A number of initialization schemes based on recursive filters are formulated and tested with a numerical weather prediction model, HIRLAM. These have an advantage over schemes which use nonrecursive filters in that they derive the initialized values from a diabatic trajectory passing through the original analysis. The changes to the analysed fields are comparable in size to typical observational errors. A non-recursive implementation of the recursive filters makes the new initialization schemes as easy to use as the original non-recursive filter schemes. It also allows use of higher order filters without added cost. An initialization method using sixth-order filter is compared to a scheme based on an non-recursive optimal filter, and is found to produce similar results for less than half the computational cost. If the sole aim is noise suppression, a filter whose output validates later than the initial time may be used. The advantage of this is that computation time is further reduced and phase error completely eliminated.