Custom node for ComfyUI that enhances PoseAndFaceDetection from ComfyUI-WanAnimatePreprocess with manual point tracking, stable face crops and temporal filtering.

Requirements

ComfyUI-WanAnimatePreprocess must be installed in custom_nodes/.

Node: Tracked Pose and Face Detection

Category: WanAnimatePreprocess

Problem solved

On nearly-static subjects with shadows, lighting changes or transient occlusions (leaves, objects passing in front), the internal YOLO tracker can lock onto a shadow instead of the real face — making the generated rig unstable.

How it works

Run once → the node saves the first frame and displays it in an interactive canvas inside the node UI
Shift+click on stable landmarks: pupils, nose, shoulders, hands, etc.
Run again → the node tracks those exact points with Lucas-Kanade optical flow across all frames, constrains YOLO+ViTPose to a ROI derived from those points, and produces stable face crops

Parameters

Parameter	Default	Description
`roi_padding`	1.3	ROI expansion factor around tracked points bbox. Increase if ViTPose cuts off body parts.
`smooth_keypoints`	0.0	EMA smoothing on ViTPose keypoints (0=off, 1=maximum). Confidence-aware: undetected keypoints hold their last value instead of pulling the rig toward zero.
`face_crop_scale`	1.5	Face crop size multiplier relative to tracked point spread. Increase if face is cut off at edges.
`face_median_window`	5	Temporal median window (frames) on face crop pixels. Removes transient occlusions (leaves, shadows). 1=disabled.

Canvas controls

Shift+click — add tracking point
Right-click on existing point — remove it
Undo — remove last point
Clear — remove all points
Points are saved in the workflow JSON and restored on reload

Outputs

Drop-in compatible with DrawViTPose, FaceBBoxToMask, WanVideoAnimateEmbeds — identical output types to the original PoseAndFaceDetection.

Output	Description
`pose_data`	ViTPose rig (smoothed if smooth_keypoints > 0)
`face_images`	Stable face crops — fixed bbox centered on tracked point median centroid + temporal median
`key_frame_body_points`	JSON keyframe data
`bboxes`	Person detection bbox
`face_bboxes`	Fixed face bboxes (same stability as face_images)

Installation

ComfyUI/custom_nodes/
├── ComfyUI-WanAnimatePreprocess/    ← required dependency
└── ComfyUI-WanAnimatePreprocess-track/  ← this package

Clone or extract this repo into ComfyUI/custom_nodes/ and restart ComfyUI.