Reference fidelity, published
Blend up to 14 reference images with up to 5 characters held consistent on the Pro tier — the strongest published character-consistency spec of the image families we run.
Nano Banana is Google's own name for Gemini native image generation — a family built around reference fidelity: blend up to fourteen images while five people stay themselves. In MML ONE, six routes from the Gemini image family answer your canvas and storyboard.
By Google DeepMind
What the Nano Banana family does best, per Google's own documentation — including the watermark it never omits.
Blend up to 14 reference images with up to 5 characters held consistent on the Pro tier — the strongest published character-consistency spec of the image families we run.
"The best model for creating images with correctly rendered and legible text," in Google's words — long copy, multilingual lettering, even translating the text inside an image.
Localized select-and-refine edits, camera-angle changes, refocus, relighting from day to night, color grading — up to 4K output across ten aspect ratios.
Gemini 3 image models reason through the composition first, and can ground the picture in live Google Search facts when the image has to be right.
From the Gemini 3 Pro Image flagship — Nano Banana Pro, GA since May 2026 — down to fast Flash tiers, one line covers final frames and coverage passes.
Exactly what our catalog serves today — the Nano Banana line under its official Gemini names.
The 2.5 / 3.0 / 3.1 rows are provider-channel routes of the Nano Banana line, not Google-announced tier names. The in-app catalog is the live truth for what each route delivers.

Frames land on the canvas and in the storyboard as versioned assets. Recurring casts, multilingual poster text, and edit-heavy sequences are what you route here.
Every generated image carries an invisible SynthID watermark — always on, by design. Auditable provenance, not optional.
Reference fidelity has hard budgets — about five people and six high-fidelity objects on the Pro tier. Past them, likeness degrades.
Best performance is limited to 15 supported languages, and the model won't always deliver the exact number of images you asked for.
Start with a premise, a screenplay, or a folder of references. We'll set up your provider keys and walk through the first scene with you.
Cookie settings
We use analytics cookies to improve MML ONE. You can decline anytime. Privacy