SkinToneNet
Estimates your Monk Skin Tone (1–10) from a photo of your face. Research demo — tell us when it's wrong.
Front-facing, face clearly visible. Even lighting helps — the model is noticeably worse in dim light.
Finding your face and running the model…
Monk
Which swatch actually matches you?
Pick the closest one. This is how we find out where the model is off.
Show what the model did
Calibrated answer
Raw argmax (uncalibrated)
Expected value over 10 classes
The network is trained on FACET, whose Monk labels sit darker than the official scale, so the raw argmax is biased and never outputs 1, 2 or 10. We decode the expected value and apply an affine correction fitted on MST-E. Held-out accuracy: 34% exact, 78% within ±1 tone.
Your photo is cropped to the face and stored, along with the swatch you pick, so we can measure the model and retrain it. Don't upload anything you wouldn't want kept.