GITS Face Tracking + Removal (Advanced)

Here's the honest summary of GITS Face Tracking + Removal (Advanced): it's the only node in this pack you can put in front of a sampler. Everything else assumes you've given up on diffusion rendering your graphic.

It does two unrelated jobs at once, which is why the outputs list is eight items long:

images -> Advanced Tracking -> preprocessing / diffusion -> Final Artwork Overlay
             |                                        ^
             +-- exact RGB overlay + alpha mask ------+

On the way in, it takes your frames, finds the faces, paints an opaque plate over each one, and draws the artwork into the image. That's guide_images - what you send to ControlNet, an img2img pass, or KSampler. On the way out it hands you the artwork separately as aligned_overlay_images plus overlay_masks, so GITS Final Artwork Overlay can re-apply the real thing after the VAE decode.

That double-apply isn't redundancy, it's the design: KSampler will approximate a logo rather than reproduce one - smeared lines, an elliptical ring. So let diffusion work on a face that's already covered, then paste the real pixels over its guess.

The inputs you'll actually touch

Ten required widgets, and they fall into three groups.

Geometry - face_scale, y_offset, rotation_speed. How big the artwork sits relative to the face, where it sits vertically (-1..1), and how fast the ring spins - negative reverses it.

Smoothing - min_cutoff, beta. It's the pack's one Euro filter, exposed raw, and the fiddliest thing here. min_cutoff is the latency/jitter tradeoff: lower is smoother and laggier, higher hugs the face but inherits detector jitter. beta is how much the filter loosens up when the face is moving fast. Overlay trailing a quick head turn? Raise min_cutoff. Vibrating on a still face? Lower it.

Track loss - hold_frames, fade_frames, tracker_confidence. When detection drops, the node coasts on motion prediction for hold_frames, then fades over fade_frames. tracker_confidence is the bar for accepting a detection at all, and it's the one people get backwards: lowering it makes tracking more willing, which is what you want on hard faces, at the cost of the odd junk detection.

Then the optional block: track_prediction, fuse_detectors (keeps YuNet faces MediaPipe's landmarks don't overlap) and small_face_boost (a second YuNet pass on a 1.5× upscale) are the three robustness toggles shipped for exactly this. effect_preset defaults to custom here rather than classic - this node expects you to drive it. start_frame and rotation_phase let you join a video mid-stream and stay in phase, which matters when you chunk a long video.

The eight outputs, sorted by what they're for

• guide_images - into your sampler, ControlNet, or whatever re-renders the scene.
• aligned_overlay_images and overlay_masks - into GITS Final Artwork Overlay, after decode. Also handy as a standalone tracked matte.
• final_images - artwork composited over the LaMa-filled frame; removed_background_images is the same plate with no artwork.
• face_occlusion_masks - the face coverage region; feeds the Final Overlay's optional face_occlusion_masks input to drive its backplate.
• tracking_preview - the guide image again, for eyeballing the track without rewiring.
• tracking_json - a JSON string (version, fps, width, height, frames[]) of the whole track. Per-frame transforms in a text output is the kind of thing you appreciate later.

The occlusion controls - occlusion_scale, occlusion_feather_px, guide_occlusion_opacity - govern the plate itself. profile_boost and partial_face_boost do narrow, literal jobs (side shots, frame-edge crops) and barely touch frontal faces by design.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Niutonian/ComfyUI-GITS-Tracked-Identity.git
cd ComfyUI-GITS-Tracked-Identity
python -m pip install -r requirements.txt
python scripts/download_face_landmarker.py

Everything goes in ComfyUI/models/gits_tracking/: face_landmarker.task (~3.6 MB), face_detection_yunet_2023mar.onnx (~0.2 MB - both from the script above, or --skip-lama) and big-lama.pt (~196 MB, from 1038lab/Lama). Restart after.

Face removal is the dependency that bites: face_removal_mode on lama needs ComfyUI-RMBG installed and loaded, because the pack borrows RMBG's AILab_LamaRemover class and points it at your local big-lama.pt. No RMBG and you get a RuntimeError naming RMBG specifically. Set face_removal_mode to disabled and everything else still works - tracking, guide, all eight outputs.

Troubleshooting

Model not found: the three files, exact filenames, gits_tracking/, restart. If you already have Big-LaMa from RMBG, copy it rather than re-downloading.

MediaPipe import error: requirements.txt into the interpreter that launches ComfyUI. On Windows embedded, python_embeded\python.exe.

Track loss or ID swaps: aggressive_track preset, lower tracker_confidence, raise hold_frames, keep fuse_detectors and small_face_boost on. For swaps specifically, locked_face and a lower max_faces - the appearance-matching layer needs contrast to work with.

LaMa flicker across frames: raise temporal_blend, keep temporal_flow on, and only lower lama_every_n_frames if you can afford the per-frame inpaint.

Inverted logo: this node has no mask_input_mode, so if the artwork comes out backwards, put an InvertMask in front of static_logo_mask or ring_logo_mask.

Tensor size mismatch at the Final Overlay: keep the decode resolution equal to the tracking resolution - match batch, height and width.

The packaged workflows/gits_core_image_batch.json is the core-nodes-only chain (Advanced → sampler → Final Overlay), and gits_video_integration_example.json shows the VideoHelperSuite setup. For long videos, load chunks.