What is deepfake video detection?
Deepfake video detection finds face swaps and other synthetic manipulation in recorded video by scoring it frame by frame and checking that the frames agree with each other over time. Raid AI submits the video as a job, reports progress while it runs, and returns one exportable verdict for the whole sequence.
A face swap that convinces at conversational speed rarely holds up when the footage is taken apart. Identity leaks at the boundaries: the edge of the face, the moments of fast motion, the frames where the model had least to work with.
Raid AI splits video into frames, scores them independently and reconciles the results, so a verdict rests on the whole sequence rather than a lucky still. Temporal inconsistency, the way a synthetic face fails to stay coherent across time, is a signal a single frame cannot give you.
Long files run as background jobs with live progress, because a forensic answer on an hour of footage is worth waiting for and a request timeout is not a useful failure mode.
The attacks this answers
- Face swaps in recorded video
- Temporal inconsistency across frames
- Fully generated video
- Footage reused from another context
From submission to verdict
Submit
Upload a file or hand over a link. The job starts immediately.
Split
The video is broken into frames and each is scored on its own.
Reconcile
Frame results are combined into a verdict for the sequence.
Report
Progress is live throughout, and the finished result is exportable.
At a glance
- Formats
- MP4, MOV, AVI, MKV, WEBM
- Execution
- Asynchronous job with live progress
- Method
- Frame split with majority vote
- Availability
- Enterprise
Questions about Video forensics
Which formats does it take?
MP4, MOV, AVI, MKV and WEBM, uploaded as a file or handed over as a link.
Why is video a job rather than an instant answer?
Because every frame is scored on its own and then the results are reconciled across the sequence. That takes longer than a single image, so the video is submitted, progress is reported live, and the finished result is stored and exportable. The SDKs poll for you.
What does it catch?
Face swaps, which hold up frame by frame but slip between frames, and other temporal inconsistency that a manipulated sequence cannot hide.
Who can use it?
Video forensics is on enterprise plans. Voice and image are available to every account.