This paper describes a method for curvature dependant skeletonisation in grey-scale images. We commence from a magnetostatic analogy, where the tangential edge flow (the cross product of the edge gradient and the image plane normal) is intepretted as a current. A vector potential is constructed by integrating the current weighted by inverse distance over the image plane. The skeleton corresponds to the location of valley lines in the vector potential. To damp noise effects we damp the current with an exponential function of the local curvature.