Abstract : We propose a new method for greatly accelerating the computation of complex, detailed shadows in a radiosity solution. Radiosity is computed using a “standard” hierarchical radiosity algorithm with clustering, but the rapid illumination variations over some large regions receiving complex shadows are computed on the fly using an efficient convolution operation, and displayed as textures. This allows the representation of complex shadowed radiosity functions on a single large polygon. We address the main issues of efficiently and consistently integrating the soft shadow calculation in the hierarchical radiosity framework. These include the identification of the most appropriate mode of calculation for each particular configuration of energy exchange, the development of adequate refinement criteria for error-driven simulation, and appropriate data structures and algorithms for radiosity representation and display. We demonstrate the efficiency of the algorithm with examples involving complex scenes, and a comparison to a clustering algorithm.