Seismic imaging helps us see underground by analyzing how waves travel through the Earth. A key challenge in this process is the lack of low-frequency (deep-penetrating) signals in recorded data, which are crucial for accurate imaging. In this work, we use a machine learning technique called a diffusion model to “fill in” the missing low-frequency signals using the high-frequency data we do have. Think of it like restoring the missing bass in a distorted audio file. What makes our method unique is that we also add physics into the process, specifically, we make sure the filled-in data follow patterns expected from wave propagation. This combination of machine learning and physical constraints leads to more realistic results. Our approach can also generate multiple possible versions of the missing data, which helps scientists understand how uncertain their final images might be. This is important because errors in low-frequency data can greatly affect the model obtained to then interpret the Earth’s interior. The results show promise for improving the accuracy and reliability of seismic imaging techniques used in energy exploration and earthquake studies.