This work complements our previous efforts in generating realistic fingerprint images for test purposes. The main variability which characterizes the acquisition of a fingerprint through an on-line sensor is modeled and a sequence of steps is defined to derive a series of impressions from the same master-fingerprint. This allows large fingerprint databases to be randomly generated according to some given parameters. The experimental results validate our technique and prove that it can be very useful for performance evaluation, learning and testing in fingerprint-based systems.