Facial expressions are crucial in human communication. Recent decades have seen growing interest in understanding the role of spatial frequencies (SFs) in emotion perception in others. While some studies have suggested a preferential treatment of low versus high SFs, the optimal SFs for recognizing basic facial expressions remain elusive. This study, conducted on Western participants, addresses this gap using two complementary methods: a data-driven method (Exp. 1) without arbitrary SF cut-offs, and a more naturalistic method (Exp. 2) simulating variations in viewing distance. Results generally showed a preponderant role of low over high SFs, but particularly stress that facial expression categorization mostly relies on mid-range SF content (i.e. ∼6-13 cycles per face), often overlooked in previous studies. Optimal performance was observed at short to medium viewing distances (1.2-2.4 m), declining sharply with increased distance, precisely when mid-range SFs were no longer available. Additionally, our data suggest variations in SF tuning profiles across basic facial expressions and nuanced contributions from low and mid SFs in facial expression processing. Most importantly, it suggests that any method that removes mid-SF content has the downfall of offering an incomplete account of SFs diagnosticity for facial expression recognition.