Existing Grid visualization systems typically focus on the distribution onto remote machines of some or all of the processes encompassing the visualization pipeline, with the aim of increasing the maximum data size, achievable frame rates or display resolution. Such systems may rely on a particular piece of visualization software, and require that the end users have some degree of knowledge in its use, and in the concepts of the Grid itself. This paper describes an architecture for Grid visualization that abstracts away from the underlying hardware and software, and presents the user with a generic interface to a range of visualization technologies, switching between hardware and software to best meet the requirements of that user. We assess the difficulties involved in creating such a system, such as selecting appropriate visualization pipelines, deciding how to distribute the description language for visualization. Finally, we describe a prototype implementation of such a system, and consider to what degree it might meet the requirements of real world visualization users. 1