RF-Diffusion is a novel generative model for radio frequency (RF) signals, inspired by the success of diffusion models in computer vision and natural language processing. It addresses the challenges of generating high-quality, time-series RF data by introducing a Time-Frequency Diffusion theory and a Hierarchical Diffusion Transformer (HDT) design. The Time-Frequency Diffusion theory enhances the original diffusion model to effectively capture information in the time, frequency, and complex-valued domains of RF signals. The HDT design incorporates a hierarchical architecture, attention-based diffusion blocks, and complex-valued operations to enable the generation of diverse, high-quality, and time-series RF data. RF-Diffusion outperforms existing generative models in synthesizing Wi-Fi and FMCW signals, and demonstrates versatility in applications such as Wi-Fi sensing systems and 5G channel estimation. The model is implemented using PyTorch and trained on 8 NVIDIA GeForce 3090 GPUs. Evaluation results show that RF-Diffusion achieves high fidelity in generating RF signals, with an average structural similarity of 81% relative to the ground truth. It also improves the accuracy of Wi-Fi gesture recognition systems by up to 11% and enhances 5G FDD channel estimation by 5.97 dB in SNR. The paper makes three main contributions: (1) proposing RF-Diffusion as the first generative diffusion model tailored for RF signals; (2) presenting the Time-Frequency Diffusion theory, an advanced evolution beyond traditional denoising-based diffusion methods; and (3) fully implementing RF-Diffusion, with extensive evaluation results from case studies showing its efficacy. The code and pre-trained model are publicly available, offering a collection of tools for both industry and academia to advance AIGC in the RF domain.RF-Diffusion is a novel generative model for radio frequency (RF) signals, inspired by the success of diffusion models in computer vision and natural language processing. It addresses the challenges of generating high-quality, time-series RF data by introducing a Time-Frequency Diffusion theory and a Hierarchical Diffusion Transformer (HDT) design. The Time-Frequency Diffusion theory enhances the original diffusion model to effectively capture information in the time, frequency, and complex-valued domains of RF signals. The HDT design incorporates a hierarchical architecture, attention-based diffusion blocks, and complex-valued operations to enable the generation of diverse, high-quality, and time-series RF data. RF-Diffusion outperforms existing generative models in synthesizing Wi-Fi and FMCW signals, and demonstrates versatility in applications such as Wi-Fi sensing systems and 5G channel estimation. The model is implemented using PyTorch and trained on 8 NVIDIA GeForce 3090 GPUs. Evaluation results show that RF-Diffusion achieves high fidelity in generating RF signals, with an average structural similarity of 81% relative to the ground truth. It also improves the accuracy of Wi-Fi gesture recognition systems by up to 11% and enhances 5G FDD channel estimation by 5.97 dB in SNR. The paper makes three main contributions: (1) proposing RF-Diffusion as the first generative diffusion model tailored for RF signals; (2) presenting the Time-Frequency Diffusion theory, an advanced evolution beyond traditional denoising-based diffusion methods; and (3) fully implementing RF-Diffusion, with extensive evaluation results from case studies showing its efficacy. The code and pre-trained model are publicly available, offering a collection of tools for both industry and academia to advance AIGC in the RF domain.