Looking through a recent thread on Reddit comparing images created with the same prompt on Nano Banana 2 and ChatGPT, I noticed an interesting trend — users seem to think that Nano Banana 2 has actually gotten worse over time.
“Nano 2 had a serious downgrade” said one user, with another replying “Yea I’m a Pro user, and the image quality seems to have been downgraded a lot. Now, almost every image looks flat or cartoonish, no matter how detailed my prompt is. It honestly feels like Google intentionally lowered the model’s performance.”
A downgrade seems like a bit of a stretch to me. For a start, Nano Banana hasn’t officially been updated since February, or at least there hasn’t been a public announcement of a change.
Latest Videos FromTechRadar
The last update to Gemini (the AI which uses Nano Banana 2) was in July when Google released Gemini 3.6 Flash. It also updated the Gemini app’s underlying model selection and orchestration, but it didn’t change the image generator.
Interestingly, once I started looking through Reddit threads I found complaints about Nano Banana 2 quality regressions dating back to May, well before the recent Gemini 3.6 Flash announcement, suggesting some users have perceived changes over time.
That doesn’t prove a regression, but it does suggest this isn’t a brand-new observation.
Google Earth integration
Google did release a new image model, Nano Banana 2 Lite, at the start of July. And more recently, Google has been integrating Nano Banana 2 into products like Google Earth, then temporarily pulling one of those features after misuse, but there’s no indication that the underlying image model itself was updated as part of any of these releases.
So, what has made Reddit users become convinced Google’s image generator has gotten worse?
One possibility is that even if the image model remained Nano Banana 2, the text model interpreting prompts may have changed. Better (or simply different) prompt interpretation can produce noticeably different images without the image generator itself changing.
I’ve always been a fan of Nano Banana 2, so I decided to recreate the Reddit comparisons myself to see how it compares to ChatGPT’s images.
House of the Dragon
The original thread was clearly written by a House of the Dragon fan, because it was comparing the AI’s ability to create a realistic image of a dragon flying overhead. I used the following prompt with Gemini and ChatGPT:
“I want an image of a photo that’s taken by somebody looking up at the sky with a dragon flying overhead. I want you to make the dragon look as realistic as possible – as if it could actually be real.”
Here’s what I got from ChatGPT:
And from Gemini:
You can vote on which one you prefer, but for me the Redditor’s claims hold up here. The ChatGPT one looks like somebody has taken a photo of a real dragon flying convincingly overhead — it feels natural and unforced. It looks like a photo taken at an odd angle, while the Gemini example has that “it looks like AI”-quality to it, with the dragon posed against a scenic background and a slightly flat quality to the image.
Next I thought I’d try them both on an image that wasn’t a fantasy animal, but something we’re all familiar with — human beings:
“I want a photo of a couple in a cafe. They are in their 50s – a man and a woman – enjoying a coffee together and chatting. Traffic is visible passing by through the windows of the cafe, and there are other people around, but they are the focus of the shot. Make it look as realistic as possible.”
Here’s what I got from ChatGPT:
And from Gemini:
Again, I think the Gemini result looks artificial. I liked the reflection of the woman’s jumper in the window, but outside those two buses look like they’re facing each other in traffic, while inside the reflection of the cafe lights in the pictures seems off. The couple also have that uncanny valley effect to them. In contrast the ChatGPT image is less detailed, but looks more realistic. Nothing in it looks unnatural.
Finally, I went for food — a full English breakfast. It’s a great final realism test because it exposes fake-looking textures, reflections, steam, crumbs, cutlery, glassware and background detail. A convincing plate of food is surprisingly hard to fake.
Here’s what I got from ChatGPT:
And from Gemini:
It’s harder to separate them here. They both do a good job at getting the textures right, and both seem to struggle with the toast, if you look closely. Gemini is slightly let down by the words on the menu, which don’t look like proper words.
I think you can tell that overall I’ve come down fairly firmly on the side of ChatGPT, but I don’t think that the “flat and cartoonish” criticism of Nano Banana 2 is fully justified when it comes to generated image quality. Rather than Nano Banana getting worse, I think something else has happened — I think ChatGPT has gotten better.
OpenAI has made enormous progress in image generation over the last few months. If ChatGPT has improved while Nano Banana 2 has stayed roughly the same, the subjective impression could easily be that Nano Banana has gotten worse, when in reality it’s just been overtaken.
ChatGPT is still noticeably slower than Gemini to generate images, but to me they look better and less AI-generated. It’s now my first choice for making AI-generated images.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
