Veo Watermark: Two Marks, One Document, and One Switch

“Veo” names a model, not a surface

The first thing the term hides is where the clip came from. Veo is a model family, and the developer documentation lists it as a set of versions with a state attached to each:

• Veo 3.1, Veo 3.1 Fast and Veo 3.1 Lite are listed as previews, with the most recent updates dated January 2026 and March 2026.
• Veo 3 and Veo 3 Fast are listed as deprecated, last updated July 2025.
• Every one of them declares the same output type: video with audio. The frame rate is 24fps, and a request returns one video.
• Duration and resolution are tied together. The documented lengths are 4, 6 and 8 seconds, the default resolution is 720p, and 1080p and 4k are available only at the 8-second length.

None of that tells you which product made a given file, and this is the second thing the term hides. The same model family is reachable through more than one surface, and the help page for video generation in Gemini Apps names a different model for that surface entirely:

You can create videos in minutes in Gemini Apps with Gemini Omni, our latest AI video generation model.

So a person typing “Veo watermark” may be describing a clip that came out of an API call, or a creative tool built on the model, or a chat app whose own help page points at a different model name. The word is a label for a family of models, not for a place where clips are made — and the answer to “can I turn this off” depends on the place, not on the family.

The one watermark the Veo documentation names

The Veo model page carries a short limitations list, and exactly one of its entries concerns watermarking:

Watermarking: Videos created by Veo are watermarked using SynthID, our tool for watermarking and identifying AI-generated content. Videos can be verified using the SynthID verification platform.

That one entry repays a second reading. First, it describes what the model does rather than offering a choice — there is no conditional anywhere in the sentence. Second, the entry names one mark. On a page that lists its limitations item by item, a visible overlay is not among them, and neither are Content Credentials.

That absence is not a gap in the page so much as a division of labour, and it is the reason this term splits the way it does. The invisible mark is a model behaviour, so it is documented where the model is documented. The visible mark is a presentation choice made by whichever product wrapped the model, so it is documented in the help page for that product — which is where the switch lives. If you have been reading a page that treats both as one item called “the Veo watermark,” this is the sentence that separates them: the model's own documentation only ever promised one.

The switch, and how far it reaches

The product documentation states its own boundary outright. That statement is the first thing worth checking, ahead of anything a tool on this term claims. The setting is called Media Watermark, and it is described as controlling visible watermarks “across all visual media you create with Gemini, including images, videos, and music.” Its tip then draws the line:

This setting only controls the visible watermark. It doesn’t affect SynthID watermarks or Content Credentials for media you create with Gemini Apps.

For video specifically, the same page describes the visible layer as a mechanism rather than a decoration:

Video: Visible watermarks appear as a persistent overlay on generated video content.

Persistent overlay is a statement about how the mark is applied, and it is meant exactly. Something applied that way runs from the first frame to the last: no moment of the clip escapes it, and nothing about the mark lapses partway through. That is why the choice concerns the file rather than one small region of one frame.

How far the switch reaches is stated twice on the same help page, and it is the detail most write-ups on this term leave out. The settings, Google writes, “apply only to Gemini Apps (web and mobile), not AI features in other Google products.” There is also a regional carve-out: in India, South Korea and Vietnam the setting is visible only on the top subscription tier, and everyone else there has the visible mark applied automatically. Both lines matter for a term built on a model name, because a clip produced somewhere other than Gemini Apps is not governed by this switch at all, however it was generated.

A short clip changes the arithmetic of the official check

This is the part of the subject where a model name actually buys you something. For an invisible mark there is a check a user can run — the verification tool inside Gemini, which covers images, videos and audio and requires a signed-in account. Three limits shape what you can put in front of it: a single file per check, a ceiling of 100 MB, and a video shorter than 90 seconds. The check is also rationed rather than open-ended:

There’s a usage quota of approximately 10 image checks, 10 video checks (up to 5 minutes total), and 10 audio checks (up to 3 hours total) within any 24-hour rolling window.

Now put the model's numbers next to those. A single Veo generation is 4, 6 or 8 seconds. At the longest of those, ten checks come to 80 seconds of video — comfortably inside the five-minute allowance, which means that for an ordinary Veo clip the limit you will actually hit is the count of ten, not the total duration. That is the opposite of the general video case, where a handful of ordinary-length clips exhausts the minutes long before the tenth check. The duration cap and the quota interact differently depending on what you are checking, and this is arithmetic from the two published figures rather than a measurement of our own.

One case escapes that arithmetic, and the number behind it sits on the same model page. Veo clips can be extended, and the documentation gives the extended output a maximum of 148 seconds. That is past the tool's 90-second ceiling, so an extended clip cannot be verified as a single file — it has to be cut down first, which means the thing you check is no longer the thing you have. For an 8-second generation the check is easy; for an extended one it is not available in the form the file takes.

The answer that comes back is not a verdict, and it is worth reading line by line:

• If a SynthID watermark is detected, Google says this means all or part of the image or video was created or edited by Google's AI models.
• If one is not detected, the file wasn’t created or edited by Google AI, but it could have been created by other AI systems.
• Some files come back unclear, for stated reasons such as Not enough details to watermark or Likely too small an edit.
• For video, the result is reported in segments: the page says Gemini will state which parts of the video a SynthID watermark was detected, and that this doesn’t mean it was detected in all parts of that portion of the video.
• And the check can currently only recognize content created by Google AI tools, even though other companies have begun adopting SynthID.

The segment line is the one that changes what a clean result means. What comes back describes the samples rather than the clip, and a hit inside one segment says nothing about the rest of the file.

Which half you have, and what each one actually needs

It is worth naming which half you are holding before you choose anything, because the two share a file and nothing else:

• The clip does not exist yet, and you want the visible mark never applied. That belongs to the product layer, which is the only layer a setting has ever governed — and the setting sits earlier in the process than any editor on this term does. It is free, and it is lossless. Where it lives and who can reach it is written up in the official setting.
• The clip exists and carries a visible mark. What you are choosing between is a framing cost and a repair. Cropping or covering is honest about what it spends — part of the picture, or a patch over it. A repair across frames needs a tool that says it does that, and the questions worth asking it are in the general case for clips.
• The clip looks clean, but something else has flagged it as AI-made. That is the model layer — the one the Veo documentation does name, and the one no control anywhere has ever offered to undo. How the mark works, and where our own tool fell short when we tested it, are on whether SynthID can be removed.
• Both are true of the same file. The usual case, and the two are independent — taking the badge off does not move the detection by so much as a step. Both are quoted side by side, without being blended, in what Google's rules and forum say.

One more practical note sits under all four. The documented output of these models is video with audio, so anything you do to a clip is a re-encode of a file that also carries a soundtrack. The picture is the part everyone discusses; the sound is one more thing that has to come through the operation unchanged.

Why this tool stops at a single frame

Our own boundary is drawn where it can be held. This tool works on one still frame at a time — PNG, JPEG or WebP, up to 60 megapixels — and it never takes a clip as input, never reads an audio track, and never implies it has done either.

The reason is structural, and it follows from what the repair actually is. The mark here is a known overlay, so the pixels beneath it can be recovered by running the compositing step in reverse: arithmetic on a template, not a guess. That works on the layer as the generator emitted it, where the blend happened once and without a codec in between. What a delivered clip hands you instead is a set of frames a codec rebuilt from neighbouring frames plus a lossy correction — an input that has already been through a transformation before you begin. That sentence is our reasoning rather than a measurement. The second reason is about consistency rather than arithmetic: one frame is judged on its own, while a clip is judged on how consecutive frames differ, and a repair that drifts by a hair between them shows up not as a slightly softer patch but as movement where there should be none.

No test of this tool against a video pipeline exists, so no page here will suggest otherwise. The routing for the two halves of this term is the one set out above: switch the visible mark off before the clip is generated if you still can; take a clip you already have to a tool that states it repairs across frames; and run the official check on whatever comes back — which, at the lengths these models produce, costs less to run than it would on almost any other clip you might be editing.