Getting Started with Google Veo

Describe a shot, get a few seconds of video with the sound already in it.

Last reviewed 2026-09-07 · for Google Veo 3.1

How it works

You describe a shot — what's in frame, how the camera moves, what you hear — and the model (the AI "brain" doing the work) returns a few seconds of video with its own sound. Each attempt is a fresh roll, not an edit, so you steer by rewriting and going again.

What it is

A text-to-video tool from Google DeepMind, reached through Google's own apps — a Google account is the only login you need.

It makes picture and sound in one pass — Google's docs list native audio as always on, where most rivals in this class hand you silent video you score yourself.

That, a daily allowance Google's own pages describe as free, and no new account or card is why you'd pick it over Runway, Pika or the Sora-class tools. It lives in Google Flow, a studio with projects and a timeline rather than a bare text box.

Before you start

Start in Google Flow (labs.google/fx/tools/flow), not the Gemini chat app — video there currently needs a paid Google AI plan.

The free allowance is small and spent per attempt — and Google is inconsistent about whether you get one. Flow's own page and its credits help article say a signed-in Google account currently gets a small daily allowance of Flow credits, the unit of spend; Flow's "Get started" article lists a paid Google AI plan as a requirement. Check which one your account lands on before planning around it. On that reading, the top-quality Veo setting costs about ten times the cheapest: a handful of short clips a day, or a single top-quality one. Credits don't roll over, and mush costs the same as your hero shot.

You'll pay past the experimenting stage, because ten attempts at one shot is a normal session. Paid Google AI plans run from a low-cost entry tier through a mainstream tier to expensive top tiers; check the plan page, these move.

First step: sign in, pick the cheapest model, and run one description several times. How much a single wording varies is day one's real lesson. You must be 18 or over and in a supported country.

Security & privacy

Flow runs on your Google account and what you type is treated like other Gemini activity: Google says a subset of conversations is read by people (disconnected from your account first) and kept up to three years, and that your activity trains its generative AI models. Turning off Keep Activity stops that and cuts retention to around 72 hours. Assume a person could read what you type.

Watermarks, and what you may not generate

Every video from Google's video models carries SynthID, an invisible watermark — a mark identifying output as AI-made. You can't turn it off, and Google's own detector reads it: disclose rather than be caught not disclosing.
A visible watermark is a separate toggle in Flow's profile menu, applied automatically for users in India, South Korea and Vietnam.
Google's Generative AI Prohibited Use Policy bans impersonating a person, living or dead, without explicit disclosure in order to deceive; non-consensual intimate imagery; and using personal data or biometrics without legally required consent. So don't animate a photo of a real colleague or public figure — expect refusals, and occasional false ones on invented characters that merely look real.

What it's good (and not good) at

Best at

Short establishing shots and B-roll where nobody speaks
Atmosphere — weather, light, texture, room tone, ambient sound
Animating a still image you already like

Okay at

Dialogue and lip-sync — it'll try; results vary shot to shot
Continuing a clip you liked — on some Flow model options, not others
On-screen text, signage, product and logo accuracy

Avoid

Surgical edits to a clip you already have — a generator, not an editor
Client work signed off without a human watching every frame

Starter prompts to try

Write like a camera operator, not a customer

Handheld medium shot, shallow depth of field: a woman in a yellow raincoat steps off a bus into heavy rain, rain drumming on her hood, bus doors hissing shut behind her.

Why this one: Veo's controls are built around camera movement and framing, so film-set words work as instruction: handheld means the camera is carried rather than fixed, shallow depth of field means only the woman is sharp. The rain and the hiss arrive as sound in the same pass.

Tweak: Swap handheld for a locked-off tripod. Same story, colder temperature.

Animate a photo you own

Animate this still: the camera drifts slowly to the right, steam rises from the mug, soft morning room tone with faint birdsong outside the window.

Why this one: Image-to-video is core to the Veo family, and because audio is native, a silent photograph comes back with a soundscape.

Tweak: Ask for a push in instead of a drift right.

Bridge a first frame and a last frame

Move from the first frame to the last frame in one continuous handheld take with no cuts, keeping the ambient street sound running throughout.

Why this one: First-and-last-frame control is a documented Veo capability — you fix where the shot starts and ends, and it solves only the middle.

Tweak: Use two near-identical frames with one moving object.

Common pitfalls

A "video" here is a handful of seconds. Flow's Veo options currently produce roughly four, six or eight seconds, the newer companion model about ten. Anything longer is you assembling clips, and continuity between them is the real work.
The model names change faster than any tutorial. Google ships new video versions constantly and is rolling out a successor, Gemini Omni, which replaces Veo in the Gemini app while both stay selectable in Flow. Trust the in-product menu over any guide, including this one.
Consistency across clips stays hard. The same character generated twice gives two similar-ish people; reference images improve the odds, not guarantee a match.

Where to next

More video tools to compare it against Make the still images you feed it

Want the full picture first? See the Google Veo 3.1 overview for pricing, the free tier and related tools.