Sharpness-Aware Minimization (SAM) has emerged as a promising alternative to stochastic gradient descent (SGD) for minimizing the loss objective in neural network training. While the motivation behind SAM is to bias models towards flatter minima that are believed to generalize better, recent studies have shown conflicting evidence on the relationship between flatness and (in-distribution) generalization, leaving the mechanism behind SAM's performance improvement unclear. In this work, we present a complementary effect that cannot be explained by in-distribution improvements alone: we argue that SAM can enhance the quality of features in datasets containing redundant or spurious features. We explain how SAM can induce feature diversity by investigating a controlled setting. Our results imply that one mechanism by which SAM improves the quality of features is by adaptively suppressing well-learned features which gives remaining features opportunity to be learned.