Google’s new AI image creator took my shirt off

I tried Google’s new Nano Banana Pro, and it immediately took my clothes off. I didn’t ask for it, but the AI ​​model obviously decided that my greeting card would look better with more skin.

Nano Banana Pro is aimed, as its name suggests, at professionals. Powered by Gemini 3, it’s actually an upgrade to the company’s popular image generation and editing tool that went viral as part of a social media trend that turned selfies into hyper-realistic 3D figurines. Google says it lets you create higher quality images that you can print, make readable text on images, and blend multiple images into a single composition. It’s also aimed at “people who want to feel like professionals,” said Naina Raisinghani, product manager at Google DeepMind. The edge. This seems fine, as I am by no means a professional. For me, the results were brilliant, but wacky. It looked good, but it seemed amateurish.

Using Nano Banana Pro is pretty simple: you go to the Gemini app, select “create images” and enable “reflection” mode. Just plug in your prompt (and image, if you’re using one) and off you go. It’s also free, although there are limits, with quotas increasing for Google AI Plus, Pro, and Ultra subscribers.

Google makes some bold claims, promising “studio-quality designs,” “impeccable text rendering” and a host of nifty and creative edits. To test them, I uploaded a simple photo of myself near The edge office in New York with the Brooklyn Bridge in the background. I had Gemini change the lighting from day to evening and it did a really good job. The result seems credible. It even handled details that often trip up image generators, like making sure cars are moving in the right direction. Adjusting the camera angle was just as easy. I asked Gemini to recreate the photo as if it were taken from a higher angle on the right and that’s what he did.

Google also says Nano Banana Pro can create infographics and diagrams to help visualize real-time information such as weather or sports. Being British, I asked about the weather for the next four days in Washington, DC and New York, where I am currently. Visually, the infographic would have been out of place on a basic forecasting site. The text and numbers looked normal – far from the garbled nonsense you often see in AI-generated images – and Gemini gave me a list of quotes at the end that helped me confirm they were accurate.

The model stumbled a bit on more complex tasks. I asked him to summarize a recent Edge story about how Europe is rolling back its AI and privacy laws in a comic book style format. The images and text were indeed rendered perfectly in a cartoonish font, but the comic didn’t summarize the story at all, instead giving a vague overview of the bloc’s AI law. The problem may be because I gave Gemini a link to the story, rather than pasting the text.

This gave me a passable comic book style summary when I did it. It communicated the gist of the real story, although I don’t think I would have been able to understand easily if I hadn’t written the source material. He also invented phrases that did not appear anywhere in my article.

To really feel like a professional designer, I tried my hand at creating greeting cards. Christmas is coming, after all. Considering I only uploaded three selfies, Gemini did a frankly amazing job of creating three full versions of myself, each in different outfits and sporting a different facial expression. He also created a realistic, snowy backdrop with Christmas trees, as I requested, and pasted “Merry Christmas!” » on top as I requested.

Gemini took liberties when I asked it to change the snowy background of the map to a summer beach for an Australian-style vacation. These freedoms were my falsified clothing: two of my clones were topless. It was weird. There were also prominent AI-generated feet and a smiling sandman to replace the snowman from the winter scene (being constructed by my topless doppelganger). There were problems, however: The Sandman was missing a shadow, unlike the other rendered objects in the photo, and the Christmas lights in the palm trees glowed magically in the bright sun. I tested his precision editing skills by asking him to add muscle to a single clone, which he did in seconds (if only it were that easy in the real world). Overall, the quality was superb and the image would have been somewhat believable (abs aside) if you didn’t know that there was a large tattoo missing from my chest.

But it wasn’t all great. The model failed to keep the exact text on my card that I asked it to have. Instead of “Merry Christmas!” » he opted for “Aussie Summer Christmas!” » He also seems to have trouble with animals: my sister’s cat is sitting in the exact same pose on stilts as the reference image I had provided in each version of the card (he did get a fancy Santa hat, though).

Overall I was impressed. Nano Banana Pro is a marked improvement over the base model. I was able to request more precise edits and it actually produces intelligible text, removing a huge barrier preventing generative AI tools like this from being usable in the real world. But alas, these characteristics were not enough to make me a good designer.