Driver monitoring is essential for intelligent transportation, yet reliably distinguishing fatigue, alcohol impairment, and cognitive distraction from visual cues remains challenging due to their subtle overlap and the severe shortage of alcohol-impaired training data. Existing studies typically focus on a single unsafe state, leading to fragmented models that do not generalize well across real driving conditions. This paper presents a unified multi-state driver monitoring framework that integrates a MobileNetV2 backbone with Squeeze-and-excitation blocks for spatial encoding and a lightweight Transformer for temporal modeling. A central contribution of this work is an identity-preserving diffusion-based augmentation pipeline designed to address the absence of suitable alcohol-impaired datasets. Using facial-landmark–guided masking and textual inversion, the pipeline generates realistic alcohol-related facial cues from fatigue images while maintaining subject identity, pose, and illumination. These class-consistent synthetic variants improve the separability of visually similar states without introducing dataset bias. Experiments on a seven-class benchmark demonstrate strong performance, achieving 97.37% test accuracy and macro-F1 \(\approx \) 0.97. The results show that combining discriminative deep learning with controlled diffusion augmentation provides an effective solution for multi-state driver monitoring, particularly in scenarios where real alcohol-impaired data are unavailable.