Diffusion model workflow: from dirty image to models outputs and uncertainty estimation. Top row: (1) Original dirty image (conditioning), with true sources circled in white. (2) The sky model showing sources detected by the Photutils localization algorithm. (3) The post-processed predicted sky model, aggregated using the median from 20 network outputs, with Photutils-localized sources. (4) Another version of the post-processed predicted sky model, this time aggregated pixel-wise using the mean from network outputs, again with Photutils-localized sources. (5) The standard deviation of multiple predictions before the post-processing step. We refer to this latter as the uncertainty estimation; it is unitless, with brighter colors (like yellow) indicating greater uncertainty. Rows 2–5 present 20 distinct outputs of the diffusion model for the corresponding dirty image, each with a different model noise initialization.