Blur Face
Evidence before adjectives
How Blur Face is tested
This page documents the checks behind product claims, expected failure cases, and the limits of the evidence. It is maintained by Will, the developer of Blur Face.
What the testing is designed to answer
The review focuses on practical questions a person needs answered before trusting a privacy workflow:
- Does the selected source file remain in the browser during editing?
- Can automatic detection miss faces under realistic conditions?
- Can the user add, move, resize, and remove redaction areas manually?
- Does the exported file contain the intended visible redactions?
- Which browsers, codecs, file sizes, and formats can prevent completion?
Local-processing verification
After page and model assets load, a test image is processed with the network disconnected. The browser Network panel is also preserved while a file is selected, edited, and exported. Asset, model, analytics, and page requests may occur; a request containing the selected source image should not.
An offline test alone is not conclusive because an application could queue a later request. Network inspection is the stronger check, and anyone can repeat it using the no-upload verification guide.
Face-detection review set
Automatic detection is evaluated as a first pass, never as the final privacy decision. Review images include:
Scale
Large portraits, small background faces, and mixed-size groups
Pose
Front-facing, profile, tilted, and partly turned heads
Visibility
Hair, hands, masks, glasses, and other occlusions
Image quality
Backlighting, shadows, motion blur, and compression
Indirect faces
Mirrors, screens, posters, and framed photographs
Coverage
Forehead, chin, ears, hairline, and identifying context
A missed face is recorded as a detection failure even when it can be covered manually. This separates detector convenience from the safety of the final export.
No invented accuracy percentage
Blur Face does not publish a universal detection percentage. A defensible number requires a defined dataset, annotation rules, face-size thresholds, device and browser versions, model version, and a repeatable protocol. Until that benchmark exists, we describe failure modes and require manual review.
Browser and file compatibility
Image editing, PDF text detection, and video decoding rely on different browser capabilities. Video support is codec-dependent: MP4 or MOV alone does not prove the browser can decode the stream. Reproducible failures are documented as limitations.
Editorial standard
Product guidance is written or reviewed by the maintainer and links to primary sources for legal, standards, and technical claims. We avoid guarantees about anonymity, compliance, detection completeness, and reversibility. Local processing reduces transfer; it does not make every publication safe or lawful.
For corrections or reproducible bugs, contact [email protected] with the browser, operating system, format or codec, and reproduction steps.