Google upgrades Gemini picture enhancing with ‘nano banana’ mannequin

Gemini is best at enhancing your pictures with a brand new mannequin replace from Google DeepMind.

The mannequin, which rose to the highest of the LMArena leaderboard below the alias “nano banana,” is definitely referred to as Gemini 2.5 Flash Picture. The Google DeepMind staff says this mannequin has been educated to make topics extra constant throughout numerous edits of AI-generated pictures. This has been a difficulty for AI picture fashions, given their unpredictable nature.

The power to add and natively edit images in Gemini has been round since April of this yr. As an alternative of studying the technical ins and outs of picture enhancing software program like Photoshop, customers can add a picture to Gemini and describe the modifications they need in easy phrases, no jargon required. With Gemini’s up to date mannequin, Google says you are able to do issues like change a topic’s outfit and placement, whereas conserving their likeness the identical.

You can too add a number of images and have the topics seem collectively in the identical picture, or add and alter particular particulars in an uploaded picture to, say, see what a room appears like with a unique shade of paint or completely different furnishings.

This is Gemini’s try at enhancing my canine into the downward canine pose and relocating her to a yoga studio. Her likeness is identical, and it efficiently edited the picture to make her eyes open, however her physique is not arched in the way in which it needs to be. (I might know, I’ve seen this playful pose from her many instances.)

This is my canine Lola, not doing yoga
Credit score: Mashable

This is the Gemini-edited model after I prompted it to open her eyes, put her within the downward canine pose, and alter the background to a yoga studio. It is shut, however not fairly proper.
Credit score: Mashable

As Google DeepMind mentioned in its announcement, the mannequin may not all the time get it proper. There may nonetheless be inaccuracies with advantageous particulars, textual content within the picture, and inconsistencies. With my experiment, my canine’s fur appears overly easy, however her total coloring, measurement, and form keep the identical. All pictures have a visual watermark and an invisible watermark referred to as SynthID to mitigate any confusion over whether or not they’re actual or AI-generated. This replace is now dwell, so you’ll be able to attempt it out for your self within the Gemini app.