Can we trust reconstructions from generative image models?

Generative models fill in missing data so convincingly that reconstructions can look plausible whether they're actually supported by measurements or just imagined by the prior. This is dangerous in medical imaging. The authors prove that a measure of how well measurements align with the prior's important directions controls reconstruction error, and use this to design better sampling patterns. On fastMRI, their adaptive measurement method outperforms standard variable-density and Poisson sampling without retraining.