Loading tool…
Street scenes, school events and protest footage often show people who never agreed to be online. Drop in a batch of photos or one video, let on-device AI find the faces, keep the ones that may stay visible, and cover the rest with blur, pixelation, a solid box or an emoji.
Processing locally on your device
Overview
Detection uses MediaPipe BlazeFace, a short-range face detector from Google released under the Apache-2.0 licence. The model is only about 230 KB and runs in the MediaPipe Tasks WebAssembly runtime inside your browser, on the graphics card when the GPU delegate is available and on the CPU otherwise. BlazeFace looks at a 128 × 128 pixel copy of the image, so on its own it misses faces that fill less than roughly a sixth of the frame. To catch crowds and distant people, the tool can also scan the image in overlapping tiles (2 × 2, 3 × 3 and, for photos, 4 × 4 grids) and merge the results, which finds much smaller faces at the cost of more work.
Every face gets a numbered box, a thumbnail in the faces list and a confidence score. Click a face in the list or on the preview to leave it visible, for example the person giving an interview, and drag on the preview to add a box by hand for anything the detector missed, such as a face in profile or on a screen. The confidence slider (30–90%) hides weak detections without running the model again. Covers can be an ellipse or a rectangle, with 0–80% extra padding to include hair, ears and chin. Photos are saved at their full resolution in the original format (JPG, PNG or WebP), and a batch downloads as one ZIP file.
Video works in two passes. First the tool checks faces on every frame, or on every second or third frame for speed, and a tracker links the detections into one track per person, interpolating the box between checked frames and bridging short gaps where a head turns away. Then each frame of the selected range is rendered with the effect at up to 30 fps and encoded to MP4 (H.264) by FFmpeg running in WebAssembly, with the original soundtrack copied or re-encoded to 192 kbps AAC. This is heavy work for a browser: expect several times the clip length on a typical laptop, so one export is limited to about 90 seconds at 720p (around 40 seconds at 1080p, over 3 minutes at 480p). Longer videos can be exported in parts using the range selector.
Choose the cover with care. A light blur or large mosaic blocks can leave a face recognisable, especially to someone who knows the person. For real anonymisation use blur or pixelation at 60% or more, or a solid box, which cannot be reversed. Always scrub through a video preview before publishing, because no detector finds every face in every frame.
| Style | Looks | Anonymisation | Best for |
|---|---|---|---|
| Blur (60%+) | Soft, natural | Strong | News-style footage, documentaries |
| Pixelate (60%+) | Classic mosaic | Strong | Clear signal that a face was hidden |
| Solid box | Hard-edged block | Complete, irreversible | Legal, safety-critical or sensitive material |
| Emoji | Playful | Complete within the emoji | Family photos, social posts |
Drop several photos, or one MP4, MOV or WebM video, into the tool.
For video, pick the range and how often to check faces, then click Detect faces.
Choose blur, pixelate, box or emoji, and set the strength, padding and shape.
Click any face you want to keep visible, and drag on the preview to cover missed areas.
Download the photo, all photos as a ZIP, or export the blurred video as MP4.
Use cases
Hide bystanders and vulnerable sources in protest photos or interview footage while keeping the speaker visible. Nothing is uploaded, so sensitive material never touches an outside server.
Cover children whose families have not given photo consent before posting event pictures or a match highlight clip on a website or newsletter.
Anonymise pedestrians in road incident clips, city vlogs or real estate walk-throughs before sharing them with insurers, forums or social media.
Strip identifiable faces from usability study recordings or screenshots of camera feeds before passing them to colleagues or vendors.
Pro tips
Answers to common questions about the face blur.
Yes. Every detected face appears in the faces list with a thumbnail. Click it, or click its box on the preview, to leave it uncovered. In a video this applies to the whole track, so the main speaker stays visible from start to finish.
BlazeFace is built for faces seen from the front at a reasonable size. Very small faces, strong profiles, heavy motion blur and faces partly covered by hands or masks can be missed. Turn on Find small faces, lower the confidence, or drag a manual box over the area.
Strong settings are. Light blur and coarse pixelation can sometimes still be recognised, so use blur or pixelation at 60% or more, or a solid box. The tool always writes new pixels into the image; the original face is not stored in the exported file.
Detection is fairly quick, but every output frame is drawn, compressed and then encoded to H.264 in WebAssembly, which is much slower than native software. On a typical laptop a 720p clip takes several times its own length. Lower resolutions are faster.
Yes. The original audio is copied unchanged when it is AAC or MP3 and the export starts at the beginning; otherwise it is re-encoded to AAC at 192 kbps and trimmed to the selected range.
No. ToolEnhancer processes files on your own device inside the browser. They are never sent to, stored on or seen by our servers.
Yes. The face blur is free to use with no sign-up, no watermarks and no daily limits.