Resting-state functional connectivity (FC) has received considerable attention in the study of brain-behavior associations. However, the low generalizability of brain-behavior studies is a common challenge due to the limited sample-to-feature ratio. In this study, we aimed to improve the generalizability of brain-behavior associations in resting-state FC by focusing on diverse and stable edges, i.e., edges that show both high between-subject and low within-subject variability. We used resting-state data from 1003 participants with multiple fMRI sessions from the Human Connectome Project to group FC edges in terms of between-subject and within-subject variability. We found that resting-state FC variability was dominated by stable individual factors and that diverse and stable edges were primarily part of heteromodal associative networks. We used canonical correlation analysis (CCA) combined with feature selection and principal component analysis (PCA) to investigate the impact of edge selection on the strength and generalizability of brain-behavior associations. Surprisingly, selection based on edge stability did not significantly affect the results, but diverse edges were more informative than uniform edges in two of the three parcellations tested. Regardless, using all edges resulted in the highest strength and generalizability of canonical correlations. The lack of improvement in generalizability with selection of stable edges may be due to unreliable estimation of within-subject edge variability or because within-subject edge variability is not related to the information value of the edges for brain-behavior associations. In other words, unstable edges may be equally informative as stable ones.