How to mosaic faces for news and journalism footage
- Step 1Pull the footage into the browser — Drop the MP4/MOV onto the tool — it stays local. Face pixelate requires the Pro-Media tier. For sensitive material, this no-upload pipeline avoids handing footage to any third party.
- Step 2Decide: auto-detect all, or manual for one source — If you want every face hidden (e.g. a protest crowd), use this auto tool. If you must perfectly conceal ONE interviewee who moves/turns, the video redactor gives a manual region you fully control.
- Step 3Set a strong pixel size for the source —
Pixel size24–40 for a concealed source so no facial structure survives. The control clamps to 4–40. - Step 4Raise sample rate if the subject moves —
Sample rate (Hz), default 4 — raise toward 8–12 for an animated interviewee or handheld camera so the track stays complete across head turns. - Step 5Pad generously for turning heads —
Padding (0–1), default 0.25 — use 0.35–0.4 so the mosaic still covers the face as the person turns toward profile within the track's union box. - Step 6Verify frame-by-frame, then protect the voice — Scrub the whole clip — a single uncovered frame breaks source protection. Then mute identifying audio with the audio region muter and strip file metadata with the metadata scrubber.
Auto-detect vs. manual for source protection
Pick the right tool for the editorial need. Both run locally.
| Need | Use | Why |
|---|---|---|
| Hide all bystanders in B-roll/crowd | Face pixelate (this tool) | Fast auto-detection of every near-camera face |
| Perfectly conceal one moving interviewee | video redactor | Manual region you control frame-to-frame; no detection gaps |
| Conceal a face the model keeps missing (profile) | video redactor | Auto-detection skips steep profiles |
| Protect the source's voice | audio region muter | Pixelation copies audio unchanged |
| Sanitise the file before sharing with an editor | metadata scrubber | Strips GPS/device/date the camera embedded |
Editorial settings
Manual starting values — the UI has no presets.
| Footage | Sample rate | Pixel size | Padding |
|---|---|---|---|
| Seated protected-source interview | 6 | 32 | 0.4 |
| Handheld vox-pop, animated subject | 10 | 28 | 0.4 |
| Protest / crowd B-roll, hide everyone | 12 | 24 | 0.3 |
| Vulnerable bystander in the background | 8 | 30 | 0.35 |
Cookbook
Editorial workflows for source and bystander protection. The mosaic is one layer; voice and metadata are the others.
Protected-source interview
A whistleblower agrees to be filmed only if unidentifiable. For a single, high-stakes face that turns and gestures, manual redaction is safer — but if the framing is stable, auto with strong settings works.
If face is stable: face-pixelate (Pixel 32, Padding 0.4, Sample 6) If face moves/turns a lot: /video-tools/video-redactor (manual) Always: /video-tools/audio-mute-region or NLE voice change Verify every single frame before export
Hide bystanders in protest B-roll
A crowd scene where you must conceal everyone who didn't consent. Auto-detection shines here — raise the sample rate for movement.
Input: protest_broll.mp4 Settings: Sample rate 12, Pixel size 24, Padding 0.3 Result: near-camera faces mosaiced (up to 12 tracks) Check: distant faces may be missed → redact manually if identifiable
Full source-protection pass
Face, voice, and metadata. Three local steps so the source is unidentifiable and the file carries no clues about where/when it was shot.
Step 1: face-pixelate → face concealed Step 2: /video-tools/audio-mute-region → remove name / alter voice region Step 3: /video-tools/metadata-scrubber → strip GPS/device/date All local; nothing uploaded
Source turns to profile mid-sentence
Auto-detection can drop a steep profile, leaving a frame or two exposed — fatal for source protection. For this, manual redaction is the safe choice.
Symptom: 1-2 frames where the profile face is visible
Reason: detector keys on frontal faces
Fix: /video-tools/video-redactor → region that covers the turn
(manual control = no detection gap)Keep reporter track, conceal interviewee
Your piece-to-camera audio and the room ambience must survive. Audio is copied untouched, so only mute the segments that identify the source.
face-pixelate → interviewee face mosaiced, all audio kept Then mute only the identifying lines: /video-tools/audio-mute-region (e.g. where the name is said)
Edge cases and what actually happens
Source turns to profile — frame exposed
May be missedThe detector keys on frontal faces, so a steep profile or a quick head-turn can leave one or two frames uncovered. For source protection a single exposed frame is unacceptable. For a must-be-perfect subject, use the manual video redactor instead of relying on auto-detection.
Voice still identifies the source
Preserved (audio copied)Pixelation conceals the face but audio is stream-copied unchanged — a recognisable voice or a spoken name still identifies the person. Always pair with the audio region muter or an NLE voice treatment for genuine source protection.
More than 12 faces in a crowd scene
Capped at 12 tracksOnly the 12 most-detected tracks are mosaiced. In a large protest or crowd, faces seen briefly may remain visible. Tighten framing, split the shot into segments, or patch with the redactor — and verify before broadcast.
Distant bystander not detected
May be missedThe short-range model targets near-camera faces; a vulnerable person far in the background may not be detected. Scrub the clip and cover anyone identifiable manually — editorial duty of care means you can't assume full auto coverage.
File metadata reveals where it was shot
Not removedPixelation doesn't touch file metadata. GPS/location and device info the camera embedded survive in the output, potentially exposing a filming location or source. Run the metadata scrubber before sharing the file with editors or publishing.
Moving subject gets a wide union box
Covered (path union)Each track is one static rectangle = the union of all positions, active for its time window. An animated interviewee gets a box covering their full range of movement. They're concealed; if the box covers distracting background, reduce Padding or switch to manual redaction for a tighter result.
No faces detected
Error: no facesIf detection returns nothing, you get No faces detected. For footage where a face is clearly present but small or angled, don't treat it as concealed — raise the sample rate or redact manually.
Output re-encoded to MP4
By designOutput is always MP4 (libx264, CRF 20). The whole frame is re-encoded near-losslessly; only face regions change. For broadcast delivery requiring a specific codec/container, transcode the result with the video transcoder.
Frequently asked questions
Is auto-detection safe enough for protecting a source?
For bystanders in B-roll, yes — fast and effective. For a single high-stakes source who must never be identifiable, be cautious: the detector keys on frontal faces and can drop a steep profile or fast head-turn for a frame or two, which would break protection. For that case, prefer the manual video redactor and verify every frame.
Does mosaicing the face protect the source completely?
No — the voice is identifying too, and this tool copies audio unchanged. Genuine source protection means concealing the face AND treating the voice (mute identifying lines with the audio region muter or pitch-shift in your NLE) AND stripping file metadata with the metadata scrubber that could reveal where/when it was shot.
Why is the mosaic the right look for news?
The blocky mosaic is the broadcast-standard, instantly-recognised signal that a person's identity has been deliberately concealed. A soft blur can read as a technical focus issue. If your house style prefers a soft look, the face blur tool offers the same auto-detection with a Gaussian-style blur instead.
Will unpublished footage be uploaded anywhere?
No. Detection runs in TensorFlow.js and encoding in FFmpeg.wasm, both in your browser tab. The footage never reaches a server — important when a leak could endanger a source. The only network call is an optional usage counter with no file content.
What if the AI misses the interviewee in profile?
It can. The detector is frontal-biased, so a steep profile or quick turn may leave a frame uncovered. For a source that must be perfectly concealed, use the manual video redactor — you draw the region and control it directly, so there are no detection gaps. Always scrub frame-by-frame before export.
Can I keep my reporter narration and ambience?
Yes. Audio is stream-copied (-c:a copy) so your piece-to-camera track and room ambience are preserved exactly. Mute only the specific segments that identify the source — for example where a name is spoken — with the audio region muter.
How many faces can it hide in a crowd?
Up to 12 tracks (the 12 most-detected). For a packed protest or crowd, faces seen only briefly may remain visible. Tighten framing, segment the shot, or patch with the video redactor, and always verify the result before broadcast given the editorial stakes.
How strong should the mosaic be for a concealed source?
Use Pixel size 24–40 and Padding 0.35–0.4 so no facial structure survives and the concealment holds up to scrutiny. The control range is 4–40; for a protected source, err toward the maximum and confirm visually.
Does the output reveal where the footage was shot?
It can if you don't strip metadata. Pixelation only changes face pixels; GPS/location and device info the camera embedded remain in the file. Run the metadata scrubber before sharing with editors or publishing to remove location clues.
What output format do I get for the edit?
An MP4 (H.264, yuv420p, CRF 20, +faststart) with audio copied — drop-in for most NLEs. The whole frame is re-encoded near-losslessly. For a specific broadcast codec/container, transcode with the video transcoder.
Which plan does this need?
Pro-Media or higher; it's gated on Free and Pro. Pro-Media: 100 GB / 50 files; Developer: 100 GB / unlimited. There's no clip-length cap on video tools — only file size and batch count — so full interviews and long B-roll are fine.
Is this a substitute for editorial/legal review?
No. The tool applies a technical anonymisation; whether it satisfies your outlet's source-protection standards and the law is an editorial and legal judgement. Use it as the production step, and run the result past your editor/legal team for high-stakes material.
Privacy first
Every JAD Video tool runs entirely in your browser via WebCodecs and FFmpeg (WebAssembly). Your video files never leave your device — verified by zero outbound network requests during processing.