Diffusion model-based methods have seen significant development in image dehazing. However, these methods still face limitations. Firstly, the heavy reliance on paired data makes it difficult for the model to adapt to complex haze distributions in real-world scenarios, significantly compromising their generalization capability. Secondly, the unclear sampling directions during the diffusion process lead to unstable dehazing performance, resulting in either residual haze or artifacts caused by excessive dehazing. To address these issues, we propose Physics-Guided Diffusion Model (PGDM) for unpaired real-world dehazing. The core innovations of the proposed model include: (1) a physics-guided conditional diffusion model, where the synergistic guidance of pseudo-clear image and transmission map enables more explicit sampling directions; (2) an unpaired cycle-consistent framework that eliminates paired-data dependency through cycle-consistency loss. Experimental results demonstrate that PGDM significantly outperforms state-of-the-art methods on multiple real-world datasets.