paddlepaddle/pp-doclayoutv3

Document layout detection only (PP-DocLayoutV3, an RT-DETR instance-segmentation plus reading-order model). It locates and classifies layout regions and predicts the logical reading order in a single forward pass; it does not recognize text. Accepts one image_url or a document_url PDF; no text-only input, and a text prompt is ignored. Default method: detect. Because it predicts multi-point boxes rather than axis-aligned rectangles, it holds up on skewed, curved, and photographed pages. Upstream it is the layout stage of PaddleOCR-VL 1.5 and GLM-OCR, so the natural pipeline is layout first (pp-doclayoutv3), then recognition on the regions you care about with an OCR model.

Choosing between this and the OCR models

You want	Use
Region boxes, labels, and reading order, no text	`paddlepaddle/pp-doclayoutv3`
Text plus line polygons	`paddleocr/pp-ocrv6` (`ocr`)
Text laid out as Markdown	`zai-org/glm-ocr`, `rednote-hilab/dots.mocr` (`markdown`)
Layout regions with their transcribed text in one call	`rednote-hilab/dots.mocr` (`parse_layout`)

Output by method

Method	Payload kind	Medium	`content.object`
`detect`	json	Image	`pp_doclayoutv3.detect.regions`, with `bbox_xywh`, optional `poly_xy`, `label`, `order`, `score`
`detect`	json	Document page	`document.page.blocks`, with the same geometry and a zero-based `index` in place of `order`

detect is the only method and the default, so method can be omitted. This is a pure layout detector with no OCR path: any other method is a 400. Text mode emits the payload alone for one image, and one <document> / <page> block per PDF with format="json" on every page; JSON mode wraps it in data. A page with no detections returns "items": [], never a string.

Region record

Key	Type	Notes
`bbox_xywh`	`[x, y, w, h]`	Normalized 0-1 box, rounded to `precision`.
`poly_xy`	`[[x, y], ...]`	Normalized 0-1 polygon, present when the model returns one; multi-point for skewed or curved regions.
`label`	`string`	Layout class (see below).
`order`	`int`	Predicted reading position, 1-based and contiguous after filtering.
`score`	`float`	Confidence 0-1, rounded to `precision`.

Coordinates are normalized against the input image or the rasterized page. See The region record. Layout classes: abstract, algorithm, aside_text, chart, content, doc_title, figure_title, footer, footnote, formula, formula_number, header, image, number, paragraph_title, reference, reference_content, seal, table, text, vision_footnote.

Method parameters

Key	Method(s)	Description
`score_threshold`	`detect`	Drop regions below this confidence (0.0-1.0). Default `0.5`. It also binarizes the instance masks that `poly_xy` is traced from, so lowering it both admits weaker regions and loosens polygons.

Request

from openai import OpenAI

client = OpenAI(
    base_url="https://gateway.vlm.run/v1/openai",
    api_key="<VLMRUN_API_KEY>",
)

response = client.chat.completions.create(
    model="paddlepaddle/pp-doclayoutv3",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://storage.googleapis.com/vlm-data-public-prod/hub/examples/document.receipt/playground/2.jpg"
                    },
                },
            ],
        }
    ],
    extra_body={"method": "detect", "method_params": {"score_threshold": 0.5}},
)

print(response.choices[0].message.content)

curl https://gateway.vlm.run/v1/openai/chat/completions \
  -X POST \
  -H "Authorization: Bearer $VLMRUN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "paddlepaddle/pp-doclayoutv3",
    "method": "detect",
    "method_params": {"score_threshold": 0.5},
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "image_url",
            "image_url": {
              "url": "https://storage.googleapis.com/vlm-data-public-prod/hub/examples/document.receipt/playground/2.jpg"
            }
          }
        ]
      }
    ]
  }'

Response

Text mode
Text mode (document)
JSON mode

A single image returns the region container alone, in reading order:

{"object": "pp_doclayoutv3.detect.regions", "items": [{"bbox_xywh": [0.0332, 0.0138, 0.1719, 0.0331], "poly_xy": [[0.0332, 0.0138], [0.2051, 0.0138], [0.2051, 0.047], [0.0332, 0.047]], "label": "doc_title", "order": 1, "score": 0.9712}, {"bbox_xywh": [0.0332, 0.0829, 0.2598, 0.0331], "label": "text", "order": 2, "score": 0.8934}]}

With a document_url PDF, one block container per page inside the <document> / <page> blocks:

<document file_name="contract.pdf" file_hash="sha256:1b7f…" file_bytes="182417" mimetype="application/pdf" num_pages="2" dpi="150">
<page page_index="0" format="json" page_width="1240" page_height="1754">
{"object": "document.page.blocks", "items": [{"index": 0, "bbox_xywh": [0.0274, 0.0114, 0.1419, 0.0274], "label": "doc_title", "score": 0.9712}]}
</page>
<page page_index="1" format="json" page_width="1240" page_height="1754">
{"object": "document.page.blocks", "items": [{"index": 0, "bbox_xywh": [0.0274, 0.0114, 0.6177, 0.2167], "label": "table", "score": 0.9302}]}
</page>
</document>

One image entry; content is the region container:

{
  "model": "paddlepaddle/pp-doclayoutv3",
  "method": "detect",
  "data": [
    {
      "object": "image",
      "image_hash": "sha256:...",
      "image_width": 1024,
      "image_height": 1448,
      "content": {
        "object": "pp_doclayoutv3.detect.regions",
        "items": [
          {
            "bbox_xywh": [0.0332, 0.0138, 0.1719, 0.0331],
            "poly_xy": [[0.0332, 0.0138], [0.2051, 0.0138], [0.2051, 0.047], [0.0332, 0.047]],
            "label": "doc_title",
            "order": 1,
            "score": 0.9712
          },
          {
            "bbox_xywh": [0.0332, 0.0829, 0.2598, 0.0331],
            "label": "text",
            "order": 2,
            "score": 0.8934
          }
        ]
      }
    }
  ]
}

For a PDF the entry is a document entry and each page’s content is that page’s document.page.blocks. See the Document OCR guide.

Aliases

paddlepaddle/pp-doclayoutv3 (preferred), pp-doclayoutv3, and PaddlePaddle/PP-DocLayoutV3_safetensors all resolve to this model.

Overview

Guides

API Reference

Misc

paddlepaddle/pp-doclayoutv3

Choosing between this and the OCR models

Output by method

Region record

Method parameters

Request

Response

Aliases

​Choosing between this and the OCR models

​Output by method

​Region record

​Method parameters

​Request

​Response

​Aliases

Choosing between this and the OCR models

Output by method

Region record

Method parameters

Request

Response

Aliases