Operator-Aware Posterior Sampling for Multivariate Weather Downscaling
This paper learns the forward operator from paired data and normalize the gradient per channel and shows that this normalization corresponds to steepest descent under a block-max norm, allowing a single guidance strength across variables.