OpenPose Rotator

⚠ Warning: This project is under active development and does not work well yet. Results may be inaccurate or unstable. Use at your own risk.

A ComfyUI custom node that rotates OpenPose figures around their torso pivot point. The node accepts an image and optional pose keypoints, detects the torso, applies Y-axis rotation (clockwise/counterclockwise), and outputs the rotated pose image.

Features

• Input: Image (required) + optional POSE_KEYPOINT from OpenPose/DWPose preprocessors
• Parameters: Direction (clockwise/counterclockwise), degrees (1-360), mode (simple/advanced)
• Fallback: When POSE_KEYPOINT is not provided, uses DWPose from comfyui-controlnet-aux to extract keypoints (requires that extension)
• Error handling: Returns input image unchanged if torso cannot be detected
• Anatomy-aware rotation: Limb-specific depth scales per OpenPose COCO body indices for more natural turns
• Adaptive depth: Depth scale adapts to shoulder width for consistent results across image sizes
• Face visibility: Face/head keypoints (nose, eyes, ears) hidden when rotation >90° (back of head)
• Occlusion: Points behind the torso are hidden; limbs drawn in depth order for correct overlap

Installation

1. Clone or copy this folder into your ComfyUI custom_nodes directory. Use a valid Python module name (underscores, not hyphens):

   ComfyUI/custom_nodes/open_pose_rotator/
   

   (If your folder is open-pose-rotator, rename it to open_pose_rotator so Python can import it.)

2. Install dependencies:

   pip install -r requirements.txt
   

3. (Optional, for image-only input) Install comfyui-controlnet-aux for DWPose fallback when pose keypoints are not connected.

Rotation Modes

• Simple (default): Orthographic projection with perspective-modulated depth inference. Best for eye-level views. Use perspective to improve results when the source image has a high or low camera angle (e.g. isometric).
• Advanced: Perspective projection after rotation; output Y can change; foreshortening applied. Use when you need more realistic 3D behavior for non–eye-level views. Requires tuning perspective and focal_length.

Parameters

| Parameter | Simple | Advanced | Description | |-----------|--------|----------|-------------| | perspective | Yes | Yes | 0 = eye level, positive = camera above (isometric), negative = camera below. Range: -0.5 to 0.5. | | focal_length | Ignored | Yes | Controls perspective strength; higher = flatter, lower = more foreshortening. Default 800, range 100–5000. |

Usage

1. Add OpenPose Rotator from the node menu under image/pose
2. Connect an OpenPose/DWPose image to the image input
3. Optionally connect POSE_KEYPOINT from DWPose Estimator or OpenPose Pose for better accuracy
4. Set mode (simple or advanced)
5. Set direction (counterclockwise or clockwise, when viewed from above)
6. Set degrees (1-360)
7. For isometric or low-angle source images, adjust perspective; for Advanced mode, use focal_length to control foreshortening

Workflow Example

• With keypoints: Load Image → DWPose Estimator → OpenPose Rotator (connect both IMAGE and POSE_KEYPOINT) → Apply ControlNet
• Image only: Load Image → OpenPose Rotator → Apply ControlNet (requires comfyui-controlnet-aux for DWPose)

Implementation Notes

Rotation uses geometry-based depth inference from 2D OpenPose keypoints (no ML dependency). Depth is inferred per OpenPose COCO body index using anatomic proportions. Simple mode uses orthographic projection with a perspective term that modulates depth by vertical position. Advanced mode applies perspective projection after rotation, allowing output Y to change and producing foreshortening. For true 3D pose lifting from 2D, libraries like VideoPose3D or MMPose exist but use different keypoint formats (e.g. Human3.6M) and would require format conversion.

License

MIT