A 3.7M-parameter D-FINE nano detector that finds layout regions on document pages: Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text and Title.
Fine-tuned from ustc-community/dfine-nano-coco on the full DocLayNet v1.2 training set, with pages containing Footnote, Formula or Title repeated 3x. 12 epochs at 1024x1024 on one A100.
It is small enough to run in a browser: onnx/layout_fp16.onnx (8 MB) runs in onnxruntime-web on WebGPU at roughly 40 ms per page.
Apache-2.0, like the D-FINE base weights. Training data: DocLayNet (CDLA-Permissive-1.0) by IBM Research.
import numpy as np, torch
from PIL import Image
from transformers import DFineForObjectDetection
model = DFineForObjectDetection.from_pretrained("avbiswas/dfine-nano-doclaynet").eval()
page = Image.open("page.png").convert("RGB") # render PDF pages at ~144 DPI
W, H = page.size
x = torch.from_numpy(np.asarray(page.resize((1024, 1024)), dtype=np.float32) / 255)
x = x.permute(2, 0, 1)[None] # plain resize + /255, no mean/std normalisation
with torch.no_grad():
out = model(pixel_values=x)
scores, labels = out.logits[0].sigmoid().max(-1) # 300 queries, one class each, no NMS
for (cx, cy, w, h), c, s in zip(out.pred_boxes[0].tolist(), labels.tolist(), scores.tolist()):
if s >= 0.5:
print(model.config.id2label[c], round(s, 2), [(cx - w/2) * W, (cy - h/2) * H, (cx + w/2) * W, (cy + h/2) * H])
mAP50-95 on a 2,500-page DocLayNet v1.2 test subset (shards 0, 2 and 4 of the official test split), at 1024 px:
| Class | This model (D-FINE nano, 3.7M) | hantian/yolo-doclaynet YOLO11s (~9.4M) |
|---|---|---|
| All | 0.652 | 0.744 |
| mAP50 (all) | 0.828 | 0.906 |
| Caption | 0.614 | 0.793 |
| Footnote | 0.454 | 0.620 |
| Formula | 0.597 | 0.714 |
| List-item | 0.706 | 0.816 |
| Page-footer | 0.666 | 0.569 |
| Page-header | 0.704 | 0.698 |
| Picture | 0.657 | 0.787 |
| Section-header | 0.620 | 0.711 |
| Table | 0.671 | 0.819 |
| Text | 0.793 | 0.852 |
| Title | 0.686 | 0.803 |
The YOLO11s numbers come from a different evaluation script (Ultralytics' validator) on the same pages, so treat the comparison as approximate. This model trades some accuracy for size and a permissive license; avbiswas/yolo11n-doclaynet scores higher (0.760) but is AGPL-3.0.
model.safetensors 15 MBonnx/layout_fp16.onnx 8 MB