[中文说明]

This project is a ComfyUI custom node implementation of SDPose-OOD.

It brings the powerful and robust pose estimation capabilities of SDPose-OOD into the ComfyUI workflow, allowing for the extraction of high-quality body or whole-body poses from images.

• Two Model Types:

  * Body: 17 keypoints, body-only pose.

  * WholeBody: 133 keypoints, including body, face, and hands.

• Automatic Model Download: Nodes will automatically download the required models on first use.

• YOLO Integration: Optional YOLO-based person detection for precise multi-person pose estimation.

• Advanced BBox Detection: In addition to YOLO, now supports precise person detection using ComfyUI-Florence2 (https://github.com/kijai/ComfyUI-Florence2) or GroundingDINO(from ComfyUI-SAM2: ) (https://github.com/neverbiasu/ComfyUI-SAM2).

• Keypoint Filtering: Options to keep or remove keypoints for the face, hands, and feet to suit different needs.

• Precision Control: Supports bf16, fp16, and fp32 inference to balance speed and VRAM usage.

• OpenPose Editor Compatibility: One-click save poses as JSON files for import and editing in the ComfyUI-OpenPose-Editor.

1. Install this node:

  * Clone this repository into your ComfyUI/custom_nodes/ directory using git clone.

  * cd ComfyUI/custom_nodes/SDPose-OOD-ComfyUI

  * pip install -r requirements.txt

2. Install Models (Automatic or Manual):

  * Automatic (Recommended): The nodes (Load SDPose Model, Load YOLO Model) will automatically download models from Hugging Face and Github on first run and place them in the correct directories.

  * Manual: You can also download the models yourself and place them in the corresponding folders:

  * SDPose Models: Download and place in ComfyUI/models/SDPose_OOD/

  * SDPose-Body: huggingface.co/teemosliang/SDPose-Body

  * SDPose-Wholebody: huggingface.co/teemosliang/SDPose-Wholebody

  * YOLO Models: Download and place in ComfyUI/models/yolo/

  * YOLOv8: yolov8n-pose.pt

  * YOLOv11: yolo11x.pt (or other YOLO models)

  * Default YOLO models (like yolov8n-pose.pt) are better suited for detecting real people.

  * SDPose-OOD itself handles anime characters well (even without YOLO).

  * If you don't want to use Florence2/GroundingDINO, you can consider this detection model which performs excellently for both real and anime characters: real_person_detection. Download your preferred model version (.pt file) from the repo and place it in the models/yolo folder.

  * groundingdino model: Download the model file and py file for either of the two models listed below, and place them in the modelsgrounding-dino directory. Usually, the model will be downloaded automatically.

  * groundingdino_swint_ogc: groundingdino_swint_ogc and GroundingDINO_SwinT_OGC.cfg.py

  * ggroundingdino_swinb_cogcoor: groundingdino_swinb_cogcoor and GroundingDINO_SwinB.cfg

3. Install Optional Dependencies (for Advanced BBox Detection):

  * For Florence2, install ComfyUI-Florence2.

• model_type: Choose Body (17 keypoints) or WholeBody (133 keypoints).

• unet_precision: Precision selection.

  * bf16: Recommended for 30-series GPUs and newer.

  * fp16: Recommended for 20-series GPUs and older.

  * fp32: For CPU inference or when maximum precision is required.

  * Note: bf16/fp16 significantly reduce VRAM usage with almost identical speed to fp32.

• device: auto will detect CUDA. CPU inference only supports fp32.

• Loads a YOLO model (e.g., yolov8n-pose.pt).

• Connect the YOLO_MODEL output of this node to the Run SDPose Estimation node.

• sdpose_model: The model from Load SDPose Model.

• image: Your input image.

• yolo_model (Optional):

  * Not connected (and bbox_input also not connected): The node will process the entire image. Suitable for single-person images.

  * Connected: The node will first use YOLO to detect all persons in the image, then process each detected person individually.

• bbox_input (Optional): New input. Can be connected to the BBox output from ComfyUI-Florence2 or GroundingDINO (ComfyUI-SAM2).

  * Note: Detection Priority: If yolo_model and (grounding_dino_model or data_from_florence2) are connected simultaneously, the node will prioritize data_from_florence2 first, followed by GroundingDINO. (This implies yolo_model is the lowest priority).

• score_threshold: Keypoint confidence threshold. Points below this score will be ignored.

• keep_face (bool): True to keep face keypoints (only effective for WholeBody).

• keep_hands (bool): True to keep hand keypoints (only effective for WholeBody).

• keep_feet (bool): True to keep feet keypoints (only effective for WholeBody).

• overlay_alpha: Controls the visibility of the output image.

  * 0.0: 100% original image.

  * 0.6: Blend of original image and pose map (original image at 60% brightness).

  * 1.0: 100% pose map (pure black background).

• save_for_editor (bool):

  * True: Saves a JSON file to the ComfyUI/output/poses/ directory with the prefix pose_edit.

  * This JSON file can be loaded by the ComfyUI-OpenPose-Editor node for further editing.

• image (Output): The final visualized image (based on overlay_alpha setting).

• pose_json (Output):

  * Important: This is a string output containing the original JSON format from the SDPose-OOD project.

  * This format is different from the JSON saved by save_for_editor and cannot be used with the OpenPose editor node.

workflow

1. Load Image to load your source image.

2. Load SDPose Model and select WholeBody and bf16.

3. Load YOLO Model and select yolov8n-pose.pt. (Or use Florence2 / GroundingDINO and connect to the appropriate input).

4. Connect all three (or replace #3 with a BBox node) to Run SDPose Estimation.

5. Set overlay_alpha to 1.0 (if you want a pure pose map).

6. Set save_for_editor to True (if you want to edit it later).

7. Connect a Preview Image node to the image output.