Tuned Reconstruction Defaults

The committed JSON registry is the inference source of truth. This page contains the complete selected-value table, model-specific exceptions, inheritance behaviour, and command for auditing the validation-only search.

Defaults and selection procedure

The authoritative registry is config/reconstruction_hparam_defaults.json. The complete corrected-validation protocol, candidate matrix, inheritance rules, and command for reproducing every tuning row are recorded in tuning/TUNING.md. The matrix resolves an exact experiment/model record first, then any configured high-view fallback, then a consensus record, and finally the implementation default. Untuned 60-view and limited-angle rows therefore use an available 20-view record or the 8/20-view consensus. Native 512 PaDIS-DPS has dedicated records.

zeta is the data-consistency coefficient, epsilon is the Langevin/DPS coefficient multiplying the sigma-dependent step, and r is predictor- corrector SNR. Values across columns are not directly comparable because each track defines the data update differently.

Method/prior	Paper-Style	Public-Compatible	LION-Physics
Whole-image DPS	zeta=0.01, epsilon=0.5	Not implemented	zeta=4.0, epsilon=0.5
PaDIS-DPS	zeta=0.0075, epsilon=0.5, noise init	zeta=0.2, epsilon=0.5, FDK init	zeta=4.25, epsilon=0.5, noise init
Patch Langevin	zeta=0.03, epsilon=0.5	zeta=0.2, epsilon=0.5	zeta=4.0, epsilon=0.5
Whole-image Langevin	Not scheduled	Not implemented	zeta=3.5, epsilon=0.5
Patch predictor-corrector	zeta=0.02, r=0.04	zeta=0.5, r=0.16	zeta=4.25, r=0.01
Whole-image predictor-corrector	Not scheduled	Not implemented	zeta=4.25, r=0.02
Patch VE-DDNM	epsilon=0.1, corrected-state clip	epsilon=0.2	epsilon=0.1, noise scale 0.5
Whole-image VE-DDNM	Not scheduled	Not implemented	epsilon=0.2
Patch averaging	Not implemented	zeta=0.3, epsilon=0.5	zeta=4.0, epsilon=0.5
Patch stitching	Not implemented	zeta=0.3, epsilon=0.5	zeta=3.0, epsilon=0.5
TV	Not implemented	Not implemented	lambda=0.001, 1000 iterations
PnP-ADMM	Not implemented	Not implemented	eta=3e-5, 60 outer iterations, at most 50 CG iterations, tolerance 1e-7

Native 512 PaDIS-DPS uses zeta=2.0, epsilon=0.5, and Gaussian initialisation for LION-Physics; Public-Compatible uses zeta=1.2, epsilon=0.5, and FDK initialisation. Both process one patch at a time and use activation checkpointing. FDK initialisation, where selected, uses a Hann filter with frequency scaling 0.3 and clips the initial image to [0,1].

Re-running the selection procedure

The generic validation matrix and its inheritance rules are documented in tuning/TUNING.md. Run it against the completed training root used for inference:

conda run -n lion-dev python -u \
  scripts/paper_scripts/PaDIS-Reproduction/tuning/PaDIS_run_reproduction_tuning.py \
  --training-root "$PADIS_TRAIN_ROOT" \
  --output-root "$LION_EXPERIMENTS_PATH/PaDIS/hparam_tuning/reproduction"

The launcher forces the min_intense_val diffusion checkpoint policy, uses the corrected validation split, and emits a manifest, exact command record, summary JSON/CSV, and logs for every sweep. Existing metrics are reused unless --rerun-existing is supplied. It is intentionally separate from the normal pipeline: the pipeline consumes the committed registry so a complete inference run cannot silently change merely because a tuning rerun produces a different finite-sample ordering.

PaDIS_hparam_defaults.py remains available for inspecting or staging a registry from completed tuner records. Do not overwrite the committed registry without reviewing its exact-model and inheritance records; that JSON, rather than an automatically exported partial sweep, is the inference source of truth.