D-FINE for document layout analysis

This is an adaptation//refactor of ArgoHA's implementation of the D-FINE object detector intended to put it in line with community standards in the (historical) document layout analysis community. It is a foundation for eventual integration into the kraken ATR engine.

Installation

Kraken is currently being rewritten to allow integration of new methods, such as the one in this repository, with plug-ins. This repository uses the architecture introduced by this rework which will eventually become kraken 7.0. The models produced by this repository are not going to be compatible with earlier kraken versions.

For the latest stable release run:

$ pip install dfine_kraken

Clone the repository and run:

$ pip install .

This will install a dfine that can be used to train models:

$ dfine ... train ...

Training

The basic syntax is very similar to kraken segmentation training using Page or ALTO XML files. During the rework many of the segmentation dataset filtering and transformation options have disappeared, being replaced by dictionaries mapping class labels to indices. As it is annoying to define mapping on the command line, mappings that do not assign one index to each class in the source data need to be defined in YAML experiment configuration files.

To train a basic model for 50 epochs from scratch:

$ dfine -d cuda:0 train *.xml

The default configuration trains lines and regions jointly. If this is not what you want take a look at the sample configuration file to see how to disable text line detection in bbox format.

Inference

Inference is integrated in kraken. You need to convert the checkpoint into weights first:

$ ketos convert -o dfine.safetensors checkpoint.ckpt

and then run kraken ... segment ... as usual:

$ kraken -i input.jpg out.xml -a segment -i dfine.safetensors

Pretrained Models

Pretrained region segmentation models trained on the LADaS dataset using the SegmOnto taxonomy (37 region types) are available for all model variants. They can be downloaded with kraken download:

Variant	Size	mAP@50	Download command
nano	~15 MB	0.2960	`kraken get 10.5281/zenodo.18715384`
small	~42 MB	0.3864	`kraken get 10.5281/zenodo.18715381`
medium	~79 MB	0.3915	`kraken get 10.5281/zenodo.18715373`
large	~126 MB	0.4308	`kraken get 10.5281/zenodo.18715364`
extra_large	~252 MB	0.4164	`kraken get 10.5281/zenodo.18715367`

The large variant achieves the best performance. The extra_large variant shows slight regression likely due to overfitting, so the large model is recommended for best accuracy.

$ kraken get 10.5281/zenodo.18715373
Processing ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 79.1/79.1 MB 0:00:00 0:00:02
Model dir: /home/mittagessen/.local/share/htrmopo/d58d541d-82cf-5ab2-b27b-a20fe2548229 (model files: ladas_m.safetensors)
$ kraken -i input.jpg out.xml -a segment -i ladas_m.safetensors
...

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

dfine_kraken-0.4.3.tar.gz (55.7 kB view details)

Uploaded Aug 25, 2026 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

dfine_kraken-0.4.3-py3-none-any.whl (66.4 kB view details)

Uploaded Aug 25, 2026 Python 3

File details

Details for the file dfine_kraken-0.4.3.tar.gz.

File metadata

Download URL: dfine_kraken-0.4.3.tar.gz
Upload date: Aug 25, 2026
Size: 55.7 kB
Tags: Source
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for dfine_kraken-0.4.3.tar.gz
Algorithm	Hash digest
SHA256	`30a3f583bde134f91ed97063b79f896c66d597f1e825aba0f09f899f07831a5e`
MD5	`47bb70cf8728180b8a095b429a21692d`
BLAKE2b-256	`37a5983c99d3478ab105a9bfd51616f7f4e9b7776b24f285088d9a377775834c`

See more details on using hashes here.

Provenance

The following attestation bundles were made for dfine_kraken-0.4.3.tar.gz:

Publisher: test.yml on mittagessen/dfine_kraken

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Statement:
- Statement type: https://in-toto.io/Statement/v1
- Predicate type: https://docs.pypi.org/attestations/publish/v1
- Subject name: dfine_kraken-0.4.3.tar.gz
- Subject digest: 30a3f583bde134f91ed97063b79f896c66d597f1e825aba0f09f899f07831a5e
- Sigstore transparency entry: 2582979456
- Sigstore integration time: Aug 25, 2026
Source repository:
- Permalink: mittagessen/dfine_kraken@9c9767af833b349b78d718f0f26cf34ba91a476a
- Branch / Tag: refs/tags/0.4.3
- Owner: https://github.com/mittagessen
- Access: public
Publication detail:
- Token Issuer: https://token.actions.githubusercontent.com
- Runner Environment: github-hosted
- Publication workflow: test.yml@9c9767af833b349b78d718f0f26cf34ba91a476a
- Trigger Event: push