How it works
Clothes Smooth does not regenerate the portrait or blur the full image. It requires one clear front-facing person, then cross-checks mirrored MoveNet Thunder shoulders and hips against separate face detection. MediaPipe six-class portrait segmentation must continuously identify the central upper garment as clothing from below the face to just above the hips. The mask is eroded before processing so the collar, sleeves, seams, silhouette, nearby skin, and background remain source-exact. Inside that safe interior, garment-normalized frequency separation measures only mid-scale luminance relief: shallow wrinkle highlights and shadows are reduced, while fine weave is retained and strong luminance or color edges from seams, prints, and logos are protected. The face, hair, skin, arms, hands, lower body, background, framing, uploaded alpha, and every pixel outside the reported clothing mask are copied unchanged into a lossless PNG. Multiple people, turned or cropped poses, covered torsos, loose or strongly patterned layers, uncertain clothing segmentation, and already-smooth fabric are refused rather than blurred or invented.
