How to blur faces for privacy before publishing
- Step 1Confirm your tier and drop the source — Face blur is Pro + Media (£19/month) and up — it needs every frame for detection and can't stream. Drop one MP4, MOV, MKV, WebM, AVI, M4V, or TS file (up to 100 GB on Pro + Media). The page reads duration and resolution locally; nothing is sent anywhere.
- Step 2Choose a sample rate that matches the scene — For a mostly static interview or panel, the default 4 Hz is fine. For footage with people moving through frame, raise toward 8–12 Hz so brief appearances are caught and blur windows stay tight. Range is 1–15 Hz; higher means a longer detection stage but fewer missed moments.
- Step 3Set a privacy-appropriate blur strength — For everyday editorial publishing, strength 25 is the default and reads naturally. For sensitive subjects (vulnerable people, whistleblowers, minors), push toward 45–50 so the result is clearly de-identified rather than merely softened. The radius is auto-clamped so small faces never reject the value.
- Step 4Pad generously for full coverage — Privacy-grade redaction usually wants padding above the 0.25 default — try 0.35–0.5 so hairlines, ears, and chins are inside the blur. Padding grows by a fraction of each face's own size, so larger near-camera faces get proportionally more margin automatically.
- Step 5Run and watch the stages — You'll see a one-time 'Loading face detector' warm-up, then 'Detecting faces · sample N/total', then 'Blurring faces · N faces · t/total'. The output is H.264 MP4 with the audio copied unchanged.
- Step 6Review every face before you hit publish — Scrub the full result. Pay attention to side profiles, fast pans, low light, and faces far from camera — the spots the detector is weakest. Patch any gap with video-redactor (manual rectangle) and strip location metadata with metadata-scrubber so the file you publish carries no GPS either.
Privacy settings cheat-sheet
Suggested control values by publishing context. All within the tool's real ranges.
| Context | Sample rate | Blur strength | Padding |
|---|---|---|---|
| Corporate / marketing video, static subjects | 4 Hz | 20–25 | 0.25 |
| Public-event footage with passers-by | 8–12 Hz | 30–40 | 0.35 |
| Sensitive subjects (vulnerable / minors) | 10–15 Hz | 45–50 | 0.45 |
| Quick social clip, one near-camera person | 4 Hz | 25 | 0.3 |
What blur does and doesn't guarantee
Be honest with yourself about what a box blur achieves so your publishing decision is sound.
| Concern | Reality |
|---|---|
| Casual recognition by a viewer | Strong box blur (35+) prevents this in normal viewing |
| Algorithmic reversal of a weak blur | Low-strength blur can be partly recovered; raise strength or use a mosaic |
| A face the detector missed | Stays fully visible — manual review is non-negotiable |
| Identifying details outside the face | Tattoos, badges, name tags, plates are not touched — use video-redactor |
| Location leakage in the file | Blur doesn't strip GPS/metadata — run metadata-scrubber |
Tier access
Face blur requires Pro + Media (full-frame, non-streamable).
| Tier | Access | Max file | Batch |
|---|---|---|---|
| Free / Pro | Blocked (upgrade prompt) | — | — |
| Pro + Media | Full access | 100 GB | 50 |
| Developer | Full access | 100 GB | Unlimited |
Cookbook
Privacy-first workflows using only the controls the tool exposes. Pair with the sibling tools where the blur alone isn't enough.
Conference panel published to YouTube
Speakers consented; the front-row audience caught on camera did not. Blur every face, then transcode if needed.
Input: panel-recording.mov (1080p, 41m) Options: sampleHz 6 · strength 30 · padding 0.35 Output: panel-recording (face-blurred).mp4 (H.264, audio copied) Follow-up (optional): metadata-scrubber to drop any camera/location tags before the upload.
Whistleblower interview — strong de-identification
High-risk subject. Maximise strength and padding; review frame by frame; consider the mosaic as a stronger alternative.
Options: sampleHz 12 · strength 50 · padding 0.5
If the subject ever turns side-on (detector weak there):
-> video-redactor: draw a rectangle over the head for
that time range as a guaranteed cover.
If a blur still feels reversible:
-> face-pixelate: a mosaic reads as deliberate redaction
and destroys more detail.Street vox-pop with named consent
Your interviewee consented; the bystanders behind them didn't. Blur all, then it's safe — but check the detector didn't miss a distant face.
Options: sampleHz 8 · strength 35 · padding 0.3 Review: scrub the background; any tiny far face the short-range model skipped gets a manual redactor pass.
Audio kept intact for an accessibility caption pass
Because audio is stream-copied, you can blur first and caption later without quality loss in the sound.
Step 1: face-blur -> MP4 with identical audio bytes
Step 2: subtitle-burner or your captioning tool on the
blurred MP4 -> publish-ready, faces hidden, captions onRemoving location data the blur doesn't touch
A blurred face doesn't help if the file still carries GPS coordinates from the phone that shot it. Chain metadata-scrubber.
face-blur (faces hidden) -> metadata-scrubber (GPS, camera make/model, date stripped) -> publish The blur handles the visual PII; the scrubber handles the embedded file PII. Both run locally, no upload.
Edge cases and what actually happens
A missed face goes live
Review requiredThe single biggest privacy risk is trusting detection blindly. Side profiles, small/distant faces, heavy backlight, and partial occlusion all reduce recall. The tool blurs only where it detected — undetected stretches are fully visible. Manual review at 1× is mandatory before publishing; patch gaps with video-redactor.
Blur is recoverable on a low strength setting
Privacy riskA weak blur can be partially reversed with image processing in some cases. For anything sensitive, use a high strength (45–50) or the mosaic in face-pixelate. Standard editorial publishing at strength 25–35 is industry-typical, but match the setting to the risk.
Non-face identifiers still visible
Out of scopeFace blur only blurs faces. Name badges, tattoos, license plates, ID cards, and on-screen text are not detected. Use video-redactor for those rectangles and audio-mute-region if someone's name is spoken aloud.
File still contains GPS / camera metadata
Out of scopeThe blur re-encodes video but doesn't guarantee a clean metadata block for privacy. Run metadata-scrubber afterward to strip GPS coordinates, camera make/model, and capture date before sharing footage that shows where it was filmed.
No faces detected at all
ErrorIf detection returns nothing across all samples, you'll see 'No faces detected. Try lowering the sample rate or use a clearer source.' Usually means faces are too small or angles too extreme for the short-range model. Manual redaction is the fallback.
Output container differs from source
By designWhatever you drop in, the published file is H.264 MP4 — there's no container/codec choice. That's fine for most platforms; if your CMS needs WebM or MOV, run video-transcoder on the blurred result.
Detection is slow on a CPU-only browser
SupportedWithout WebGPU or WebGL, detection falls back to CPU and runs slower. It still works and still stays local. Lower the sample rate to reduce the number of inferences if speed matters more than catching brief appearances.
Crowd exceeds the twelve-track cap
By designOnly the twelve most-detected face clusters are blurred. In a dense crowd, low-hit background faces may be skipped. For full-crowd anonymisation, raise the sample rate so more faces accumulate hits, and review the result.
Frequently asked questions
Is local face blur enough for GDPR?
Processing the footage locally means the redaction step doesn't send personal data to a third-party processor, which simplifies your GDPR position for that step. But GDPR judges anonymisation by whether re-identification is reasonably possible — so use a strong blur (or mosaic) for sensitive subjects, review for missed faces, and remember you still control storage and publishing of the source.
Does the footage leave my computer?
No. The MediaPipe detector (TensorFlow.js) and the FFmpeg.wasm blur both run in your browser. No frame is uploaded. Only an anonymous run counter is recorded server-side for dashboard stats.
Can a blur be reversed?
A low-strength blur can sometimes be partially recovered. Raise the strength toward 50, or use the mosaic in face-pixelate, which destroys more detail and reads as deliberate redaction. For everyday editorial use a strong box blur is the industry norm.
Will it catch every face automatically?
No tool catches 100%. The short-range model is strong on near-camera frontal/3-4 faces and weaker on profiles, tiny/distant faces, and heavy occlusion. Always review the export and patch gaps with manual redaction before publishing.
What about names spoken in the audio or badges on screen?
Face blur only touches faces. Mute spoken names with audio-mute-region, and blur badges, plates, or on-screen text with video-redactor.
Does it strip GPS/location data from the file?
Not as a guarantee — its job is the visual blur. Run metadata-scrubber afterward to remove GPS coordinates, camera info, and capture date before you share footage that reveals where it was shot.
What strength should I use for sensitive footage?
45–50, with padding around 0.45 and a higher sample rate (10–15 Hz) so brief glimpses are also covered. For lower-risk content, the defaults (25 / 0.25 / 4 Hz) are appropriate.
Is the audio affected?
No. Audio is stream-copied with -c:a copy — byte-identical to the source. Only the video is re-encoded to bake in the blur.
What format do I get to publish?
An H.264 MP4 with +faststart, ready for direct upload to most platforms and CMSs. For a different container, follow with video-transcoder.
Which tier is required?
Pro + Media (£19/month) or higher. The tool isn't streamable and needs full-frame access, so Free and Pro show an upgrade prompt.
Can I blur faces and keep the original too?
Yes — the tool produces a new MP4 and never modifies your source file. Keep the original for your records (securely) and publish only the blurred copy.
How big a file can I process?
Up to 100 GB per file on Pro + Media (and Developer). Larger files just take longer because detection and the CRF-20 re-encode both run in your browser.
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.