Paper submitted to IMAGE 2026, by Chaoshun Hu, Nizar Chemingui, Raafat Abdul Alim (TGS).
Abstract
We present a seismic inversion framework that combines poststack or post-fwi inversion with pre-trained diffusion generative models. Although diffusion models provide a powerful prior for noisy inverse problems, diffusion posterior sampling is often constrained by the high cost of repeated likelihood-score evaluations through Tweediebased gradients. To address this limitation, we replace the reverse-time stochastic differential equation with the deterministic probability flow ODE (PFODE). This formulation enables larger time steps, faster inference, and better scalability while maintaining high inversion quality. Results on the Marmousi model and the Volve field dataset show that the PFODE approach preserves structural continuity, improves subsurface recovery, and lowers computational cost. Overall, the proposed method provides an efficient, robust, and high-fidelity framework for seismic reconstruction.
Read the full article here.