Most existing fast out-of-core simplification methods are based only on uniform or approximated uniform sampling and thus hard to perform adaptive sampling according to the detail distribution of the models. The loss of visual fidelity may be large when processing models with non-uniform distribution of detail. In this paper, an efficient high quality out-of-core simplification method is presented. It can deal with models that are too large to be loaded into main memory. This method greatly improves the simplification result of the models whose detail is non-uniform distributed. In this paper, an initial simplified model is first generated by uniform sampling. A statistical analysis of the original model is also performed at the same time, based on which detail shifting and locally refined sampling to the model can be performed. The algorithm greatly improves the quality of the result model while the processing time is still linear to the model size and the memory cost is also small.