Please don't. "Raw" R^2 of different estimation models cannot be compared at all (you are talking about OLS here, are you?). If you wanna compare two models, use an F-Test: F-test - Wikipedia, the free encyclopedia
When there is a linear model, y = a + bx, and a log-log model, ln y = a + b ln x, how can the raw R^2 be setup to be compared?
Do you just compute the linear model's raw R^2 with y natural logged and then exponentiate the y in the log log model?