The image model that finally does consistent characters

Google shipped an image generation and editing model — Gemini 2.5 Flash Image, which spent its preview period under the codename "nano-banana" — that maintains subject identity across edits.

The internet used it to put people in figurine boxes for two weeks. Underneath the meme is a capability that unlocks actual products.

what identity preservation means#

Take a photo of a person. Ask for the same person in a different setting, at a different angle, in a different style. Previous models produced someone who looked similar. This one produces someone who looks like the same person.

The same holds for products, characters, and objects across a sequence of edits. Edit, then edit the edit, then edit that, and the subject stays coherent instead of drifting into a different thing over four generations.

That drift was the blocking problem for every serious use of image generation.

why it matters commercially#

Consistency is the difference between a toy and a tool:

• Product photography. One studio shot, then the product in twenty contexts. The product has to be the same product or it is fraud.
• Character work. Comics, storyboards, game assets, explainer videos. Everything narrative requires the character to persist across frames.
• Brand assets. A mascot that changes subtly in every image is not an asset.
• Virtual try-on and staging. The room has to stay the same room.

Every one of those was demonstrated in 2023, was impressive in a demo, and did not ship, because the drift made it useless past three images.

the multi-turn editing model#

The interaction is conversational. Generate, then refine in natural language, with the model maintaining state:

> a photo of a ceramic mug on a wooden desk
> now make the desk marble
> add steam
> shoot it from a lower angle
> put the same mug in a cafe window

Each step preserves what came before. That is a fundamentally better interaction than regenerating from a longer prompt, which is what the previous generation forced, and which lost everything you liked about the previous image.

the provenance question#

Outputs carry SynthID watermarking. Google has been consistent about including it and it is genuinely better than nothing.

It is also not a solution to what this capability enables. A model that can put any identifiable person in any scene, photorealistically, with the identity preserved, is a harassment and disinformation tool with a creative-tools UI on top.

Watermarks help platforms detect at scale. They do not help the person whose face is in an image they did not consent to, and they survive exactly as long as nobody is motivated to strip them.

I do not have a good answer here and I am suspicious of anyone who claims to. The capability exists, it will exist in open weights within a year, and the mitigations are all downstream — platform policy, legal remedies, and social norms that have not formed yet.

for developers#

The API is straightforward and the pricing is per-image. Two practical notes:

Include provenance metadata in anything you generate and pass it through your pipeline. It costs nothing and the regulatory direction is clear.

Build a consent and review step into any product where a user uploads someone else's likeness. It is not required yet. It will be, and retrofitting it after launch is much harder than designing it in.