Abstract Seismic data are often contaminated by random noise, which obscures weak reflections and complicates seismic interpretation. In this work, we propose a Vision Transformer-based diffusion model for seismic random noise attenuation. The denoising task is formulated within a diffusion framework, transforming the ill-posed problem into a sequence of stable and progressive denoising steps. This generative approach effectively alleviates the mean-collapse and over-smoothing issues inherent in traditional CNN and Transformer-based regression methods, thereby facilitating the recovery of high-frequency seismic details. We adopt a Transformer-based architecture with strong global modeling capability as the backbone of the diffusion model, enabling a global receptive field through self-attention mechanisms. Comparative experiments conducted on synthetic and field 2D seismic datasets confirm the clear performance advantage of the proposed method over f-k filtering, UNet, and standard diffusion models, achieving superior signal preservation with substantially reduced noise residue and signal leakage.
Zhao et al. (Sat,) studied this question.