This depends on which type of noise you are adding. I will presuming you are adding zero-mean Gaussian noise; firstly because that's mostly likely what you are using and secondly because its nice and easy to work with.
If we have a quick look on
Wikipedia[
^] for PSNR we get a whole bunch of useful stuff. First lets consider the formula for the MSE; the two images are the same and will cancel each other out, just leaving the Gaussian noise. So you need to set the mean square of the Gaussian noise to the right level.
Rearranging the second (will become clear why I chose that one shortly) PSNR formula to find out the desired MSE should result in this, assuming I haven't made any foolish maths errors:
sqrt(MSE) = MAXi/(10^(PSNR/20))
Now since we are just looking at the noise (since the two actual images cancelled out) and the mean of our Gaussian is zero, then the sqrt(MSE) is just the
RMS[
^] of the noise and by a happy coincidence the RMS is simply the variance of the Gaussian.
Probably worth looking for a good text book on signals, unfortunately I can't think of an optimum choice off the top of my head.
Charles Keepax