This study compares the empirical performance of different continuous time models for the dynamics of nine implied volatility indices for both stocks and commodities. The models include linear, quadratic, and non-linear drift specifications with affine, constant elasticity of variance (CEV), and stochastic elasticity of variance (SEV) diffusions. Overall, we find that a non-linear drift specification is important when imposing an affine structure on the diffusion, whereas a simple linear drift is adequate with a CEV and SEV specification, of which the SEV is dominant. For all but two of the indices we investigate, the best specification is an SEV diffusion with linear drift. For gold and the US dollar/euro exchange rate, there is little difference between a CEV and SEV diffusion with linear drift.