Blur faces for privacy before publishing

JAD uses TensorFlow.js with a lightweight face detection model optimised for browser inference. Detection accuracy is high for front-facing subjects in reasonable lighting (>90% recall). It may miss faces that are partially occluded, very small in frame (below ~30px), angled away, or in very dark/backlit conditions. Review the output at 1x speed before publishing for any compliance-sensitive use case.

Yes — the detection runs independently on each frame and repositions the blur region per frame. Tracking is based on detection position each time, not motion-based interpolation, so if a face is momentarily undetected (occluded, fast motion) the blur may disappear for that frame. For critical compliance use, supplement with manual review.

GDPR defines anonymisation as rendering re-identification impossible. Standard Gaussian blur can be partially reversed with image processing techniques in some cases. For high-risk contexts (research, journalism involving vulnerable subjects), consider higher blur intensity or pixelation instead. For everyday editorial use (social publishing, corporate video), strong Gaussian blur is generally considered sufficient and is industry-standard practice.