This paper examines the role of visuospatial resources in the generation of predictive and bridging inferences. It is suggested that predictive inferences need visuospatial resources but bridging inferences do not, and that these inferences may be generated at different levels of text representation.