Abstract: Using the historical Beijing "7.21" extreme precipitation event as an example, the six ensemble schemes (the initial condition (IC), multi-physics (MULTI), 3 stochastic physics as well as a combination of IC and stochastic-physics (COM) were compared in the following three aspects of heavy precipitation forecasts: the performance of ensemble means, ensemble ranges and probabilities with respect to the control forecast, characteristics of ensemble spreads, and spread-forecast error relations. The results show that: (1) In spite of the existence of large systematic forecast error, all the ensembles, especially the IC, MULTI and COM, are able to noticeably improve torrential-rain prediction over the control forecast in both intensity and location, and provide more complete information including the forecast uncertainty for users to make a better decision. (2) Forecast spreads of the three stochastic physics ensembles are similar to each other, generally much less than those of the IC and MULTI ensembles and mainly concentrated near the center of the severe rainfall area. As a result, the ensemble spread is enhanced in the vicinity of the severe precipitation area but little is changed elsewhere after stochastic physics is employed in addition to IC perturbations, which leads to virtually no improvement to the overall spread over a larger domain comparing to that of the IC ensemble. By decomposing spread over the spatial scales, it further shows that the forecast diversity contributed by the stochastic physics is mainly in the smaller-scale ( 1000 km; at smaller scales (< 500 km),multi-physics technique could produce larger precipitation spread than IC perturbation does, another advantage of multi-physics approach over other approaches is that it could partially reduce forecast bias. And, (3) the spread spectrum is similar to the forecast error spectrum over spatial scales for all the ensembles, i.e., decreasing with the increase of the spatial scale. However, the magnitude of the spread spectrum is smaller than that of the forecast error spectrum (indicating under-dispersion), this departure increases rapidly with the decrease of spatial scale and becomes large over the small scales.