Abstract: As the complexity of virtual environments increases, it becomes a critical issue to maintain the quality of a walkthrough experience. In this paper, we propose an effective data management scheme to address this issue in client-server architecture. First, we propose using real-time scene management to manage the computing resources on the client side by reducing the amount of transmitted geometry data. Second, we propose a prioritized most likelihood movement model to prefetch potential future objects based on the user's current motion intention. Lastly, a hybrid coherence cache model is proposed to take advantages of both the temporal and spatial localities of the walkthrough process. We have done extensive experiments to demonstrate how these techniques can improve the effectiveness of walkthrough in a large virtual environment.