We compare different past sea level reconstructions over the 1950-2009 time span using the Empirical Orthogonal Function (EOF) approach. The reconstructions are based on 91 long (up to 60 years) but sparsely distributed tide-gauge records and gridded sea level data from two numerical ocean models over 1958-2007 (the DRAKKAR/NEMO model without data assimilation and the simple ocean data assimilation ocean reanalysis-SODA-) and satellite altimetry data over 1993-2009. We find that the reconstructed global mean sea level computed over the ~60-year-long time span little depends on the input spatial grids. This is unlike the regional variability maps that appear very sensitive to the considered input spatial grids. Using the DRAKKAR/NEMO model, we test the influence of the period covered by the input spatial grids and the number of EOFs modes used to reconstruct sea level. Comparing with tide-gauge records not used in the reconstruction, we determine optimal values for these two parameters. As suggested by previous studies, the longer the time span covered by the spatial grids, the better the fit with unused tide gauges. Comparison of the reconstructed regional trends over 1950-2009 based on the two ocean models and satellite altimetry grids shows good agreement in the tropics and substantial differences in the mid and high latitude regions, and in western boundary current areas as well. The reconstructed spatial variability seems very sensitive to the input spatial information. No clear best case emerges. Thus, using the longest available model-based spatial functions will not necessarily give the most realistic results as it will be much dependent on the quality of the model (and its associated forcing). Altimetry-based reconstructions (with 17-year long input grids) give results somewhat similar to cases with longer model grids. It is likely that better representation of the sea level regional variability by satellite altimetry compensates the shorter input grids length. While waiting for much longer altimetry records, improved past sea level reconstructions may be obtained by averaging an ensemble of different model-based reconstructions, as classically done in climate modelling. Here, we present such a 'mean' reconstruction (with associated uncertainty) based on averaging the three individual reconstructions discussed above.