Abstract: This paper presents a new method of precomputing high-quality soft shadows that can be cast on a static scene as well as on dynamic objects added to that scene. The method efficiently merges the visibility computed from many shadow maps into a penumbra deep shadow map (PDSM) structure. The resulting structure effectively captures the changes of attenuation in each PDSM pixel, and therefore constitutes an accurate representation of light attenuation. By taking advantage of the visibility coherence, the method is able to store a compact representation of the visibility for every location within the field of view of the PDSM. Modern programmable graphics hardware technology is used by the method to cast real-time complex soft shadows.