Face Fix - Prepare Full Face Video (Shot Aware)

Consider this the little sibling of Face Fix - Prepare Full Video With Shot Tracking. Face Fix - Prepare Full Face Video (Shot Aware) does the same core job - find the primary face, track it through the video, cut a shot-aware 512×512 face sequence for LTX, and keep the mapping needed to paste repaired faces back - but it skips the Z-Image anchor machinery entirely. No anchor_images, no anchor_count, no distance-based repair filtering. It's the version you use when your repair plan is pure LTX: feed it the face video, let LTX regenerate it, composite back.

It exists because the full prepare node is designed around a hybrid workflow where Z-Image sets the identity and LTX does motion. If you're not using Z-Image anchors - maybe you're on a clean LTX-only pipeline, or you want fewer moving parts while you debug - this node hands you the same tracking quality without the anchor bookkeeping.

How it works

Same detection engine as the full version: OpenCV DNN face detection with the model files bundled in the pack's assets/ folder, per-shot tracking, hard-cut detection that resets the track at scene changes, and crop smoothing so the box doesn't jitter. The difference is in scope, not mechanism.

The inputs

• video_frames - the complete input frame sequence (VHS Load Video's IMAGE output).
• detection_confidence - default 0.4, notably more permissive than the full node's 0.7, because there's no Z-Image anchor step to filter through. Lower = more detections, more false positives.
• crop_padding - extra space around the face crop, default 0.08.
• minimum_face_pixels - reject faces smaller than this many source pixels (default 12).
• rotation_assist - Off, Auto, or Force. Rotated detection for tilted faces; Auto only does the extra work when it thinks it's needed.
• short_gap_tracking - bridge up to this many consecutive missed detections (default 2, range 0–12). Higher tolerates longer blurs; 0 repairs only freshly detected frames.
• cut_sensitivity - hard-cut threshold. Default 0.15, so it's tuned to catch subtler cuts than the full node.
• crop_smoothing - stabilizes crop position/size within a shot (default 0.85).

The outputs

Only two, and that's the whole point:

• face_video_512 - the shot-aware 512×512 face sequence for LTX.
• face_fix_context - tracked crop boxes and original frames. Wire this into a Composite node at the end, or the repaired faces can't find their way home.

Installing it

Same pack install as everything else here: Manager → search vrgamedev, or:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git

Restart, hard-refresh. Detection models are bundled. The rest of the Face Fix workflow (LTX models, the pack's heavy pip requirements) still applies - this node just drops the Z-Image requirement.

Common issues

• More false detections than the full node. Expected - detection_confidence starts at 0.4. Raise it toward 0.6–0.7 if junk faces are getting tracked.
• Face pasted where there is none. Lower short_gap_tracking. At 12 it will happily bridge a long stretch.
• Jittery crops. Raise crop_smoothing toward 0.9, or check cut_sensitivity if it's resetting mid-scene.

Honestly, if you never use Z-Image anchors, reach for this one first - it's the same tracking smarts with less to misconfigure.