Shijun Cheng, Randy Harsuko, and Tariq Alkhalifah

 GitHub: https://github.com/DeepWave-KAUST/DiffVMB-pub

A shallow-to-deep velocity model building approach that uses known shallow information as a prior and progressively generates deeper layers, while flexibly fusing well data and structural information. Compared with one-shot or simple tiling, it stitches layers more smoothly, reduces boundary artifacts, and provides uncertainty estimates. It performs well on standard in-distribution benchmarks and shows solid generalization to a large out-of-distribution case (e.g., Marmousi II) when well constraints are available. Overall, combining shallow priors with well and structural guidance in a depth-progressive diffusion framework yields more stable and geologically plausible large-scale velocity models.