CSI-based activity recognition has emerged as a promising field in IoT, enabling precise localization, behavior analysis, and health monitoring. However, collecting comprehensive datasets that cover a wide range of environmental conditions and specific actions-such asfalls—presents notable challenges. This is particularly true when dealing with rare or complex events, where diverse and extensive datasets are essential. Traditional generative models, such as GANs and VAEs, often struggle to generate high-quality, diverse, and realistic data. To overcome these challenges, this study introduces the use of diffusion generative models as a novel approach for synthesizing raw CSI data. This method generates synthetic samples with corresponding labels, aiding the training of robust classifiers. Diffusion models, offer a more effective solu tion by progressively denoising random noise into realistic samples. Our experimental results confirm that this approach significantly improves classifier accuracy, even with limited real-world data, thereby reducing the time and resources typically required for traditional data collection. Specifically, our findings support the hypothesis that diffusion-generated data can greatly enhance model performance in scenarios with scarce real world data. The proposed TCN-T model, an enhanced version of the Temporal Convolutional Network (TCN), demonstrated a marked improvement in accuracy—from 91.97% to 94.74%. Similarly, other baseline models also showed notable improvements: CNN1D increased from 83.74% to 89.7%, CNN2D from 81.96% to 88.71%, and CNNLSTM from 82.39% to 90.18%. Furthermore, when working with sufficient original data, the TCN-T model achieved an accuracy increase from 94.74% to 96.59%.