opera-repro-data (private)

Training and test data used by the OPERA re-implementation (huangpengtao00-dotcom/opera-repro). For collaborators only; the images come from research datasets with their own licenses (DIV2K, Flickr2K, MiOIR / MiO100, AgenticIR test split) — do not redistribute.

train/ (synth_5k_135: 1,000 clean images x 5 degraded versions = 5,000 LQ)

• HQ/<id>.png clean reference, DIV2K+Flickr2K crops (<= 1 MP, 64-multiple), 1,000 images
• depth/<id>.mat depth maps used for haze synthesis
• LQ/<id>+<deg1>+<deg2>+...png degraded input; the file name lists the degradations in the order they were applied (tokens: noise, rain, haze, motionblur, defocusblur, lr, jpeg, dark). single:dual:triple = 1:3:5, 8 degradation types, 145 legal ordered sequences (OPERA paper setting), synthesized with the AgenticIR/OPERA pipeline (overlay/synth_smoke.sh).
• The verl training parquet is not included; regenerate with overlay/gen/process_train_data_patched.py --image-dir train/LQ.

test/ (MiO100 benchmark, AgenticIR split, 1,440 images)

• LQ/Group {A,B,C}/<combo>/<id>.png identical (md5) to the AgenticIR / 4KAgent release; HarnessIR uses the same 1,440 list
• HQ64/<id>.png MiO100 clean originals cropped (top-left) to a multiple of 64 so they match the LQ size (x4 for the LR group)
• Evaluation protocol: PSNR/SSIM on the Y channel, per group means, "Avg" = mean of the three groups (overlay/mio_table.py). LQ-input row on this protocol: 19.86 / 0.57 / 0.58 (matches HarnessIR).

dev_mio/ , dev_mio_b/ (development sets; failure analysis and data selection happen here, never on test/)

Same 16 degradation combos as MiO100 Groups A/B/C, synthesized with the training pipeline on MiOIR GT images that neither the training set (GT 1-1000) nor MiO100 uses: dev_mio = GT 2001+ (16 x 20 = 320 LQ), dev_mio_b = GT 2329+ (16 x 50 = 800 LQ). HQ/<id>.png, depth/<id>.mat, LQ/<id>+<tok>+<tok>.png (degradations in applied order), manifest.json (combo -> group, ids). Built by overlay/make_dev_set.sh; the enumeration oracle on them (work/enum_dev_mio*/candidates.jsonl) is in the model repos' train/.

focus/ (targeted training data, 16 x 50 = 800 LQ, MiOIR GT 1001-1808)

Same layout and combos as the dev sets; used for the SFT rounds 3+ labels (opera-planner-sft-r* train/ files reference it).

dev_mio/ , dev_mio_b/ (development sets; failure analysis and data selection happen here, never on test/)

Same 16 degradation combos as MiO100 Groups A/B/C, synthesized with the training pipeline on MiOIR GT images that neither the training set (GT 1-1000) nor MiO100 uses: dev_mio = GT 2001+ (16 x 20 = 320 LQ), dev_mio_b = GT 2329+ (16 x 50 = 800 LQ). HQ/<id>.png, depth/<id>.mat, LQ/<id>+<tok>+<tok>.png (degradations in applied order), manifest.json (combo -> group, ids). Built by overlay/make_dev_set.sh; the enumeration oracle on them (work/enum_dev_mio*/candidates.jsonl) is in the model repos' train/.

focus/ (targeted training data, 16 x 50 = 800 LQ, MiOIR GT 1001-1808)

Same layout and combos as the dev sets; used for the SFT rounds 3+ labels (opera-planner-sft-r* train/ files reference it).

dev_mio/ , dev_mio_b/ (development sets; failure analysis and data selection happen here, never on test/)

Same 16 degradation combos as MiO100 Groups A/B/C, synthesized with the training pipeline on MiOIR GT images that neither the training set (GT 1-1000) nor MiO100 uses: dev_mio = GT 2001+ (16 x 20 = 320 LQ), dev_mio_b = GT 2329+ (16 x 50 = 800 LQ). HQ/<id>.png, depth/<id>.mat, LQ/<id>+<tok>+<tok>.png (degradations in applied order), manifest.json (combo -> group, ids). Built by overlay/make_dev_set.sh; the enumeration oracle on them (work/enum_dev_mio*/candidates.jsonl) is in the model repos' train/.

focus/ (targeted training data, 16 x 50 = 800 LQ, MiOIR GT 1001-1808)

Same layout and combos as the dev sets; used for the SFT rounds 3+ labels (opera-planner-sft-r* train/ files reference it).