ComfyUI-AutoTrackFaces

Use RetinaFace to detect and automatically track faces in video sequences with smooth camera movement effects

Custom Node

Auto Track Faces (Video)

Automatically track faces in video sequences with smooth camera movement effects. This node processes a batch of video frames and creates a cinematic tracking effect that follows the largest face in the scene.

• images: Input video frames as a batch of images with shape (batch, height, width, channels)

• smoothness: Controls camera movement smoothness (default: 0.85, range: 0.0-0.98). 0=no smoothing, 0.85=recommended, 0.95=ultra smooth (eliminates jitter)

• responsiveness: Controls reaction speed to face movement (default: 0.3, range: 0.1-1.0). 0.1=very stable, 0.3=recommended, 1.0=fast response (may jitter)

• detect_interval: Face detection frequency for performance optimization (default: 1, range: 1-30). 1=every frame (most accurate), 5=5x speed boost

• aspect_ratio: Output aspect ratio for the tracked video (default: 16:9)

• output_width: Output video width, 0=auto calculate based on aspect ratio (default: 0, range: 0-8192)

• output_height: Output video height, 0=auto calculate based on aspect ratio (default: 0, range: 0-8192)

• scale_factor: How much padding to add around detected faces (default: 8.5, range: 1.0-10.0)

• shift_factor: Vertical positioning of face in frame (default: 0.33, range: 0-1). 0=top edge, 0.5=center, 1.0=bottom edge

Returns:

• tracked_video: The processed video with smooth face tracking
• tracking_info: Detailed tracking report including statistics and frame-by-frame information

Algorithm Features:

• Intelligent keyframe selection for optimal performance
• Exponential Moving Average (EMA) smoothing with adaptive speed limits
• Subpixel smoothing to eliminate tiny camera jitter
• Automatic resolution calculation based on input and aspect ratio
• Smooth interpolation between detection frames
• Comprehensive tracking statistics and reporting

Performance Tips:

• Higher detect_interval values provide significant speed boosts (e.g., 5x faster with interval=5)
• The algorithm automatically balances detection accuracy with performance
• Use reasonable scale_factor values (0.8-0.9) for best results
• For videos with fast-moving subjects, increase responsiveness slightly

---

Forked and modified from liusida/ComfyUI-AutoCropFaces