<p>Pre-stack seismic inversion is a fundamental tool for estimating subsurface elastic parameters, but it constitutes a severely ill-posed inverse problem due to factors such as band-limited data, acquisition noise, and limited angular coverage. To address this challenge, in this paper, we propose a novel inversion framework that integrates the plug-and-play (PnP) paradigm with a denoising diffusion model. Our approach formulates the inversion within a half-quadratic splitting (HQS) optimization scheme, which decouples the problem into a physics-based data fidelity term and a prior regularization term. The core of our contribution is to reframe the generative reverse process of diffusion models as a principled, learned prior within the PnP framework. This method reinterprets the generative process as a principled regularizer within a physics-constrained loop, enabling robust and high-fidelity reconstructions. Extensive experiments demonstrate that our proposed method significantly outperforms established regularization techniques in both quantitative metrics and the preservation of fine-scale geological structures.</p>