The speed advantage of neural representation has had a profound impact on scene reconstruction. However, it often involves collecting a large number of points along a ray, even if these points can be filtered. These points consume significant resources, and processing them also requires a considerable amount of time. We propose a new framework for scene representation and a unique training strategy. Specifically, we use a set of low-resolution grids to guide the sampling of the current grid-based model. Initially, we evenly sample points along rays and query their volume density using the low-resolution grid. Then, with our improved hierarchical sampling strategy, we concentrate on sampling near points with higher volume density. Subsequently, we query their volume density using the high-resolution grid. We optimize both low and high-resolution grids jointly in the first stage and only optimize the high-resolution grid in the second stage. Experiments show that we only need to collect about one-tenth of the points compared to traditional methods based on display grids, saving multiple times the GPU resources. Additionally, we further improve training time and rendering speed by around 30%, with more pronounced benefits at higher grid resolutions.