Arousal is fundamental for affective experience and, together with valence, defines the core affective space. Precise brain models of affective arousal are lacking, leading to continuing debates of whether the neural systems generalize across valence domains and are separable from those underlying autonomic arousal or wakefulness. Here, we combine naturalistic fMRI with predictive modeling to develop a brain affective arousal signature (BAAS, discovery-validation design, n = 60, 36). We demonstrate its (1) sensitivity and generalizability across mental processes, valence, and stimulation modality and (2) neural distinction from autonomic arousal and wakefulness (24 studies, n = 868). Affective arousal is encoded in distributed cortical-subcortical (e.g., prefrontal, periaqueductal gray) systems with local similarities in thalamo-amygdala-insula systems between affective and autonomous arousal. We demonstrate application of the BAAS to improve specificity of established valence-specific neuromarkers. Our study provides a biologically plausible model for affective arousal that aligns with the affective space and has a high application potential.