Scenes have become more and more complex and difficult to render at interactive rates. As a result, many algorithms have been developed to simplify scenes to meet this requirement. These are: visibility culling, level-of-detail and image-based rendering. The visibility-culling algorithms do not work well when a great number of primitives are still visible from current viewpoint. Level-of-detail (LOD) algorithms cannot easily simplify scenes with a large number of visible objects. Image-based rendering algorithms require a vast amount of storage and cannot deal with dynamic scenes. We attempt to develop a new approach to render complex scenes at interactive rates by combining the above three techniques. We need to develop new methods to meet the requirements: develop an efficient partition scheme; determine the criteria of replacing the primitives with texture; reconstruct at 3D hierarchical model from the texture; develop a forecasting technique to exploit temporal coherence.