Observation-Aligned Mask Priors for Learning Physical Dynamics from Authentic Occlusions

In three linesA framework learns authentic occlusion mask distributions using Bayesian Flow Networks to train diffusion-based reconstruction models on incomplete observations. Tested on oceanographic satellite data (256×256), it improves MSE and PSNR over diffusion baselines by preventing zero-query dead zones.

Summary generated by Claude — human-verified