1Get a photo in
One face, looking at the camera, evenly lit. After taking a photo you can crop it down to the face. The photo is processed in your browser and never leaves your device.
The camera is off.
No photo yet
2Pick the set of composites
The study built separate composite faces for men and women. Choose the set the photo should be compared with.
Each composite averages just 15 faces of students around 20 years old who, judging by the images, look white (the paper does not report ethnicity). The further a face is from that group, the less the comparison means.
3Develop
Add a photo and pick a composite set first.
4The proof sheet
Each row is one trait. The bar shows the gap in cosine similarity between your face and the two composites: it points toward the composite your face is closer to. The gap is the only hint about how strong the evidence is, and all rows share one scale.
These labels describe which blurry average face a neural network finds more similar to yours. They are not a personality assessment.
Frame by frame
How it works
- Find the face. A face detector locates the face, rotates it so the eyes are level, and crops it.
- Embed it. FaceNet-512, the same model the original Python script uses through the DeepFace library, turns the crop into a vector of 512 numbers.
- Compare. The same is done for the composite faces. For each trait, cosine similarity to the "high" and "low" composite is computed, and the higher one wins.
Everything runs locally in your browser; the model (about 45 MB) is downloaded once and then cached. This is a browser port of the original Python script by Luděk Stehlík. It uses a different face detector than the script, so numbers can differ slightly.