FB-GDM Removes Hyperparameter Tuning in Guided Diffusion

FB-GDM introduces a fully-Bayesian method for high-dimensional linear inverse problems that eliminates the need for task-specific hyperparameter tuning. By treating precision parameters as latent variables via variational inference, it achieves full image resolution scaling at a computational cost comparable to standard pseudoinverse-guided diffusion.