CoachBate H3 Subject Track + Crop

Start here, because this is the node the whole pack exists for

Every video model has the same embarrassing failure: a face that occupies 90 pixels of a 1080p frame comes back as a smear. Not because the model is bad, but because the latent has no budget to spend on that region. The answer has been solved in images for years - crop the region, render it big, paste it back - and ComfyUI Impact Pack's detectors and detailers made it an industry standard. Video is harder: the box has to move, and a box that jitters makes the whole thing worse than doing nothing.

CoachBate H3 Subject Track + Crop is the front half of that loop for MiniMax H3: detect the subject on every frame, follow one subject through the clip, crop it so it fills a 768 canvas, and emit both a constant-size crop batch for H3 and the transform the stitch needs to put the result back. It's a fork of ComfyUI-H3-FaceRefine (MIT, credited in source) with the face-only machinery stripped out, so "subject" can be a face, a hand, or anything a YOLO model finds.

How the tracking actually works

Pick a rule with select (largest, left_most, centre_most, closest_to_xy, …) and the subject is chosen once per shot - on the first frame that holds it - after which continuity follows that same subject, never re-ranked per frame. That matters: a subject who walks across frame while someone else crosses the other way is still tracked correctly.

Two things smooth the box, and they want different amounts:

• smooth_window (default 21 frames, ~0.9s at 24fps) smooths the crop centre. Too much and the box lags fast movement; too little and it shivers.
• size_smooth_window (default 51) smooths the crop size, and genuinely wants more. Size jitter makes the crop breathe, which changes the resample factor every frame, which reads as shimmer.

smooth_method is gaussian by default (best jitter rejection), savgol if you want a push-in to keep its shape at large windows. There's also cut_detection, off by default and correct that way for H3, since H3 generates single-shot clips - leave it off unless you're refining an edited video with real cuts, because with it on each shot picks its own subject again by select, which is exactly what you don't want mid-take.

The inputs that matter, and the crop geometry

Required: images, detector, confidence, crop_factor, canvas_width, canvas_height, canvas_mode, smooth_window, size_smooth_window, smooth_method, size_mode.

Three of those decide the result:

• crop_factor (default 2.5) - crop side as a multiple of the detected box. 2.5 puts the subject at roughly 40% of the crop. Bigger means more context so the seam lands in hair or background, at the cost of magnification. 2.0–3.0 is the useful band; 1.5 is available and more aggressive.
• canvas_mode / canvas_width / canvas_height - the resolution H3 generates the crops at. 768 is H3's native short edge and the default. manual uses your numbers exactly; auto_no_downscale sizes the canvas from the largest crop so nothing is ever downscaled (expensive on clips with close-ups); auto_capped_768 clamps that to 768. Cost scales with area. size_mode decides whether the crop keeps a constant subject fraction per frame (per_frame, right for push-ins) or one size for the whole clip (max_of_clip, static shots only).

Optional but worth knowing: select_index (which subject in the ranking to follow - feed it from CoachBate H3 Loop Subject), select_override (a string name that overrides select), and crop_size_from (height or longest_side, the latter so a subject lying sideways isn't cut off).

Outputs: crops (the IMAGE batch), transform (the geometry the inject/denoise/stitch nodes all need - don't try to reconstruct it), preview (the tracked box drawn on each frame; green means detected, red means interpolated), report, canvas_w, canvas_h, frame_count.

Install

ComfyUI Manager → search "CoachBate", or:

cd ComfyUI/custom_nodes
git clone https://github.com/CoachBate/ComfyUI-CoachBate.git

Restart ComfyUI. You also need ultralytics installed - the node imports it to run the detector

• and detector weights in models/ultralytics/segm/ or models/ultralytics/bbox/, registered by ComfyUI-Impact-Subpack (the pack looks there directly if it isn't installed). A bbox model tracks fine but pastes as an ellipse; use a -seg model if the seam matters.

Where people get burned

• The detector isn't listed - it's not in models/ultralytics/segm or bbox, or ultralytics isn't installed. The error names the folders it searched.
• The box jumps to the wrong person - your select rule picks the wrong subject at the lock-on frame. Set select_override or select_index explicitly rather than tuning windows.
• A hard cut lands on someone else - cut_detection = none doing its job: everyone is renumbered at a cut, so the tracker picks the nearest subject to the last position, whoever that is. Turn cut detection on if the clip is an edit rather than a take.