How it works
Posture Correct does not regenerate the portrait or infer an unseen 3D body pose. It requires one clear front-facing person, then cross-checks mirrored MoveNet Thunder shoulder, elbow, wrist, and hip landmarks against separate face detection. The head must remain centered over a nearly level shoulder line, while the pelvis may show only a mild measurable sideways offset. Six-class portrait segmentation must continuously trace the fitted clothing, both complete arms and hands, and clean low-detail background gaps between each arm and the torso or hips. Editing starts below the sleeves: a bounded source-pixel field moves only the verified torso and pelvis, peaks at the hips, and fades to zero through the upper thighs. Each row expands or compresses only the verified source background gap beside the torso; both arms and hands are copied from the source exactly instead of being curved by the correction. The face, head, shoulder line, outer background, and every pixel outside the reported torso-and-gap field are also unchanged in the lossless PNG. If an arm touches the body, the destination-side gap is too narrow, or the source background is busy, the tool refuses the photo rather than bending an arm, revealing invented pixels, or regenerating anatomy. Already-upright photos, forward or backward slouch, strong lean, turned poses, missing joints, transparency, and ambiguous silhouettes are refused for the same reason.
