Hugging Face's logo

Datasets:
PaddlePaddle
/
Real5-OmniDocBench

Modalities:
Image
Dataset card Data Studio Files
xet
Community
zhouchangda commited on
Commit
0cf9f18
·
1 Parent(s): fae9f4e

update benchmark

Browse files
Files changed (1) hide show
  1. README.md +935 -1
README.md CHANGED
@@ -59,27 +59,961 @@ To illustrate the impact of real-world visual distortions on document parsing, w
59
 
60
  ### 1. Scanning
61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
62
  *In the scanning scenario, the models’ overall performance remains high, with PaddleOCR-VL-1.5 typically achieving the best results.*
63
 
64
  ---
65
 
66
  ### 2. Curving
67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
68
  *Curved documents introduce moderate challenges, causing a slight performance drop, but specialized VLMs maintain competitive accuracy.*
69
 
70
  ---
71
 
72
  ### 3. Screen-photo
73
 
74
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75
 
76
  ### 4. Lighting-variation
77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
78
  *Severe lighting variations present significant challenges, resulting in a noticeable performance drop for most models.*
79
 
80
  ---
81
 
82
  ### 5. Skewing-variation
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
83
  *Skewed documents also impact accuracy, but top-performing models still provide reasonable results.*
84
 
85
 
 
59
 
60
  ### 1. Scanning
61
 
62
+ <table border="1" style="width:100%; border-collapse: collapse;">
63
+ <tr>
64
+ <th><strong>Model Type</strong></th>
65
+ <th><strong>Methods</strong></th>
66
+ <th><strong>Parameters</strong></th>
67
+ <th><strong>Overall&#x2191</strong></th>
68
+ <th><strong>Text<sup>Edit</strong>&#x2193</sup></th>
69
+ <th><strong>Formula<sup>CDM</strong>&#x2191</sup></th>
70
+ <th><strong>Table<sup>TEDS</strong>&#x2191</sup></th>
71
+ <th><strong>Reading Order<sup>Edit</strong>&#x2193</sup></th>
72
+ </tr>
73
+ <tr>
74
+ <td rowspan="2"><strong>Pipeline Tools</strong></td>
75
+ <td>Maker-1.8.2</td>
76
+ <td>-</td>
77
+ <td>70.27</td>
78
+ <td>0.223</td>
79
+ <td>77.03</td>
80
+ <td>56.05</td>
81
+ <td>0.238</td>
82
+ </tr>
83
+ <tr>
84
+ <td>PP-StructureV3</td>
85
+ <td>-</td>
86
+ <td>84.68</td>
87
+ <td>0.094</td>
88
+ <td>84.34</td>
89
+ <td>79.06</td>
90
+ <td>0.092</td>
91
+ </tr>
92
+ <tr>
93
+ <td rowspan="5">General VLMs</td>
94
+ <td>GPT-5.2</td>
95
+ <td>-</td>
96
+ <td>84.43</td>
97
+ <td>0.142</td>
98
+ <td>85.68</td>
99
+ <td>81.78</td>
100
+ <td>0.109</td>
101
+ </tr>
102
+ <tr>
103
+ <td>Qwen2.5-VL-72B</td>
104
+ <td>72B</td>
105
+ <td>86.19</td>
106
+ <td>0.110</td>
107
+ <td>86.14</td>
108
+ <td>83.41</td>
109
+ <td>0.114</td>
110
+ </tr>
111
+ <tr>
112
+ <td>Gemini-2.5 Pro</td>
113
+ <td>-</td>
114
+ <td>89.25</td>
115
+ <td>0.073</td>
116
+ <td>87.44</td>
117
+ <td>87.62</td>
118
+ <td>0.098</td>
119
+ </tr>
120
+ <tr>
121
+ <td>Qwen3-VL-235B-A22B-Instruct</td>
122
+ <td>235B</td>
123
+ <td>89.43</td>
124
+ <td>0.059</td>
125
+ <td>89.01</td>
126
+ <td>85.19</td>
127
+ <td>0.066</td>
128
+ </tr>
129
+ <tr>
130
+ <td>Gemini-3 Pro</td>
131
+ <td>-</td>
132
+ <td>89.47</td>
133
+ <td>0.071</td>
134
+ <td>88.16</td>
135
+ <td>87.37</td>
136
+ <td>0.078</td>
137
+ </tr>
138
+ <tr>
139
+ <td rowspan="12">Specialized VLMs</td>
140
+ <td>Dolphin</td>
141
+ <td>322M</td>
142
+ <td>72.16</td>
143
+ <td>0.154</td>
144
+ <td>64.58</td>
145
+ <td>67.27</td>
146
+ <td>0.130</td>
147
+ </tr>
148
+ <tr>
149
+ <td>Dolphin-1.5 </td>
150
+ <td>0.3B</td>
151
+ <td>83.39</td>
152
+ <td>0.097</td>
153
+ <td>76.25</td>
154
+ <td>83.65</td>
155
+ <td>0.090</td>
156
+ </tr>
157
+ <tr>
158
+ <td>MinerU2-VLM</td>
159
+ <td>0.9B</td>
160
+ <td>83.60</td>
161
+ <td>0.094</td>
162
+ <td>79.76</td>
163
+ <td>80.44</td>
164
+ <td>0.091</td>
165
+ </tr>
166
+ <tr>
167
+ <td>MonkeyOCR-pro-1.2B </td>
168
+ <td>1.9B</td>
169
+ <td>84.64</td>
170
+ <td>0.123</td>
171
+ <td>84.17</td>
172
+ <td>82.13</td>
173
+ <td>0.145</td>
174
+ </tr>
175
+ <tr>
176
+ <td>MonkeyOCR-3B </td>
177
+ <td>3.7B</td>
178
+ <td>84.65</td>
179
+ <td>0.100</td>
180
+ <td>84.16</td>
181
+ <td>79.81</td>
182
+ <td>0.143</td>
183
+ </tr>
184
+ <tr>
185
+ <td>Nanonets-OCR-s</td>
186
+ <td>3B</td>
187
+ <td>85.52</td>
188
+ <td>0.106</td>
189
+ <td>88.09</td>
190
+ <td>79.11</td>
191
+ <td>0.106</td>
192
+ </tr>
193
+ <tr>
194
+ <td>Deepseek-OCR</td>
195
+ <td>3B</td>
196
+ <td>86.17</td>
197
+ <td>0.078</td>
198
+ <td>83.59</td>
199
+ <td>82.69</td>
200
+ <td>0.085</td>
201
+ </tr>
202
+ <tr>
203
+ <td>dots.ocr</td>
204
+ <td>3B</td>
205
+ <td>86.87</td>
206
+ <td>0.083</td>
207
+ <td>83.27</td>
208
+ <td>85.68</td>
209
+ <td>0.081</td>
210
+ </tr>
211
+ <tr>
212
+ <td>MonkeyOCR-pro-3B </td>
213
+ <td>3.7B</td>
214
+ <td>86.94</td>
215
+ <td>0.103</td>
216
+ <td>86.29</td>
217
+ <td>84.86</td>
218
+ <td>0.141</td>
219
+ </tr>
220
+ <tr>
221
+ <td>MinerU2.5</td>
222
+ <td>1.2B</td>
223
+ <td>90.06</td>
224
+ <td>0.052</td>
225
+ <td>88.22</td>
226
+ <td>87.16</td>
227
+ <td>0.050</td>
228
+ </tr>
229
+ <tr>
230
+ <td>PaddleOCR-VL</td>
231
+ <td>0.9B</td>
232
+ <td><u>92.11</u></td>
233
+ <td><u>0.039</u></td>
234
+ <td><u>90.35</u></td>
235
+ <td><u>89.90</u></td>
236
+ <td><u>0.048</u></td>
237
+ </tr>
238
+ <tr>
239
+ <td><strong>PaddleOCR-VL-1.5</strong></td>
240
+ <td>0.9B</td>
241
+ <td><strong>92.92</strong></td>
242
+ <td><strong>0.040</strong></td>
243
+ <td><strong>91.53</strong></td>
244
+ <td><strong>91.26</strong></td>
245
+ <td><strong>0.050</strong></td>
246
+ </tr>
247
+ </table>
248
+
249
  *In the scanning scenario, the models’ overall performance remains high, with PaddleOCR-VL-1.5 typically achieving the best results.*
250
 
251
  ---
252
 
253
  ### 2. Curving
254
 
255
+ <table border="1" style="width:100%; border-collapse: collapse;">
256
+ <tr>
257
+ <th><strong>Model Type</strong></th>
258
+ <th><strong>Methods</strong></th>
259
+ <th><strong>Parameters</strong></th>
260
+ <th><strong>Overall&#x2191</strong></th>
261
+ <th><strong>Text<sup>Edit</strong>&#x2193</sup></th>
262
+ <th><strong>Formula<sup>CDM</strong>&#x2191</sup></th>
263
+ <th><strong>Table<sup>TEDS</strong>&#x2191</sup></th>
264
+ <th><strong>Reading Order<sup>Edit</strong>&#x2193</sup></th>
265
+ </tr>
266
+ <tr>
267
+ <td rowspan="2"><strong>Pipeline Tools</strong></td>
268
+ <td>Maker-1.8.2</td>
269
+ <td>-</td>
270
+ <td>58.98</td>
271
+ <td>0.349</td>
272
+ <td>72.71</td>
273
+ <td>39.08</td>
274
+ <td>0.390</td>
275
+ </tr>
276
+ <tr>
277
+ <td>PP-StructureV3</td>
278
+ <td>-</td>
279
+ <td>59.34</td>
280
+ <td>0.376</td>
281
+ <td>68.22</td>
282
+ <td>47.40</td>
283
+ <td>0.261</td>
284
+ </tr>
285
+ <tr>
286
+ <td rowspan="5">General VLMs</td>
287
+ <td>GPT-5.2</td>
288
+ <td>-</td>
289
+ <td>76.26</td>
290
+ <td>0.239</td>
291
+ <td>80.90</td>
292
+ <td>71.80</td>
293
+ <td>0.165</td>
294
+ </tr>
295
+ <tr>
296
+ <td>Gemini-2.5 Pro</td>
297
+ <td>-</td>
298
+ <td>87.63</td>
299
+ <td>0.092</td>
300
+ <td>86.50</td>
301
+ <td>85.59</td>
302
+ <td>0.109</td>
303
+ </tr>
304
+ <tr>
305
+ <td>Qwen2.5-VL-72B</td>
306
+ <td>72B</td>
307
+ <td>87.77</td>
308
+ <td>0.086</td>
309
+ <td>88.85</td>
310
+ <td>83.06</td>
311
+ <td>0.102</td>
312
+ </tr>
313
+ <tr>
314
+ <td>Gemini-3 Pro</td>
315
+ <td>-</td>
316
+ <td>88.90</td>
317
+ <td>0.086</td>
318
+ <td>88.10</td>
319
+ <td><u>87.20</u></td>
320
+ <td>0.087</td>
321
+ </tr>
322
+ <tr>
323
+ <td>Qwen3-VL-235B-A22B-Instruct</td>
324
+ <td>235B</td>
325
+ <td><u>89.99</u></td>
326
+ <td><u>0.051</u></td>
327
+ <td><u>89.06</u></td>
328
+ <td>85.95</td>
329
+ <td><u>0.064</u></td>
330
+ </tr>
331
+ <tr>
332
+ <td rowspan="12">Specialized VLMs</td>
333
+ <td>Dolphin-1.5 </td>
334
+ <td>0.3B</td>
335
+ <td>50.50</td>
336
+ <td>0.383</td>
337
+ <td>47.24</td>
338
+ <td>42.52</td>
339
+ <td>0.309</td>
340
+ </tr>
341
+ <tr>
342
+ <td>Dolphin</td>
343
+ <td>322M</td>
344
+ <td>60.35</td>
345
+ <td>0.316</td>
346
+ <td>61.06</td>
347
+ <td>51.58</td>
348
+ <td>0.247</td>
349
+ </tr>
350
+ <tr>
351
+ <td>Deepseek-OCR</td>
352
+ <td>3B</td>
353
+ <td>67.20</td>
354
+ <td>0.328</td>
355
+ <td>73.59</td>
356
+ <td>60.80</td>
357
+ <td>0.226</td>
358
+ </tr>
359
+ <tr>
360
+ <td>MinerU2-VLM</td>
361
+ <td>0.9B</td>
362
+ <td>73.73</td>
363
+ <td>0.202</td>
364
+ <td>77.72</td>
365
+ <td>63.65</td>
366
+ <td>0.173</td>
367
+ </tr>
368
+ <tr>
369
+ <td>MonkeyOCR-pro-1.2B </td>
370
+ <td>1.9B</td>
371
+ <td>76.59</td>
372
+ <td>0.196</td>
373
+ <td>78.85</td>
374
+ <td>70.52</td>
375
+ <td>0.221</td>
376
+ </tr>
377
+ <tr>
378
+ <td>MonkeyOCR-3B </td>
379
+ <td>3.7B</td>
380
+ <td>77.27</td>
381
+ <td>0.164</td>
382
+ <td>79.08</td>
383
+ <td>69.18</td>
384
+ <td>0.211</td>
385
+ </tr>
386
+ <tr>
387
+ <td>MonkeyOCR-pro-3B </td>
388
+ <td>3.7B</td>
389
+ <td>78.90</td>
390
+ <td>0.168</td>
391
+ <td>79.55</td>
392
+ <td>73.94</td>
393
+ <td>0.212</td>
394
+ </tr>
395
+ <tr>
396
+ <td>Nanonets-OCR-s</td>
397
+ <td>3B</td>
398
+ <td>83.56</td>
399
+ <td>0.121</td>
400
+ <td>86.24</td>
401
+ <td>76.57</td>
402
+ <td>0.124</td>
403
+ </tr>
404
+ <tr>
405
+ <td>MinerU2.5</td>
406
+ <td>1.2B</td>
407
+ <td>83.76</td>
408
+ <td>0.154</td>
409
+ <td>85.92</td>
410
+ <td>80.71</td>
411
+ <td>0.104</td>
412
+ </tr>
413
+ <tr>
414
+ <td>PaddleOCR-VL </td>
415
+ <td>0.9B</td>
416
+ <td>85.97</td>
417
+ <td>0.093</td>
418
+ <td>85.45</td>
419
+ <td>81.77</td>
420
+ <td>0.092</td>
421
+ </tr>
422
+ <tr>
423
+ <td>dots.ocr</td>
424
+ <td>3B</td>
425
+ <td>86.01</td>
426
+ <td>0.087</td>
427
+ <td>85.03</td>
428
+ <td>81.74</td>
429
+ <td>0.093</td>
430
+ </tr>
431
+ <tr>
432
+ <td><strong>PaddleOCR-VL-1.5</strong></td>
433
+ <td>0.9B</td>
434
+ <td><strong>90.92</strong></td>
435
+ <td><strong>0.061</strong></td>
436
+ <td><strong>91.42</strong></td>
437
+ <td><strong>87.41</strong></td>
438
+ <td><strong>0.078</strong></td>
439
+ </tr>
440
+ </table>
441
+
442
  *Curved documents introduce moderate challenges, causing a slight performance drop, but specialized VLMs maintain competitive accuracy.*
443
 
444
  ---
445
 
446
  ### 3. Screen-photo
447
 
448
+ <table border="1" style="width:100%; border-collapse: collapse;">
449
+ <tr>
450
+ <th><strong>Model Type</strong></th>
451
+ <th><strong>Methods</strong></th>
452
+ <th><strong>Parameters</strong></th>
453
+ <th><strong>Overall&#x2191</strong></th>
454
+ <th><strong>Text<sup>Edit</strong>&#x2193</sup></th>
455
+ <th><strong>Formula<sup>CDM</strong>&#x2191</sup></th>
456
+ <th><strong>Table<sup>TEDS</strong>&#x2191</sup></th>
457
+ <th><strong>Reading Order<sup>Edit</strong>&#x2193</sup></th>
458
+ </tr>
459
+ <tr>
460
+ <td rowspan="2"><strong>Pipeline Tools</strong></td>
461
+ <td>Maker-1.8.2</td>
462
+ <td>-</td>
463
+ <td>63.65</td>
464
+ <td>0.290</td>
465
+ <td>72.73</td>
466
+ <td>47.21</td>
467
+ <td>0.325</td>
468
+ </tr>
469
+ <tr>
470
+ <td>PP-StructureV3</td>
471
+ <td>-</td>
472
+ <td>66.89</td>
473
+ <td>0.204</td>
474
+ <td>73.26</td>
475
+ <td>47.82</td>
476
+ <td>0.165</td>
477
+ </tr>
478
+ <tr>
479
+ <td rowspan="5">General VLMs</td>
480
+ <td>GPT-5.2</td>
481
+ <td>-</td>
482
+ <td>76.75</td>
483
+ <td>0.208</td>
484
+ <td>79.27</td>
485
+ <td>71.73</td>
486
+ <td>0.148</td>
487
+ </tr>
488
+ <tr>
489
+ <td>Qwen2.5-VL-72B</td>
490
+ <td>72B</td>
491
+ <td>86.48</td>
492
+ <td>0.100</td>
493
+ <td>87.46</td>
494
+ <td>82.00</td>
495
+ <td>0.102</td>
496
+ </tr>
497
+ <tr>
498
+ <td>Gemini-2.5 Pro</td>
499
+ <td>-</td>
500
+ <td>87.11</td>
501
+ <td>0.103</td>
502
+ <td>85.30</td>
503
+ <td>86.31</td>
504
+ <td>0.117</td>
505
+ </tr>
506
+ <tr>
507
+ <td>Gemini-3 Pro</td>
508
+ <td>-</td>
509
+ <td>88.86</td>
510
+ <td>0.084</td>
511
+ <td>87.33</td>
512
+ <td><u>87.65</u></td>
513
+ <td>0.087</td>
514
+ </tr>
515
+ <tr>
516
+ <td>Qwen3-VL-235B-A22B-Instruct</td>
517
+ <td>235B</td>
518
+ <td><u>89.27</u></td>
519
+ <td><u>0.068</u></td>
520
+ <td><u>88.72</u></td>
521
+ <td>85.85</td>
522
+ <td><u>0.071</u></td>
523
+ </tr>
524
+ <tr>
525
+ <td rowspan="12">Specialized VLMs</td>
526
+ <td>Dolphin</td>
527
+ <td>322M</td>
528
+ <td>64.29</td>
529
+ <td>0.232</td>
530
+ <td>58.66</td>
531
+ <td>57.38</td>
532
+ <td>0.195</td>
533
+ </tr>
534
+ <tr>
535
+ <td>Dolphin-1.5 </td>
536
+ <td>0.3B</td>
537
+ <td>69.76</td>
538
+ <td>0.205</td>
539
+ <td>61.80</td>
540
+ <td>68.00</td>
541
+ <td>0.177</td>
542
+ </tr>
543
+ <tr>
544
+ <td>Deepseek-OCR</td>
545
+ <td>3B</td>
546
+ <td>75.31</td>
547
+ <td>0.220</td>
548
+ <td>77.68</td>
549
+ <td>70.26</td>
550
+ <td>0.169</td>
551
+ </tr>
552
+ <tr>
553
+ <td>MinerU2-VLM</td>
554
+ <td>0.9B</td>
555
+ <td>78.77</td>
556
+ <td>0.139</td>
557
+ <td>79.02</td>
558
+ <td>71.17</td>
559
+ <td>0.123</td>
560
+ </tr>
561
+ <tr>
562
+ <td>MonkeyOCR-pro-1.2B </td>
563
+ <td>1.9B</td>
564
+ <td>80.24</td>
565
+ <td>0.148</td>
566
+ <td>80.78</td>
567
+ <td>74.74</td>
568
+ <td>0.179</td>
569
+ </tr>
570
+ <tr>
571
+ <td>MonkeyOCR-3B </td>
572
+ <td>3.7B</td>
573
+ <td>80.71</td>
574
+ <td>0.122</td>
575
+ <td>81.33</td>
576
+ <td>73.04</td>
577
+ <td>0.177</td>
578
+ </tr>
579
+ <tr>
580
+ <td>MonkeyOCR-pro-3B </td>
581
+ <td>3.7B</td>
582
+ <td>82.44</td>
583
+ <td>0.124</td>
584
+ <td>81.55</td>
585
+ <td>78.13</td>
586
+ <td>0.177</td>
587
+ </tr>
588
+ <tr>
589
+ <td>PaddleOCR-VL </td>
590
+ <td>0.9B</td>
591
+ <td>82.54</td>
592
+ <td>0.103</td>
593
+ <td>83.58</td>
594
+ <td>74.36</td>
595
+ <td>0.107</td>
596
+ </tr>
597
+ <tr>
598
+ <td>Nanonets-OCR-s</td>
599
+ <td>3B</td>
600
+ <td>84.86</td>
601
+ <td>0.112</td>
602
+ <td>86.65</td>
603
+ <td>79.09</td>
604
+ <td>0.117</td>
605
+ </tr>
606
+ <tr>
607
+ <td>dots.ocr</td>
608
+ <td>3B</td>
609
+ <td>87.18</td>
610
+ <td>0.081</td>
611
+ <td>85.34</td>
612
+ <td>84.26</td>
613
+ <td>0.079</td>
614
+ </tr>
615
+ <tr>
616
+ <td>MinerU2.5</td>
617
+ <td>1.2B</td>
618
+ <td>89.41</td>
619
+ <td>0.062</td>
620
+ <td>87.55</td>
621
+ <td>86.83</td>
622
+ <td>0.053</td>
623
+ </tr>
624
+ <tr>
625
+ <td><strong>PaddleOCR-VL-1.5</strong></td>
626
+ <td>0.9B</td>
627
+ <td><strong>90.09</strong></td>
628
+ <td><strong>0.061</strong></td>
629
+ <td><strong>90.78</strong></td>
630
+ <td><strong>85.97</strong></td>
631
+ <td><strong>0.070</strong></td>
632
+ </tr>
633
+ </table>
634
 
635
  ### 4. Lighting-variation
636
 
637
+ <table border="1" style="width:100%; border-collapse: collapse;">
638
+ <tr>
639
+ <th><strong>Model Type</strong></th>
640
+ <th><strong>Methods</strong></th>
641
+ <th><strong>Parameters</strong></th>
642
+ <th><strong>Overall&#x2191</strong></th>
643
+ <th><strong>Text<sup>Edit</strong>&#x2193</sup></th>
644
+ <th><strong>Formula<sup>CDM</strong>&#x2191</sup></th>
645
+ <th><strong>Table<sup>TEDS</strong>&#x2191</sup></th>
646
+ <th><strong>Reading Order<sup>Edit</strong>&#x2193</sup></th>
647
+ </tr>
648
+ <tr>
649
+ <td rowspan="2"><strong>Pipeline Tools</strong></td>
650
+ <td>Maker-1.8.2</td>
651
+ <td>-</td>
652
+ <td>66.31</td>
653
+ <td>0.259</td>
654
+ <td>74.80</td>
655
+ <td>50.03</td>
656
+ <td>0.337</td>
657
+ </tr>
658
+ <tr>
659
+ <td>PP-StructureV3</td>
660
+ <td>-</td>
661
+ <td>73.38</td>
662
+ <td>0.158</td>
663
+ <td>77.75</td>
664
+ <td>58.19</td>
665
+ <td>0.126</td>
666
+ </tr>
667
+ <tr>
668
+ <td rowspan="5">General VLMs</td>
669
+ <td>GPT-5.2</td>
670
+ <td>-</td>
671
+ <td>80.88</td>
672
+ <td>0.191</td>
673
+ <td>84.41</td>
674
+ <td>77.37</td>
675
+ <td>0.134</td>
676
+ </tr>
677
+ <tr>
678
+ <td>Qwen2.5-VL-72B</td>
679
+ <td>72B</td>
680
+ <td>87.25</td>
681
+ <td>0.087</td>
682
+ <td>86.44</td>
683
+ <td>84.03</td>
684
+ <td>0.097</td>
685
+ </tr>
686
+ <tr>
687
+ <td>Gemini-2.5 Pro</td>
688
+ <td>-</td>
689
+ <td>87.97</td>
690
+ <td>0.083</td>
691
+ <td>86.13</td>
692
+ <td>86.11</td>
693
+ <td>0.103</td>
694
+ </tr>
695
+ <tr>
696
+ <td>Qwen3-VL-235B-A22B-Instruct</td>
697
+ <td>235B</td>
698
+ <td>89.27</td>
699
+ <td>0.060</td>
700
+ <td>87.81</td>
701
+ <td>86.05</td>
702
+ <td>0.070</td>
703
+ </tr>
704
+ <tr>
705
+ <td>Gemini-3 Pro</td>
706
+ <td>-</td>
707
+ <td>89.53</td>
708
+ <td>0.073</td>
709
+ <td>87.78</td>
710
+ <td>88.14</td>
711
+ <td>0.080</td>
712
+ </tr>
713
+ <tr>
714
+ <td rowspan="12">Specialized VLMs</td>
715
+ <td>Dolphin</td>
716
+ <td>322M</td>
717
+ <td>67.29</td>
718
+ <td>0.197</td>
719
+ <td>61.42</td>
720
+ <td>60.10</td>
721
+ <td>0.173</td>
722
+ </tr>
723
+ <tr>
724
+ <td>Dolphin-1.5 </td>
725
+ <td>0.3B</td>
726
+ <td>75.61</td>
727
+ <td>0.159</td>
728
+ <td>70.04</td>
729
+ <td>72.69</td>
730
+ <td>0.133</td>
731
+ </tr>
732
+ <tr>
733
+ <td>Deepseek-OCR</td>
734
+ <td>3B</td>
735
+ <td>78.10</td>
736
+ <td>0.192</td>
737
+ <td>81.71</td>
738
+ <td>71.81</td>
739
+ <td>0.156</td>
740
+ </tr>
741
+ <tr>
742
+ <td>MinerU2-VLM</td>
743
+ <td>0.9B</td>
744
+ <td>80.51</td>
745
+ <td>0.135</td>
746
+ <td>80.72</td>
747
+ <td>74.29</td>
748
+ <td>0.123</td>
749
+ </tr>
750
+ <tr>
751
+ <td>MonkeyOCR-pro-1.2B </td>
752
+ <td>1.9B</td>
753
+ <td>82.11</td>
754
+ <td>0.144</td>
755
+ <td>82.07</td>
756
+ <td>78.67</td>
757
+ <td>0.172</td>
758
+ </tr>
759
+ <tr>
760
+ <td>MonkeyOCR-3B </td>
761
+ <td>3.7B</td>
762
+ <td>83.16</td>
763
+ <td>0.118</td>
764
+ <td>83.63</td>
765
+ <td>77.62</td>
766
+ <td>0.168</td>
767
+ </tr>
768
+ <tr>
769
+ <td>MonkeyOCR-pro-3B </td>
770
+ <td>3.7B</td>
771
+ <td>84.71</td>
772
+ <td>0.120</td>
773
+ <td>84.13</td>
774
+ <td>82.02</td>
775
+ <td>0.171</td>
776
+ </tr>
777
+ <tr>
778
+ <td>Nanonets-OCR-s</td>
779
+ <td>3B</td>
780
+ <td>85.01</td>
781
+ <td>0.099</td>
782
+ <td>87.94</td>
783
+ <td>76.96</td>
784
+ <td>0.112</td>
785
+ </tr>
786
+ <tr>
787
+ <td>dots.ocr</td>
788
+ <td>3B</td>
789
+ <td>87.57</td>
790
+ <td>0.068</td>
791
+ <td>85.07</td>
792
+ <td>84.44</td>
793
+ <td>0.076</td>
794
+ </tr>
795
+ <tr>
796
+ <td>MinerU2.5</td>
797
+ <td>1.2B</td>
798
+ <td>89.57</td>
799
+ <td>0.065</td>
800
+ <td><u>88.36</u></td>
801
+ <td>86.87</td>
802
+ <td>0.062</td>
803
+ </tr>
804
+ <tr>
805
+ <td>PaddleOCR-VL </td>
806
+ <td>0.9B</td>
807
+ <td><u>89.61</u></td>
808
+ <td><u>0.049</u></td>
809
+ <td>86.66</td>
810
+ <td><u>87.02</u></td>
811
+ <td><u>0.055</u></td>
812
+ </tr>
813
+ <tr>
814
+ <td><strong>PaddleOCR-VL-1.5</strong></td>
815
+ <td>0.9B</td>
816
+ <td><strong>92.24</strong></td>
817
+ <td><strong>0.047</strong></td>
818
+ <td><strong>92.47</strong></td>
819
+ <td><strong>88.96</strong></td>
820
+ <td><strong>0.059</strong></td>
821
+ </tr>
822
+ </table>
823
+
824
  *Severe lighting variations present significant challenges, resulting in a noticeable performance drop for most models.*
825
 
826
  ---
827
 
828
  ### 5. Skewing-variation
829
+
830
+ <table border="1" style="width:100%; border-collapse: collapse;">
831
+ <tr>
832
+ <th><strong>Model Type</strong></th>
833
+ <th><strong>Methods</strong></th>
834
+ <th><strong>Parameters</strong></th>
835
+ <th><strong>Overall&#x2191</strong></th>
836
+ <th><strong>Text<sup>Edit</strong>&#x2193</sup></th>
837
+ <th><strong>Formula<sup>CDM</strong>&#x2191</sup></th>
838
+ <th><strong>Table<sup>TEDS</strong>&#x2191</sup></th>
839
+ <th><strong>Reading Order<sup>Edit</strong>&#x2193</sup></th>
840
+ </tr>
841
+ <tr>
842
+ <td rowspan="2"><strong>Pipeline Tools</strong></td>
843
+ <td>PP-StructureV3</td>
844
+ <td>-</td>
845
+ <td>37.98</td>
846
+ <td>0.557</td>
847
+ <td>44.37</td>
848
+ <td>25.27</td>
849
+ <td>0.417</td>
850
+ </tr>
851
+ <tr>
852
+ <td>Maker-1.8.2</td>
853
+ <td>-</td>
854
+ <td>41.27</td>
855
+ <td>0.536</td>
856
+ <td>60.16</td>
857
+ <td>17.23</td>
858
+ <td>0.543</td>
859
+ </tr>
860
+ <tr>
861
+ <td rowspan="5">General VLMs</td>
862
+ <td>GPT-5.2</td>
863
+ <td>-</td>
864
+ <td>75.00</td>
865
+ <td>0.257</td>
866
+ <td>80.27</td>
867
+ <td>70.47</td>
868
+ <td>0.167</td>
869
+ </tr>
870
+ <tr>
871
+ <td>Qwen3-VL-235B-A22B-Instruct</td>
872
+ <td>235B</td>
873
+ <td>86.56</td>
874
+ <td><u>0.077</u></td>
875
+ <td>83.96</td>
876
+ <td>83.41</td>
877
+ <td><u>0.091</u></td>
878
+ </tr>
879
+ <tr>
880
+ <td>Qwen2.5-VL-72B</td>
881
+ <td>72B</td>
882
+ <td>86.90</td>
883
+ <td><u>0.077</u></td>
884
+ <td>87.26</td>
885
+ <td>81.14</td>
886
+ <td><u>0.091</u></td>
887
+ </tr>
888
+ <tr>
889
+ <td>Gemini-2.5 Pro</td>
890
+ <td>-</td>
891
+ <td>89.07</td>
892
+ <td><u>0.077</u></td>
893
+ <td>87.89</td>
894
+ <td>86.99</td>
895
+ <td>0.104</td>
896
+ </tr>
897
+ <tr>
898
+ <td>Gemini-3 Pro</td>
899
+ <td>-</td>
900
+ <td><u>89.45</u></td>
901
+ <td>0.080</td>
902
+ <td><u>88.33</u></td>
903
+ <td><u>88.06</u></td>
904
+ <td>0.092</td>
905
+ </tr>
906
+ <tr>
907
+ <td rowspan="12">Specialized VLMs</td>
908
+ <td>Dolphin-1.5 </td>
909
+ <td>0.3B</td>
910
+ <td>28.16</td>
911
+ <td>0.553</td>
912
+ <td>25.60</td>
913
+ <td>14.18</td>
914
+ <td>0.419</td>
915
+ </tr>
916
+ <tr>
917
+ <td>Dolphin</td>
918
+ <td>322M</td>
919
+ <td>44.83</td>
920
+ <td>0.500</td>
921
+ <td>51.34</td>
922
+ <td>33.22</td>
923
+ <td>0.321</td>
924
+ </tr>
925
+ <tr>
926
+ <td>MonkeyOCR-pro-1.2B </td>
927
+ <td>1.9B</td>
928
+ <td>62.18</td>
929
+ <td>0.292</td>
930
+ <td>66.25</td>
931
+ <td>49.46</td>
932
+ <td>0.317</td>
933
+ </tr>
934
+ <tr>
935
+ <td>Deepseek-OCR</td>
936
+ <td>3B</td>
937
+ <td>63.01</td>
938
+ <td>0.327</td>
939
+ <td>73.27</td>
940
+ <td>48.48</td>
941
+ <td>0.231</td>
942
+ </tr>
943
+ <tr>
944
+ <td>MonkeyOCR-pro-3B </td>
945
+ <td>3.7B</td>
946
+ <td>64.47</td>
947
+ <td>0.251</td>
948
+ <td>69.06</td>
949
+ <td>49.42</td>
950
+ <td>0.301</td>
951
+ </tr>
952
+ <tr>
953
+ <td>MonkeyOCR-3B </td>
954
+ <td>3.7B</td>
955
+ <td>65.67</td>
956
+ <td>0.248</td>
957
+ <td>69.23</td>
958
+ <td>52.59</td>
959
+ <td>0.300</td>
960
+ </tr>
961
+ <tr>
962
+ <td>MinerU2-VLM</td>
963
+ <td>0.9B</td>
964
+ <td>68.16</td>
965
+ <td>0.230</td>
966
+ <td>74.45</td>
967
+ <td>53.07</td>
968
+ <td>0.191</td>
969
+ </tr>
970
+ <tr>
971
+ <td>MinerU2.5</td>
972
+ <td>1.2B</td>
973
+ <td>75.24</td>
974
+ <td>0.305</td>
975
+ <td>81.78</td>
976
+ <td>74.39</td>
977
+ <td>0.151</td>
978
+ </tr>
979
+ <tr>
980
+ <td>PaddleOCR-VL </td>
981
+ <td>0.9B</td>
982
+ <td>77.47</td>
983
+ <td>0.192</td>
984
+ <td>78.81</td>
985
+ <td>72.83</td>
986
+ <td>0.193</td>
987
+ </tr>
988
+ <tr>
989
+ <td>Nanonets-OCR-s</td>
990
+ <td>3B</td>
991
+ <td>81.98</td>
992
+ <td>0.121</td>
993
+ <td>85.78</td>
994
+ <td>72.22</td>
995
+ <td>0.133</td>
996
+ </tr>
997
+ <tr>
998
+ <td>dots.ocr</td>
999
+ <td>3B</td>
1000
+ <td>84.27</td>
1001
+ <td>0.087</td>
1002
+ <td>85.73</td>
1003
+ <td>75.74</td>
1004
+ <td>0.094</td>
1005
+ </tr>
1006
+ <tr>
1007
+ <td><strong>PaddleOCR-VL-1.5</strong></td>
1008
+ <td>0.9B</td>
1009
+ <td><strong>91.08</strong></td>
1010
+ <td><strong>0.054</strong></td>
1011
+ <td><strong>89.24</strong></td>
1012
+ <td><strong>89.36</strong></td>
1013
+ <td><strong>0.081</strong></td>
1014
+ </tr>
1015
+ </table>
1016
+
1017
  *Skewed documents also impact accuracy, but top-performing models still provide reasonable results.*
1018
 
1019