The advent of "OCR 2.0" and large-scale vision-language models (VLMs) has set new benchmarks in text recognition. However, these unified architectures often come with significant computational demands, challenges in precise text localization within complex layouts, and a propensity for textual hallucinations. Revisiting the prevailing notion that model scale is the sole path to high accuracy, this paper introduces PP-OCRv5, a meticulously optimized, lightweight OCR system with merely 5 million parameters. We demonstrate that PP-OCRv5 achieves performance competitive with many billion-parameter VLMs on standard OCR benchmarks, while offering superior localization precision and reduced hallucinations. The cornerstone of our success lies not in architectural expansion but in a data-centric investigation. We systematically dissect the role of training data by quantifying three critical dimensions: data difficulty, data accuracy, and data diversity. Our extensive experiments reveal that with a sufficient volume of high-quality, accurately labeled, and diverse data, the performance ceiling for traditional, efficient two-stage OCR pipelines is far higher than commonly assumed. This work provides compelling evidence for the viability of lightweight, specialized models in the large-model era and offers practical insights into data curation for OCR. The source code and models are publicly available at https://github.com/PaddlePaddle/PaddleOCR.
Figures & tables
Figure 1: PP-OCRv5 vs. Vision-Language Models: 1-EditDist Accuracy and Parameter Size Comparison.
Figure 2: Overview of the PP-OCRv5 Framework: (a) Lightweight model architecture design; (b) Large-scale data construction and optimization strategies; (c) Distribution of training data across 22.6M samples; (d) Comparative performance of the PP-OCR series across multiple scenarios.
Figure 3: Comprehensive analysis of training data characteristics. Our key findings include: (a) A “sweet spot” of medium-difficulty data is most effective; (b) The model exhibits remarkable robustness to label noise; (c) Data diversity is paramount for performance; (d) Model accuracy scales nearly linearly with data quantity.
Score Range
Ratio
Description
< 0.80
3.2%
Hard / Ambiguous
0.80-0.90
10.4%
Challenging
0.90-0.95
25.1%
High Quality
0.95-0.97
48.5%
Sweet Spot (Optimal)
> 0.97
12.8%
Simple (Stability)
Table 1: Confidence Score Distribution of the 22.6M Training Set.
Category
Size (10,000s)
1. Printed & General Scenes
Printed Chinese
5,000
Printed English
2,390
General Scenes
4,010
Single Characters
540
Special Formats
50
Table 2: Statistics of the PP-OCRv5 training dataset. The data is categorized into four major groups, with a total of approximately 22.6M samples.
Model
Weighted Acc.
Scenario-specific Accuracy (%)
Handwritten
Printed
TC-CN
Ancient
JP
Confus.
Gen.
Pinyin
Vert.
Art.
CN
EN
CN
EN
PP-OCRv3
42.5
12.5
22.2
72.9
77.0
13.5
17.1
27.6
33.2
73.1
21.3
44.5
42.9
PP-OCRv4
53.0
29.8
25.5
83.9
66.0
32.2
25.9
47.2
46.0
81.1
25.9
59.2
55.6
PP-OCRv5
80.1
41.7
49.4
86.1
87.5
72.0
57.9
75.8
55.7
77.0
72.5
80.9
54.0
Table 3: Text recognition accuracy (%) of PP-OCR series models on our in-house benchmark, a comprehensive multi-scenario benchmark. The weighted accuracy (Weighted Acc.) is the primary overall metric. PP-OCRv5 demonstrates superior robustness across most scenarios. Red and blue cells indicate the best and second-best results per column, respectively. Bold values highlight PP-OCRv5 results.
Model Type
Model
Params
ALL_avg ↓
Language
Text Background
Text Rotate
EN ↓
ZH ↓
Mixed ↓
White ↓
Single ↓
Multi ↓
Normal ↓
Rotate90 ↓
Horizontal ↓
Rotate270 ↓
Vision Language Models
Qwen3-VL
235B
0.026
0.016
0.026
0.034
0.023
0.023
0.030
0.023
0.046
0.029
0.170
Qwen2-VL
72B
0.173
0.072
0.274
0.286
0.234
0.155
0.148
0.223
0.273
0.067
0.721
InternVL2
76B
0.115
0.074
0.155
0.242
0.113
0.352
0.269
0.132
0.610
0.595
0.907
GPT4o
-
0.122
0.020
0.224
0.125
0.167
0.140
0.220
0.168
0.115
0.132
0.718
Pipeline Tools & Expert Vision Models
Tesseract-OCR
-
0.324
0.096
0.551
0.250
0.439
0.328
0.331
0.426
0.117
0.984
0.969
Table 4: Component-level OCR text recognition evaluation on OmniDocBench. Results are reported by OmniDocBench [ 19 ] except PP-OCRv5 and Qwen3-VL. Red and blue cells indicate the best and second-best results per column, respectively. Bold values highlight PP-OCRv5 results.