How it works
Height Adjust does not send the portrait through a full-frame generator. It requires one clear head-to-toe person and cross-checks mirrored MoveNet Thunder shoulder, elbow, wrist, hip, knee, and ankle landmarks against separate face detection. Six-class portrait segmentation must then trace one continuous shoulder-to-shoe envelope, independently verify both separated leg and complete shoe silhouettes, and confirm clean background beside the body and below both feet. One vertical source-pixel mesh begins about 10% below the shoulder line and proportionally lengthens the arms, torso, and legs through the ankles by up to a restrained 5.5% at maximum intensity. Complete shoes are translated downward without intentional scaling, nearby source background is compressed only inside the measured side-and-floor envelope, and displacement fades to zero at every boundary. The head, face, hair, upper shoulder line, distant background, and every pixel outside the envelope are copied unchanged into a lossless PNG. Cropped, raised-arm, bent, touching, occluded, transparent, or ambiguous full-body poses are refused instead of guessed.
