This paper revisits the classical Kennicutt method for inferring the stellar IMF from the integrated light properties of galaxies. The large-size, uniform high-quality data set from the SDSS DR4 is combined with more in-depth modeling and quantitative statistical analysis to search for systematic IMF variations as a function of galaxy luminosity. Galaxy Ha equivalent widths are compared to a broadband color index to constrain the IMF. This parameter space is useful for breaking degeneracies that are traditionally problematic. Age and dust corrections are largely orthogonal to IMF variations. In addition, the effects of metallicity and smooth SFH e-folding times are small compared to IMF variations. We find that for the sample as a whole the best-fitting IMF slope above 0.5 M? is G=1.4535, with a negligible random error of ±0.0004 and a systematic error of ±0.1. Galaxies brighter than around Mr,0.1=-20 (including galaxies like the Milky Way, which has Mr,0.1~-21) are well fitted by a universal G~1.4 IMF, similar to Salpeter, and smooth, exponential SFHs. Fainter galaxies prefer steeper IMFs, and the quality of the fits reveals that for these galaxies a universal IMF with smooth SFHs is actually a poor assumption. Several sources of sample bias are ruled out as the cause of these luminosity-dependent IMF variations. Analysis of bursting SFH models shows that an implausible coordination of burst times is required to fit a universal IMF to the Mr,0.1=-17 galaxies. This leads to the conclusions that the IMF in low-luminosity galaxies has fewer massive stars, by either steeper slope or lower upper mass cutoff, and is not universal.