FaceRuler
A privacy-first browser-based facial assessment application that performs inference on the client side using MediaPipe FaceLandmarker. It maps geometric measurements of facial landmarks to PSL scale ratings and does not upload photos at any point.
Timeline (last 90 days)
The FaceRuler app performs face inference on the browser side, using MediaPipe FaceLandmarker (478 keypoints) compiled to WASM, with a model of about 3.6 MB; the inference stage generates no third-party requests, and photos never leave the browser tab.
FaceRuler's scoring engine is composed entirely of deterministic functions, contains no LLM, and the same keypoints always produce the same report.
FaceRuler's eight sub-metrics are aggregated into five dimensions, with fixed weights: symmetry 0.2, proportion 0.25, structure 0.25, eyes 0.15, harmony 0.15.
FaceRuler uses a logistic curve to map the weighted total score (0-100) onto the PSL 0-8 scale, with the function overallToPsl(overall) = 8 / (1 + exp(-(overall - 50) / 13)), anchored so that a total score of 50 corresponds exactly to PSL 4.0.
The FaceRuler developer calibrated the FWHR center value at 2.75 from keypoint batches and documented the mismatch with the values in the papers.
When handling head rotation, FaceRuler measures the midline angle from the top of the forehead to the chin and, before computing any metric, rotates all keypoints into the corrected coordinate system.
FaceRuler's 'harmony' dimension is derived from the standard deviation of the sub-scores rather than the average, in order to reward consistency and penalize features that drag it down.
All Facts (7)
The FaceRuler app performs face inference on the browser side, using MediaPipe FaceLandmarker (478 keypoints) compiled to WASM, with a model of about 3.6 MB; the inference stage generates no third-party requests, and photos never leave the browser tab.
50%UnverifiedFaceRuler's scoring engine is composed entirely of deterministic functions, contains no LLM, and the same keypoints always produce the same report.
50%UnverifiedFaceRuler's eight sub-metrics are aggregated into five dimensions, with fixed weights: symmetry 0.2, proportion 0.25, structure 0.25, eyes 0.15, harmony 0.15.
50%UnverifiedFaceRuler uses a logistic curve to map the weighted total score (0-100) onto the PSL 0-8 scale, with the function overallToPsl(overall) = 8 / (1 + exp(-(overall - 50) / 13)), anchored so that a total score of 50 corresponds exactly to PSL 4.0.
50%UnverifiedThe FaceRuler developer calibrated the FWHR center value at 2.75 from keypoint batches and documented the mismatch with the values in the papers.
50%UnverifiedWhen handling head rotation, FaceRuler measures the midline angle from the top of the forehead to the chin and, before computing any metric, rotates all keypoints into the corrected coordinate system.
50%UnverifiedFaceRuler's 'harmony' dimension is derived from the standard deviation of the sub-scores rather than the average, in order to reward consistency and penalize features that drag it down.
50%