How it works
Hip Adjust does not regenerate the portrait or send it through a full-frame image model. It first requires one clear front-facing person and cross-checks mirrored MoveNet Thunder shoulder and hip landmarks against separate face detection. Six-class portrait segmentation must then isolate a continuous lower hip and pelvis clothing silhouette and independently confirm clean background gaps before both arms or hands. One horizontal source-pixel mesh begins below the natural waist, reaches its maximum near the hip line, and fades out just below it. At maximum intensity, the verified clothing outline can broaden by up to a restrained 12% per side; nearby verified background is redistributed only inside the measured gap, and displacement tapers to zero at every boundary. The face, upper body, natural waist, arms, hands, lower legs, distant background, and every pixel outside the local envelope are copied unchanged into a lossless PNG. Bare or hidden hip outlines, loose clothing, touching hands, cropped bodies, transparent cutouts, and ambiguous poses are refused instead of guessed.

























