About Detect Face Shape

Who made this

I am Michael Machatschek. I build small, free web tools at 1000 Ventures, and this is one of them. You can write to me at michael@1000ventures.io.

I am a developer, not a stylist or a doctor. Face shape is a styling guide, not a medical measurement.

How the model works

A scan has three steps. First, a face detector finds and crops the face. Second, an image model scores it against the shapes it knows. Third, a geometry check measures face proportions and adjusts the diamond score.

• The classifier is a Swin Transformer (the small Swin-Tiny variant), fine-tuned to tell five shapes apart: oval, round, square, oblong and heart.
• Diamond is not a class the model learned. It comes from measured landmark proportions (MediaPipe Face Mesh), which shift probability toward diamond when the cheekbones are clearly the widest part.
• The result is a score for all six shapes, so you can see when you sit between two.

How it was tested

The code has automated tests for the face finder, the cropping, the diamond logic and the full scan, including photos with no face, tilted faces and several faces. They check that the pipeline behaves, not that every verdict is right.

I do not publish an accuracy percentage, because I have not measured one on an independent, labeled set of photos. Face shape categories are also fuzzy: two stylists can disagree on the same face. Treat the result as a good starting point and check it against a mirror.

What happens to your photo

Your photo is used to produce your result and is not stored. The details are in the privacy policy.