SynthID on every Gemini Omni Flash video: check it after you edit

Google says SynthID is applied to all Gemini Omni Flash output. A short plan for confirming the mark and recording it before and after a trim or re-encode.

4 min readSume
All posts

Google's Gemini Omni docs state that SynthID is applied to all Omni Flash video output. The page read here does not say whether the mark survives a trim, a re-encode, or a platform upload, so the safe plan is to test the file you actually deliver.

What the page states

The same page lists 3 to 10 second generations, up to 40 seconds total when extended, and 16:9 and 9:16 output. SynthID is one line in that list, and it describes what Google adds at generation, not what remains after your editing chain.

Where editing enters

Sume's video trim endpoint produces a new MP4. Its default precision is exact, a frame-accurate re-encode with libx264, while keyframe uses a stream copy. The docs describe the container and codec behavior; they say nothing about watermarks, so do not assume either mode preserves one.

Pipeline steps that can change a file after generation
StepWhat changes in the fileCheck
Download from SumeSume-hosted copy of the provider outputStore checksum and job id
Trim, exact precisionFrame-accurate re-encodeTest the trimmed file
Trim, keyframe precisionStream copy; cut may start earlyTest the trimmed file
Platform uploadPlatform may re-encodeTest what the platform serves

A test plan

Treat detection as a measured step, with one result per stage. Use whichever detector Google or your compliance team names, and record the outcome rather than assuming it.

  • Keep the original download and its checksum.
  • Run the detector on the original, then on each edited version.
  • Write the result next to the job id and model id.
  • Re-test after any change to the editing chain, including codec settings.

Disclosure is a separate question

A watermark that machines can detect is not the same as a label viewers can see. Platforms and regulators ask about disclosure of AI-generated content in different ways, and a hidden mark may not satisfy a rule that expects a visible label.

Decide separately what your channel requires: a visible label, a platform toggle, or a metadata flag. Then record the choice per clip, next to the detection result.

  • Hidden mark: a provenance signal, not a label.
  • Visible label: decided by channel rules.
  • Platform toggle: set at upload.
  • Record all three per clip.

Related checks

The stored posts on watermarks after re-encode and credentials after download cover the related checks for provenance data. The video docs show which fields to save per job.

Sources

Related posts

More in Use cases

All Use cases posts

Written by Sume