Abstract Arousal and valence are fundamental dimensions of affective experience signifying levels of activation and pleasantness, respectively. These dimensions play a crucial role in shaping emotional responses and behaviors, with significant implications for psychopathology. Previous machine learning studies had some success decoding these states from brain activation patterns observed during task-based functional magnetic resonance imaging (fMRI), but the results have varied across studies. Moreover, prior studies have often been limited by small sample sizes, weak decoding performance, and non-whole-brain analyses, leaving the neural representations of arousal and valence largely unresolved. Here we successfully decoded arousal and valence from whole-brain task-fMRI data collected from 132 participants during exposure to 300 unique emotional stimuli, including 150 movie clips and 150 text scenarios that reliably induced a wide range of arousal and valence states. Mass univariate general linear models identified block-level activation (emotion stimuli > washout) from all gray matter voxels. Multivariate regression analysis predicted arousal and valence ratings based on these gray matter activations. Patterns in the fMRI data underlying arousal and valence were robust, as they were successfully decoded across both induction modalities using five different linear multivariate regression models. Although significant, decoding from scenarios was less successful than from movies, likely due to their more imaginative nature. In particular, decoding arousal from scenarios only showed low predictive utility. Representations of arousal and valence were widespread throughout the brain, and we reveal cerebellar and brainstem contributions that have largely been absent in past fMRI decoding studies. These findings clarify the distributed neural basis of arousal and valence and provide a foundation for future clinical research on the role of these constructs in affective dysregulation.