Gemini Omni Flash: Google's Fast Video Model

TL;DR. Gemini Omni Flash is a Google video model with four modes instead of the usual two: from text, from a picture, from a reference and editing a finished video. The last one is rare, and it is the reason to open this model. The model is fast, the duration is set explicitly, it works in Pixyn, and the cost starts from 21 tokens. At Google it is in preview status.

---

Four modes

Mode	What you give	What for
From text	a prompt	a scene from scratch
From a picture	an image + a prompt	animate a frame
From a reference	a sample image + a prompt	hold a style or an object
Video editing	a finished clip + a prompt	rework what is already shot

The fourth mode is rare. Generative models usually make video but do not rework existing footage. Here you can take your own clip and ask to change it with text.

Replace the background with an evening city, keep the person and their movement unchanged.
Warm light, backlit from behind.

Don't do this: "make it nicer", "improve the quality".

The model reworks what you described, not what you implied. It will understand "nicer" its own way, and the result will differ from the source unpredictably. Describe a specific change: what to replace, what to keep, what the light should become.

How to phrase an edit of a finished clip

The editing mode behaves not like a filter but like a reshot scene: the model rebuilds the frames, relying on the source motion. Two practical consequences follow.

First: always say what to keep. If you write only "replace the background", the model is entitled to change everything else too: the pose, the clothes, the light. The wording "replace the background, keep the person and their movement unchanged" holds what you took your clip for.

Second: one edit at a time. Three changes in one prompt almost always give a result where one worked and two drifted. It is simpler to run two takes in a row than to figure out which wording broke the frame.

Keep the person, their pose and the camera movement unchanged.
Replace the background with an empty hall with concrete walls.
The light is cold, from the windows on the left, with soft shadows.

How Omni Flash differs from its catalog neighbors

Task	What to take
Rework an already finished video	Omni Flash
Hold a specific character across several references	Seedance 2.5
Extend a good but short clip	Grok Imagine
Raise the resolution of footage	video upscaling
Make a character talk	lip sync

This is not a "who is better" ranking: the models cover different steps, and one project often needs two or three.

What matters about duration and price

The duration is set explicitly, and the price depends on it: the base includes a few seconds, and each next second adds to it. The number on the button is recalculated as soon as you move the slider, so check it before the run.

Where to use it

1. Editing footage you already shot. Change the background, the time of day or the setting without reshooting.
2. Quick variants of one scene. The model is fast, so trying variants costs less time.
3. Animating stills. A product or interior photo turned into short motion.

Practical habits

• Save the sources and the prompts. To reproduce a result, you need both the files and the exact wording.
• Download finished clips right away. Links to results are not permanent with any provider.

---

Try it. Create an account: new accounts get 6 tokens. That is enough for a picture; video with Omni Flash starts from 21 tokens and grows with the duration. Plans are on the pricing page.

FAQ

How is Omni Flash different from Veo? They are different Google models for different tasks. Omni Flash is faster and can edit a finished video; Veo is stronger in picture quality and native sound. A breakdown of Veo is in a separate article.

Can I upload my own video and rework it? Yes, there is a separate editing mode for that.

What does preview mean? That the model is not in its final status at the vendor. It works, and Google may change its behavior over time.

Why are there no duration and format settings in editing mode? Because the source video sets them: the input there is only your clip and the prompt. The length and the frame proportions are inherited. In the other three modes, the duration and format are set explicitly.

How many references does the reference mode take? The field is designed for several images at once, not one: you can show an object from different sides or set a hero and a style at the same time.

How many times can I rework one clip? As many as you like, but each edit is a new generation and a new charge. It is cheaper to get what you need in two or three runs with precise wording than in ten with general ones.

Can I upload a long video and rework it whole? The practical scope is one scene. A long recording makes sense to cut into fragments and process in parts, so the model holds coherence. How spending works: the card about tokens.

Source: the Gemini Omni Flash announcement on Google's blog.