How to actually use Gemini Omni

Gemini Omni is a video model that edits, transforms, and generates clips by conversation. You hand it a starting point (a video you already have, a still image, or just a sentence), describe what you want changed, and it produces a new clip. Then you keep talking. The same character holds across edits, the lighting follows, the physics work. Google calls it "create anything from anything" and that line lands once you have used it for ten minutes.

It launched on May 19, 2026 at Google I/O 2026, built by Google DeepMind, and is rolling out inside the Gemini app right now. The full model name is Gemini Omni Flash. For the purposes of this guide, "Gemini Omni" means that.

Who it is for

Anyone who needs to produce or modify short video and does not want to learn an editor. The shape of the work is different from Veo 3 or Sora. Those tools take one prompt and give one clip back. Omni takes a clip and a conversation, and gives you a clip that responds to what you actually said. That conversational loop is the unlock.

• Creators making short-form: punch up a take, swap a background, restage a shot, all without re-recording.
• Marketers and PMs: stand up a 10-second product visual without booking a video team.
• Educators and explainer-makers: ask for a claymation about protein folding and the result is scientifically correct, not just stylised.
• Anyone with a sketch or a still that wants to see it move.

There is no separate "Omni" product. The model lives inside surfaces you may already use. Pick the one that matches what you are trying to do.

• Gemini app (gemini.google.com or the mobile app): the default. Full chat interface, drag a clip in, type, iterate. Plus / Pro / Ultra only.
• Google Flow: workflow-style canvas for longer-form video work. Same model underneath.
• YouTube Shorts Remix: in-app, free. Tap "Remix" on a Short, describe the change, watch it apply.
• YouTube Create app (18+): free on phone. Built for vertical short-form from the start.

There are three ways to begin. The right one depends on how much you already have.

Start from words

Type a description. Omni generates a short clip from scratch. This is the closest to how Veo 3 or Sora feel. Best for moments where you have no reference footage and need something to react to.

Start from an existing clip

Upload a video you already have (or one Omni made on a previous turn) and ask for changes. This is where Omni is genuinely different from the rest of the field. The same scene, restaged, with the same character. The same character, moved to a new venue. The lighting still holds.

Start from a reference

Drop a photo, a sketch, or another clip in alongside your prompt. Omni reads the reference as style, mood, or content and applies it. This is the underused mode. A pencil sketch becomes live footage. A still becomes a moving scene. A photo of a wooden puppet becomes the new look for your character.

Once you have a starting clip, the multi-turn loop is where you spend most of your time. Treat it like talking to an editor, not searching a stock library. One change per turn. Look at the result. Decide what to push on.

The model carries context across turns. That means the character you set up on turn one is still recognisably the same character on turn five. Lighting, camera language, and palette tend to carry too. You can lean on that.

What the prompts should sound like

Direct, concrete, one change at a time. The shape that works best for me:

[change]    What is different now? ("Move her to a sunset beach.")
[constraint] What must stay the same? ("Keep her face, posture, and the violin.")
[style]     What should it feel like? ("Warm light, low contrast, slight handheld.")

The constraint line is the one most people skip. Without it, the model may swap out things you wanted to keep. Naming what should not change is half the prompt.

When to restart vs keep going

If a turn ends up two or three steps away from where you wanted, do not keep correcting. Go back to a known-good earlier turn and branch from there. Trying to claw a drifted clip back through more turns usually compounds the drift.

The fastest way to get a clip that looks intentional is to anchor it in something real. Omni accepts photos, paintings, frame grabs, other videos, even rough sketches. The reference does not have to match your scene. It has to communicate the look.

• Visual references: drop a single frame from a film, a portrait, a product photo. "Match this lighting." "Match this palette." "Make her look like she belongs in this world."
• Aesthetic references in words: "Wes Anderson palette." "Eames-era industrial." "Stripe Press but darker." A real, specific reference beats five abstract adjectives.
• Functional references: feed it a clip that has the camera move you want and say "this camera language, my subject." It will do the right thing more often than you expect.

Most generative video tools are pattern matchers wearing a physics costume. Omni has more of the physics under the hood, plus a real model of the world (the same one that makes Gemini good at answering questions). That shows up in two places worth knowing about.

It models physics, not vibes

Hand it a marble track. The bounces land in the right spots. The audio (when audio is included) matches the impact. Drop a ball and it falls correctly. This is not a complete physics engine, but it is enough that you can describe a scene mechanically and trust the result.

It knows things

Because Omni is built on Gemini, it carries world knowledge through to the visual output. Ask for a claymation explainer on protein folding and you get one that is scientifically right, not just stylish. Ask for the inside of a cell, the structure of a violin, the gait of a horse, and the output respects how those things actually look.

It syncs on-screen text and action

You can ask for text that appears in time with what is happening on screen. Lower thirds that land on the beat, labels that follow a moving object, captions that read in time with a gesture. This works well enough to make titled explainers without an editor in the loop.

A short, honest list. None of these are deal-breakers, but knowing them up front saves you a frustrating turn.

• No audio editing. You can keep native audio that came in with your input clip, and the model can produce synced sound on a generated scene. You cannot ask it to "give her a different voice" or "swap out the soundtrack." Google held this back at launch over deepfake risk.
• Video only, for now. The Omni family will grow, but at launch it is video generation and video editing. Not stills, not audio, not 3D.
• No API at launch. The blog post says "coming in the coming weeks". If you want to build with it, watch the rollout.
• Quota burns fast on complex multi-turn jobs. Two heavy reference-driven prompts can chew through a meaningful slice of your weekly Pro quota. Plan your shots.
• Every output carries SynthID and C2PA Content Credentials. This is a feature, not a limit, but worth knowing if you are publishing.

Concrete is better than abstract. Here is the exact sequence I would run on a fresh account to feel the loop in under ten minutes.

1. Open the Gemini app, attach one short clip

Any 5 to 15 second clip you already have. Something with a person or a clear subject and decent lighting. If you do not have one, skip to step 2 and start from words instead.

2. Type the first turn

Take this clip and move the same person to a sunset beach.

Keep their face, clothes, and posture the same. Keep the camera framing the same.

Lighting should feel warm, low contrast, slight handheld camera. Closer to a Sofia Coppola short than a music video.

Wait for the result. Watch what holds and what shifts. Make a note of the one thing you would change.

3. The second turn

One change only. "Pull the camera a little wider." "Make the sun lower in the frame." "Take the wind out of her hair." Resist stacking three asks into one turn. The loop works better when each turn is a single, named edit.

4. Save the version you like, branch from it

When a turn lands, say so. "Keep this exact look. Now show the same character walking down the shoreline at the same time of day." Naming the version anchors the model to it before you push on.

If you have no clip to start with

Start from words. Try the marble one, because it is the cleanest demo of the physics that makes Omni feel different.

A glass marble rolling down a wooden chain-reaction track. Three clean bounces, then a small brass bell at the end that rings when the marble strikes it. Top-down lighting, shallow depth of field. Native sound: the bounces and the bell.

Then ask it to "do the same scene, but the marble is now made of mercury" and watch the physics still hold.

Three things, in order, once you have spent fifteen minutes inside the tool.

1. Run the marble prompt and one multi-turn prompt on a clip you already own. That is the whole loop. Everything else is variation.
2. Try a reference-driven transform. A photo, a sketch, a film still. The first time a reference clicks, the model stops feeling like a video toy and starts feeling like a collaborator.
3. Set yourself a constraint: one finished short, made entirely inside Omni, by the end of the weekend. The fastest way to learn the edges of the model is to actually finish something with it.

• Google: Introducing Gemini Omni (blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni)
• Google DeepMind: Gemini Omni model page (deepmind.google/models/gemini-omni)
• Google: Google I/O 2026, all the announcements (blog.google/innovation-and-ai/technology/ai/google-io-2026-all-our-announcements)
• 9to5Google: Gemini Omni rollout (9to5google.com/2026/05/19/gemini-omni-create-anything-model-video)