How it works
Arm Adjust does not send the portrait through a full-frame generator. It requires one clear person and cross-checks mirrored MoveNet Thunder shoulder, elbow, wrist, and hip landmarks against separate face detection. Six-class portrait segmentation must then isolate both arms as separate near-vertical silhouettes with verified background immediately beside each edge. Two independent horizontal source-pixel meshes narrow only the safe middle portions of those arm bands—up to a restrained 14% per side at maximum intensity—while fading to zero below the shoulders, before the wrists, and at every side boundary. Existing skin, sleeve, and nearby background pixels are redistributed only inside the measured envelopes; the face, shoulders, central torso, hands, lower body, and every pixel outside both bands are copied unchanged into a lossless PNG. Bent, touching, occluded, cropped, transparent, or ambiguous arms are refused instead of guessed.

























