How to anonymise faces in video for court / legal evidence
- Step 1Work on a working copy on Pro + Media — Keep your evidentiary master untouched and load a working copy. Face blur is Pro + Media tier (it needs full-frame access and isn't streamable). Accepted inputs: MP4, MOV, MKV, WebM, AVI, M4V, TS, up to 100 GB. Nothing uploads.
- Step 2Set a high sample rate for evidence — For legal redaction, favour thoroughness over speed: use 10–15 Hz so brief appearances and fast movements are captured and blur windows are tight. The default is 4 Hz; the range is 1–15.
- Step 3Use maximum-confidence strength and padding — Set strength near 50 and padding near 0.45–0.5 so the redaction is unambiguous and covers hairline, ears, and chin. Both stay within the tool's clamps, so even small faces are handled.
- Step 4Run the detect + blur pass — JAD warms the detector once, samples faces across the timeline, then runs the single-pass encode to H.264 MP4. The progress dashboard labels each stage so you can note what the tool reported.
- Step 5Verify every frame, then patch gaps — Step through the redacted copy and confirm each face stays covered for its entire on-screen time. Wherever the AI missed (profile, distance, occlusion), draw a manual rectangle with video-redactor over that exact time range. This makes the redaction defensible rather than 'best-effort'.
- Step 6Strip metadata and prepare the disclosure copy — Run metadata-scrubber so the disclosed file carries no incidental GPS/device data, and keep your unredacted master archived separately. Mute any spoken identifying details with audio-mute-region if the audio is in scope.
Evidence-grade settings
Conservative values chosen for thoroughness, all within the tool's real ranges.
| Control | Evidence setting | Default | Why |
|---|---|---|---|
| Sample rate | 10–15 Hz | 4 Hz | Catch brief appearances and fast motion; tighten blur windows |
| Blur strength | ~50 | 25 | Unambiguous, harder-to-reverse redaction |
| Padding | 0.45–0.5 | 0.25 | Cover hairline/ears/chin so no identifying edge shows |
Defensible-redaction workflow
Why each step matters for a disclosure bundle.
| Step | Tool | Purpose |
|---|---|---|
| 1 · Work on a copy | — | Preserve the unredacted master / chain of custody |
| 2 · Auto-blur faces | face-blur | Cover the faces the AI reliably finds |
| 3 · Verify frame-by-frame | manual review | Confirm coverage; identify gaps |
| 4 · Patch gaps | video-redactor | Guaranteed cover for missed faces |
| 5 · Mute spoken names | audio-mute-region | Redact identifying audio if in scope |
| 6 · Scrub metadata | metadata-scrubber | Remove incidental GPS/device data from the disclosed file |
Tier access
Face blur requires Pro + Media; it is not streamable.
| Tier | Access | Max file | Batch |
|---|---|---|---|
| Free / Pro | Blocked | — | — |
| Pro + Media | Full | 100 GB | 50 |
| Developer | Full | 100 GB | Unlimited |
Cookbook
Redaction recipes for evidentiary footage. The throughline: auto-blur first, verify everything, guarantee the gaps manually.
CCTV exhibit — blur uninvolved parties
A camera captured the incident plus uninvolved members of the public. Auto-blur, then verify and patch.
Input: exhibit-cam3.mkv (the working copy)
Options: sampleHz 12 · strength 50 · padding 0.5
Verify: step through; a person at the far end of the
corridor (small/distant) was missed.
Patch: video-redactor over that figure's time range.
Result: defensible redaction, master untouched.Witness statement video — protect the witness
Single near-camera subject who must be anonymised. Strong settings; confirm continuous coverage even when they turn their head.
Options: sampleHz 15 · strength 50 · padding 0.45
Watch: the moment they look down at notes (chin-tuck)
and any side turn — verify the blur holds.
If it drops for a frame: raise sampleHz already maxed,
so patch that instant with video-redactor.Bodycam disclosure — many transient faces
A bodycam clip can exceed the twelve-track cap as officers and members of the public move through. Auto-handles the most-detected; verify the rest.
Console may show: capped to 12 most-detected Approach: blur covers the 12 highest-hit faces; for the background figures with few hits, redact manually per time range. Document which were auto vs manual.
Preserve evidentiary audio
The spoken record is evidence. Because audio is copied unchanged, the blur doesn't touch it.
face-blur -> video re-encoded (blur baked in)
-> audio: -c:a copy (identical bytes)
If a name must be redacted from the audio too:
-> audio-mute-region over that timestamp.Clean disclosure file
Strip incidental metadata so the disclosed copy reveals only what's intended.
Master (unredacted) -> archive securely, do not disclose
Working copy -> face-blur -> verify -> patch
-> metadata-scrubber (GPS/device/date)
-> disclosed exhibitEdge cases and what actually happens
AI misses a face in the exhibit
Verify requiredDetection is not exhaustive — profiles, small/distant faces, and occlusion can be skipped, and the tool blurs only where it detected. For legal work this is unacceptable if left unchecked: verify every frame and patch gaps with video-redactor. Document which faces were auto-blurred and which were redacted manually.
Blur strength too low for a defensible redaction
RiskA weak blur can be partially reversed. For evidence use the maximum strength (50) and high padding, or use the mosaic in face-pixelate for a harder, clearly-deliberate redaction that destroys more detail.
Twelve-track cap in a busy scene
By designOnly the twelve most-detected face clusters are auto-blurred. In a crowded exhibit, low-hit background faces may be skipped. Raise the sample rate so more faces accumulate hits, then verify and manually redact any remaining figures.
No faces detected
ErrorIf the model finds nothing across all samples, you'll get 'No faces detected. Try lowering the sample rate or use a clearer source.' For poor-quality CCTV this is common; fall back to manual redaction with video-redactor.
Container changes to MP4
By designOutput is always H.264 MP4 — there is no codec/container option. If the court requires the original container, transcode the redacted MP4 with video-transcoder and note the conversion.
Identifying details beyond faces
Out of scopeFaces only. Plates, ID cards, tattoos, screen content, and on-screen text are not detected. Use video-redactor for those, and audio-mute-region for spoken identifiers.
Disclosed file still carries device/location metadata
Out of scopeThe blur re-encodes video but doesn't guarantee a sanitised metadata block. Run metadata-scrubber on the disclosure copy to remove GPS, camera make/model, and capture date.
Long exhibit takes significant time
ExpectedDetection and the CRF-20 re-encode run in your browser and scale with duration and resolution. A multi-hour exhibit is allowed (up to 100 GB) but takes real time. This is the trade-off for keeping the footage in your custody.
Frequently asked questions
Does processing footage locally help with chain of custody?
Yes. Because detection and blur run in your browser, the evidentiary footage never passes through a third-party server during redaction, so you don't hand custody to an outside processor for that step. You still control storage and handling of the master and the disclosed copy.
Can I describe exactly what the tool did to the exhibit?
Yes. It samples faces at the rate you set, groups them by overlap into regions, and applies an FFmpeg boxblur to those regions in a single re-encode to H.264 MP4 with the audio copied unchanged. The behaviour is deterministic for a given input and settings.
Is auto-detection reliable enough for legal redaction on its own?
No tool's detection is exhaustive — profiles, distant faces, and occlusion can be missed. The defensible approach is auto-blur, then verify every frame, then guarantee any gaps with manual redaction via video-redactor.
What settings should I use for evidence?
High sample rate (10–15 Hz), maximum strength (~50), and high padding (0.45–0.5) for thorough, unambiguous coverage. Then verify and patch.
Will the original file be altered?
No. The tool outputs a new MP4 and never modifies your source. Keep the unredacted master archived; disclose only the redacted, verified copy.
Is a blur reversible?
A low-strength blur can be partially recovered. Use the maximum strength, or the mosaic in face-pixelate, for a harder redaction in high-stakes disclosure.
How do I handle spoken names and on-screen text?
Mute spoken identifiers with audio-mute-region and blur on-screen text, badges, or plates with video-redactor. Face blur covers faces only.
Should I strip metadata from the disclosure copy?
Usually yes — run metadata-scrubber to remove GPS, camera, and date tags so the disclosed file reveals only the intended content. Keep that decision documented.
What output format do I get?
Always H.264 MP4 with +faststart. There's no container choice; transcode afterward with video-transcoder if a specific format is required.
Which tier do I need?
Pro + Media (£19/month) or higher. The tool needs full-frame access and isn't streamable, so it's blocked on Free and Pro.
Can I batch-redact multiple exhibits?
Yes — Pro + Media allows up to 50 files per job (Developer unlimited). Each file is detected and blurred independently. Verify each output individually.
What if the AI groups two close faces into one region?
Overlap-based clustering can merge faces that are very close in frame into a single union box. That's safe for redaction (both stay covered) but blurs a larger area. If you need tighter, separate coverage, reduce padding or redact manually.
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.