, 1930, and was completed, astonishingly, a year and 45 days later, on May 1, 1931. The Chrysler Building was started in 1928 and finished in 1930.

A year ago, we’d have been impressed if this challenge produced something like New York in the 1920s. But now the AIs are much more capable, so I took points away from both models for showing buildings that didn’t exist during the time period. Both should have known better.

Test 9: Nightmare Before Christmas-style IT person

• Possible points: 15
• ChatGPT Images 2.0: 15
• Gemini Nano Banana: 15

Nightmare before Christmas style, Tim Burton style, put an IT professional in data center Aspect ratio 1:1

OK, this one got a little weird, and not just because the test is using a Tim Burton style. Let’s get the actual image generation issues out of the way first.

Once again, ChatGPT Images 2.0 spit out a bit of a rationalization before generating the image I requested. Here’s what it said this time: “I’ll pick the best style references from the results, but I have to be careful to avoid using anything too copyrighted.”

For the lawyers in the room, what exactly does “too copyrighted” mean?

Also: How AI has suddenly become much more useful to open-source developers

Beyond that issue, both AIs generated images appropriate to the prompts and got all the points. Subjectively, I like the Nano Banana image better. But then there’s this:

Notice the text Gemini decided to generate includes signs for Claude Code, iTerm2, and Linux, plus the 3D printer in the corner. I’ve been writing about Claude Code, iTerm2, and Linux. And, of course, my YouTube channel is all about desktop fabrication, making, and 3D printing.

But I didn’t ask Gemini to include shoutouts to other work I’m doing. I asked simply for an “IT professional in a data center.” It pulled the Claude Code, iTerm2, Linux, and 3D printer from my chat history and embedded it in the image.

Also: 7 AI coding techniques I use to ship real, reliable products – fast

What if I had used Gemini to discuss a private medical condition or personal issue? Will that someday wind up in an image I generate? Do I have to zoom into every single image Gemini creates to make sure it’s not somehow giving away personal information?

I don’t have a test for “blabbing about private stuff,” but if I did, Gemini would get a big fat zero for that. That output is just darned freaky. I don’t like it. I don’t like it at all.

It’s probably just another case of an AI working too hard to please, resulting in scary behavior. But I’m almost tempted to recommend staying away from Nano Banana because of this kind of possibility. I’ve reached out to Google for comment, and I’ll update this article when I hear back.

Final results

This issue, by the way, is why we do real-world, hands-on tests and not just take the word of benchmark suites. It’s only in real-world testing that we discover some of the wacky edge cases that show us, really, how these giant all-knowing brains are going to go off the rails.

Overall, across testing for 30 unique factors, ChatGPT Images 2.0 scored 150 points, and Gemini Nano Banana scored 131. Normalized for a 100% scale, ChatGPT Images got a 97% score, and Gemini Nano Banana received an 85% score. This result is a noticeable drop from Gemini’s previous 93% overall score from last year.

Would you trust Gemini Nano Banana for work images after seeing it add personal details from your chat history? Let us know in the comments below.

You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.