PaDIS Concepts

Patch diffusion prior

PaDIS trains a denoiser over randomly located image patches. Optional position channels encode each patch location in the full image. At inference, shifted patch partitions produce a whole-image score without requiring a whole-image diffusion network.

Noise schedule

Most reported reconstruction experiments use 100 geometrically spaced noise levels from sigma_max=10 to sigma_min=0.002 and 10 inner updates per level. Method-specific exceptions are recorded in the checked-in hyperparameter registry and the PaDIS-Reproduction README.

Implementation conventions

paper

Follows the CT schedule and equations described by Hu et al.

public_repo

Reproduces the released repository mechanics where these differ from the publication.

lion_physics

Uses LION CT operators and operator-normalised data consistency without fitted public-repository scaling constants.

Corrected identifiers

The Chambolle–Pock TV implementation is stored as cp_tv and displayed as CP with prior TV. The limited-angle acquisition is stored as ct_20_limited_angle_120. Legacy admm_tv and ct_fanbeam_180 values remain accepted when reading commands and historical results.