What actually happens
to your selfie
No invented numbers, no vague "AI magic." Here's the real pipeline, in the order it runs.
Face detection & alignment
Google's MediaPipe FaceMesh finds 468 landmark points on your face. If your head is tilted, the image is rotated back level using your eye positions before anything else runs — a tilted photo otherwise throws off every zone that comes after it. The photo must show one face, filling at least 5% of the frame, at a small enough angle that the read stays reliable.
Five zones, not a grid
Early versions mapped a fixed grid onto the photo. That broke at even a 5° head tilt — grid cells drifted onto the wrong part of the face. Zones are now built from the landmarks themselves: T-zone, left cheek, right cheek, chin, and under-eye, each a real region anchored to your actual bone structure, not a rectangle guessed from the frame.
Skin segmentation & scoring
A MediaPipe segmentation model separates skin from hair, eyes, lips, and background. Within each zone, small patches of real skin are scored across five traits — oiliness, dryness, acne, pigmentation, sensitivity — using computer-vision measurements (texture, color variance, light reflectance), not a single black-box number.
Acne gets a second, dedicated model
Acne severity is refined by a MobileNetV2 classifier — a small neural network trained specifically to grade acne lesions by severity, run on cropped patches from each zone. It overwrites the general texture-based estimate with a purpose-built one.
Facial hair is excluded, not guessed through
A separate classifier detects facial hair regions and excludes them from acne, dryness, pigmentation, and sensitivity scoring in that zone, rather than scoring stubble as if it were skin texture. A zone under facial hair is honestly reported as unmeasured, not silently guessed.
A confidence score that can downgrade the routine
Photo quality, blur, cross-zone consistency, and per-zone detector reliability combine into one confidence score. If it falls below a threshold, every score is zeroed and the routine trims to maintenance-only — the system is built to say 'we're not sure' rather than hand out a confident-sounding routine from an unreliable photo.
Ingredients matched to what was actually measured
Your five axis scores are matched against a database of ingredients tagged by which concern they address, ranked, and filtered for conflicts (see the ingredient guide) before being assembled into an AM/PM routine with real products available in India.
Questions people actually ask
Is Skinzy a medical diagnosis?
No. Skinzy is a cosmetic skin-analysis tool, not a diagnostic device. It scores five visible skin traits from a photo and suggests skincare ingredients — it doesn't diagnose, treat, or prevent any medical condition. See a dermatologist for anything that looks or feels like more than a cosmetic concern.
What does Skinzy actually measure?
Five zones on your face — T-zone, both cheeks, chin, and under-eye — each scored on five traits: oiliness, dryness, acne, pigmentation, and sensitivity. MediaPipe FaceMesh finds 468 facial landmarks to locate the zones precisely; separate models then score each trait.
How accurate is the confidence score?
The confidence score (shown as a percentage, typically 70-81%) blends photo quality, blur, how consistent the scores are across zones, and how reliable each detector was on your specific photo. Below a threshold, Skinzy trims the routine to maintenance-only rather than guessing.
Does Skinzy work for all skin tones?
Pigmentation scoring is calibrated per zone using LAB lightness variance and weighted by an estimated Fitzpatrick tone, but it's still the weakest of the five axes — lightness-based measurement can conflate shadow and pore texture with actual melanin, which affects darker skin tones more. This is a known limitation, not a solved problem.