Enhancing Seismic Post-stack Reconstruction with Diffusion Models: Addressing Uncertainty and Structural Complexity

Seismic post-stack data often contains missing traces, affecting processing and interpretation. Deep learning methods, both supervised and unsupervised, have advanced reconstruction but face challenges in generalization and computational cost. We propose a reconstruction method that integrates a pre-trained generative diffusion model with Deep Image Prior (DIP) to ensure data consistency. The diffusion model captures the prior distribution of seismic images, generating realistic reconstructions, while DIP enforces structural consistency using the inherent properties of seismic data. Additionally, diffusion models enable uncertainty quantification by generating multiple samples, providing insights into reliability. Our approach demonstrates strong reconstruction performance on both field and synthetic seismic images, even in structurally complex and previously unseen scenarios. Compared to existing diffusion-based methods, our technique reduces neural function evaluations by up to 4x, significantly improving efficiency. This highlights its ability to handle high geological variability across different exploration targets while maintaining computational feasibility. Our implementation is available at https://github.com/PAULGOYES/CDDIP.git