Gemini video SynthID removal

Video generated inside the Gemini app is produced by Veo, Google's video generation model, underneath. The output can carry two distinct signals: a visible mark on the footage, and SynthID, an invisible watermark embedded in the pixel data itself. Conflating the two leads to the wrong expectations about what removal actually involves.

Both are addressed inside our own media workflow. The upload control on this page opens it: the visible mark is edited directly across frames, and SynthID is cleared through our in-house regeneration technique, which produces a new rendering of the video rather than editing the marked pixels.

A visible mark behaves as expected: it is present on screen, and its position across the clip is not necessarily fixed. SynthID operates on an entirely different layer: a sub-perceptual signal extended per frame through the pixel data, engineered specifically to survive cropping, resizing and re-encoding.

Visible mark

On-screen, editable directly, position not guaranteed constant across frames.

SynthID

Sub-perceptual, in-pixel, a Google DeepMind technology used across multiple products. Cleared through regeneration rather than a direct edit.

The operation runs through the upload control on this page.

Step 1

Retain the original, first-generation export rather than a re-share or screen recording. Each additional step degrades quality.

Step 2

Drop the file into the box above. The visible mark is edited, and regeneration produces a new rendering that clears SynthID.

Step 3

Review the result against the original. It is a new render built from your video, not the source file with a mark peeled off.

SynthID's resilience to crop, resize and re-encode is not incidental, it is the design goal. That property is precisely why editing the marked file directly does not work against it, on this network or anywhere else.

Regeneration answers that constraint by not competing with SynthID's durability on the original pixels at all. It produces a new rendering of the video instead, which is how the original file's embedded signal stops applying to the output.

The verification side carries its own honest limitation: only the issuer's own detector can authoritatively confirm SynthID's presence. We make no claim about what that detector, or any other third-party checker, would report on a regenerated file.

The same resilience explains why isolating a short segment from a longer generation does not reliably remove the signal from the original file either. SynthID is embedded frame by frame across the full duration of the video, not applied once at the start, so a clip cut down to a few seconds can still carry a detectable version of it in its unedited form, consistent with the source it was taken from. Regeneration addresses this by rendering new output rather than trimming the marked one.

Any container-level metadata a Gemini video carries sits in the file wrapper, distinct from both the visible mark and SynthID. Extracting a still frame, screenshotting the clip, or re-encoding it typically discards that container metadata, with no effect on SynthID, which resides in the pixel data rather than the wrapper.

The visible mark sits somewhere between the two in durability. Rendered into the frame data rather than the wrapper, it survives the same operations that discard container metadata, yet it remains an ordinary edited element of the image, unlike SynthID, so a direct edit targeting its actual location removes it from a given copy. SynthID has no equivalent location to target, which is the entire reason regeneration exists as a separate operation.

No unconditional guarantee is made. For the visible mark: a genuine editing task with real limits. For SynthID: regeneration is a real, applied capability, not a theoretical one, but no result is claimed to be perfect on every clip, and no specific verdict from any third-party detector, including Google's own, is promised.

A limited set of legitimate situations.

Creators finishing their own output

Clearing a visible mark and SynthID from footage they generated and hold the rights to.

Teams evaluating Gemini's video capability

Reviewing generated clips internally before further use.

Anyone assessing a clip's origin

Working out whether footage came from Gemini or Veo, using what is actually verifiable.