How it works
Perspective Face Correction cross-checks a 478-point face mesh against independently detected eyes, nose, and mouth. It compares three separate pose cues—the nose offset, projected cheek widths, and mesh cheek depth—and proceeds only when they agree on a mild turn. One continuous horizontal source warp then reduces that foreshortening inside the verified face oval. Sampling stays within existing facial pixels and fades to zero before the oval boundary, so hair, clothing, background, and the facial outline remain unchanged. It never generates the hidden side of a profile, and an already front-facing face remains pixel-exact.

























