| ComfyUI-AnimePose |
| Licensed under the GNU Affero General Public License v3.0 (see LICENSE). |
|
|
| ================================================================================ |
| Upstream work |
| ================================================================================ |
|
|
| This project is a derivative of: |
|
|
| bizarre-pose-estimator |
| https://github.com/ShuhongChen/bizarre-pose-estimator |
| Copyright (c) Shuhong Chen and Matthias Zwicker |
| Licensed under the GNU Affero General Public License v3.0 |
|
|
| Everything that makes pose estimation on illustrated characters work here -- |
| the model architecture, the trained weights, the segmentation model, the crop |
| geometry and the keypoint schema -- is their research and their work. This |
| project contributes packaging: a ComfyUI node, a Detectron2-free reimplementation |
| of the inference path, and a confidence gate on top of the model's heatmaps. |
|
|
| If you use this, cite their paper: |
|
|
| @inproceedings{chen2022bizarre, |
| title={Transfer Learning for Pose Estimation of Illustrated Characters}, |
| author={Chen, Shuhong and Zwicker, Matthias}, |
| booktitle={Proceedings of the IEEE/CVF Winter Conference on |
| Applications of Computer Vision}, |
| year={2022} |
| } |
|
|
| The published weight files are a repackaging of their released checkpoints: |
| same tensors, converted to safetensors, with the unused Detectron2 branch |
| removed. `scripts/verify_weights.py` checks this claim tensor-by-tensor |
| against the official release. |
|
|
| Because the upstream work is AGPL-3.0, so is this, and so is anything derived |
| from it. In particular, if you run this as part of a network service, the |
| AGPL's section 13 requires you to offer the corresponding source to the users |
| of that service. |
|
|
| ================================================================================ |
| Other components |
| ================================================================================ |
|
|
| The OpenPose-18 skeleton renderer in anime_pose/pose_convert.py reproduces the |
| drawing conventions (limb pairs, colours, stick width) of `draw_bodypose` from |
| comfyui_controlnet_aux's DWPose module, so output is interchangeable with it: |
|
|
| https://github.com/Fannovel16/comfyui_controlnet_aux |
|
|
| The keypoint prior is torchvision's keypointrcnn_resnet50_fpn, standing in for |
| the Detectron2 model the original uses: |
|
|
| https://github.com/pytorch/vision (BSD-3-Clause) |
|
|
