DocLayout-YOLO — Model Checkpoints

Official checkpoints for DocLayout-YOLO, a document-specialized real-time detector built on YOLOv12 with Page-Axis Layout Attention (PALA) and mesh-guided synthetic pretraining.

Checkpoints

Pretrained models (for fine-tuning on your own layout datasets), trained on large-scale synthetic document pretraining data:

• DocLayout-YOLO-pretrain-N.pt
• DocLayout-YOLO-pretrain-S.pt
• DocLayout-YOLO-pretrain-M.pt
• DocLayout-YOLO-pretrain-L.pt
• DocLayout-YOLO-pretrain-X.pt

Downstream models (fine-tuned on four public document-layout benchmarks, five model sizes each), under downstream/:

• downstream/D4LA/ · downstream/M6Doc/ · downstream/HJDataset/ · downstream/PRIMA-LAD/

General-purpose model (22 fine-grained layout categories, drop-in front-end for document parsing pipelines):

• DocLayout-YOLO-universal.pt

Quick inference:

from ultralytics import YOLO

model = YOLO("DocLayout-YOLO-universal.pt")  # download from the link above
results = model.predict("page.png", imgsz=1024, conf=0.25, iou=0.5, verbose=False)
results[0].save(filename="result.jpg")

Or try it online without any setup: Online Demo (ModelScope).

Main Results (mAP50-95)

Model	D4LA	M6Doc	HJDataset	PRIMA-LAD
n	67.2	66.2	93.6	67.7
s	70.6	71.1	94.8	70.3
m	70.7	73.3	95.3	71.3
l	72.4	75.1	95.7	73.4
x	72.4	75.5	96.0	73.1

Click a score to download the corresponding checkpoint.

Datasets

The four downstream benchmarks are released as separate dataset repositories, each a YOLO-layout dataset directory (images/, labels/, split lists train.txt / val.txt / test.txt):

dataset	repository	pages
D4LA	juliozhao/D4LA	22,184
M6Doc	juliozhao/M6Doc	17,252
HJDataset	juliozhao/HJDataset	4,012
PRIMA-LAD	juliozhao/PRIMA-LAD	956

Download

# pip install huggingface_hub
from huggingface_hub import snapshot_download

for repo, local in [
    ("juliozhao/D4LA",      "D4LA"),        # keep these exact directory names
    ("juliozhao/M6Doc",     "m6doc"),       # (the val/train scripts depend on them)
    ("juliozhao/HJDataset", "HJDataset"),
    ("juliozhao/PRIMA-LAD", "prima-lad"),
]:
    snapshot_download(repo, repo_type="dataset", local_dir=local)

Each repository is ready to use as-is: the split lists point to the actual file locations, so no post-processing is needed after download.

Usage

The val.py / train.py scripts referenced below ship in the release package. Typical flows:

Validation (dataset root = the parent directory holding the four dataset folders):

python val.py --model downstream/PRIMA-LAD/s_DocLayout-YOLO.pt \
  --data prima-lad --dataset-root /absolute/path/to/datasets

--data accepts one of the dataset aliases d4la | m6doc | hjdataset | prima-lad.

Fine-tuning from a pretrained checkpoint:

python train.py --data ultralytics/cfg/datasets/D4LA.yaml \
  --pretrain DocLayout-YOLO-pretrain-X.pt \
  --epoch 500 --image-size 1600 --batch-size 64 --project runs

Prediction:

from ultralytics import YOLO

model = YOLO("downstream/M6Doc/m_DocLayout-YOLO.pt")
results = model.predict("page.png", imgsz=1600, conf=0.25, device="cuda:0")
results[0].save(filename="result.jpg")

SHA256 (first 16 hex chars)

repo path	sha256 (16)	size (MB)
DocLayout-YOLO-pretrain-N.pt	178a9ff51e3cd848	9.6
DocLayout-YOLO-pretrain-S.pt	f203c9d02740a549	35.8
DocLayout-YOLO-pretrain-M.pt	bfbec6286ef2f659	79.3
DocLayout-YOLO-pretrain-L.pt	da336fad052f7ff2	107.0
DocLayout-YOLO-pretrain-X.pt	fe0d4daef3d14c8a	238.6
downstream/D4LA/n_DocLayout-YOLO.pt	ddc1189beed91e27	5.0
downstream/D4LA/s_DocLayout-YOLO.pt	a1d0f8efe22580f3	18.2
downstream/D4LA/m_DocLayout-YOLO.pt	ecff5e042e03626b	39.9
downstream/D4LA/l_DocLayout-YOLO.pt	826ec4281d1064a2	53.9
downstream/D4LA/x_DocLayout-YOLO.pt	fa1d38605ee26115	119.8
downstream/M6Doc/n_DocLayout-YOLO.pt	62e3c9352a050769	5.0
downstream/M6Doc/s_DocLayout-YOLO.pt	7ce579090a2a0d51	18.1
downstream/M6Doc/m_DocLayout-YOLO.pt	e0b3f0356ed4bb6e	40.0
downstream/M6Doc/l_DocLayout-YOLO.pt	f4cf7b8e56978ade	53.9
downstream/M6Doc/x_DocLayout-YOLO.pt	bddd9e3ebb10a412	119.8
downstream/HJDataset/n_DocLayout-YOLO.pt	1099870f0b82c01f	5.0
downstream/HJDataset/s_DocLayout-YOLO.pt	6044020c8f13d5de	18.2
downstream/HJDataset/m_DocLayout-YOLO.pt	65cf6375760cb08b	39.9
downstream/HJDataset/l_DocLayout-YOLO.pt	5878bc467353c810	53.9
downstream/HJDataset/x_DocLayout-YOLO.pt	7dd575f4115537ea	119.7
downstream/PRIMA-LAD/n_DocLayout-YOLO.pt	f21f9a798a021374	5.0
downstream/PRIMA-LAD/s_DocLayout-YOLO.pt	4e9177b75a3c825c	18.2
downstream/PRIMA-LAD/m_DocLayout-YOLO.pt	1fcccb0f9273d83a	39.9
downstream/PRIMA-LAD/l_DocLayout-YOLO.pt	38fb188c3567d033	53.9
downstream/PRIMA-LAD/x_DocLayout-YOLO.pt	d79b8c6f9af4e47a	119.7
universal/DocLayout-YOLO-universal.pt	5c733afe7481aea1	119.6

License

The source code is distributed under the GNU Affero General Public License v3.0 (AGPL-3.0).

Citation

@misc{zhao2024doclayoutyoloenhancingdocumentlayout,
      title={DocLayout-YOLO: Enhancing Document Layout Analysis through Diverse Synthetic Data and Global-to-Local Adaptive Perception},
      author={Zhiyuan Zhao and Hengrui Kang and Bin Wang and Conghui He},
      year={2024},
      eprint={2410.12628},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2410.12628},
}