PaddleOCR-VL Introduction

PaddleOCR-VL is a SOTA and resource-efficient model tailored for document parsing. Its core component is PaddleOCR-VL-0.9B, a compact yet powerful vision-language model (VLM) that integrates a NaViT-style dynamic resolution visual encoder with the ERNIE-4.5-0.3B language model to enable accurate element recognition. This innovative model efficiently supports 109 languages and excels in recognizing complex elements (e.g., text, tables, formulas, and charts), while maintaining minimal resource consumption. Through comprehensive evaluations on widely used public benchmarks and in-house benchmarks, PaddleOCR-VL achieves SOTA performance in both page-level document parsing and element-level recognition. It significantly outperforms existing solutions, exhibits strong competitiveness against top-tier VLMs, and delivers fast inference speeds. These strengths make it highly suitable for practical deployment in real-world scenarios.

1. Environment Preparation

Install PaddlePaddle and PaddleOCR:

python -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
python -m pip install -U "paddleocr[doc-parser]"
python -m pip install https://paddle-whl.bj.bcebos.com/nightly/cu126/safetensors/safetensors-0.6.2.dev0-cp38-abi3-linux_x86_64.whl

For Windows users, please use WSL or a Docker container.

2. Quick Start

PaddleOCR-VL supports two usage methods: CLI command line and Python API. The CLI command line method is simpler and suitable for quickly verifying functionality, while the Python API method is more flexible and suitable for integration into existing projects.

2.1 Command Line Usage

Run a single command to quickly test the PaddleOCR-VL ：

paddleocr doc_parser -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/paddleocr_vl_demo.png

# Use --use_doc_orientation_classify to enable document orientation classification
paddleocr doc_parser -i ./paddleocr_vl_demo.png --use_doc_orientation_classify True

# Use --use_doc_unwarping to enable document unwarping module
paddleocr doc_parser -i ./paddleocr_vl_demo.png --use_doc_unwarping True

# Use --use_layout_detection to enable layout detection
paddleocr doc_parser -i ./paddleocr_vl_demo.png --use_layout_detection False

Command line supports more parameters. Click to expand for detailed parameter descriptions

Parameter	Description	Type	Default
input	Data to be predicted, required. For example, the local path of an image file or PDF file: /root/data/img.jpg;Such as a URL link, for example, the network URL of an image file or PDF file:Example;Such as a local directory, which should contain the images to be predicted, for example, the local path: /root/data/(Currently, prediction for directories containing PDF files is not supported. PDF files need to be specified with a specific file path).	str
save_path	Specify the path where the inference result file will be saved. If not set, the inference results will not be saved locally.	str
layout_detection_model_name	Name of the layout area detection and ranking model. If not set, the default model of the production line will be used.	str
layout_detection_model_dir	Directory path of the layout area detection and ranking model. If not set, the official model will be downloaded.	str
layout_threshold	Score threshold for the layout model. Any floating-point number between 0-1. If not set, the parameter value initialized by the production line will be used.float	layout_nms	layout_threshold	Score threshold for the layout model. Any value between 0-1. If not set, the default value is used, which is 0.5.
Whether to use post-processing NMS for layout detection. If not set, the parameter value initialized by the production line will be used, with a default initialization of	True .bool	layout_unclip_ratio
Expansion coefficient for the detection boxes of the layout area detection model. Any floating-point number greater than	0 . If not set, the parameter value initialized by the production line will be used.float	layout_merge_bboxes_mode
Merging mode for the detection boxes output by the model in layout detection.	large , when set to large, it means that among the detection boxes output by the model, for overlapping and contained boxes, only the outermost largest box is retained, and the overlapping inner boxes are deleted;small , when set to small, it means that among the detection boxes output by the model, for overlapping and contained boxes, only the innermost small box is retained, and the overlapping outer boxes are deleted;union , no filtering of boxes is performed, and both inner and outer boxes are retained;If not set, the parameter value initialized by the production line will be used. str	vl_rec_model_name
Name of the multimodal recognition model. If not set, the default model of the production line will be used.	str	vl_rec_model_dir
Directory path of the multimodal recognition model. If not set, the official model will be downloaded.	str	vl_rec_backend
Inference backend used by the multimodal recognition model.	str	vl_rec_server_url
If the multimodal recognition model uses an inference service, this parameter is used to specify the server URL.	str	vl_rec_max_concurrency
If the multimodal recognition model uses an inference service, this parameter is used to specify the maximum number of concurrent requests.	str	doc_orientation_classify_model_name
Name of the document orientation classification model. If not set, the default model of the production line will be used.	str	doc_orientation_classify_model_dir
Directory path of the document orientation classification model. If not set, the official model will be downloaded.	str	doc_unwarping_model_name
Name of the text image rectification model. If not set, the default model of the production line will be used.	str	doc_unwarping_model_dir
Directory path of the text image rectification model. If not set, the official model will be downloaded.	str	use_doc_orientation_classify
Whether to load and use the document orientation classification module. If not set, the parameter value initialized by the production line will be used, with a default initialization of	False .bool	use_doc_unwarping
Whether to load and use the text image rectification module. If not set, the parameter value initialized by the production line will be used, with a default initialization of	False .bool	use_layout_detection
Whether to load and use the layout area detection and ranking module. If not set, the parameter value initialized by the production line will be used, with a default initialization of	True .bool	use_chart_recognition
Whether to load and use the chart parsing module. If not set, the parameter value initialized by the production line will be used, with a default initialization of	False .bool	bool
format_block_content	Controls whether to format the content in block_contentas Markdown. If not set, the parameter value initialized by the production line will be used, which is initially set to Falseby default.	bool
use_queues	Used to control whether to enable internal queues. When set to True, data loading (such as rendering PDF pages as images), layout detection model processing, and VLM inference will be executed asynchronously in separate threads, with data passed through queues, thereby improving efficiency. This approach is particularly efficient for PDF documents with a large number of pages or directories containing a large number of images or PDF files.	bool
prompt_label	The prompt type setting for the VL model, which takes effect only when use_layout_detection=False.	str
repetition_penalty	The repetition penalty parameter used for VL model sampling.	float
temperature	The temperature parameter used for VL model sampling.	float
top_p	The top-p parameter used for VL model sampling.	float
min_pixels	The minimum number of pixels allowed when the VL model preprocesses images.	int
max_pixels	The maximum number of pixels allowed when the VL model preprocesses images.	int
device	The device used for inference. Supports specifying specific card numbers: CPU: For example, cpuindicates using the CPU for inference; GPU: For example, gpu:0indicates using the first GPU for inference; NPU: For example, npu:0indicates using the first NPU for inference; XPU: For example, xpu:0indicates using the first XPU for inference; MLU: For example, mlu:0indicates using the first MLU for inference; DCU: For example, dcu:0indicates using the first DCU for inference; If not set, the parameter value initialized by the production line will be used by default. During initialization, the local GPU device 0 will be used preferentially. If it is not available, the CPU device will be used.	str
enable_hpi	Whether to enable high-performance inference.	bool	False
use_tensorrt	Whether to enable the TensorRT subgraph engine of Paddle Inference. If the model does not support acceleration via TensorRT, acceleration will not be used even if this flag is set. For PaddlePaddle with CUDA 11.8, the compatible TensorRT version is 8.x (x>=6). It is recommended to install TensorRT 8.6.1.6.	bool	False
precision	Computational precision, such as fp32, fp16.	str	fp32
enable_mkldnn	Whether to enable MKL-DNN accelerated inference. If MKL-DNN is not available or the model does not support acceleration via MKL-DNN, acceleration will not be used even if this flag is set.	bool	True
mkldnn_cache_capacity	MKL-DNN cache capacity.	int	10
cpu_threads	The number of threads used for inference on the CPU.	int	8
paddlex_config	The file path of the PaddleX production line configuration.	str

The inference result will be printed in the terminal. The default output of the PP-StructureV3 pipeline is as follows:

👉Click to expand

  
{'res': {'input_path': 'paddleocr_vl_demo.png', 'page_index': None, 'model_settings': {'use_doc_preprocessor': False, 'use_layout_detection': True, 'use_chart_recognition': False, 'format_block_content': False}, 'layout_det_res': {'input_path': None, 'page_index': None, 'boxes': [{'cls_id': 6, 'label': 'doc_title', 'score': 0.9636914134025574, 'coordinate': [np.float32(131.31366), np.float32(36.450516), np.float32(1384.522), np.float32(127.984665)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9281806349754333, 'coordinate': [np.float32(585.39465), np.float32(158.438), np.float32(930.2184), np.float32(182.57469)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9840355515480042, 'coordinate': [np.float32(9.023666), np.float32(200.86115), np.float32(361.41583), np.float32(343.8828)]}, {'cls_id': 14, 'label': 'image', 'score': 0.9871416091918945, 'coordinate': [np.float32(775.50574), np.float32(200.66502), np.float32(1503.3807), np.float32(684.9304)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9801855087280273, 'coordinate': [np.float32(9.532196), np.float32(344.90594), np.float32(361.4413), np.float32(440.8244)]}, {'cls_id': 17, 'label': 'paragraph_title', 'score': 0.9708921313285828, 'coordinate': [np.float32(28.040405), np.float32(455.87976), np.float32(341.7215), np.float32(520.7117)]}, {'cls_id': 24, 'label': 'vision_footnote', 'score': 0.9002962708473206, 'coordinate': [np.float32(809.0692), np.float32(703.70044), np.float32(1488.3016), np.float32(750.5238)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9825374484062195, 'coordinate': [np.float32(8.896561), np.float32(536.54895), np.float32(361.05237), np.float32(655.8058)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9822263717651367, 'coordinate': [np.float32(8.971573), np.float32(657.4949), np.float32(362.01715), np.float32(774.625)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9767460823059082, 'coordinate': [np.float32(9.407074), np.float32(776.5216), np.float32(361.31067), np.float32(846.82874)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9868153929710388, 'coordinate': [np.float32(8.669495), np.float32(848.2543), np.float32(361.64703), np.float32(1062.8568)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9826608300209045, 'coordinate': [np.float32(8.8025055), np.float32(1063.8615), np.float32(361.46588), np.float32(1182.8524)]}, {'cls_id': 22, 'label': 'text', 'score': 0.982555627822876, 'coordinate': [np.float32(8.820602), np.float32(1184.4663), np.float32(361.66394), np.float32(1302.4507)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9584776759147644, 'coordinate': [np.float32(9.170288), np.float32(1304.2161), np.float32(361.48898), np.float32(1351.7483)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9782056212425232, 'coordinate': [np.float32(389.1618), np.float32(200.38202), np.float32(742.7591), np.float32(295.65146)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9844875931739807, 'coordinate': [np.float32(388.73303), np.float32(297.18463), np.float32(744.00024), np.float32(441.3034)]}, {'cls_id': 17, 'label': 'paragraph_title', 'score': 0.9680547714233398, 'coordinate': [np.float32(409.39468), np.float32(455.89386), np.float32(721.7174), np.float32(520.9387)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9741666913032532, 'coordinate': [np.float32(389.71606), np.float32(536.8138), np.float32(742.7112), np.float32(608.00165)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9840384721755981, 'coordinate': [np.float32(389.30988), np.float32(609.39636), np.float32(743.09247), np.float32(750.3231)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9845995306968689, 'coordinate': [np.float32(389.13272), np.float32(751.7772), np.float32(743.058), np.float32(894.8815)]}, {'cls_id': 22, 'label': 'text', 'score': 0.984852135181427, 'coordinate': [np.float32(388.83267), np.float32(896.0371), np.float32(743.58215), np.float32(1038.7345)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9804865717887878, 'coordinate': [np.float32(389.08478), np.float32(1039.9119), np.float32(742.7585), np.float32(1134.4897)]}, {'cls_id': 22, 'label': 'text', 'score': 0.986461341381073, 'coordinate': [np.float32(388.52643), np.float32(1135.8137), np.float32(743.451), np.float32(1352.0085)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9869391918182373, 'coordinate': [np.float32(769.8341), np.float32(775.66235), np.float32(1124.9813), np.float32(1063.207)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9822869896888733, 'coordinate': [np.float32(770.30383), np.float32(1063.938), np.float32(1124.8295), np.float32(1184.2192)]}, {'cls_id': 17, 'label': 'paragraph_title', 'score': 0.9689218997955322, 'coordinate': [np.float32(791.3042), np.float32(1199.3169), np.float32(1104.4521), np.float32(1264.6985)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9713128209114075, 'coordinate': [np.float32(770.4253), np.float32(1279.6072), np.float32(1124.6917), np.float32(1351.8672)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9236552119255066, 'coordinate': [np.float32(1153.9058), np.float32(775.5814), np.float32(1334.0654), np.float32(798.1581)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9857938885688782, 'coordinate': [np.float32(1151.5197), np.float32(799.28015), np.float32(1506.3619), np.float32(991.1156)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9820687174797058, 'coordinate': [np.float32(1151.5686), np.float32(991.91095), np.float32(1506.6023), np.float32(1110.8875)]}, {'cls_id': 22, 'label': 'text', 'score': 0.9866049885749817, 'coordinate': [np.float32(1151.6919), np.float32(1112.1301), np.float32(1507.1611), np.float32(1351.9504)]}]}}}

For explanation of the result parameters, refer to 2.2 Python Script Integration.

Note: The default model for the production line is relatively large, which may result in slower inference speed. It is recommended to use inference acceleration frameworks to enhance VLM inference performance for faster inference.

2.2 Python Script Integration

The command line method is for quick testing and visualization. In actual projects, you usually need to integrate the model via code. You can perform pipeline inference with just a few lines of code as shown below:

from paddleocr import PaddleOCRVL

pipeline = PaddleOCRVL()
# pipeline = PaddleOCRVL(use_doc_orientation_classify=True) # Use use_doc_orientation_classify to enable/disable document orientation classification model
# pipeline = PaddleOCRVL(use_doc_unwarping=True) # Use use_doc_unwarping to enable/disable document unwarping module
# pipeline = PaddleOCRVL(use_layout_detection=False) # Use use_layout_detection to enable/disable layout detection module
output = pipeline.predict("./paddleocr_vl_demo.png")
for res in output:
    res.print() ## Print the structured prediction output
    res.save_to_json(save_path="output") ## Save the current image's structured result in JSON format
    res.save_to_markdown(save_path="output") ## Save the current image's result in Markdown format

For PDF files, each page will be processed individually and generate a separate Markdown file. If you want to convert the entire PDF to a single Markdown file, use the following method:

from pathlib import Path
from paddleocr import PaddleOCRVL

input_file = "./your_pdf_file.pdf"
output_path = Path("./output")

pipeline = PaddleOCRVL()
output = pipeline.predict(input=input_file)

markdown_list = []
markdown_images = []

for res in output:
    md_info = res.markdown
    markdown_list.append(md_info)
    markdown_images.append(md_info.get("markdown_images", {}))

markdown_texts = pipeline.concatenate_markdown_pages(markdown_list)

mkd_file_path = output_path / f"{Path(input_file).stem}.md"
mkd_file_path.parent.mkdir(parents=True, exist_ok=True)

with open(mkd_file_path, "w", encoding="utf-8") as f:
    f.write(markdown_texts)

for item in markdown_images:
    if item:
        for path, image in item.items():
            file_path = output_path / path
            file_path.parent.mkdir(parents=True, exist_ok=True)
            image.save(file_path)

Note:

• In the example code, the parameters use_doc_orientation_classify and use_doc_unwarping are all set to False by default. These indicate that document orientation classification and document image unwarping are disabled. You can manually set them to True if needed.

The above Python script performs the following steps:

(1) Instantiate the production line object. Specific parameter descriptions are as follows:

Parameter	Parameter Description	Parameter Type	Default Value
layout_detection_model_name	Name of the layout area detection and ranking model. If set to None, the default model of the production line will be used.	str|None	None
layout_detection_model_dir	Directory path of the layout area detection and ranking model. If set to None, the official model will be downloaded.	str|None	None
layout_threshold	Score threshold for the layout model. float: Any floating-point number between0-1 ;dict :{0:0.1} The key is the class ID, and the value is the threshold for that class;None : If set toNone	, the parameter value initialized by the production line will be used.	float|dict|None
None	layout_nms Whether to use post-processing NMS for layout detection. If set toNone	, the parameter value initialized by the production line will be used.	bool|None
None	layout_unclip_ratio Expansion coefficient for the detection box of the layout area detection model.float : Any floating-point number greater than0 ;Tuple[float,float] : The respective expansion coefficients in the horizontal and vertical directions;dict, where the key of the dict is ofint type, representingcls_id, and the value is oftuple type, such as{0: (1.1, 2.0)} , indicating that the center of the detection box for class 0 output by the model remains unchanged, with the width expanded by 1.1 times and the height expanded by 2.0 times;None : If set toNone	, the parameter value initialized by the production line will be used.	float|Tuple[float,float]|dict|None
None	layout_merge_bboxes_mode Method for filtering overlapping boxes in layout area detection.str :large ,small ,union , indicating whether to retain the large box, small box, or both during overlapping box filtering;dict: The key of the dict is ofint type, representingcls_id, and the value is ofstr type, such as{0: "large", 2: "small"} , indicating that the large mode is used for the detection box of class 0, and the small mode is used for the detection box of class 2;None : If set toNone	, the parameter value initialized by the production line will be used.	str|dict|None
None	vl_rec_model_name Name of the multimodal recognition model. If set toNone	, the default model of the production line will be used.	str|None
None	vl_rec_model_dir Directory path of the multimodal recognition model. If set toNone	, the official model will be downloaded.	str|None
None	vl_rec_backend	Inference backend used by the multimodal recognition model.	int|None
None	vl_rec_server_url	If the multimodal recognition model uses an inference service, this parameter is used to specify the server URL.	str|None
None	vl_rec_max_concurrency	If the multimodal recognition model uses an inference service, this parameter is used to specify the maximum number of concurrent requests.	str|None
None	doc_orientation_classify_model_name Name of the document orientation classification model. If set toNone	, the default model of the production line will be used.	str|None
None	doc_orientation_classify_model_dir Directory path of the document orientation classification model. If set toNone	, the official model will be downloaded.	str|None
doc_unwarping_model_name	Name of the text image rectification model. If set to None, the default model of the production line will be used.	str|None	None
doc_unwarping_model_dir	Directory path of the text image rectification model. If set to None, the official model will be downloaded.	str|None	None
use_doc_orientation_classify	Whether to load and use the document orientation classification module. If set to None, the parameter value initialized by the production line will be used, and it is initialized to False by default.	bool|None	None
use_doc_unwarping	Whether to load and use the text image rectification module. If set to None, the parameter value initialized by the production line will be used, and it is initialized to False by default.	bool|None	None
use_layout_detection	Whether to load and use the layout area detection and sorting module. If set to None, the parameter value initialized by the production line will be used, and it is initialized to True by default.	bool|None	None
use_chart_recognition	Whether to load and use the chart parsing module. If set to None, the parameter value initialized by the production line will be used, and it is initialized to False by default.	bool|None	None
format_block_content	Controls whether to format the content in block_content into Markdown format. If set to None, the parameter value initialized by the production line will be used, and it is initialized to False by default.	bool|None	None
device	Device used for inference. Supports specifying specific card numbers: CPU: For example, cpu indicates using the CPU for inference; GPU: For example, gpu:0 indicates using the first GPU for inference; NPU: For example, npu:0 indicates using the first NPU for inference; XPU: For example, xpu:0 indicates using the first XPU for inference; MLU: For example, mlu:0 indicates using the first MLU for inference; DCU: For example, dcu:0 indicates using the first DCU for inference; None: If set to None, during initialization, the local GPU device 0 will be used preferentially. If not available, the CPU device will be used.	str|None	None
enable_hpi	Whether to enable high-performance inference.	bool	False
use_tensorrt	Whether to enable the TensorRT subgraph engine of Paddle Inference. If the model does not support acceleration via TensorRT, acceleration will not be used even if this flag is set. For PaddlePaddle with CUDA 11.8, the compatible TensorRT version is 8.x (x>=6), and it is recommended to install TensorRT 8.6.1.6.	bool	False
precision	Computational precision, such as fp32, fp16.	str	"fp32"
enable_mkldnn	Whether to enable MKL-DNN accelerated inference. If MKL-DNN is not available or the model does not support acceleration via MKL-DNN, acceleration will not be used even if this flag is set.	bool	True
mkldnn_cache_capacity	MKL-DNN cache capacity.	int	10
cpu_threads	Number of threads used for inference on the CPU.	int	8
paddlex_config	Path to the PaddleX production line configuration file.	str|None	None

(2) Call the predict()method of the PaddleOCR-VL production line object for inference prediction. This method will return a list of results. Additionally, the production line also provides the predict_iter()Method. The two are completely consistent in terms of parameter acceptance and result return. The difference lies in that predict_iter()returns a generator, which can process and obtain prediction results step by step. It is suitable for scenarios involving large datasets or where memory conservation is desired. You can choose either of these two methods based on actual needs. Below are the parameters of the predict()method and their descriptions:

Parameter	Parameter Description	Parameter Type	Default Value
input	Data to be predicted, supporting multiple input types. Required. Python Var: such as numpy.ndarrayrepresenting image data str: such as the local path of an image file or PDF file: /root/data/img.jpg;such as a URL link, such as the network URL of an image file or PDF file:Example;such as a local directory, which should contain the images to be predicted, such as the local path: /root/data/(Currently, prediction for directories containing PDF files is not supported. PDF files need to be specified with a specific file path) list: List elements should be of the aforementioned data types, such as [numpy.ndarray, numpy.ndarray], ["/root/data/img1.jpg", "/root/data/img2.jpg"], ["/root/data1", "/root/data2"].	Python Var|str|list
use_doc_orientation_classify	Whether to use the document orientation classification module during inference. Setting it to Nonemeans using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
use_doc_unwarping	Whether to use the text image rectification module during inference. Setting it to Nonemeans using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
use_layout_detection	Whether to use the layout region detection and sorting module during inference. Setting it to Nonemeans using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
use_chart_recognition	Whether to use the chart parsing module during inference. Setting it to Nonemeans using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
layout_threshold	The parameter meaning is basically the same as the instantiation parameter. Setting it to Nonemeans using the instantiation parameter; otherwise, this parameter takes precedence.	float|dict|None	None
layout_nms	The parameter meaning is basically the same as the instantiation parameter. Setting it to Nonemeans using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
layout_unclip_ratio	The parameter meaning is basically the same as the instantiation parameter. Setting it to Nonemeans using the instantiation parameter; otherwise, this parameter takes precedence.	float|Tuple[float,float]|dict|None	None
layout_merge_bboxes_mode	The parameter meaning is basically the same as the instantiation parameter. Setting it to Nonemeans using the instantiation parameter; otherwise, this parameter takes precedence.	str|dict|None	None
use_queues	Used to control whether to enable internal queues. When set to True, data loading (such as rendering PDF pages as images), layout detection model processing, and VLM inference will be executed asynchronously in separate threads, with data passed through queues, thereby improving efficiency. This approach is particularly efficient for PDF documents with many pages or directories containing a large number of images or PDF files.	bool|None	None
prompt_label	The prompt type setting for the VL model, which takes effect only when use_layout_detection=False.	str|None	None
format_block_content	The parameter meaning is basically the same as the instantiation parameter. Setting it to Nonemeans using the instantiation parameter; otherwise, this parameter takes precedence.	bool|None	None
repetition_penalty	The repetition penalty parameter used for VL model sampling.	float|None	None
temperature	Temperature parameter used for VL model sampling.	float|None	None
top_p	Top-p parameter used for VL model sampling.	float|None	None
min_pixels	The minimum number of pixels allowed when the VL model preprocesses images.	int|None	None
max_pixels	The maximum number of pixels allowed when the VL model preprocesses images.	int|None	None

(3) Process the prediction results: The prediction result for each sample is a corresponding Result object, supporting operations such as printing, saving as an image, and saving as a jsonfile:

Method	Method Description	Parameter	Parameter Type	Parameter Description	Default Value
print()	Print results to the terminal	format_json	bool	Whether to format the output content using JSONindentation.	True
		indent	int	Specify the indentation level to beautify the output JSONdata, making it more readable. Only valid when format_jsonis True.	4
		ensure_ascii	bool	Control whether non- ASCIIcharacters are escaped as Unicode. When set to True, all non- ASCIIcharacters will be escaped; Falseretains the original characters. Only valid when format_jsonis True.	False
save_to_json()	Save the results as a json format file	save_path	str	The file path for saving. When it is a directory, the saved file name will be consistent with the input file type naming.	None
		indent	int	Specify the indentation level to beautify the output JSONdata, making it more readable. Only valid when format_jsonis True.	4
		ensure_ascii	bool	Control whether non- ASCIIcharacters are escaped as Unicode. When set to True, all non- ASCIIcharacters will be escaped; Falseretains the original characters. Only valid when format_jsonis True.	False
save_to_img()	Save the visualized images of each intermediate module in png format	save_path	str	The file path for saving, supporting directory or file paths.	None
save_to_markdown()	Save each page in an image or PDF file as a markdown format file separately	save_path	str	The file path for saving. When it is a directory, the saved file name will be consistent with the input file type naming	None
		pretty	bool	Whether to beautify the markdownoutput results, centering charts, etc., to make the markdownrendering more aesthetically pleasing.	True
		show_formula_number	bool	Control whether to retain formula numbers in markdown. When set to True, all formula numbers are retained; Falseretains only the formulas	False
save_to_html()	Save the tables in the file as html format files	save_path	str	The file path for saving, supporting directory or file paths.	None
save_to_xlsx()	Save the tables in the file as xlsx format files	save_path	str	The file path for saving, supporting directory or file paths.	None

- Calling the `print()` method will print the results to the terminal. The content printed to the terminal is explained as follows: - `input_path`: `(str)` The input path of the image or PDF to be predicted. - `page_index`: `(Union[int, None])` If the input is a PDF file, it indicates the current page number of the PDF; otherwise, it is `None`. - `model_settings`: `(Dict[str, bool])` Model parameters required for configuring the production line. - `use_doc_preprocessor`: `(bool)` Controls whether to enable the document preprocessing sub-production line. - `use_seal_recognition`: `(bool)` Controls whether to enable the seal text recognition sub-production line. - `use_table_recognition`: `(bool)` Controls whether to enable the table recognition sub-production line. - `use_formula_recognition`: `(bool)` Controls whether to enable the formula recognition sub-production line. - `doc_preprocessor_res`: `(Dict[str, Union[List[float], str]])` A dictionary of document preprocessing results, which exists only when `use_doc_preprocessor=True`. - `input_path`: `(str)` The image path accepted by the document preprocessing sub-production line. When the input is a `numpy.ndarray`, it is saved as `None`. Here, it is `None`. - `page_index`: `None`. Since the input here is a `numpy.ndarray`, the value is `None`. - `model_settings`: `(Dict[str, bool])` Model configuration parameters for the document preprocessing sub-production line. - `use_doc_orientation_classify`: `(bool)` Controls whether to enable the document image orientation classification sub-module. - `use_doc_unwarping`: `(bool)` Controls whether to enable the text image distortion correction sub-module. - `angle`: `(int)` The prediction result of the document image orientation classification sub-module. When enabled, the actual angle value is returned. - `parsing_res_list`: `(List[Dict])` A list of parsing results, where each element is a dictionary. The order of the list is the reading order after parsing. - `block_bbox`: `(np.ndarray)` The bounding box of the layout area. - `block_label`: `(str)` The label of the layout area, such as `text`, `table`, etc. - `block_content`: `(str)` The content within the layout area. - `block_id`: `(int)` The index of the layout area, used to display the layout sorting results. - `block_order` `(int)` The order of the layout area, used to display the layout reading order. For non-sorted parts, the default value is `None`. - `overall_ocr_res`: `(Dict[str, Union[List[str], List[float], numpy.ndarray]])` A dictionary of global OCR results. - `input_path`: `(Union[str, None])` The image path accepted by the image OCR sub-production line. When the input is a `numpy.ndarray`, it is saved as `None`. - `page_index`: `None`. Since the input here is a `numpy.ndarray`, the value is `None`. - `model_settings`: `(Dict)` Model configuration parameters for the OCR sub-production line. - `dt_polys`: `(List[numpy.ndarray])` A list of polygonal bounding boxes for text detection. Each detection box is represented by a numpy array consisting of the coordinates of 4 vertices, with an array shape of (4, 2) and a data type of int16. - `dt_scores`: `(List[float])` A list of confidence scores for text detection boxes. - `text_det_params`: `(Dict[str, Dict[str, int, float]])` Configuration parameters for the text detection module. - `limit_side_len`: `(int)` The side length limit value during image preprocessing. - `limit_type`: `(str)` The processing method for the side length limit. - `thresh`: `(float)` The confidence threshold for text pixel classification. - `box_thresh`: `(float)` The confidence threshold for text detection boxes. - `unclip_ratio`: `(float)` The expansion coefficient for text detection boxes. - `text_type`: `(str)` The type of text detection, currently fixed as "general". - `text_type`: `(str)` The type of text detection, currently fixed as "general". - `textline_orientation_angles`: `(List[int])` The prediction results of text line orientation classification. When enabled, the actual angle values are returned (e.g., [0,0,1]). - `text_rec_score_thresh`: `(float)` The filtering threshold for text recognition results. - `rec_texts`: `(List[str])` A list of text recognition results, containing only texts with confidence scores exceeding `text_rec_score_thresh`. - `rec_scores`: `(List[float])` List of confidence scores for text recognition, filtered by `text_rec_score_thresh` - `rec_polys`: `(List[numpy.ndarray])` List of text detection boxes filtered by confidence, with the same format as `dt_polys` - `formula_res_list`: `(List[Dict[str, Union[numpy.ndarray, List[float], str]]])` List of formula recognition results, with each element being a dict - `rec_formula`: `(str)` Formula recognition result - `rec_polys`: `(numpy.ndarray)` Formula detection box, with shape (4, 2) and dtype int16 - `formula_region_id`: `(int)` Region number where the formula is located - `seal_res_list`: `(List[Dict[str, Union[numpy.ndarray, List[float], str]]])` List of seal text recognition results, with each element being a dict - `input_path`: `(str)` Input path of the seal image - `page_index`: `None`, as the input here is `numpy.ndarray`, so the value is `None` - `model_settings`: `(Dict)` Model configuration parameters for the seal text recognition sub-pipeline - `dt_polys`: `(List[numpy.ndarray])` List of seal detection boxes, with the same format as `dt_polys` - `text_det_params`: `(Dict[str, Dict[str, int, float]])` Configuration parameters for the seal detection module, with specific parameter meanings as above - `text_type`: `(str)` Type of seal detection, currently fixed as "seal" - `text_rec_score_thresh`: `(float)` Filtering threshold for seal text recognition results - `rec_texts`: `(List[str])` List of seal text recognition results, containing only texts with confidence exceeding `text_rec_score_thresh` - `rec_scores`: `(List[float])` List of confidence scores for seal text recognition, filtered by `text_rec_score_thresh` - `rec_polys`: `(List[numpy.ndarray])` List of seal detection boxes filtered by confidence, with the same format as `dt_polys` - `rec_boxes`: `(numpy.ndarray)` Array of rectangular bounding boxes for detection boxes, with shape (n, 4) and dtype int16. Each row represents a rectangle - `table_res_list`: `(List[Dict[str, Union[numpy.ndarray, List[float], str]]])` List of table recognition results, with each element being a dict - `cell_box_list`: `(List[numpy.ndarray])` List of bounding boxes for table cells - `pred_html`: `(str)` HTML format string of the table - `table_ocr_pred`: `(dict)` OCR recognition results for the table - `rec_polys`: `(List[numpy.ndarray])` List of detection boxes for cells - `rec_texts`: `(List[str])` Recognition results for cells - `rec_scores`: `(List[float])` Recognition confidence scores for cells - `rec_boxes`: `(numpy.ndarray)` Array of rectangular bounding boxes for detection boxes, with shape (n, 4) and dtype int16. Each row represents a rectangle - Calling the `save_to_json()` method will save the above content to the specified `save_path`. If a directory is specified, the saved path will be `save_path/{your_img_basename}_res.json`. If a file is specified, it will be saved directly to that file. Since json files do not support saving numpy arrays, the `numpy.array` types within will be converted to list form. - Calling the `save_to_img()` method will save the visualization results to the specified `save_path`. If a directory is specified, visualized images for layout region detection, global OCR, layout reading order, etc., will be saved. If a file is specified, it will be saved directly to that file. (Production lines typically contain many result images, so it is not recommended to directly specify a specific file path, as multiple images will be overwritten, retaining only the last one.) - Calling the `save_to_markdown()` method will save the converted Markdown file to the specified `save_path`. The saved file path will be `save_path/{your_img_basename}.md`. If the input is a PDF file, it is recommended to directly specify a directory; otherwise, multiple markdown files will be overwritten. Additionally, it also supports obtaining visualized images and prediction results with results through attributes, as follows:

Attribute	Attribute Description
json	Obtain the prediction jsonresult in the format
img	obtain in the format of dictvisualized image
markdown	obtain in the format of dictmarkdown result

- The prediction result obtained through the `json` attribute is data of dict type, with relevant content consistent with that saved by calling the `save_to_json()` method. - The prediction result returned by the `img` attribute is data of dict type. The keys are `layout_det_res`, `overall_ocr_res`, `text_paragraphs_ocr_res`, `formula_res_region1`, `table_cell_img`, and `seal_res_region1`, with corresponding values being `Image.Image` objects: used to display visualized images of layout region detection, OCR, OCR text paragraphs, formulas, tables, and seal results, respectively. If optional modules are not used, the dict only contains `layout_det_res`. - The prediction result returned by the `markdown` attribute is data of dict type. The keys are `markdown_texts`, `markdown_images`, and `page_continuation_flags`, with corresponding values being markdown text, images displayed in Markdown (`Image.Image` objects), and a bool tuple used to identify whether the first element on the current page is the start of a paragraph and whether the last element is the end of a paragraph, respectively.

3. Enhancing VLM Inference Performance Using Inference Acceleration Frameworks

The inference performance under the default configuration has not been fully optimized and may not meet actual production requirements. PaddleOCR supports enhancing the inference performance of VLM through inference acceleration frameworks such as vLLM and SGLang, thereby accelerating the inference speed in production lines. The usage process mainly consists of two steps:

1. Start the VLM inference service;
2. Configure the PaddleOCR production line to invoke the VLM inference service as a client.

3.1 Starting the VLM Inference Service

3.1.1 Using Docker Images

PaddleOCR provides Docker images for quickly starting the vLLM inference service. The service can be started using the following command:

docker run \
    -it \
    --rm \
    --gpus all \
    --network host \
    ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddlex-genai-vllm-server

The service listens on port 8080 by default.

When starting the container, you can pass in parameters to override the default configuration. The parameters are consistent with the paddleocr genai_server command (see the next subsection for details). For example:

docker run \
    -it \
    --rm \
    --gpus all \
    --network host \
    ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddlex-genai-vllm-server \
    paddlex_genai_server --model_name PaddleOCR-VL-0.9B --host 0.0.0.0 --port 8118 --backend vllm

3.1.2 Installation and Usage via PaddleOCR CLI

Since the inference acceleration framework may have dependency conflicts with the PaddlePaddle framework, it is recommended to install it in a virtual environment. Taking vLLM as an example:

# Create a virtual environment
python -m venv .venv
# Activate the environment
source .venv/bin/activate
# Install PaddleOCR
python -m pip install "paddleocr[doc-parser]"
# Install dependencies for inference acceleration service
paddleocr install_genai_server_deps vllm

Usage of the paddleocr install_genai_server_deps command:

paddleocr install_genai_server_deps <name of the inference acceleration framework>

The currently supported frameworks are named vllm and sglang, corresponding to vLLM and SGLang, respectively.

After installation, you can start the service using the paddleocr genai_server command:

paddleocr genai_server --model_name PaddleOCR-VL-0.9B --backend vllm --port 8118

The parameters supported by this command are as follows:

Parameter	Description
--model_name	Model name
--model_dir	Model directory
--host	Server hostname
--port	Server port number
--backend	Backend name, i.e., the name of the inference acceleration framework used. Options are vllm or sglang.
--backend_config	A YAML file can be specified, which contains backend configurations.

3.2 How to Use the Client

After starting the VLM inference service, the client can invoke the service through PaddleOCR.

3.2.1 CLI Invocation

The backend type (vllm-server or sglang-server) can be specified via --vl_rec_backend, and the service address can be specified via --vl_rec_server_url. For example:

paddleocr doc_parser --input paddleocr_vl_demo.png --vl_rec_backend vllm-server --vl_rec_server_url http://127.0.0.1:8118/v1

3.2.2 Python API Invocation

Pass the vl_rec_backend and vl_rec_server_url parameters when creating the PaddleOCRVL object:

pipeline = PaddleOCRVL(vl_rec_backend="vllm-server", vl_rec_server_url="http://127.0.0.1:8118/v1")

3.2.3 Service-Oriented Deployment

The fields VLRecognition.genai_config.backend and VLRecognition.genai_config.server_url can be modified in the configuration file, for example:

VLRecognition:
  ...
  genai_config:
    backend: vllm-server
    server_url: http://127.0.0.1:8118/v1

3.3 Performance Tuning

The default configuration is tuned on a single NVIDIA A100 and assumes exclusive client service, so it may not be suitable for other environments. If users encounter performance issues during actual use, they can try the following optimization methods.

3.3.1 Server-side Parameter Adjustment

Different inference acceleration frameworks support different parameters. Refer to their respective official documentation to learn about available parameters and when to adjust them:

• vLLM Official Parameter Tuning Guide
• SGLang Hyperparameter Tuning Documentation

The PaddleOCR VLM inference service supports parameter tuning through configuration files. The following example demonstrates how to adjust the gpu-memory-utilization and max-num-seqs parameters of the vLLM server:

1. Create a YAML file named vllm_config.yaml with the following content:

gpu-memory-utilization: 0.3
   max-num-seqs: 128


2. Specify the configuration file path when starting the service, for example, using the `paddleocr genai_server` command:

```bash
paddleocr genai_server --model_name PaddleOCR-VL-0.9B --backend vllm --backend_config vllm_config.yaml


If you are using a shell that supports process substitution (such as Bash), you can also pass configuration items directly when starting the service without creating a configuration file:

```bash
paddleocr genai_server --model_name PaddleOCR-VL-0.9B --backend vllm --backend_config <(echo -e 'gpu-memory-utilization: 0.3\nmax-num-seqs: 128')

3.3.2 Client-Side Parameter Adjustment

PaddleOCR groups sub-images from single or multiple input images and initiates concurrent requests to the server. Therefore, the number of concurrent requests significantly impacts performance.

• For the CLI and Python API, the maximum number of concurrent requests can be adjusted using the vl_rec_max_concurrency parameter.
• For service-based deployment, modify the VLRecognition.genai_config.max_concurrency field in the configuration file.

When there is a one-to-one correspondence between the client and the VLM inference service, and the server-side resources are sufficient, the number of concurrent requests can be appropriately increased to enhance performance. If the server needs to support multiple clients or has limited computational resources, the number of concurrent requests should be reduced to prevent service abnormalities caused by resource overload.

3.3.3 Recommendations for Performance Tuning on Common Hardware

The following configurations are tailored for scenarios with a one-to-one correspondence between the client and the VLM inference service.

NVIDIA RTX 3060

• Server-Side
• vLLM: gpu-memory-utilization=0.8

Comments