Mixed real and AI footage: what Gemini's SynthID check reports
A video that mixes camera footage and Gemini Omni clips shows SynthID in some segments only. What Google says it reports, and a test edit on Sume.

If you cut Gemini Omni clips together with real footage, Google's Gemini verification reports which segments carry SynthID rather than labeling the whole file. Google's post gives audio at a specific second range as an example of the output, so a mixed edit should show marks only where the generated clips sit.
Facts come from Google's Gemini app verification post, the Gemini API Omni docs and DeepMind's Omni page, read on 2026-10-03. Google says all Omni videos carry SynthID; DeepMind's page adds C2PA metadata.
Why this matters for editors
A newsroom or brand team may need to say which parts of a piece are generated. Segment reporting supports that. It also means the order and timing of your edit are your own record: keep a note of which seconds are generated, so you can compare to what the check reports.
| What to log | Why |
|---|---|
| Seconds where Omni clips sit | Compare against reported segments |
| Whether audio is native Omni audio | Marks are scanned in audio and visual tracks separately |
| Trim and re-encode steps | Re-encoding may change a mark's strength |
| Clip length under 90 seconds and 100 MB | Gemini's upload limits |
Build a test edit on Sume
Timeline 1.0 joins ordered clips on one audio spine and returns a single MP4. Put a generated 8 second Omni clip between two real clips, render, and upload the result to Gemini. The timeline docs price a render at $0.10 per output minute rounded up, so a one minute test costs $0.10. Slot start times are explicit, which gives you the seconds log for free.
curl -X POST https://api.sume.com/v1/timeline-1.0/render \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: mixed-synthid-test-001" \
-d '{
"audio": { "mode": "silence", "duration_seconds": 24 },
"video": [
{ "source_url": "https://media.sume.com/artifacts/artf_demo/real-a.mp4", "start": 0, "duration": 8 },
{ "source_url": "https://media.sume.com/artifacts/artf_demo/omni.mp4", "start": 8, "duration": 8 },
{ "source_url": "https://media.sume.com/artifacts/artf_demo/real-b.mp4", "start": 16, "duration": 8 }
]
}'Read the result
Expect the check to point at the middle window. If it also flags the outer windows or misses the middle, record that and try a version with the original Omni audio. Do not publish a claim about detection reliability from one run; the Timeline docs cover the render options.
Sources
Related posts
More in Use cases
- Mobile game ad cutdowns: one master to 6, 15, 30 and 60 seconds
Mintegral recommends 6/15/30/60 s, Unity and Axon cap at 60 s, Liftoff at 180 s. Cut one portrait master to each length with Sume's trim API for $0.02 a job.
- Move an object in a photo with AI: erase, then place with two masks
No Sume image model has a move operation. Move an object by erasing it with one mask and placing it with a second, using GPT Image 2.5 and Python.
- Movember 2026 team update videos from phone photos, one per week
Movember asks people to talk about men's health. Make four weekly update clips for a fundraising team page from phone photos, with caption cues you wrote.
- Movie and TV clip Shorts: what YouTube's help calls original
Unedited movie or TV clips are ineligible on Shorts. What YouTube's pages allow instead, and how to cut trailer-style Shorts from your own film or series.
Written by Sume