IBB_POSE — ComfyUI Pose Estimation Node

A ComfyUI custom node for robust human pose estimation.
Inspired by SDPose-OOD.

Compatibility: Python 3.13 · torch 2.10+ · CUDA 12/13 · ComfyUI Portable

---

Features

Feature	Details
Skeleton types	Body (17 kpts COCO), OpenPose (18 kpts), WholeBody (133 kpts)
Detection modes	Single person · Multi-person
Output	Pose image · OpenPose JSON · Both
Auto-download	Models fetched automatically on first use
Compatibility	Python 3.13, torch 2.x, CUDA 12/13, CPU

---

Installation

1. Clone into your ComfyUI/custom_nodes/ directory:

   cd ComfyUI/custom_nodes
   git clone https://github.com/IBB666/IBB_POSE
   

2. Install base dependencies:

   cd IBB_POSE
   pip install -r requirements.txt
   

3. Install only the backend you plan to use:

   # Choose ONE backend path based on your workflow / hardware:
   
   # Body / OpenPose
   pip install ultralytics
   
   # WholeBody (CPU)
   pip install onnxruntime
   
   # WholeBody (GPU, optional replacement for CPU package)
   # python -m pip uninstall -y onnxruntime
   # pip install onnxruntime-gpu
   

4. Start ComfyUI – models download automatically on first use.

Windows / ComfyUI install note

• IBB_POSE no longer requires groundingdino-py, chumpy or huggingface_hub during node installation.
• If your embedded ComfyUI Python had previous failed installs cached, remove those partial packages before retrying, for example. Only do this if you know other nodes do not rely on those packages:

  python -m pip uninstall -y groundingdino-py chumpy huggingface_hub
  python -m pip cache purge
  

• The node now keeps heavy pose backends optional and only imports them when the selected mode needs them.
• The ComfyUI runtime no longer needs the OpenMMLab stack just to load the SDPose decoder. mmpose / mmcv / mmengine are only needed for the standalone evaluation scripts under /eval.

Manual model placement

Mode	Model	Destination
Body / OpenPose	yolo11n-pose.pt (small), yolo11m-pose.pt (medium), yolo11x-pose.pt (large)	ComfyUI/models/IBB_POSE/
WholeBody	dw-ll_ucoco_384.onnx + yolox_l.onnx	ComfyUI/models/IBB_POSE/

WholeBody models: yzd-v/DWPose

---

Nodes

IBB Pose — Load Model

Loads and caches the pose estimation model.

Parameter	Options	Description
skeleton_type	Body / WholeBody / OpenPose	Pose skeleton format
model_size	small / medium / large	Model size / accuracy tradeoff
device	auto / cuda / cpu	Inference device
precision	fp32 / fp16 / bf16	Numeric precision (fp32 for CPU)
auto_download	True / False	Fetch model files automatically

Output: IBB_POSE_MODEL

---

IBB Pose — Run Estimation

Runs pose estimation on one or more images.

Parameter	Options	Description
ibb_pose_model	IBB_POSE_MODEL	Model from Load node
image	IMAGE	Input image(s)
detection_mode	single / multi	Person detection strategy
output_type	PoseImage / JSON / Both	What to compute
score_threshold	0.05–0.95	Minimum keypoint confidence
overlay_alpha	0.0–1.0	0=original, 1=pure pose map
pose_scale	0.1–5.0	Line / dot thickness multiplier
keep_face	bool	Include face keypoints (WholeBody)
keep_hands	bool	Include hand keypoints (WholeBody)
keep_feet	bool	Include foot keypoints (WholeBody)
save_json	bool	Save JSON to ComfyUI/output/
filename_prefix	string	Output JSON file prefix

Outputs: pose_image (IMAGE) · pose_json (STRING)

---

IBB Pose — JSON to Image

模型目录

• SDPose 模型：ComfyUI/models/IBB_POSE/
• YOLO 模型：ComfyUI/models/yolo/
• GroundingDINO 模型：ComfyUI/models/grounding-dino/

首次运行时会自动下载以下资源：

• teemosliang/SDPose-Body
• teemosliang/SDPose-Wholebody
• GroundingDINO 配置和权重

仓库已内置 empty_text_encoder/empty_embedding.safetensors，也保留了 generate_empty_embedding.py 以便重新生成。

工作流

示例工作流位于：

• /workflow/ibb_pose.json

兼容说明

• 依赖 ComfyUI 的 folder_paths / model_management
• Florence2 检测为可选能力，需要额外安装对应 ComfyUI 节点
• OpenPose 编辑 JSON 可与 ComfyUI-OpenPose-Editor-jd 配合使用

致谢

• 上游实现：judian17/ComfyUI-SDPose-OOD
• 原始项目：T-S-Liang/SDPose-OOD