Volume rendering is a visualizing technique for 3D volume data. The pixel values of a rendered image are determined by accumulating sampled values form volume data. Usually, the product of opacity and shading values is used as a sampled value. However, the size of the volume data is usually too big to handle in real time. Therefore, a control measure, which changes the level of detail (LOD) of the rendered image, may be introduced for obtaining a reasonable rendering speed. In this paper, we introduce a new criterion for controlling the LOD of the rendered image, and a new octree-based rendering method using this criterion efficiently. As the new criterion, the variance of the opacity and normal vector product is adopted and used to classify volume blocks into an octree structure. In the rendering stage, normal blocks are rendered by using the shear-warp factorization and single-valued blocks by using a template, while zero blocks are skipped. By performing in this fashion, the prosed scheme can reduce the overall rendering time. The scheme is evaluated by rendering a skull volume data ste obtained from an x-ray CT system.