<p>Deconvolution is a critical technique for improving seismic data resolution. It eliminates the filtering effects of the seismic wavelet to obtain reflection coefficients. However, deconvolution is a highly ill-posed problem, making accurate recovery of reflection coefficients challenging. Traditional methods obtain a stable solution to ill-posed problems by incorporating explicit model-based priors based on simplified assumptions. However, they often struggle to produce satisfactory results in complex geological environments or with noisy data. Although deep learning methods can effectively enhance deconvolution results by learning implicit priors from training data, they also suffer from certain limitations, such as generalization problems. Diffusion models offer a novel approach to acquiring prior information, known as generative priors. They directly model the approximate probability distribution of reflection coefficients by training a denoiser and learning to generate reflection coefficients from random noise. This denoiser provides the diffusion model with inherent robustness to random noise. We propose a 2D seismic deconvolution diffusion model (SeisDPS) that combines generative priors with model-based priors to achieve more accurate and high-fidelity reflection coefficients. However, existing diffusion-based deconvolution methods typically focus on processing 2D data, and even a few recent 3D approaches are computationally expensive. Then, building upon SeisDPS, we propose an algorithm for 3D seismic deconvolution (SeisDPS-3D). This 3D algorithm applies SeisDPS to one direction (e.g., inline profiles using SeisDPS) while applying unconditional sampling to the other (e.g., crossline profiles for continuous sampling correction) and vice versa. This unconditional sampling uses the prior information learned by the 2D diffusion model as 3D lateral constraints in orthogonal directions, thereby effectively improving the spatial continuity of the 3D deconvolution results. Under the same conditions, compared to the 2D diffusion model using slice processing, our method only increases the memory consumption by 5%. Our method is trained solely on 2D synthetic reflection coefficients. It demonstrates superior results to the sparse-spike inversion method on 3D synthetic and field datasets. Experimental results show that our method exhibits greater robustness to noise and improved spatial continuity.</p>