Gemini Omni Flash

What it is

Gemini Omni Flash is a generative model from Google DeepMind for creating content from many kinds of input, "starting with video". ^[2] The team sums it up in a single comparison: it is "like Nano Banana, but for video" — a general generation and editing model whose first application is moving images rather than still ones. ^[2]

Video creation with the model launched in the Gemini app on 19 May 2026, for Google AI subscribers aged 18 and older. ^[1][2]

Working by conversation

Users do not assemble a video by dragging clips onto a timeline. They describe what they want in prompts that can mix text, photos and video, and the model produces a clip. ^[1][2] Editing works the same way: a new instruction builds on the earlier ones, so a scene stays consistent as it is revised rather than being rebuilt from nothing each time. ^[1][2]

There are three ways in. A user can start from scratch, pull material from the camera roll, or begin with a preset template. ^[1]

What it accepts and what it can change

The model takes references in several forms: an image, text, a video of up to three seconds, or audio. ^[2] From those it can change the style of a shot, alter what the characters are doing, shift the camera angle, swap one character for another, carry motion from a reference clip onto a new subject, or turn a sketch into footage. ^[2]

Google DeepMind lists four modes for Gemini Omni Flash: video editing, text-to-video, image-to-video and reference-to-video — the last meaning a clip generated to match material supplied as a reference. ^[2]

While work is in progress, drafts render at 360p, a low resolution that keeps iteration quick; a finished piece can then be upscaled to 1080p or 4K. ^[2]

Versions and the API

The model reached developers through the Gemini API in stages. gemini-omni-flash-preview appeared on 30 June 2026, gemini-omni-1.1-flash on 27 August 2026, and gemini-omni-1.2-flash on 5 November 2026. ^[3]

Each step was measurable. In side-by-side ratings, Omni 1.2 was preferred over 1.1 in 62% of comparisons on visual consistency in text-to-video, among other categories. ^[2]

Google recommended gemini-omni-1.2-flash as the replacement for the Veo 3.1 API previews, which were shut down on 22 October 2026. ^[3] In other words, developers who had built on that earlier preview were pointed toward the Omni line.

Where it runs

Beyond the Gemini app, the model is available in Google Flow, Google AI Studio and the Gemini API, in YouTube Shorts and in Google Vids. ^[2] That spread matters: the same model sits behind a consumer app, a developer endpoint and editing tools inside products people already use for short video and workplace clips.

Outputs carry two kinds of provenance marking. SynthID embeds an imperceptible watermark directly into generated media so it can later be identified as machine-made, and C2PA Content Credentials attach tamper-evident metadata describing how a piece of media was produced. ^[2] Both are meant to keep synthetic video traceable once it leaves the tool that made it.

Why it matters

The interesting shift is the interface rather than any single output. Making video used to mean operating editing software; here it means holding a conversation with a model that remembers the previous turns, so style, cast and camera decisions persist across revisions. ^[1][2] Adding audio and short video references, motion transfer and sketch-to-video to the same model puts several jobs that were once separate tools behind one prompt box. ^[2] The staged API releases and the preference ratings suggest the capability is being pushed forward version by version rather than delivered once. ^[2][3]