Determining an accurate initial native-like protein fold is one of the most important and time-consuming steps of de novo NMR structure determination. Here we demonstrate that high-quality native-like models can be rapidly generated from initial structures obtained using limited NOE assignments, through replica exchange molecular dynamics refinement with a generalized Born implicit solvent (REX/GB). Conventional structure calculations using an initial sparse NOE set were unable to identify a unique topology for the zinc-bound C-terminal domain of E. coli chaperone Hsp33, due to a lack of unambiguous long range NOEs. An accurate overall topology was eventually obtained through laborious hand identification of long range NOEs. However we were able to obtain high-quality models with backbone RMSD values of about 2 angstroms with respect to the final structures, using REX/GB refinement with the original limited set of initial NOE restraints. These models could then be used to make further assignments of ambiguous NOEs and thereby speed up the structure determination process. The ability to calculate accurate starting structures from the limited unambiguous NOE set available at the beginning of a structure calculation offers the potential of a much more rapid and automated process for NMR structure determination.