{
  "dataset": "AVT-VQDB-UHD-1 public test-1 subset",
  "videos": 120,
  "sources": {
    "bigbuck_bunny_8bit": "Animation",
    "cutting_orange_tuil": "Cutting orange",
    "vegetables_tuil": "Vegetables",
    "water_netflix": "Water"
  },
  "selection": "All120 public clips; all30 variants of all4 publicly available sources. No metric-based subset selection.",
  "sampling": {
    "bigbuck_bunny_8bit": [
      0,
      54,
      109,
      163,
      218,
      272,
      327,
      381,
      436,
      490,
      545,
      599
    ],
    "cutting_orange_tuil": [
      0,
      54,
      109,
      163,
      218,
      272,
      327,
      381,
      436,
      490,
      545,
      599
    ],
    "vegetables_tuil": [
      0,
      54,
      109,
      163,
      218,
      272,
      327,
      381,
      436,
      490,
      545,
      599
    ],
    "water_netflix": [
      0,
      54,
      109,
      163,
      217,
      272,
      326,
      381,
      435,
      489,
      544,
      598
    ]
  },
  "input": "Full field of view; bicubic common3840x2160 canvas then Lanczos448x252 RGB; same input frames for all metrics.",
  "alignment": "Paired decoded-frame indices at matching frame rate, ignoring container timestamp origin. No optical flow or content alignment.",
  "model": "Unchanged DINOv2-S/14 block1, CLS-target lens64, sigma2 chroma preprocessing; spatial image JLD only, no CLS output term or temporal encoder.",
  "raw_baseline": "Same block and chroma preprocessing as JLD; omit projection only.",
  "pooling": "RMS over patches and sampled frames for pixel, raw and JLD; mean frame LPIPS AlexNet.",
  "labels": "Unmodified laboratory MOS and CI from metadata/test_1_mos_ci.csv; larger MOS better.",
  "pair_rule": "All within-source pairs with nonoverlapping published MOS+-CI intervals; ties from a metric count as incorrect.",
  "visual_rule": "Per source, closest pixel PSNR among eligible pairs with MOS gap at least0.75; break ties by IDs. No raw,LPIPS or JLD score used.",
  "limitations": "Four source scenes only; pair observations share videos and are dependent; no statistical significance claim. Resizing and sparse frame sampling differ from subjective display conditions. No regional human ratings."
}