Google Flow

Google Flow is a Google creative environment for developing images, clips and video scenes with generative models. It supports audiovisual creation and editing within a composition workflow.

In brief

Video generation and avatars

Official website

Contents
  1. What Google Flow is
  2. Origins and evolution from VideoFX
  3. Current models and announced retirements
  4. Text, ingredients and frames
  5. Images, versions and visual preparation
  6. Clip editing and scene construction
  7. The agent as production support
  8. Tools: small custom workflows
  9. Library, downloads and preservation
  10. Access, credits and regional differences
  11. Avatars and representing people
  12. Data, provenance and publication
  13. Google Flow Music within the ecosystem
  14. Flow TV and the trademark dispute
  15. Practical applications in OSH
  16. Worked example and result checks
  17. Limitations and conditions of use
  18. When to choose it and how to complement it
  19. Frequently asked questions
  20. Related profiles
  21. On the Sabentis blog
  22. Sources and documentation

Back to Tools

What Google Flow is

Google Flow is a Google Labs creative environment for developing images, video and scenes with generative AI. It combines models, references, editing and asset organisation, alongside a conversational agent and custom tools. It is not the name of a single model: choosing Flow does not by itself determine the technology generating each asset.

The practical distinction is between producing a clip, revising it and constructing a complete piece. A story needs shot selection, continuity, sound and editorial judgement. This profile explains those layers separately and describes documentation reviewed on 8 November 2026, because models, plans and availability change frequently.

Reference documentation: Flow May 2026 updates.

Origins and evolution from VideoFX

Google introduced Flow on 20 May 2025 as an evolution of VideoFX, a 2024 experiment. The offering combined Veo, Imagen and Gemini for audiovisual creation. In February 2026 it announced the integration of Whisk and ImageFX capabilities, with voluntary project transfers beginning in March.

The chronology explains the changing focus: from experimenting with video generation to bringing image preparation, animation and scene revision into one workspace. A launch-era guide may therefore describe different models or controls. Retain the date of internal production instructions and review them when the interface changes, especially when a team needs to reproduce an established workflow.

Reference documentation: Flow launch, 20 May 2025 · Flow February 2026 redesign.

Current models and announced retirements

Current documentation lists Omni Flash 1.2, Veo 3.1 Lite and Nano Banana variants. Veo Extend retired on 5 November 2026. Retirement of Veo 3.1 Fast, Quality and Omni 1.1 is announced for 5 December; that remains a future change at this review date.

Omni 1.2 Flash, introduced on 5 November, extends native generation to fifteen seconds and combines visual, motion and audio references. Google acknowledges remaining difficulties with complex physics and fast action. Compare models using the same narrative objective, assessing both the accepted shot and required revisions. Higher resolution does not correct a wrongly depicted action or make an implausible sequence physically valid.

Reference documentation: Google Flow — Models and supported features · Gemini Omni 1.2 Flash, 5 November 2026.

Text, ingredients and frames

Flow supports text generation, images used as ingredients and transitions guided by starting or ending frames, depending on the model. Each input has a different purpose: text describes intent, a reference guides appearance and a frame establishes a starting or finishing composition.

A useful brief separates subject, location, action and camera. For a scene about keeping a walkway clear, specify what must remain constant and what changes. Avoid requesting many independent actions at once. Save a reference version before varying lighting or framing, so the reviewer can identify whether an aesthetic improvement also changed the message or the relationship between objects.

Reference documentation: Google Flow — Create videos.

Images, versions and visual preparation

Images can be final assets or starting material for videos. Flow provides generation, cropping and instruction-based editing, retaining versions and prompts in its history. The agent can also work with variations or several selected images.

For a series, first prepare a small library: protagonist, setting and recognisable objects. A wide view helps agree the layout; a close view supports detail review. Keep approved images separate from rejected trials when selecting references. If an illustration moves a barrier or walkway, reassess the entire spatial relationship before animating it. Approval of the original still image should not automatically carry over to every edited version.

Reference documentation: Image creation and editing.

Clip editing and scene construction

Generative editing modifies clips through instructions. Scenebuilder can order, trim and download a sequence, while saved frames can guide later generations. Current extension requires Omni Flash 1.2 and an eligible subscription; it is distinct from the former Veo Extend feature.

Separate correction from assembly. Changing lighting regenerates content; trimming selects what viewers see. In a three-shot story, review every entrance and exit, the character’s orientation and persistent objects. A cut can remove information explaining a decision even when both clips seem correct individually. Keep the source clips and approved sequence, so later revisions can be assessed against the same narrative rather than against memory.

Reference documentation: Editing and Scenebuilder.

The agent as production support

Flow Agent can help brainstorm, prepare visual scripts, choose models, generate variations and organise assets. The current guide places it on the web and computers. Queries have their own quota; media generation consumes the relevant credits.

Its value depends on a verifiable task. Ask for three alternatives of the same shot with explicit framing differences rather than an undefined improvement. Before commissioning a batch, agree what counts as acceptance. In shared work, retain the script and decisions outside the conversation: the team should understand why a version was selected and which changes remain outstanding. A proposed production plan still needs an owner and a defined final output.

Reference documentation: Flow Agent.

Tools: small custom workflows

Tools builds utilities with generated code and interfaces. Creating tools requires an eligible plan; free users can use prepared tools. Sharing exposes code and other details. A shared link preserves a version, so later edits do not automatically propagate.

A reasonable application is a template requesting asset variations through format and tone fields. Before sharing it, test empty inputs, long instructions and credit consumption. Do not embed keys, internal data or confidential references in its configuration. Retain a version identifier and explain what users supply, what the tool produces and which review is required before using the output. Reusable automation deserves its own acceptance check.

Reference documentation: Flow Tools.

Library, downloads and preservation

Flow organises assets into projects and collections and supports reusable characters. Shared links can be used by anyone receiving them. Google Takeout exports media and metadata about prompts, models and organisation, including JSON files for reconstructing relationships within the downloaded archive.

Separate delivering a film from preserving its project. Delivery needs the final video, version identification and approved text. The working archive should identify references, decisions and assets used. Before deleting a project, check that its export opens and that necessary pieces can be located. A folder containing hundreds of variants without editorial selection remains difficult to maintain, even when the download itself is complete.

Reference documentation: Projects, collections and sharing · Google Takeout export.

Access, credits and regional differences

Flow requires adulthood and regional availability; Spain appears among supported countries. Credit documentation provides for daily free access and subscription allocations, with consumption varying by model and operation. Agent, Tools and other features do not necessarily have equivalent computer and mobile availability.

Budget for a reviewed piece rather than a single generation. Allow room for trials and corrections and verify the conditions of the account actually being used. A business producing several language versions must account for review, assembly and archiving as well as credits. Avoid building a production schedule around an announced feature that has not yet appeared in that account or region.

Reference documentation: Flow credits · Flow availability · Get started and watermarking.

Avatars and representing people

The avatar feature creates a visual and audio representation of the user. At the review date it is unavailable in the European Economic Area, United Kingdom and Switzerland. Enrolment requires a mobile device, and avatar data cannot be shared through Tools or public project links.

This restriction matters to teams in Spain: production should not depend on that feature. A fictional character and an avatar of a real person have different requirements. Define who appears, how the content is identified and who authorises each version. Representing someone should not attribute statements or decisions they have not approved, especially in communications presented as organisational instructions.

Reference documentation: Avatars.

Data, provenance and publication

The “Help improve Google Flow” control can disable use of interactions to improve Google products and technologies. Help documentation warns against entering confidential information. Veo, Omni and Nano Banana outputs include SynthID, and a visible watermark setting is also available.

Synthetic provenance and accuracy are different questions: a mark identifies generation rather than validating content. For internal communications, use authorised references and remove unnecessary data before uploading them. Retain the approved version and its editorial owner. If depicting a workplace recreation, identify it as an illustration and separate fictional details from the real instructions accompanying the video. Publication review should cover both the picture and its context.

Reference documentation: Data controls · Get started and watermarking.

Google Flow Music within the ecosystem

Google Flow Music originated as ProducerAI, which joined Google Labs in February 2026. It focuses on creating and refining music, lyrics and instruments. A Believe and TuneCore collaboration was announced in May, and Google confirmed Lyria 3.5 availability in the service in September.

Distinguish a standalone song from a scene’s sound. A safety communication may need intelligible speech, pauses and silence even when a generated soundtrack is available. Prepare the message first and add music for a specific purpose. Google states that it does not claim ownership of original content generated with Flow Music; that does not by itself settle rights in user-supplied material.

Reference documentation: ProducerAI joins Google Labs · Flow Music and Believe · Lyria 3.5 availability.

Flow TV and the trademark dispute

Flow TV presents channels and short films made with these technologies and exposes prompts for some clips. It can support analysis of staging decisions; a selected showcase does not reveal all the generations required to obtain a result.

In February 2026 Reuters reported Autodesk’s lawsuit over the Flow name and noted that Google had no immediate comment. The USPTO record lists a September opposition concerning GOOGLE FLOW. These are distinct proceedings, not an infringement judgment. The dispute also helps distinguish this tool from similarly named Autodesk audiovisual products, which belong to another company and have different functions. Product identity should remain explicit when comparing documentation.

Reference documentation: Get started and watermarking · Reuters, Autodesk trademark complaint · USPTO opposition 91310468.

Practical applications in OSH

Its potential occupational safety and health role is supporting visual stories and fictional awareness scenarios. The safety content must be designed first: which hazard the audience should recognise, which action the piece should explain and which boundaries the depiction must respect. A polished output does not establish that these learning goals have been achieved.

It can develop a connected illustrative story: noticing a condition, considering it and using a communication channel. It can also support a visual script through which specialists and communicators agree what to show before filming.

In OSH, the sequence should preserve the logic of the approved action. If a character identifies an unsafe condition, the next scene should not imply they alter it without authorisation simply to achieve a quicker visual ending. A clear decision sequence matters more than cinematic transitions.

Worked example and result checks

Example: prepare a short story about reporting an unsafe condition, showing its detection and notification to the responsible person. The team validates the scene and accompanies it with an explanation of the actual reporting channels available in the organisation.

  1. Write the actual reporting route and separate actions to be shown from information explained through text.
  2. Prepare authorised references for style and necessary objects. Avoid readable internal documents and recognisable people.
  3. Create clips with one action per scene. Keep names and terminology consistent without inventing contacts or channels.
  4. Review clips and joins: check what the character knows, what they decide and what actually follows the report.
  5. Add the workplace’s authentic channel and verify that viewers can locate it. Identify the story as a recreation.

This is a proposed learning scenario. A fictional report being sent does not prove receipt or the operation of the actual channel.

Limitations and conditions of use

Reusing references does not ensure every detail remains constant. Check equipment, surroundings, clothing and visible messages between shots. A small change can alter the interpretation of an element with preventive significance.

Start and end frames do not validate the transition between them. Visual quality is also distinct from physical simulation. Review usage requirements and feature availability before planning deliverables that depend on many regenerations. Keep the approved script available to assess differences between versions.

Reference documentation: Gemini Omni 1.2 Flash, 5 November 2026.

When to choose it and how to complement it

Flow fits work requiring references and related scenes to be organised. Runway offers another approach to motion control. VEED can support editing and caption review for the distributed material.

A selection trial should cover two consecutive scenes, a correction and the final file. This evaluates continuity and maintenance rather than only an isolated clip. Check the result with the person responsible for the workplace instruction.

Frequently asked questions

Does an ingredient guarantee an object stays unchanged?

No. It is a reference; review the result and continuity between clips.

Does every model provide every feature?

No. Documentation distinguishes capabilities and conditions. Check the active model and account options.

Can it replace explanation of the actual reporting channel?

No. Connect the story to the workplace’s genuine procedure, responsible people and resources.

Request a Demo

Discover all that Sabentis can do for your organization.

Try Sabentis

request a demo
stars 5
GetApp Software Advice Capterra