In network-based walkthroughs the server transmits on-line only the primitives which are potentially visible from the client's current viewpoint (visibility streaming). To reduce the bandwidth requirements it is necessary to minimize the set of view-dependent potentially visible primitives using occlusion culling techniques. However, even a slight change in the viewpoint might require the computation and transmission of many new primitives and thus latency is inevitable. In this paper we present an algorithm for determining a conservative superset of an ε-neighborhood of a given viewpoint. Having such an ε-superset, the client is free to render the model independently of the server as long as he moves within that ε-neighborhood.