Groundbreaking applications such as ChatGPT have heightened research interest in generative artificial intelligence (GAI). Essentially, GAI excels not only in content generation but also in signal processing, offering support for wireless sensing. Hence, we introduce a novel GAI-assisted human flow detection system (G-HFD). Rigorously, G-HFD first uses channel state information (CSI) to estimate the velocity and acceleration of propagation path length change of the human-induced reflection (HIR). Then, given the strong inference ability of the diffusion model, we propose a unified weighted conditional diffusion model (UW-CDM) to denoise the estimation results, enabling the detection of the number of targets. Next, we use the CSI obtained by a uniform linear array with wavelength spacing to estimate the HIR's time of flight and direction of arrival (DoA). In this process, UW-CDM solves the problem of ambiguous DoA spectrum, ensuring accurate DoA estimation. Finally, through clustering, G-HFD determines the number of subflows and the number of targets in each subflow, i.e., the subflow size. The evaluation based on practical downlink communication signals shows G-HFD's accuracy of subflow size detection can reach 91%. This validates its effectiveness and underscores the significant potential of GAI in the context of wireless sensing.