Google has updated its Gemini AI image generation tool with a build that caused a stir after it was released under the code name Nano Bananas.
The upgrade, technically called Gemini 2.5 Flash Image, lets users generate images through voice and text prompts, including swapping out participants in a photo, changing what they are wearing, or merging people from real images with new backgrounds. Google formally released it on Tuesday, though only via the Gemini mobile app, with the web version not getting all the new capabilities yet.
We’ve been testing the new engine and the results are impressive. For example, Reg US Editor Avram Piltch took a photo of just his torso and a separate photo of two chairs. When he uploaded both photos to Gemini on his phone, he asked the engine to draw him seated in the red chair.
Gemini not only placed Piltch in the red chair, but drew arms and legs for him that weren’t in the original torso picture. It even completed the logo on his t-shirt that was only half visible in the original image. The only inaccurate thing about the merged photo was that his pants were black when, in real life, he was wearing blue jeans. He asked Gemini to change the pants to light blue jeans, and it did so without issue.
Gemini created this image from two separate photos. The arms and legs weren’t in either original shot. – Click to enlarge
“Just give Gemini a photo to work with, and tell it what you’d like to change to add your unique touch. Gemini lets you combine photos to put yourself in a picture with your pet, change the background of a room to preview new wallpaper or place yourself anywhere in the world you can imagine — all while keeping you, you,” the Chocolate Factory said.
“Once you’re done, you can even upload your edited image back into Gemini to turn your new photo into a fun video.”
In other tests, Piltch was took a picture of his daughter and asked that two statues next to her be removed. The statues disappeared with the shadow from a nearby tree extending to where they were before. He then asked that his daughter appear in front of the pyramids and Gemini obliged, even changing her posture so she was standing straighter.
One major improvement users will notice right off the bat is how fast it is. In tests, images were done in seconds, with all the work taking place in the cloud. (Admittedly, we were using an older Pixel.) We even made the bananas cover art for this piece using Gemini.
In a move that should have Adobe worried, Gemini shows real skill in letting image editors use AI to replace in seconds what might have taken a graphic designer hours, or at least minutes. Where you used to have to Photoshop someone into a picture, now you can just ask the tool to do it for you.
Google has included a SynthID watermark to allow someone to identify the AI-generated images, which should be a big help in cutting down on fake pictures for spam, incitement, or for other purposes.
That’s not going to stop a wave of AI-generated spam shortly to be hitting your inboxes, but it will at least provide some safety checks. There are still some guardrails in Gemini 2.5 Flash Image, although they are somewhat limited. Generating pictures of Hitler, for example, is difficult but not impossible. If you want a celebrity such as Taylor Swift or Donald Trump, you won’t have any problems, though. It does at least have safeguards against generating images of a pornographic bent, thankfully.
Overall, while some images came out less than perfect, it’s still a worthy rival to other LLM image designers from OpenAI or xAI’s Grok.
Google is rolling out the new system for Gemini API, Google AI Studio for developers, and Vertex AI, at a cost of $30 per one million output tokens, with each image being 1290 output tokens ($0.039 per image).
This is very much an interim build, Google said, with more improvements coming down the line. It’s also partnered with OpenRouter.ai and fal.ai to make the technology more accessible. It’ll now be up to other AI companies to match Google’s very compelling new feature set. ®