Scientists often use seismic waves to create images of what lies beneath the Earth's surface, but the process called full waveform inversion is tricky because limited and noisy data can lead to uncertain results. Rather than producing one single map, a Bayesian approach tries to show a range of possible underground structures along with how likely each is. However, traditional ways of doing this require huge amounts of computing time because they keep solving complex wave equations repeatedly. A new preprint posted on arXiv introduces a technique that blends Stein variational gradient descent with the alternating direction method of multipliers in a dual augmented Lagrangian setup. By keeping the wave operator fixed to an updated background model between frequency batches, the method only needs to factor the equations once per particle per frequency. This brings the total cost down to roughly that of running a few standard deterministic inversions. The researchers tested their approach on the standard Marmousi II synthetic model and found it produced well-calibrated uncertainty estimates with image quality similar to more expensive methods. This is a preprint (https://arxiv.org/abs/2603.24751) and has not yet been peer-reviewed. The study is purely computational with no real-world field data, and the authors note that high-dimensional generalization remains a challenge in the field.