Institute for Advanced Study/Princeton University Early Universe/Cosmology Lunch Discussion

In astrophysics, we often aim to disentangle independent underlying source distributions from incomplete and noisy observations. Examples include deblending galaxies in a crowded field, disentangling the emissions of dark matter candidates from standard astrophysical backgrounds, and separating telluric lines from stellar variability. Recent advances have shown that diffusion models can learn complex prior distributions directly from corrupted data. Building on this insight, we introduce a novel diffusion-based method for source separation. We show that the independence of sources allows us to directly sample from the joint posterior of all the sources given an observation. We leverage these samples within an expectation‐maximization framework to simultaneously infer the prior distributions of multiple sources. Our method accommodates variable-resolution observations and can recover distinct sources even on challenging imaging problems. Our results highlight the potential to extract data-driven priors in the era of large-scale sky surveys.