Introduction to PP-StructureV3

PP-StructureV3 pipeline, based on the Layout Parsing v1 pipeline, has strengthened the ability of layout detection, table recognition, and formula recognition. It has also added the ability to understand charts and restore reading order, as well as the ability to convert results into Markdown files. In various document data, it performs excellently and can handle more complex document data. This pipeline also provides flexible service-oriented deployment methods, supporting the use of multiple programming languages on various hardware. Moreover, it also provides the ability for secondary development. You can train and optimize on your own dataset based on this pipeline, and the trained model can be seamlessly integrated.

Key Metrics

Method Type	Methods	OverallEdit↓		TextEdit↓		FormulaEdit↓		TableEdit↓		Read OrderEdit↓
		EN	ZH	EN	ZH	EN	ZH	EN	ZH	EN	ZH
Pipeline Tools	PP-structureV3	0.145	0.206	0.058	0.088	0.295	0.535	0.159	0.109	0.069	0.091
	MinerU-0.9.3	0.15	0.357	0.061	0.215	0.278	0.577	0.18	0.344	0.079	0.292
	MinerU-1.3.11	0.166	0.310	0.0826	0.2000	0.3368	0.6236	0.1613	0.1833	0.0834	0.2316
	Marker-1.2.3	0.336	0.556	0.08	0.315	0.53	0.883	0.619	0.685	0.114	0.34
	Mathpix	0.191	0.365	0.105	0.384	0.306	0.454	0.243	0.32	0.108	0.304
	Docling-2.14.0	0.589	0.909	0.416	0.987	0.999	1	0.627	0.81	0.313	0.837
	Pix2Text-1.1.2.3	0.32	0.528	0.138	0.356	0.276	0.611	0.584	0.645	0.281	0.499
	Unstructured-0.17.2	0.586	0.716	0.198	0.481	0.999	1	1	0.998	0.145	0.387
	OpenParse-0.7.0	0.646	0.814	0.681	0.974	0.996	1	0.284	0.639	0.595	0.641
Expert VLMs	GOT-OCR	0.287	0.411	0.189	0.315	0.36	0.528	0.459	0.52	0.141	0.28
	Nougat	0.452	0.973	0.365	0.998	0.488	0.941	0.572	1	0.382	0.954
	Mistral OCR	0.268	0.439	0.072	0.325	0.318	0.495	0.6	0.65	0.083	0.284
	OLMOCR-sglang	0.326	0.469	0.097	0.293	0.455	0.655	0.608	0.652	0.145	0.277
	SmolDocling-256M_transformer	0.493	0.816	0.262	0.838	0.753	0.997	0.729	0.907	0.227	0.522
General VLMs	Gemini2.0-flash	0.191	0.264	0.091	0.139	0.389	0.584	0.193	0.206	0.092	0.128
	Gemini2.5-Pro	0.148	0.212	0.055	0.168	0.356	0.439	0.13	0.119	0.049	0.121
	GPT4o	0.233	0.399	0.144	0.409	0.425	0.606	0.234	0.329	0.128	0.251
	Qwen2-VL-72B	0.252	0.327	0.096	0.218	0.404	0.487	0.387	0.408	0.119	0.193
	Qwen2.5-VL-72B	0.214	0.261	0.092	0.18	0.315	0.434	0.341	0.262	0.106	0.168
	InternVL2-76B	0.44	0.443	0.353	0.29	0.543	0.701	0.547	0.555	0.317	0.228

The above data is from: * OmniDocBench * OmniDocBench: Benchmarking Diverse PDF Document Parsing with Comprehensive Annotations

End to End Benchmark

The performance of PP-StructureV3 and MinerU with different configurations under different GPU environments are as follows.

Requirements: * Paddle 3.0 * PaddleOCR 3.0.0 * MinerU 1.3.10 * CUDA 11.8 * cuDNN 8.9

Local inference

Local inference was tested with both V100 and A100 GPU, evaluating the performance of PP-StructureV3 under 6 different configurations. The test data consists of 15 PDF files, totaling 925 pages, including elements such as tables, formulas, seals, and charts.

In the following PP-StructureV3 configuration, please refer to PP-OCRv5 for OCR model details, see Formula Recognition for formula recognition model details, and refer to Text Detection for the max_side_limit setting of the text detection module.

Env: NVIDIA Tesla V100 + Intel Xeon Gold 6271C

Methods	Configurations				Average time per page (s)	Average CPU （%）	Peak RAM Usage （GB）	Average RAM Usage （GB）	Average GPU （%）	Peak VRAM Usage （GB）	Average VRAM Usage （GB）
PP-StructureV3	OCR Models	Formula Recognition Model	Chart Recognition Model	text detection module max_side_limit
	Server	PP-FormulaNet-L	✗	4096	1.77	111.4	6.7	5.2	38.9	17.0	16.5
	Server	PP-FormulaNet-L	✔	4096	4.09	105.3	5.5	4.0	24.7	17.0	16.6
	Mobile	PP-FormulaNet-L	✗	4096	1.56	113.7	6.6	4.9	29.1	10.7	10.6
	Server	PP-FormulaNet-M	✗	4096	1.42	112.9	6.8	5.1	38	16.0	15.5
	Mobile	PP-FormulaNet-M	✗	4096	1.15	114.8	6.5	5.0	26.1	8.4	8.3
	Mobile	PP-FormulaNet-M	✗	1200	0.99	113	7.0	5.6	29.2	8.6	8.5
MinerU	-				1.57	142.9	13.3	11.8	43.3	31.6	9.7

NVIDIA A100 + Intel Xeon Platinum 8350C

Methods	Configurations				Average time per page (s)	Average CPU （%）	Peak RAM Usage （GB）	Average RAM Usage （GB）	Average GPU （%）	Peak VRAM Usage （GB）	Average VRAM Usage （GB）
PP-StructureV3	OCR Models	Formula Recognition Model	Chart Recognition Model	text detection module max_side_limit
	Server	PP-FormulaNet-L	✗	4096	1.12	109.8	9.2	7.8	29.8	21.8	21.1
	Server	PP-FormulaNet-L	✔	4096	2.76	103.7	9.0	7.7	24	21.8	21.1
	Mobile	PP-FormulaNet-L	✗	4096	1.04	110.7	9.3	7.8	22	12.2	12.1
	Server	PP-FormulaNet-M	✗	4096	0.95	111.4	9.1	7.8	28.1	21.8	21.0
	Mobile	PP-FormulaNet-M	✗	4096	0.89	112.1	9.2	7.8	18.5	11.4	11.2
	Mobile	PP-FormulaNet-M	✗	1200	0.64	113.5	10.2	8.5	23.7	11.4	11.2
MinerU	-				1.06	168.3	18.3	16.8	27.5	76.9	14.8

Serving Inference

The serving inference test is based on the NVIDIA A100 + Intel Xeon Platinum 8350C environment, with test data consisting of 1500 images, including tables, formulas, seals, charts, and other elements.

Instances Number	Concurrent Requests Number	Throughput	Average Latency (s)	Success Number/Total Number
4 GPUs ✖️ 1 instance/gpu	4	1.69	2.36	100%
4 GPUs ✖️ 4 instances/gpu	16	4.05	3.87	100%

Pipeline benchmark data

Click to expand/collapse the table

Pipeline configuration	Hardware	Avg. inference time (s)	Peak CPU utilization (%)	Avg. CPU utilization (%)	Peak host memory (MB)	Avg. host memory (MB)	Peak GPU utilization (%)	Avg. GPU utilization (%)	Peak device memory (MB)	Avg. device memory (MB)
PP_StructureV3-default	Intel 8350C + A100	1.38	1384.60	113.26	5781.59	3431.21	100	32.79	37370.00	34165.68
	Intel 6271C + V100	2.38	608.70	109.96	6388.91	3737.19	100	39.08	26824.00	24581.61
	Intel 8563C + H20	1.36	744.30	112.82	6199.01	3865.78	100	43.81	35132.00	32077.12
	Intel 8350C + A10	1.74	418.50	105.96	6138.25	3503.41	100	48.54	18536.00	18353.93
	Intel 6271C + T4	3.70	434.40	105.45	6865.87	3595.68	100	71.92	13970.00	12668.58
PP_StructureV3-pp	Intel 8350C + A100	3.50	679.30	105.96	13850.20	5146.50	100	14.01	37656.00	34716.95
	Intel 6271C + V100	5.03	494.20	105.63	13542.94	4833.55	100	20.36	29402.00	26607.92
	Intel 8563C + H20	3.17	481.50	105.13	14179.97	5608.80	100	19.35	35454.00	32512.19
PP_StructureV3-full	Intel 8350C + A100	8.92	697.30	102.88	13777.07	4573.65	100	18.39	38776.00	37554.09
	Intel 6271C + V100	13.12	437.40	102.36	13974.00	4484.00	100	17.50	29878.00	28733.59
PP_StructureV3-seal	Intel 8350C + A100	1.39	747.50	112.55	5788.79	3742.03	100	33.81	38966.00	35832.44
	Intel 6271C + V100	2.44	630.10	110.18	6343.39	3725.98	100	42.23	28078.00	25834.70
	Intel 8563C + H20	1.40	792.20	113.63	6673.60	4417.34	100	46.33	35530.00	32516.87
	Intel 8350C + A10	1.75	422.40	106.08	6068.87	3973.49	100	50.12	19630.00	18374.37
	Intel 6271C + T4	3.76	400.30	105.10	6296.28	3651.42	100	72.57	14304.00	13268.36
PP_StructureV3-chart	Intel 8350C + A100	7.70	746.80	102.69	6355.58	4006.48	100	22.38	37380.00	36730.73
	Intel 6271C + V100	10.58	599.20	102.51	5754.14	3333.78	100	21.99	26820.00	26253.70
	Intel 8350C + A10	8.03	413.30	101.31	6473.29	3689.84	100	26.19	18540.00	18494.69
	Intel 6271C + T4	11.69	460.90	101.85	6503.12	3524.06	100	46.81	13966.00	12481.94
PP_StructureV3-notable	Intel 8350C + A100	1.24	738.30	110.45	5638.16	3278.30	100	35.32	30320.00	27026.17
	Intel 6271C + V100	2.24	452.40	107.79	5579.15	3635.95	100	43.00	23098.00	20684.43
	Intel 8563C + H20	1.18	989.00	107.71	6041.76	4024.76	100	50.67	33780.00	29733.15
	Intel 8350C + A10	1.58	225.00	102.56	5518.10	3333.08	100	49.90	21532.00	18567.99
	Intel 6271C + T4	3.40	413.30	103.58	5874.88	3662.49	100	76.82	13764.00	11890.62
PP_StructureV3-noformula	Intel 6271C	7.85	1172.50	964.70	17739.00	11101.02	N/A	N/A	N/A	N/A
	Intel 8350C	8.83	1053.50	970.64	15463.48	9408.19	N/A	N/A	N/A	N/A
	Intel 8350C + A100	0.84	788.60	124.25	6246.39	3674.32	100	30.57	40084.00	37358.45
	Intel 6271C + V100	1.42	606.20	115.53	7015.57	3707.03	100	35.63	29540.00	27620.28
	Intel 8563C + H20	0.87	644.10	119.23	6895.76	4222.85	100	50.00	36878.00	34104.59
	Intel 8350C + A10	1.03	377.50	106.87	5819.88	3830.19	100	42.87	19340.00	17550.94
	Intel 6271C + T4	2.02	430.20	109.21	6600.62	3824.18	100	65.75	14332.00	12712.18
PP_StructureV3-lightweight	Intel 6271C	4.36	1189.70	995.78	14000.50	9374.97	N/A	N/A	N/A	N/A
	Intel 8350C	3.74	1049.60	967.77	12960.96	7644.25	N/A	N/A	N/A	N/A
	Hygon 7490 + P800	0.86	572.20	120.84	8290.49	3569.44	N/A	N/A	N/A	N/A
	Intel 8350C + A100	0.61	823.40	126.25	9258.22	3776.63	52	18.95	7456.00	7131.95
	Intel 6271C + V100	1.07	686.80	116.70	9381.75	4126.28	58	22.92	8450.00	8083.30
	Intel 8563C + H20	0.46	999.00	122.21	9734.78	4516.40	61	24.41	7524.00	7167.52
	Intel 8350C + A10	0.70	355.40	111.51	9415.45	4094.06	89	30.85	7248.00	6927.58
	M4	12.22	223.60	107.35	9531.22	7884.61	N/A	N/A	N/A	N/A
	Intel 6271C + T4	1.13	461.40	112.16	7923.09	3837.31	85	41.67	8218.00	7902.04

Pipeline configuration	description
PP_StructureV3-default	Default configuration
PP_StructureV3-pp	Based on the default configuration, document image preprocessing is enabled
PP_StructureV3-full	Based on the default configuration, document image preprocessing and chart parsing are enabled
PP_StructureV3-seal	Based on the default configuration, seal text recognition is enabled
PP_StructureV3-chart	Based on the default configuration, chart parsing is enabled
PP_StructureV3-notable	Based on the default configuration, table recognition is disabled
PP_StructureV3-noformula	Based on the default configuration, formula recognition is disabled
PP_StructureV3-lightweight	Based on the default configuration, all task models are replaced with lightweight versions

• Test environment:
  • PaddlePaddle 3.1.0、CUDA 11.8、cuDNN 8.9
  • PaddleX @ develop (f1eb28e23cfa54ce3e9234d2e61fcb87c93cf407)
  • Docker image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddle:3.1.0-gpu-cuda11.8-cudnn8.9
• Test data:
  • Test data containing 280 images including tables, seals, formulas, and charts.
• Test strategy:
  • Warm up with 20 samples, then repeat the full dataset once for performance testing.
• Note:
  • Since we did not collect device memory data for NPU and XPU, the corresponding entries in the table are marked as N/A.

PP-StructureV3 Demo

More Demos

FAQ

Q: What is the default model configuration? If I need higher accuracy, faster speed, or lower GPU memory usage, which parameters should I adjust or which models should I switch to? How significant is the impact on results?

A: By default, the largest models for each module are used. Section 3.3 demonstrates how different model selections affect GPU memory consumption and inference speed. You can choose an appropriate model based on your device capabilities and the complexity of your samples. Additionally, in the Python API or CLI, you can set the device parameter as <device_type>:<device_id1>,<device_id2>... (e.g., gpu:0,1,2,3) to enable multi-GPU parallel inference. If the built-in multi-GPU parallel inference does not meet your speed requirements, you may refer to the example code for multi-process parallel inference and further optimize it for your specific scenario: Multi-process Parallel Inference.

---

Q: Can PP-StructureV3 run on CPU?

A: While PP-StructureV3 is recommended to run on GPU for optimal performance, it also supports CPU inference. Thanks to a variety of configuration options and sufficient optimization for lightweight models, users can refer to section 3.3 to select lightweight configurations for CPU-only environments. For example, on an Intel 8350C CPU, the inference time per image is about 3.74 seconds.

---

Q: How can I integrate PP-StructureV3 into my own project?

A:
- For Python projects, you can directly integrate using the PaddleOCR Python API.
- For projects in other programming languages, it is recommended to use service-based deployment. PaddleOCR supports client-side integration in multiple languages, including C++, C#, Java, Go, and PHP. Please refer to the official documentation for details.
- If you need to interact with large language models, PaddleOCR also provides the MCP service. For more information, please refer to the MCP Server documentation.

---

Q: Can the serving handle concurrent requests?

A: In the basic service deployment solution, the service processes only one request at a time, which is mainly intended for quick validation, development pipeline integration, or scenarios that do not require concurrent processing. In the high-stability serving solution, the service also processes one request at a time by default, but users can achieve horizontal scaling and concurrent processing by adjusting the configuration as outlined in the service deployment guide.

---

Q: How can I reduce latency and increase throughput in serving?

A: PaddleOCR offers two types of service deployment solutions. Regardless of the solution used, enabling high-performance inference plugins can accelerate model inference and thus reduce latency. For the high-stability deployment solution, throughput can be further increased by adjusting the service configuration to run multiple instances, making full use of your hardware resources. Please refer to the documentation for more details on configuring high-stability serving.