Anbinh93/image_restoration

SCDC image restoration data bundle

Exactly the files used by the SCDC project (blind degradation graph encoder + unified two-step ResShift restorer): https://github.com/anbinh93/SCDC

Archive	Content
Rain13K.tar	Rain13K pairs referenced by the Phase-1 splits (train / val / locked Test100+Test1200) and the Phase-2 manifest
RESIDE_used.tar	RESIDE hazy/clear files used for training/validation plus the 300 held-out evaluation scenes
LOLv2.tar	LOLv2-real, train (689) + test (100)
SIDD_Small.tar	SIDD Small sRGB, all 160 scenes (121 train / 19 val / 20 locked = scene 008)
Snow100K_gt.tar, Snow100K_synthetic.tar	Snow100K training gt (clean pool) and synthetic snow (real-unknown pool)
manifests/	train_val_manifest_v1.json (Phase-2 train/val, sha digests), rain13k_quarantine_v1.json, per-split file lists, path_roots.json

Paths inside the archives are relative to two roots (see manifests/path_roots.json): /home/admin_mcn/data/ -> data/ and /home/admin_mcn/.cache/kagglehub/ -> kagglehub/. To reuse the manifests on another machine, extract everything under one directory and rewrite the two absolute prefixes in the JSON files to that directory.

All datasets belong to their original authors and keep their original licenses (research use): Rain13K (Jiang et al. / MPRNet compilation), RESIDE (Li et al.), LOLv2 (Yang et al.), SIDD (Abdelhamed et al.), Snow100K (Liu et al.).

Completeness patches (2026-09-05)

• Rain13K_missing.tar: the rest of the Kaggle rain13k-resized folder (Rain100H/L, Test2800, quarantined train pairs) — the split loader validates the whole folder structure, extract it over Rain13K.tar.
• RESIDE_fix.tar: 5 hazy files of the first-5000 RESIDE window that the Phase-1 split drops (their absence shifts the window and leaks 5 held-out scenes into training). Extract both patches before building the splits.