How does Google Veo 3 work according to the public docs?

The public Gemini API docs describe Veo 3.1 as an async video generation model: you send a prompt, optionally guide the result with images or frames, poll the operation until it is done, and then download the generated video.

What does Google publicly confirm about Veo 3?

Google publicly confirms 8-second 720p, 1080p, or 4k generation with native audio, portrait video support, video extension, frame-specific generation, and up to three reference images.

Does Google publish the full Veo 3 architecture and training pipeline?

Not in the public Gemini API documentation. The public docs explain the generation flow and user-facing controls, but they do not publish a full transformer diagram, weight architecture, or complete training pipeline.

How Veo 3 works, step by step

How Google Veo 3 works: the public flow, controls, and limits

The public Gemini API docs describe an async video generation flow, not a fully published internal architecture. You send a prompt, optionally add images or frames, poll the operation until it finishes, and download an 8-second 720p, 1080p, or 4k video with native audio. Not sure which route you have yet? See how to access Veo 3.

Quick answer

The current public docs confirm the generation flow and controls: prompt input, async job polling, 8-second output, native audio, portrait support, video extension, frame-specific generation, and up to three reference images. They do not publish the full internal model diagram, training pipeline, or a detailed transformer stack explanation in the API docs.

Open Veo docs Read the prompt guide Check Veo pricing

Need the public source first? Use the official Gemini API video docs.

Veo 3 guides

Choose the Veo guide you need

Find practical help with access, video length, prompts, pricing, API setup, and how Veo works.

Open Veo Studio

Access How to access Veo 3 Free access, rollout routes, and first-use path. Length How long are Veo videos? 8s max, 4s/6s options, and extension rules. Prompts Veo prompt guide Official structure, examples, audio, and negatives. Mechanics How Veo 3 works Async flow, controls, public docs, and limits. Pricing Veo pricing and credits Budget route before repeated renders. API Veo 3.1 API docs Endpoints, params, status flow, and examples.

Generation shape

8 seconds

The current public Veo 3.1 docs describe high-fidelity 8-second generation.

Output modes

720p, 1080p, 4k

The public docs list 720p, 1080p, and 4k output support, with native audio generation.

Control layer

Up to 3 images

Public image-based direction supports up to three reference images, plus first/last-frame workflows.

Job model

Async polling

You submit a generation request, poll the operation, and download the generated file when it is ready.

Public generation flow

This is the part of “how Veo 3 works” that the public docs actually show.

Step 1

Write the prompt and optional controls

Start with a text prompt, then optionally add reference images, first/last frames, portrait orientation, or extension settings depending on the workflow.

Step 2

Submit a long-running generation request

The official examples use async video generation calls. The generation does not return instantly as a finished inline response.

Step 3

Poll the operation until it is done

The public code samples repeatedly check the operation status. That is the documented flow for waiting on the finished video.

Step 4

Download the generated file

Once the job is complete, the examples download the resulting video file rather than treating the response as a synchronous final asset.

More Veo guides

What the public docs confirm vs what they do not

Topic	Publicly confirmed	Not publicly documented in detail
Generation flow	Prompt in, long-running operation, poll status, download output.	Internal scheduler design, cluster orchestration, or exact serving stack.
Capabilities	8-second video, native audio, portrait mode, extension, frame-specific generation, up to three reference images.	A complete internal breakdown of which submodels handle audio, motion, or image conditioning.
Model internals	Google describes Veo publicly as a state-of-the-art video generation model.	The full transformer topology, weight layout, training corpus, and exact pipeline internals.
Prompt control	Prompting, reference images, first/last frames, and extension are publicly described control layers.	A complete public spec for every latent control or motion-planning subsystem.

What “how it works” means in practice

Practical signal 1

Native audio is first-class

The public docs frame Veo 3.1 as a video model with natively generated audio, not as a silent clip generator that always needs a separate audio pass.

Practical signal 2

Reference images shape the result

Up to three reference images, plus first and last frame control, show that Veo is not just prompt-only text-to-video in the current public workflow.

Practical signal 3

Extension is built into the workflow

Video extension means the generation process can continue an earlier output instead of starting from zero every time.

Practical signal 4

The controls matter more than the model internals

For a usable result, focus on the prompt, reference image, aspect ratio, duration and audio settings. Google does not publish every internal implementation detail.

Generate a clip or open the guide you need

Use the API docs for integration, the prompt guide for better shot descriptions, pricing for a cost estimate, or the access guide to create an account.

Open Veo docs Read prompt guide Check pricing

Veo is the video model; Flow is one creation interface

Veo generates the video, while Flow organises filmmaking and scene work around supported Google models. They are related but not interchangeable product names, so access and controls can differ.

In UlazAI, choose the visible Veo mode, provide the supported prompt or references and review duration, audio and cost before generating. Treat unofficial “Flow character” field names as unverified unless they appear in the API contract.