Generative Artificial Intelligence (GAI) is being explored for enhancing wireless sensing, particularly in human flow detection. This paper introduces a GAI-assisted human flow detection system (G-HFD) that leverages channel state information (CSI) and a unified weighted conditional diffusion model (UW-CDM) to improve signal processing and detection accuracy. The system first estimates the velocity and acceleration of propagation path length changes caused by human-induced reflections (HIR) using CSI. It then uses UW-CDM to denoise the estimation results, enabling accurate detection of the number of targets. Next, it estimates the direction of arrival (DoA) and time of flight (ToF) of HIR using a uniform linear array (ULA) with wavelength spacing. UW-CDM is used to resolve ambiguous DoA spectra, ensuring accurate DoA estimation. Finally, clustering is applied to determine the number of subflows and the size of each subflow. Evaluation using practical downlink communication signals shows that G-HFD achieves a subflow size detection accuracy of 91%, validating its effectiveness and highlighting the potential of GAI in wireless sensing. The system design includes velocity and acceleration estimation, DoA and ToF estimation, and flow detection through clustering. The proposed UW-CDM is trained to denoise V-A spectra and generate clear DoA spectra, enabling accurate human flow detection. The system is evaluated in practical communication scenarios, demonstrating its performance in detecting the number of human targets and subflows. The results show that G-HFD outperforms existing methods in accuracy and effectiveness.Generative Artificial Intelligence (GAI) is being explored for enhancing wireless sensing, particularly in human flow detection. This paper introduces a GAI-assisted human flow detection system (G-HFD) that leverages channel state information (CSI) and a unified weighted conditional diffusion model (UW-CDM) to improve signal processing and detection accuracy. The system first estimates the velocity and acceleration of propagation path length changes caused by human-induced reflections (HIR) using CSI. It then uses UW-CDM to denoise the estimation results, enabling accurate detection of the number of targets. Next, it estimates the direction of arrival (DoA) and time of flight (ToF) of HIR using a uniform linear array (ULA) with wavelength spacing. UW-CDM is used to resolve ambiguous DoA spectra, ensuring accurate DoA estimation. Finally, clustering is applied to determine the number of subflows and the size of each subflow. Evaluation using practical downlink communication signals shows that G-HFD achieves a subflow size detection accuracy of 91%, validating its effectiveness and highlighting the potential of GAI in wireless sensing. The system design includes velocity and acceleration estimation, DoA and ToF estimation, and flow detection through clustering. The proposed UW-CDM is trained to denoise V-A spectra and generate clear DoA spectra, enabling accurate human flow detection. The system is evaluated in practical communication scenarios, demonstrating its performance in detecting the number of human targets and subflows. The results show that G-HFD outperforms existing methods in accuracy and effectiveness.