← All articlesTools

Veo 3.1: a practical guide to audio, references, Flow and SynthID

Native audio, reference images, first and last frames, extension, access and limits: how to use Google Veo 3.1 in November 2026, and when to switch to Gemini Omni.

Veo 3.1 is Google DeepMind’s video model, released on 15 October 2025: it generates clips of 4, 6 or 8 seconds at 24 frames per second, with synchronised audio, in 16:9 or 9:16 and up to 4K. It takes up to three reference images or a first and last frame, and a clip can be extended in 7-second steps. As of November 2026 it is still the latest version of Veo on the platforms that offer it, from Google Flow to Adobe Firefly.

This guide gets into the practical detail: settings, audio, references, extension, access, limits and SynthID. The technical data comes from the documentation of the platforms that integrate Veo 3.1. Where they disagree, we say so, because that is exactly where a project can trip up.

What Veo 3.1 can do: the specs that matter

Veo 3.1 builds on Veo 3, the model with which Google introduced synchronised audio in May 2025: dialogue, sound effects and ambience. According to Wikipedia’s entry on Google Flow, version 3.1 added richer dialogue, environmental effects and music, along with new editing tools.

SettingVeo 3.1Note
Length4, 6 or 8 secondsPer generation
Frame rate24 fpsFixed
Resolution720p, 1080p, 4K4K is not everywhere
Formats16:9 and 9:16Only these two
AudioDialogue, effects, ambience, musicCan be switched off
ReferencesUp to 3 imagesOr first and last frame
Extension+7 seconds per step, up to 20 stepsAbout 148 seconds

The table combines fal’s model page, which documents 4K, 24 fps and extension, with Higgsfield’s, which lists lengths and references. The model comes in two main versions: Standard, for quality, and Fast, quicker and cheaper. Fal also mentions a lighter tier that produces silent 720p clips, which it does not offer.

Native audio: when to keep it and when to switch it off

Audio is the reason many people choose Veo. With it switched on, you get lip-synced dialogue, effects, ambience and music in one pass. In Firefly Boards, as of August 2026, Veo 3.1 and Veo 3.1 Fast were the only video models with audio.

But audio costs. On fal, as of November 2026, one second of Standard at 1080p costs $0.20 without audio and $0.40 with it; on Fast it goes from $0.10 to $0.15. An 8-second clip with audio therefore costs $3.20 on Standard. When a scene will end up under music and sound design, generating it silent halves the cost of every attempt.

  • Keep it on for close-ups with lines, where the lip sync has to be born with the picture.
  • Switch it off for mood shots, packshots and scenes that will sit on a music track.
  • Test your language before you rely on it: fal mentions dialogue in multiple languages, but none of the sources we checked lists which.

Even when generated audio works, finished work still goes through a proper mix. We cover this in our guide to AI audio for video.

Reference images, first and last frame, extension

Veo 3.1 gives you three ways to take chance out of the generation. With reference images, up to three, you lock a character, an object or a style. On Higgsfield, subject consistency from 1 to 3 images is reserved for the Standard version. With a first and last frame, you give the model two images and it writes the movement that joins them.

Extension, finally, adds up to 7 seconds at a time from the previous clip, for a maximum of 20 steps. The more steps you stack, the further any flaw spreads: plan your cuts and restart from an approved frame rather than stretching a shot indefinitely.

A dependable workflow for a controlled shot starts from images, not from the prompt:

  1. Draw or generate the start and end keyframes and get them approved.
  2. Write the prompt like a storyboard line: shot size, camera move, action, light, sound.
  3. Only choose 8 seconds if you need it: shorter clips cost less and cut together better.
  4. Run tests on Fast or at 720p, then regenerate the chosen take on Standard at the final resolution.
  5. Extend only if the action has to carry on without a cut.

Where to use Veo 3.1: Flow, Gemini, Firefly and other platforms

Veo’s home is Google Flow, the tool unveiled at Google I/O on 20 May 2025 and renamed Google Flow at I/O on 19 May 2026. According to Wikipedia, users had generated more than 100 million videos there by August 2025. Flow is built for longer projects with continuity between scenes; the Gemini app for short, quick clips. Veo is paid for through several subscription tiers and Google’s AI credits.

Outside Google, Adobe’s partner models page, updated on 23 October 2026, lists Veo 3.1 in Firefly’s text-to-video and image-to-video, in Boards, in the Firefly video editor and even in Lightroom. In Firefly, according to Adobe’s guide, the resolutions are 720p and 1080p. Higgsfield offers Standard and Fast, also at 720p and 1080p. For 4K you go to fal, through the API. Kult Studios is a Higgsfield and Adobe partner; bear that in mind.

One detail worth noting: fal states that its Veo 3.1 endpoint will be deprecated on 17 November 2026, without naming a replacement. If you build an automated pipeline on a provider, check end-of-service dates before you commit to a delivery.

Limits, rules and SynthID

The most obvious limit is length: 8 seconds per generation. Veo 3.1 is designed for shots, not whole scenes, and the script should treat it that way. Then there are the filters: according to fal, the model blocks graphic violence, explicit content, content that promotes harm and depictions of real people and celebrities. A real-life spokesperson is not something you generate with Veo; you shoot them, or work with written consent through a different pipeline.

Every video carries SynthID, Google’s invisible watermark: according to fal it survives most re-encoding and cannot be disabled. In Firefly, Content Credentials are added on top, and according to Adobe they also record which model was used. We explain how they work in our piece on Content Credentials and C2PA; the transparency obligations are covered in our guide to the AI Act.

Filters alone, though, do not guarantee responsible use. In July 2025 Media Matters found racist and antisemitic videos made with Veo 3 on TikTok, as Wikipedia records. Content review is still human work.

Veo 3.1 or Gemini Omni?

At Google I/O on 19 May 2026 Google introduced Gemini Omni, a multimodal model that generates video from any kind of input. On fal the version available in November 2026 is Gemini Omni Flash 1.2, with dedicated functions for editing, extending and upscaling existing video. In Firefly Boards it takes one video plus three images, or up to four images.

For now the division of labour is clear. Veo 3.1 remains the engine for planned shots, with keyframes, references and audio. Omni makes sense when you start from footage that already exists and want to change it one instruction at a time.

Veo 3.1 works when the scene is already decided

Veo 3.1 performs best with an approved storyboard, fixed keyframes and a shot list of 4 to 8-second shots. Without them, you end up generating dozens of attempts to stumble on what should have been decided beforehand.

It is how we approach our brand films: the right engine for every shot, client approval at every step, usage rights written down before anything is generated.

Frequently asked questions

How long can a Veo 3.1 video be?

Each generation lasts 4, 6 or 8 seconds. With extension you can add up to 7 seconds at a time, for a maximum of 20 steps and roughly 148 seconds in total.

Where can I use Veo 3.1 outside Google?

As of November 2026 Veo 3.1 is available in Adobe Firefly, including Boards and the video editor, on Higgsfield and through API providers such as fal. Resolutions differ: Firefly and Higgsfield offer 720p and 1080p, while fal also offers 4K.

Can you remove the SynthID watermark from Veo 3.1?

No. SynthID is an invisible watermark applied to every Veo 3.1 video and, according to fal’s documentation, it cannot be disabled. Stripping it with third-party tools is a poor idea for any brand that needs to be open about its use of AI.

Sources

  1. Veo 3.1 | Text to Video, Image to Video, Reference to Video, Extend Video · fal, 2026-11
  2. Veo 3.1 on Higgsfield · Higgsfield, 2026-11
  3. AI Video Generator: Kling, Veo, Seedance and more · Higgsfield, 2026-11
  4. Partner models in Adobe products · Adobe Help Center, 2026-10-23
  5. Generate videos using partner models · Adobe Help Center, 2026-08-07
  6. Partner models to generate videos (Firefly Boards) · Adobe Help Center, 2026-08-18
  7. Adobe Firefly expands video and image creation with new AI capabilities and custom models · Adobe, 2026-03-19
  8. Kling 3.0 & Omni AI Video Generator now in Adobe Firefly (Content Credentials FAQ) · Adobe, 2026
  9. Veo (text-to-video model) · Wikipedia, 2026-11
  10. Google Flow · Wikipedia, 2026-11
  11. Google Gemini · Wikipedia, 2026-11
  12. Google models on fal (Gemini Omni Flash 1.2) · fal, 2026-11
  13. Gemini Omni on fal · fal, 2026-11

Got an idea? Let’s make it a film.

We’ll reply with a concrete proposal, built around your brief.