zjkhurry/comfyui-docres-weights

DocRes / DDC dewarping weights

Inference-only weights for the comfyui-docres ComfyUI nodes. These are not trained here — they are conversions of the checkpoints released by the original authors, kept in safetensors form so they can be downloaded without a pickle loader.

Files

file	size	used by
docres.safetensors	58 MB	DocRes Restore — dewarping, deshadowing, appearance, deblurring
mbd.safetensors	227 MB	DocRes Restore — page mask for the dewarping task only
ddc_fiducial1024_v1.safetensors	51 MB	DDC Predict Points — control-point grid

Sources and citations

docres.safetensors, mbd.safetensors

From DocRes — A Generalist Model Toward Unifying Document Image Restoration Tasks by Jiaxin Zhang, Dezhi Peng, Chongyu Liu, Peirong Zhang, Lianwen Jin, CVPR 2024. MIT licensed, © 2024 Jiaxin Zhang. Paper: https://arxiv.org/abs/2405.04408

Original checkpoints: checkpoints/docres.pkl and data/MBD/checkpoint/mbd.pkl from the authors' release. They were published as full training checkpoints containing an optimizer_state alongside the model weights; those optimizer tensors are stripped here, which is what accounts for the size difference. Tensor values are unmodified apart from removal of the DataParallel module. prefix.

The page-mask network (mbd) comes from MBD — Mask-Based Deepwarping (Yuefei Gu et al.).

Backbone: Restormer (Zamir et al., 2021).

ddc_fiducial1024_v1.safetensors

From Document-Dewarping-with-Control-Points — Document Dewarping with Control Points by Guo-Wang Xie, Fei Yin, Xu-Yao Zhang, Cheng-Lin Liu, 2022. Paper: https://arxiv.org/abs/2203.10543 Code: https://github.com/gwxie/Document-Dewarping-with-Control-Points (MIT)

This is the authors' released ICDAR 2021 competition checkpoint (2021-02-03 16_15_55flat_img_by_fiducial_points-fiducial1024_v1.pkl, from Source/ICDAR2021/.../143/model_parameter). Same treatment as above: the optimizer_state is dropped and the module. prefix removed. Verified to produce bit-identical predictions to the original .pkl (max difference 0.000000).

Loading

from safetensors.torch import load_file
state = load_file("docres.safetensors")   # keys have no `module.` prefix

Intended use

Document image restoration and page dewarping of scanned or photographed documents. Intended for research and general use; verify results before relying on them for archival or legal purposes.

License

MIT, following the upstream projects. See the linked repositories for the authoritative terms.