{
  "paper": {
    "title": "PhyAVBench: A Challenging Audio Physics-Sensitivity Benchmark for Physically Grounded Text-to-Audio-Video Generation",
    "authors_core": [
      "Tianxin Xie",
      "Wentao Lei",
      "Kai Jiang",
      "Guanjie Huang",
      "Pengfei Zhang"
    ],
    "corresponding": "Li Liu",
    "affiliations": [
      "HKUST(GZ)",
      "Tencent"
    ],
    "year": 2026,
    "arxiv": "2512.23994",
    "arxiv_version_in_source": "v4 [cs.SD] 18 May 2026",
    "pdf": "https://arxiv.org/pdf/2512.23994",
    "html": "https://arxiv.org/html/2512.23994",
    "code": "https://github.com/imxtx/PhyAVBench"
  },
  "claim": {
    "one_sentence": "Previous benchmarks primarily focus on audio-video temporal synchronization, while largely overlooking explicit evaluation of audio-physics grounding.",
    "source": "Abstract"
  },
  "dataset": {
    "name": "PhyAV-Sound-11K",
    "hours": 25.5,
    "videos": 11605,
    "participants": 184,
    "prompt_groups": 337,
    "avg_videos_per_prompt": 17,
    "gt_videos_per_prompt_range": "2∼40",
    "dimensions": 6,
    "test_points": 41,
    "sources": [
      "Abstract",
      "Section III-A",
      "Table I",
      "Contribution 4"
    ]
  },
  "numbers": [
    {
      "value": "25.5 hours / 11,605 videos / 184 participants",
      "what": "PhyAV-Sound-11K scale",
      "where": "Abstract; Section III-A"
    },
    {
      "value": "337 paired-prompt groups",
      "what": "Human-refined prompt groups",
      "where": "Abstract; Section III-A"
    },
    {
      "value": "average of 17 videos per prompt",
      "what": "Contribution 4 clarifies per prompt, not per group. 11605/(337×2)≈17.22",
      "where": "Abstract (ambiguous 'each'); Contribution 4"
    },
    {
      "value": "2∼40 GT videos per prompt",
      "what": "Table I range",
      "where": "Table I"
    },
    {
      "value": "6 dimensions / 41 test points",
      "what": "Audio-physics hierarchy",
      "where": "Abstract; Section III-A; Table IV"
    },
    {
      "value": "17 SOTA models",
      "what": "Abstract and Conclusion say 17; Section I says over 17; Table II has 17 rows / 15 unique systems (Ovi, LTX appear in T2AV and I2AV)",
      "where": "Abstract; Section I; Section V; Table II"
    },
    {
      "value": "74 participants",
      "what": "Human study headcount",
      "where": "Abstract; Section IV-B; Appendix B-B"
    },
    {
      "value": "20 prompt groups × 10 T2AV systems incl. GT × 3 questions = 600; 419 valid responses after QC",
      "what": "Human eval protocol",
      "where": "Appendix B-B"
    },
    {
      "value": "Sora 2 CPRS 0.4512",
      "what": "Best commercial T2AV CPRS on main table (CLAP embedding)",
      "where": "Section I; Table II; Section IV-C2"
    },
    {
      "value": "JavisDiT++ CPRS 0.2844",
      "what": "Open-source T2AV example the authors contrast with Sora 2",
      "where": "Table II; Section IV-C2"
    },
    {
      "value": "MMAudio CPRS 0.4003",
      "what": "Best V2A CPRS",
      "where": "Table II; Section IV-C2"
    },
    {
      "value": "closed-source: 100 random prompt pairs, T2AV only; open-source: all 337 groups",
      "what": "Evaluation budget split",
      "where": "Section IV-A"
    },
    {
      "value": "k=5",
      "what": "Gaussian kernel width in CPRS magnitude term",
      "where": "Section IV-B eq. 6"
    },
    {
      "value": "Table III P.C. vs PVR-MOS: FAD 0.51, TA-IB 0.58, CLAP 0.84, AV-IB 0.76, DeSync 0.62, AVH 0.72, Javis 0.71, CPRS 0.92",
      "what": "Automatic metrics vs human physical-variation MOS",
      "where": "Table III"
    },
    {
      "value": "Table VIII: LAION-CLAP 0.9230, ImageBind 0.9457, CAV-MAE-Sync 0.8927",
      "what": "CPRS–PVR-MOS correlation under three embeddings",
      "where": "Appendix B-A Table VIII"
    },
    {
      "value": "8,808 per-sample CPRS records",
      "what": "Difficulty distribution input; text says 'Table 1 of the main paper' but the eval table is Table II",
      "where": "Appendix A-C"
    },
    {
      "value": "m01 63.284k s; m02 16.009k; m03 6.058k; m05 3.404k; m06 2.234k; m04 1.148k",
      "what": "Duration by major dimension (sum 92.137k s ≈ 25.59 h, rounded to 25.5 h in Abstract)",
      "where": "Appendix A-B; Figure 7"
    },
    {
      "value": "Mean CPRS by dimension 0.3302–0.3558; m03 0.3302 hardest; m04 0.3558 easiest among the six",
      "what": "Difficulty at major-dimension level",
      "where": "Appendix A-C"
    },
    {
      "value": "c12 0.2972; c10 0.2976; c09 0.4072; c19 0.4097",
      "what": "Hardest / easiest named subcategories",
      "where": "Appendix A-C"
    },
    {
      "value": "t05 Shape 0.2854; t26 Diffraction 0.2891; t47 Freezing/Shattering 0.4752",
      "what": "Named test-point mean CPRS",
      "where": "Appendix A-C"
    },
    {
      "value": "Fig. 3 Hunyuan-Foley CPRS 0.799 and 0.642; JavisDiT 0.262 / 0.270; LTX 0.253 / 0.267",
      "what": "Qualitative examples (Helmholtz / occlusion), not Table II averages",
      "where": "Figure 3"
    },
    {
      "value": "Sora 2 CPRS 0.4512 (CLAP) / 0.4288 (ImageBind) / 0.3479 (CAV-MAE-Sync)",
      "what": "Same model, three embeddings; under CAV-MAE-Sync Wan 2.6 0.3595 ranks above Sora 2",
      "where": "Tables V–VII"
    }
  ],
  "table2_cprs": {
    "T2AV": {
      "Sora 2": 0.4512,
      "Wan 2.6": 0.412,
      "Seedance 1.5 Pro": 0.4044,
      "Kling v2.6": 0.3718,
      "Veo 3.1": 0.3563,
      "Ovi": 0.329,
      "LTX": 0.3235,
      "JavisDiT": 0.2938,
      "JavisDiT++": 0.2844
    },
    "I2AV": {
      "MOVA": 0.3613,
      "LTX": 0.3285,
      "Ovi": 0.3198,
      "UniVerse-1": 0.2729
    },
    "V2A": {
      "MMAudio": 0.4003,
      "HunyuanVideo-Foley": 0.3896,
      "ThinkSound": 0.3186,
      "FoleyCrafter": 0.3019
    }
  },
  "table2_full": {
    "note": "Copied from Table II / arXiv HTML. Closed-source T2AV on 100 pairs; open-source on 337 groups.",
    "rows": [
      {
        "model": "Sora 2",
        "task": "T2AV-C",
        "FVD": 578.7,
        "KVD": 72.04,
        "FAD": 4.25,
        "TV-IB": 0.3415,
        "TA-IB": 0.1465,
        "CLIP": 0.2713,
        "CLAP": 0.0314,
        "AV-IB": 0.2137,
        "DeSync": 0.031,
        "AVH": 0.2118,
        "Javis": 0.1803,
        "CPRS": 0.4512
      },
      {
        "model": "Wan 2.6",
        "task": "T2AV-C",
        "FVD": 656.62,
        "KVD": 108.25,
        "FAD": 1.48,
        "TV-IB": 0.3434,
        "TA-IB": 0.1467,
        "CLIP": 0.2792,
        "CLAP": 0.0313,
        "AV-IB": 0.2143,
        "DeSync": 0.237,
        "AVH": 0.2141,
        "Javis": 0.1846,
        "CPRS": 0.412
      },
      {
        "model": "Seedance 1.5 Pro",
        "task": "T2AV-C",
        "FVD": 968.03,
        "KVD": 315.38,
        "FAD": 3.09,
        "TV-IB": 0.3274,
        "TA-IB": 0.1836,
        "CLIP": 0.2663,
        "CLAP": 0.0186,
        "AV-IB": 0.2794,
        "DeSync": 0.0095,
        "AVH": 0.2754,
        "Javis": 0.2412,
        "CPRS": 0.4044
      },
      {
        "model": "Kling v2.6",
        "task": "T2AV-C",
        "FVD": 737.52,
        "KVD": 162.64,
        "FAD": 1.48,
        "TV-IB": 0.3373,
        "TA-IB": 0.1652,
        "CLIP": 0.2725,
        "CLAP": 0.0261,
        "AV-IB": 0.2654,
        "DeSync": 0.1995,
        "AVH": 0.2756,
        "Javis": 0.2355,
        "CPRS": 0.3718
      },
      {
        "model": "Veo 3.1",
        "task": "T2AV-C",
        "FVD": 755.51,
        "KVD": 189.83,
        "FAD": 3.48,
        "TV-IB": 0.3394,
        "TA-IB": 0.1123,
        "CLIP": 0.2723,
        "CLAP": 0.0376,
        "AV-IB": 0.2136,
        "DeSync": 0.031,
        "AVH": 0.2211,
        "Javis": 0.195,
        "CPRS": 0.3563
      },
      {
        "model": "Ovi",
        "task": "T2AV-O",
        "FVD": 529.44,
        "KVD": 238.87,
        "FAD": 2.7,
        "TV-IB": 0.3125,
        "TA-IB": 0.0536,
        "CLIP": 0.2632,
        "CLAP": 0.0021,
        "AV-IB": 0.1509,
        "DeSync": 0.3421,
        "AVH": 0.1578,
        "Javis": 0.1355,
        "CPRS": 0.329
      },
      {
        "model": "LTX",
        "task": "T2AV-O",
        "FVD": 682.38,
        "KVD": 175.86,
        "FAD": 5.76,
        "TV-IB": 0.312,
        "TA-IB": 0.115,
        "CLIP": 0.2618,
        "CLAP": 0.0021,
        "AV-IB": 0.1178,
        "DeSync": 0.2488,
        "AVH": 0.1332,
        "Javis": 0.1039,
        "CPRS": 0.3235
      },
      {
        "model": "JavisDiT",
        "task": "T2AV-O",
        "FVD": 656.82,
        "KVD": 244.63,
        "FAD": 3.29,
        "TV-IB": 0.3131,
        "TA-IB": 0.1387,
        "CLIP": 0.2594,
        "CLAP": -0.0167,
        "AV-IB": 0.1293,
        "DeSync": 0.2502,
        "AVH": 0.132,
        "Javis": 0.1148,
        "CPRS": 0.2938
      },
      {
        "model": "JavisDiT++",
        "task": "T2AV-O",
        "FVD": 474.62,
        "KVD": 76.89,
        "FAD": 4.55,
        "TV-IB": 0.3303,
        "TA-IB": 0.1182,
        "CLIP": 0.2693,
        "CLAP": 0.0072,
        "AV-IB": 0.1553,
        "DeSync": 0.4903,
        "AVH": 0.1776,
        "Javis": 0.1498,
        "CPRS": 0.2844
      },
      {
        "model": "MOVA",
        "task": "I2AV",
        "FVD": 157.43,
        "KVD": 14.17,
        "FAD": 1.37,
        "TV-IB": 0.3116,
        "TA-IB": 0.1204,
        "CLIP": 0.2676,
        "CLAP": 0.0135,
        "AV-IB": 0.2116,
        "DeSync": 0.7788,
        "AVH": 0.211,
        "Javis": 0.1679,
        "CPRS": 0.3613
      },
      {
        "model": "LTX",
        "task": "I2AV",
        "FVD": 254.35,
        "KVD": 45.99,
        "FAD": 7.53,
        "TV-IB": 0.3094,
        "TA-IB": 0.098,
        "CLIP": 0.2641,
        "CLAP": 0.0201,
        "AV-IB": 0.0964,
        "DeSync": 0.291,
        "AVH": 0.101,
        "Javis": 0.0772,
        "CPRS": 0.3285
      },
      {
        "model": "Ovi",
        "task": "I2AV",
        "FVD": 122.79,
        "KVD": 7.46,
        "FAD": 3.15,
        "TV-IB": 0.3049,
        "TA-IB": 0.0493,
        "CLIP": 0.2645,
        "CLAP": -0.0008,
        "AV-IB": 0.1526,
        "DeSync": 0.3372,
        "AVH": 0.1579,
        "Javis": 0.1373,
        "CPRS": 0.3198
      },
      {
        "model": "UniVerse-1",
        "task": "I2AV",
        "FVD": 340.2,
        "KVD": 32.14,
        "FAD": 5.84,
        "TV-IB": 0.2573,
        "TA-IB": 0.064,
        "CLIP": 0.2338,
        "CLAP": -0.0076,
        "AV-IB": 0.0792,
        "DeSync": 0.705,
        "AVH": 0.0806,
        "Javis": 0.0565,
        "CPRS": 0.2729
      },
      {
        "model": "MMAudio",
        "task": "V2A",
        "FVD": null,
        "KVD": null,
        "FAD": 1.36,
        "TV-IB": null,
        "TA-IB": 0.1588,
        "CLIP": null,
        "CLAP": 0.0265,
        "AV-IB": 0.2684,
        "DeSync": 0.2896,
        "AVH": 0.2697,
        "Javis": 0.22,
        "CPRS": 0.4003
      },
      {
        "model": "HunyuanVideo-Foley",
        "task": "V2A",
        "FVD": null,
        "KVD": null,
        "FAD": 1.64,
        "TV-IB": null,
        "TA-IB": 0.1475,
        "CLIP": null,
        "CLAP": 0.0272,
        "AV-IB": 0.2721,
        "DeSync": 0.307,
        "AVH": 0.2727,
        "Javis": 0.226,
        "CPRS": 0.3896
      },
      {
        "model": "ThinkSound",
        "task": "V2A",
        "FVD": null,
        "KVD": null,
        "FAD": 1.6,
        "TV-IB": null,
        "TA-IB": 0.1161,
        "CLIP": null,
        "CLAP": 0.0228,
        "AV-IB": 0.1874,
        "DeSync": 0.7138,
        "AVH": 0.1941,
        "Javis": 0.1458,
        "CPRS": 0.3186
      },
      {
        "model": "FoleyCrafter",
        "task": "V2A",
        "FVD": null,
        "KVD": null,
        "FAD": 1.88,
        "TV-IB": null,
        "TA-IB": 0.1457,
        "CLIP": null,
        "CLAP": 0.0175,
        "AV-IB": 0.252,
        "DeSync": 0.8878,
        "AVH": 0.258,
        "Javis": 0.2101,
        "CPRS": 0.3019
      }
    ]
  },
  "unverified_or_inconsistent": [
    "Abstract 'each grounded with an average of 17 videos' is ambiguous (group vs prompt). Contribution 4: each prompt, average 17. 11605/337≈34.4 per group; 11605/(337×2)≈17.22 per prompt.",
    "Section I: 'over 17 SOTA models' vs Abstract/Conclusion '17'. Table II has 17 rows; unique systems are 15 (Ovi and LTX listed under both T2AV and I2AV).",
    "Section IV-C: 'the TV-IB score suggests that Seedance 1.5 Pro outperforms the other models' — Table II Seedance TV-IB 0.3274 is the lowest commercial T2AV TV-IB; its TA-IB 0.1836 is the highest. Likely a TV-IB/TA-IB slip.",
    "Section IV-C numbers the AV-sync paragraph as '3) Audio-Visual Semantic Alignment', repeating the previous heading; the paragraph is about synchronization.",
    "Equation (1) writes ground-truth pairs as (a_gt,i, a_gt,i); e_b,gt uses b_gt,i. Typo for (a_gt,i, b_gt,i).",
    "Taxonomy (Section III-B, Table I): sixth dimension 'Complex Coupling'. Table IV header: 'Complex Phenomena' (m06), with a subcategory also named Complex Coupling (c21).",
    "Appendix A-A: 'each comprising 2–4 subcategories'. Table IV lists 19 subcategory codes (c01–c13, c15, c17–c21). Appendix A-C says '21 divisions'.",
    "Table IV t-codes are not contiguous: missing t30, t35–t37, t40–t41, t43, t45, t49; also missing c14, c16. Count of listed t-codes is 41.",
    "Appendix A-C: 'evaluated in Table 1 of the main paper' — main eval table is Table II (Table I is the benchmark comparison).",
    "Closed-source scores are on a random 100-pair subset; open-source on all 337. Table II prints them in one table without a subset tag.",
    "Main-table Sora 2 CPRS 0.4512 uses CLAP. Under CAV-MAE-Sync (Table VII) Sora 2 is 0.3479 and Wan 2.6 0.3595 ranks first in T2AV. Authors say rankings remain largely consistent; the T2AV top-1 does change.",
    "Table III CPRS 0.92 vs Table VIII LAION-CLAP 0.9230 — rounding, not a contradiction.",
    "Appendix duration sum 92.137k s = 25.5936 h vs Abstract 25.5 h — rounding.",
    "Human MOS absolute scores appear only in Figure 5; running text gives ranks (Wan 2.6 first, then Sora 2 and Kling v2.6), not Arabic MOS values. This reading does not invent those pixels.",
    "'Zero data leakage' is a collection claim (newly recorded). No audit against commercial training corpora is reported.",
    "Appendix A-B t24 duration 0.088k s is named as an example of a rare test point."
  ]
}