The information values of specific facial features differ depending on the emotional expression. The eyes are more diagnostic of fear, while the mouth area is more relevant for recognizing happy emotions. Previous research found sensitivity to the distribution of such diagnostic features in reflexive eye movements, but results are partly inconsistent. Additionally, it is unclear whether autonomic activity reflects visual processing of these features. We briefly (50 or 150ms) presented emotional faces (fearful, happy, and neutral) either in an upright or inverted fashion to elicit reflexive saccades. By vertically shifting faces relative to the previous central fixation cross, participants were also foveated on either the eye or the mouth region. We analysed the proportion of saccades towards facial features in the visual periphery as well as pupil size. Although participants' classification of facial expressions was generally very accurate and fast, the initial focus on diagnostic features had no effect on their performance. However, we replicate the previously observed preferential visual orienting towards diagnostic features even for very brief presentation times. Pupil size was not modulated by the experimental manipulations. These findings demonstrate that facial expression categorization can be accomplished with very restricted visual input.