Description
Street footage, vlogs and meeting recordings often need bystanders' and colleagues' faces hidden before they go out. Doing it by hand in an editor means dragging a mosaic box along frame by frame, and the moment someone turns their head the face slips out. video-face-anonymizer is a Skill for AI agents: hand it a video and it finds every face and follows it through the whole clip, covering it with mosaic, blur or a solid black box while keeping the audio intact.
It does more than per-frame detection. Face boxes are smoothed over time so the mask doesn't jitter when the camera shakes, and when a face is missed for a few frames the mask stays put for a while, so a quick turn or a glance down doesn't expose anyone. The output is always an H.264 + yuv420p MP4 that plays in Telegram and common players.
Two detectors: InsightFace SCRFD by default, which handles distant small faces, side profiles, low light and crowds better but is slower on CPU; switch to YuNet for well-lit close and medium shots when speed matters.
Three masking styles:
Smoothed tracking: Boxes are matched across frames by IoU and blended over time;
Hold on misses:
Tunable detection: Raise
Audio and compatibility: FFmpeg muxes the original audio back in and encodes H.264, yuv420p with faststart so the file streams in a browser.
Built-in validation: The Skill tells the agent to check codec and audio with ffprobe and compare at least six timestamps — start, middle, end, side profiles, nearest and farthest faces — before delivering, and to reject the result if any face is still recognizable.
Setup:
It does more than per-frame detection. Face boxes are smoothed over time so the mask doesn't jitter when the camera shakes, and when a face is missed for a few frames the mask stays put for a while, so a quick turn or a glance down doesn't expose anyone. The output is always an H.264 + yuv420p MP4 that plays in Telegram and common players.
Features
Two detectors: InsightFace SCRFD by default, which handles distant small faces, side profiles, low light and crowds better but is slower on CPU; switch to YuNet for well-lit close and medium shots when speed matters.
Three masking styles:
--mode mosaic (default), blur and black. Use black for sensitive material — mosaic and blur are visual obfuscation, not irreversible anonymization.Smoothed tracking: Boxes are matched across frames by IoU and blended over time;
--smooth sets how steady the mask follows.Hold on misses:
--hold keeps a mask in place for a set number of frames after a face drops out, covering brief turns and occlusions.Tunable detection: Raise
--det-size to catch smaller, farther faces, increase --margin to cover face edges, and adjust --score for detection confidence.Audio and compatibility: FFmpeg muxes the original audio back in and encodes H.264, yuv420p with faststart so the file streams in a browser.
Built-in validation: The Skill tells the agent to check codec and audio with ffprobe and compare at least six timestamps — start, middle, end, side profiles, nearest and farthest faces — before delivering, and to reject the result if any face is still recognizable.
Setup:
pip install opencv-python numpy insightface onnxruntime plus ffmpeg on the system; YuNet-only use needs just opencv-python and numpy. SCRFD downloads the roughly 275 MB buffalo_l model on first use. Run it with python scripts/face_mosaic.py input.mp4 output.mp4.