I think it is depends what you wanted to do with the generated structures. It might because the native model has big clash so that it causes the decoy to break down a lot. Generally, the more decoys are generated, the better coverage you will get.
If you want to keep the bond angles of native structure, you can try to use the flag -use_input_bond, which “Minimize or refold using the bond lengths and angles from the starting structure, useful for minimizing a structure without idealizing it first.”
> I noticed that when refining the crude ab initio rosetta model with -relax and -nstruct higher than 1 the decoys you might obtain quite different structures (e.g. there might be quite significant changes in relative orientation of helices).
> Is it common in the “official” rosetta ab initio protocols (for instance that used in CASP7) to generate multiple decoys with -relax starting from the same crude model? Or you rather generate single decoy from each crude ab initio model?