Class imbalance complicates image-based driver behavior recognition when hazardous behaviors are underrepresented. This study evaluates how synthetic-data quantity affects classification performance across convolutional and transformer architectures. Using the six-class DMD-TFIW0 dataset, FLUX.1-dev with a pretrained realism-oriented low-rank adaptation module generates images of drinking, yawning, sleepy driving, and distracted driving. Generated images undergo visual screening and semantic assessment using Contrastive Language–Image Pretraining (CLIP). A ResNet50-based classifier, ResNet-18, and Swin Transformer are compared under three training conditions: oversampling with conventional augmentation, and conventional augmentation supplemented with moderate or extensive synthetic data. Synthetic images are used exclusively for training; validation and testing use fixed partitions containing only real images.Relative to baseline, extensive synthetic supplementation increases macro F1-score from 94.61% to 95.53% for ResNet-18 and from 94.82% to 95.96% for Swin Transformer, which achieves the highest accuracy of 96.72%. However, the ResNet50-based classifier performs best with moderate supplementation, reaching a macro F1-score of 94.74\%, compared with 93.34\% at baseline and 94.40\% with extensive supplementation. Moderate supplementation also lowers Swin Transformer's macro F1-score to 94.23\%, demonstrating a non-monotonic response to synthetic-data quantity. These findings show that targeted diffusion-based augmentation can improve recognition on real test images within this benchmark, but its benefits depend on both architecture and augmentation level. Synthetic-data quantity therefore requires architecture-specific evaluation.