Principled Design of Diffusion-based Optimizers for Inverse Problems
Score-based diffusion models are used for inverse problems but suffer from slow inference and difficult hyperparameter tuning. OptDiff introduces principled reparameterizations for noise schedule and sampler hyperparameters to ensure invariance across problem configurations. This allows for efficient, decoupled tuning using a small set of samples, resulting in significant speedups and improved image quality in tasks like MRI reconstruction.