Level sets give a nice implicit way of defining shapes. Instead of having to work with complicated maps f : ℝm → ℝn, we reexpress them in the form g : ℝn → ℝ where g−1(c) = Im(f), for some chosen c (usually picked to be zero).
Except we can't usually do this. The codomain usually can't be made to just be scalars; in general you need it to be ℝn − m. But you also usually can't get the full domain — you can only do this locally (because you want g to be a smooth function, you can't just arbitrarily define it, so sometimes there's no way to make it capture all of the image of f).
But notice something else: in doing this we are not capturing all the data of f
If f : ℝm → ℝm is a smooth embedding then for every p ∈ Im(f), there is an open neighbourhood p ∈ U ⊆ ℝn and smooth map gp : U → ℝn − m such that