---
license: agpl-3.0
tags:
  - pose-estimation
  - comfyui
  - anime
library_name: pytorch
---

# ComfyUI-AnimePose — weights

Weight files for the [ComfyUI-AnimePose](https://github.com/dalai2/ComfyUI-AnimePose)
custom node. The node downloads them automatically on first run; you do not
need to fetch them by hand.

| File | Size | What it is |
|---|---|---|
| `anime_pose_head.safetensors` | 37.5 MB | ResNet backbone + keypoint head |
| `character_bg_seg.safetensors` | 233 MB | Character/background segmenter |

## Provenance

These are **a repackaging of someone else's released checkpoints**, not newly
trained weights. The source is:

> **bizarre-pose-estimator** — https://github.com/ShuhongChen/bizarre-pose-estimator
> Copyright (c) Shuhong Chen and Matthias Zwicker, AGPL-3.0

Two changes were made, both mechanical:

1. Converted from `.ckpt` to safetensors, so loading does not unpickle.
2. Dropped the `rcnn.*` Detectron2 R101 keypoint branch from the pose
   checkpoint (313 MB), which the node never loads — torchvision's
   `keypointrcnn_resnet50_fpn` stands in for it. 351 MB → 37.5 MB.

No tensor was retrained, fine-tuned, quantized or otherwise altered.
`scripts/verify_weights.py` in the node repo checks that claim
tensor-by-tensor against the official release: 336/336 pose tensors and
690/690 segmentation tensors bit-identical, 573 `rcnn.*` tensors dropped on
purpose.

## Citation

If you use this, cite the original paper:

```bibtex
@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}
}
```

## License

AGPL-3.0, inherited from the upstream work. Anything derived from these
weights is AGPL-3.0 too. In particular, AGPL section 13 requires that if you
run this as part of a network service, you offer the corresponding source to
the users of that service.