Google has just introduced the new “Nano Banana” model in the Gemini app (code name of the Gemini 2.5 Flash Image model), developed by Deepmind (the company controlled by Alphabet that develops solutions to ai). Its peculiarity is not so much in the ability to generate realistic images, as in the consistency that manages to give to the latter. To translate the matter into small terms, with this new model if you retain the face of a friend or dog to insert it in different contexts, the system will be able to preserve its distinctive features in a stable and recognizable way, minimizing the possibility of running into unrealistic results. This opens unpublished creative possibilities: from simple changes of looks, to the merger of multiple photographs to give life to new scenes. All without losing similarity with the original subjects, one of the most difficult aspects to be obtained with previous generations of models.
You can thus imagine yourself in another decade, transform a bare room into a furnished environment with few controls or even combine the graphic motifs of one object with another, such as applying the pattern of the wings of a butterfly on a dress. Editing is no longer a technical process reserved for professionals, but becomes something extremely democratic and accessible to everyone. And, as a guarantee of transparency, all images bring with them visible and digital invisible digital that certify the intervention of the AI.
What can Nano Banana do, the new Google AI model
The new Nano Banana model (Code Name of Gemini 2.5 Flash Image), an architecture designed to guarantee substantial results. One of the historical challenges of the generative systems, in fact, is their non -deterministic nature: the same prompt, that is, the text instructions given to a chatbot AI, can produce very different results depending on the randomness within the model. This phenomenon made it difficult to keep delicate details as intact such as the features of the face, proportions of the body or characteristic features of a pet. With Nano Banana, the contextual memory allows the system to “remember” the details already developed, offering a more uniform result between one modification and another.
This means that if you decide to give the model a photo of you, asking him to transform you into a matadorin a painter or in the protagonist of a sitcom of the 90s, your face will always remain recognizable, as can be seen from the example in the following video.
Another innovative aspect is the possibility of blending multiple photographs. It is possible to combine two subjects present in two distinct and separate images and both combine in an output image that goes far beyond a simple collage: the photo returned as a final result is an image consistent with the original content, able to look authentic. In the example present in the following video, you can see how the image of a girl and that of a dog were merged so that the girl caressed her pet in the output image. It is probably the most clear example of how powerful and precise this model is.
The model also supports the so-called Multi-Turn Editing, i.e. a sequence of subsequent changes that accumulate coherent transformations. You can start with an empty room, change the color of the walls, insert a table, add bookcases or paintings, without the previous steps being lost or distorted. This incremental approach is much more similar to human design work and allows you to build complex scenes in several phases.
If you want to test these characteristics of the Banana Nano model, know that you can do it, since in an official note Google has confirmed its availability in Gemini’s app both to users with subscription and the free ones, and has announced that the Nano Banana model will soon be accessible to developers through dedicated bees and in professional environments such as study and vertex AI.
Each modified image has a watermark, or rather two
Given the results to say the least excellent that it is possible to achieve with such a powerful tool. As for safety and transparency, every output produced with Gemini is marked by a visible watermark that clearly shows the artificial origin of the image, and by an invisible Synthid digital watermark. The latter is substantially a marker that remains detectable even after cutouts or changes and is designed to combat any improper uses of AI.
