Comparison between the Wiener filter (top left) and our score-based MAP (not shown) and posterior mean reconstructions (top right, average of 1500 samples), assuming a Gaussian prior. The MAP reconstruction matches the Wiener filter, both in terms of absolute pixel error and power spectrum ratio (bottom right, blue line). Our posterior mean is also in excellent agreement with the Wiener filter, as illustrated by the pixel absolute error (bottom left) and the power spectrum relative error (bottom right, orange line). The bottom panel shows the relative error between the Gaussian posterior samples’ power spectra and the theoretical power spectrum.