Mask Tracker — Motion/Propagate/Anchor/Consistency

This pack ships a lot of video-mask machinery, and Mask Tracker (MEC) is the node that pulls four of those engines under one roof. Pick a mode, and the corresponding engine runs: per-frame motion detection, seed-mask propagation, keyframe interpolation, or a flicker consistency check. Same (mask, video) input pair for all of them, same five-port output schema. It's the "unified video-mask tracker" and it's worth reaching for before you go hunting through the individual nodes.

The four modes

• motion (default) - per-frame motion mask from four independent detectors: pixel_diff, optical_flow (Farneback or phase correlation), background_sub, and histogram_diff. camera_compensation (default on) subtracts global camera motion via stabilization_method (homography/affine/translation) so a panning shot doesn't trigger motion everywhere. combine_method (union/intersection) merges the active detectors, then grow_pixels and min_region_size clean up the result. The output score is motion intensity. This is the mode for "what actually moved in this clip."
• propagate - seed mask on one source_frame, push it to all frames. Methods: static (same mask everywhere), optical_flow, sam2_video (needs a wired sam_model + optional points_json), fade, scale_linear. bidirectional propagates both forward and backward from the seed. This is the rotoscope-helper.
• anchor - SDF-based interpolation between keyframe masks. Give anchor_frames as a CSV ("0,10,30") and one mask per anchor in the mask input; it morphs smoothly between them across total_frames with your choice of easing and sdf_iterations. Optional flow_refinement warps the SDF to follow actual motion (needs video). The output score is mean confidence.
• consistency_check - flicker scoring between consecutive frames, via metric (mask_iou / pixel_diff / flow_warp). This is the QA mode: run it on a propagated batch to find the frames that break.

Outputs

All modes share: masks (per-frame batch), preview, score (mode-specific scalar), info_json (the mode's diagnostic payload), and metric (a label string saying what produced the result). The mask input is required by propagate/anchor/consistency but optional for motion.

Where it fits

For a video cutout, propagate + SAM2 is the quality path (SAM2's video memory is genuinely good at holding identity). For change detection, motion with camera compensation is the fast, model-free path. For a clean keyframe morph without any tracking, anchor is overkill-free interpolation. And after any of them, consistency_check tells you if it actually held together - pair that with MaskTemporal for the stabilization + warning report.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git
pip install opencv-python>=4.7.0 scipy>=1.10.0

or ComfyUI Manager → search "CustomNodePacks", restart, confirm [MEC] Loaded .... The only heavy dependency is sam2_video propagate mode, which needs a SAM 2.1 checkpoint in ComfyUI/models/sam2/. Everything else is OpenCV/torch and works on CPU. When a mode silently does less than you expected, read info_json - the node reports what actually ran.