In this paper we address the problem of creating accurate joint models from real motions while allowing scalability. We propose an automatic method to model, scale and simulate non-idealized joints from the external motion of markers. We demonstrate the method on the human knee joint modeling for musculoskeletal analysis and for character animation. The resulting joints, called correlative joints, are character and motion independent and rely on linear combinations of degrees of freedom calculated from multiple regression laws. We show that by using such models, inverse kinematics solvers find better solutions when tracking motions and solving constraints. Importing correlative joints into new models involves only minimal requirements on landmarks locations and no costly additional computations.