Image editing in Gemini just got a major upgrade

Google DeepMind dropped a new image editing model into the Gemini app today, and the pitch is pretty specific: it's supposed to nail likeness. Not a vague resemblance, but the actual person or the actual dog, recognizable after you've swapped their outfit, their decade, or their entire backdrop.

That focus on consistency is the real story here. Anyone who's played with AI image tools knows the annoying drift that happens across edits — your face slowly turns into someone else's face by the third revision. Google says this update, built by the Gemini Apps and DeepMind teams under David Sharon and Nicole Brichtova, was tuned specifically to fix that. Give it a photo of your chihuahua, ask for a tutu, and the dog is supposed to still look like your dog, not a generic stock chihuahua in a tutu.

The feature set goes beyond simple retouching. You can blend two separate photos into one scene — say, you and your dog on a basketball court, stitched together from two unrelated pictures. There's multi-turn editing too, so you can paint a room, then add a bookshelf, then drop in a coffee table, one instruction at a time, without Gemini rebuilding the whole image from scratch each round. And there's a style-transfer trick where you lift a pattern or texture from one photo — butterfly wings, flower petals — and apply it to an object in another, like rain boots or a dress.

Google is also leaning on this as a bridge into video: once you've got an edited image you like, you can feed it back into Gemini and turn it into a short clip. Every output, whether generated or edited, carries a visible watermark plus Google's invisible SynthID marker, which is the company's now-standard answer to the