Spaces:
Sleeping
Sleeping
File size: 6,897 Bytes
91d49eb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 | {
"model": "split",
"paper": "SPLIT: Training-Free AI-Generated and Partially Edited Video Detection via Spatial Patch-Level Incoherence and Temporal Roughness (Hyun & Kim, ECCV 2026)",
"arxiv_id": "2607.02886",
"official_repo": "https://github.com/mldljyh/SPLIT",
"training_set_in_paper": "NONE - SPLIT is training-free (frozen encoder, real-only threshold calibration).",
"encoder_in_paper": "X-CLIP B/16 (microsoft/xclip-base-patch16) patch tokens; CLIP, DINO and CNN encoders are ablations.",
"paper_protocol": "Official video2frame.py: ffmpeg JPEG frames, 8 fps over 3 s (random start for longer clips); read_video: 16 frames (8 if fewer than 16), cv2 BGR, 10% crop of the longer side, resize 224, ImageNet normalization. Score = TTR**8 * LSMI. Metric: fake recall at FPR 0.1/1/5%, threshold = (1-FPR) quantile of the real scores (eval.py evaluate_split).",
"synthscope_encoder": "X-CLIP B/16, as in the paper. SynthScope's own implementation (the official repo has no license) returns the official SPLIT_model.forward_score on the same frames (relative difference <= 7e-6, job 269966).",
"benchmarks": [
{
"name": "GenVideo (GenVideo-Val, MSR-VTT real calibration, supp. Table 18)",
"datasets": [
"GenVideo-Val"
],
"subsets_or_generators": [
"Crafter",
"Gen2",
"HotShot",
"Lavie",
"ModelScope",
"MoonValley",
"MorphStudio",
"Show_1",
"Sora",
"WildScrape"
],
"metric": "fake recall @ FPR 0.1/1/5% (Overall = all fakes pooled)",
"paper_number": {
"fake_recall_fpr0.1": 82.13,
"fake_recall_fpr1": 91.69,
"fake_recall_fpr5": 97.2
},
"paper_per_generator": {
"Crafter": {
"fpr_0.1": 88.84,
"fpr_1": 96.78,
"fpr_5": 99.64
},
"Gen2": {
"fpr_0.1": 53.26,
"fpr_1": 77.61,
"fpr_5": 94.71
},
"HotShot": {
"fpr_0.1": 99.86,
"fpr_1": 100,
"fpr_5": 100
},
"Lavie": {
"fpr_0.1": 92.5,
"fpr_1": 96.36,
"fpr_5": 98.79
},
"ModelScope": {
"fpr_0.1": 86.71,
"fpr_1": 93.57,
"fpr_5": 97.86
},
"MoonValley": {
"fpr_0.1": 49.36,
"fpr_1": 76.36,
"fpr_5": 95.37
},
"MorphStudio": {
"fpr_0.1": 94.71,
"fpr_1": 98.43,
"fpr_5": 99.71
},
"Show_1": {
"fpr_0.1": 96.86,
"fpr_1": 99.86,
"fpr_5": 100
},
"Sora": {
"fpr_0.1": 87.5,
"fpr_1": 98.21,
"fpr_5": 100
},
"WildScrape": {
"fpr_0.1": 71.66,
"fpr_1": 79.76,
"fpr_5": 85.89
}
},
"official_eval_protocol": "All 10,000 MSR-VTT reals calibrate the threshold; all fakes of the 10 generators scored (official eval.py, no max_len).",
"our_status": "reproduced",
"our_number": {
"official_code_our_data": {
"fpr_0.1": 82.22,
"fpr_1": 90.17,
"fpr_5": 96.89
},
"synthscope": {
"fpr_0.1": 98.87,
"fpr_1": 99.86,
"fpr_5": 99.98
},
"note": "Official code on our data (JPEG frame round-trip, job 270323) reproduces the paper: Overall fake recall 82.22/90.17/96.89 vs 82.13/91.69/97.20. SynthScope decodes frames directly (no JPEG) and reaches 98.87/99.86/99.98 (job 270384; mAP 99.79, AUROC 99.97, acc 99.41 at the calibrated threshold)."
},
"official_code_per_generator": {
"Crafter": {
"fpr_0.1": 89.99,
"fpr_1": 96.21,
"fpr_5": 99.64
},
"Gen2": {
"fpr_0.1": 56.16,
"fpr_1": 76.3,
"fpr_5": 94.64
},
"HotShot": {
"fpr_0.1": 99.86,
"fpr_1": 100.0,
"fpr_5": 100.0
},
"Lavie": {
"fpr_0.1": 93.21,
"fpr_1": 96.07,
"fpr_5": 98.71
},
"ModelScope": {
"fpr_0.1": 87.29,
"fpr_1": 93.14,
"fpr_5": 97.86
},
"MoonValley": {
"fpr_0.1": 52.88,
"fpr_1": 74.76,
"fpr_5": 95.37
},
"MorphStudio": {
"fpr_0.1": 95.0,
"fpr_1": 98.0,
"fpr_5": 99.71
},
"Show_1": {
"fpr_0.1": 97.86,
"fpr_1": 99.71,
"fpr_5": 100.0
},
"Sora": {
"fpr_0.1": 91.07,
"fpr_1": 98.21,
"fpr_5": 100.0
},
"WildScrape": {
"fpr_0.1": 73.27,
"fpr_1": 79.74,
"fpr_5": 86.42
}
},
"our_per_generator": {
"Crafter": {
"fpr_0.1": 100.0,
"fpr_1": 100.0,
"fpr_5": 100.0
},
"Gen2": {
"fpr_0.1": 100.0,
"fpr_1": 100.0,
"fpr_5": 100.0
},
"HotShot": {
"fpr_0.1": 100.0,
"fpr_1": 100.0,
"fpr_5": 100.0
},
"Lavie": {
"fpr_0.1": 99.93,
"fpr_1": 99.93,
"fpr_5": 100.0
},
"ModelScope": {
"fpr_0.1": 99.57,
"fpr_1": 99.86,
"fpr_5": 100.0
},
"MoonValley": {
"fpr_0.1": 100.0,
"fpr_1": 100.0,
"fpr_5": 100.0
},
"MorphStudio": {
"fpr_0.1": 100.0,
"fpr_1": 100.0,
"fpr_5": 100.0
},
"Show_1": {
"fpr_0.1": 100.0,
"fpr_1": 100.0,
"fpr_5": 100.0
},
"Sora": {
"fpr_0.1": 100.0,
"fpr_1": 100.0,
"fpr_5": 100.0
},
"WildScrape": {
"fpr_0.1": 89.94,
"fpr_1": 98.92,
"fpr_5": 99.78
}
},
"dataset_source": "genvideo loader (ModelScope cccnju/Gen-Video GenVideo-Val.zip); reals = MSR-VTT."
},
{
"name": "FakeParts (partial edits; Tables 1, 15-16)",
"datasets": [
"FakeParts"
],
"metric": "fake recall @ FPR",
"paper_number": {
"fake_recall_fpr0.1": 74.45
},
"our_status": "missing",
"dataset_source": "not in SynthScope yet"
},
{
"name": "ViF-Bench (Tables 3, 19)",
"datasets": [
"ViF-Bench"
],
"metric": "fake recall @ FPR",
"paper_number": {
"fake_recall_fpr0.1": 84.52
},
"our_status": "missing",
"dataset_source": "not in SynthScope yet"
}
],
"summary": "Training-free; exact parity with the official model. The official pipeline on our GenVideo copy reproduces the paper's MSR-VTT-calibrated table; SynthScope's direct decoding (no JPEG round-trip) is much stronger at low FPR."
} |