Extensions/comfyui-gits-tracked-identity
ComfyUI Extension

comfyui-gits-tracked-identity

Temporally tracked identity graphic overlays for ComfyUI (single image, video batch, webcam)

By Niutonian·Created 2 months ago·Updated about a month ago· 0
Niutonian/ComfyUI-GITS-Tracked-Identity
Nodes4
On cloudLocal install
CategoryGITS/Compositing, GITS
Stars0
Updatedabout a month ago

Nodes (4)

GITS Final Artwork Overlay
GITS/Compositing
GITS Face Replacement (Simple)
GITS
GITS Face Tracking + Removal (Advanced)
GITS/Tracking
GITS Webcam Face Replacement
GITS/Live
Readme

ComfyUI GITS Tracked Identity

A ComfyUI custom-node package that places a tracked, camera-facing identity graphic over one or more faces — for single images, ordered video batches, and live webcam. The name describes the visual technique; this repository contains no copyrighted franchise artwork. Bundled cyan placeholders are original geometric art.

Version 0.2.0 — multi-face IDs, Temporal LaMa, webcam, glitch/signal effects, appearance matching, motion prediction, and improved temporal fill.

Everything runs locally. Dependencies and models use MIT/Apache-style licenses only (see Licenses).

Why the overlay is applied twice

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

GITS Face Tracking + Removal (Advanced) puts an opaque proxy plate plus the graphic into the images sent toward diffusion. It also returns the exact aligned RGB graphic and alpha separately. GITS Final Artwork Overlay applies those exact pixels after VAE decoding so letters, circles, thin lines, and ring rotation do not deform. Do not expect KSampler to reproduce logo art.

For a fast path with no diffusion, use Simple or Webcam: they track, optionally remove the face with LaMa, and composite the exact graphic in one node.

Artwork is optional on every node. If static_logo, static_logo_mask, ring_logo, and ring_logo_mask are all unconnected, tracking and LaMa face removal still run. The outputs are then the reconstructed background (face removed) with no graphic overlay.

Installation

Reference environment: Python 3.10 or 3.11.

Clone into ComfyUI

cd ComfyUI\custom_nodes
git clone https://github.com/Niutonian/ComfyUI-GITS-Tracked-Identity.git
cd ComfyUI-GITS-Tracked-Identity

Or copy this folder to ComfyUI/custom_nodes/ComfyUI-GITS-Tracked-Identity.

Dependencies

Use the same interpreter that starts ComfyUI:

# Embedded ComfyUI Easy Install (Windows example)
..\..\python_embeded\python.exe -m pip install -r requirements.txt

# Or a normal venv
python -m pip install -r requirements.txt

OpenCV and PyTorch are normally already provided by ComfyUI. Face removal also needs ComfyUI-RMBG (or an equivalent Big-LaMa provider) loaded in ComfyUI.

Models (explicit download only)

Nodes never download weights at queue time. Install models once, then restart ComfyUI. Search the node menu for GITS.

Where to place them

Create this folder if it does not exist:

ComfyUI/models/gits_tracking/

Expected files (exact names):

| File | Purpose | Approx. size | |------|---------|-------------:| | face_landmarker.task | MediaPipe face landmarks (primary tracker) | ~3.6 MB | | face_detection_yunet_2023mar.onnx | OpenCV YuNet fallback (small / hard faces) | ~0.2 MB | | big-lama.pt | Big-LaMa face removal / inpaint | ~196 MB |

Full example on a typical Windows Easy-Install layout:

ComfyUI/
  models/
    gits_tracking/
      face_landmarker.task
      face_detection_yunet_2023mar.onnx
      big-lama.pt
  custom_nodes/
    ComfyUI-GITS-Tracked-Identity/

If you already have Big-LaMa from ComfyUI-RMBG at ComfyUI/models/RMBG/Lama/big-lama.pt, you can copy that file into gits_tracking/ instead of downloading it again. GITS prefers models/gits_tracking/big-lama.pt and can fall back to the RMBG path when configured through the LaMa remover node.

Manual download URLs

Download each file in a browser (or with curl / wget) and save it into ComfyUI/models/gits_tracking/ with the exact filename above:

  1. Face Landmarker → save as face_landmarker.task
    https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/latest/face_landmarker.task

  2. YuNet face detector → save as face_detection_yunet_2023mar.onnx
    https://github.com/opencv/opencv_zoo/raw/main/models/face_detection_yunet/face_detection_yunet_2023mar.onnx

  3. Big-LaMa → save as big-lama.pt
    https://huggingface.co/1038lab/Lama/resolve/main/big-lama.pt
    (model card / files page: https://huggingface.co/1038lab/Lama)

PowerShell example:

$dest = "C:\path\to\ComfyUI\models\gits_tracking"
New-Item -ItemType Directory -Force -Path $dest | Out-Null
Invoke-WebRequest -Uri "https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/latest/face_landmarker.task" -OutFile "$dest\face_landmarker.task"
Invoke-WebRequest -Uri "https://github.com/opencv/opencv_zoo/raw/main/models/face_detection_yunet/face_detection_yunet_2023mar.onnx" -OutFile "$dest\face_detection_yunet_2023mar.onnx"
Invoke-WebRequest -Uri "https://huggingface.co/1038lab/Lama/resolve/main/big-lama.pt" -OutFile "$dest\big-lama.pt"

Automatic installer script

From the custom node folder (after it is inside custom_nodes, the script writes into ComfyUI/models/gits_tracking/):

python scripts\download_face_landmarker.py

Use --skip-lama to install only the two face-detection models.

Optional: sync from a development checkout

powershell -ExecutionPolicy Bypass -File scripts\deploy_to_comfyui.ps1 `
  -Target "C:\path\to\ComfyUI\custom_nodes\ComfyUI-GITS-Tracked-Identity"

Or set GITS_COMFYUI_CUSTOM_NODES and run the script without -Target.

Nodes

| Node | Role | |------|------| | GITS Face Replacement (Simple) | One-node image/video path: track → LaMa → exact overlay | | GITS Webcam Face Replacement | Stateful live path; keeps tracker + LaMa cache across prompts | | GITS Face Tracking + Removal (Advanced) | Full outputs for diffusion / ControlNet / diagnostics | | GITS Final Artwork Overlay | Exact post-decode composite |

Effect presets

Available on Simple, Advanced, and Webcam (effect_preset):

| Preset | Intent | |--------|--------| | custom | Use widget values only | | classic | Default censor look, mild signal flicker | | subtle | Softer edges, slower ring, more smoothing | | aggressive_track | Longer hold, lower confidence, more prediction | | live_balanced | Webcam-friendly hold/fade and Temporal LaMa | | cinematic_glitch | Scanlines / RGB shift / stronger flicker |

Face modes

  • locked_face — keep the first selected identity
  • largest_face — always follow the largest face
  • all_faces — process up to max_faces identities
  • single_face — legacy Advanced one-tracker path

Tracking robustness (v0.2)

  • MediaPipe FaceLandmarker (VIDEO mode) + OpenCV YuNet fallback
  • Detector fusion: YuNet faces that do not overlap landmarks are kept
  • Small-face boost: extra YuNet pass on a 1.5× upscale
  • Appearance matching: local RGB histogram signatures reduce ID swaps
  • Bipartite association for multi-face frames
  • Motion prediction during hold/fade after brief occlusions
  • One Euro smoothing + configurable hold/fade

Temporal face removal

  • Cropped Big-LaMa at 256 / 384 / 512
  • lama_every_n_frames + optical flow (OpenCV DIS when available, else Farneback)
  • temporal_blend cross-fades fresh LaMa with the warped prior (less flash)
  • Webcam: fast (blur), balanced (interval LaMa), quality (every frame)
  • Color-safe composite: Big-LaMa rewrites the full frame (or crop). After inpaint, the node blends the result only under the face-remove mask and keeps original RGB outside that mask. Unmasked background colors match the input (no global cast from the LaMa provider’s full-frame rewrite).

Visual options

  • glitch_intensity — scanlines, channel shift, block noise on artwork
  • signal_flicker — digital opacity noise when tracking is weak
  • edge_aware_mask — grow removal mask along image edges (hair)
  • yaw_foreshorten — cheap horizontal squash from estimated head yaw
  • profile_boost — extra face-mask / LaMa expansion on side shots only (0 = mild auto, 1 = aggressive profile coverage; frontal almost unchanged)
  • partial_face_boost — extra expansion when the face is cropped by the frame edge or only partly visible (0 = mild auto, 1 = aggressive; full faces almost unchanged)

Workflows

Packaged under workflows/:

  • gits_simple_demo.json — Simple node, single image or batch
  • gits_core_image_batch.json — Advanced + Final Overlay (core nodes only)
  • gits_integrated_working_demo.json / gits_working_demo.json — LaMa path
  • gits_video_integration_example.json — VideoHelperSuite template
  • gits_webcam_template.json — live template

For long videos, load chunks with VideoHelperSuite to limit RAM.

Performance tips

| Use case | Suggestion | |----------|------------| | Single image | Simple + classic, LaMa on, Temporal LaMa irrelevant for B=1 | | Video batch | Simple/Advanced, temporal_lama=true, interval 2–3, blend ~0.35 | | Webcam live | Webcam node, live_balanced, balanced mode, resolution 256–384 | | Many faces | Lower max_faces, raise interval, or use fast mode live |

MediaPipe is CPU-side. Unchanged logos are converted once per call.

Troubleshooting

  • Model not found: download the three files into ComfyUI/models/gits_tracking/ (see Models), or run scripts/download_face_landmarker.py, then restart ComfyUI
  • MediaPipe import error: install requirements.txt into ComfyUI’s Python
  • Logo inverted: Simple/Webcam default comfy_load_image fixes core Load Image masks; Advanced may need InvertMask
  • Track loss: try aggressive_track, lower tracker_confidence, raise hold_frames, keep small_face_boost / fuse_detectors on
  • ID swaps: locked_face + appearance matching; avoid extreme motion blur
  • LaMa flicker: raise temporal_blend, keep temporal_flow on, lower interval only if quality allows
  • Whole image looks recolored / different “color profile”: fixed in current builds — full-frame LaMa used to return the provider’s rewritten tensor for every pixel. Output now restores original RGB outside the face mask. If you still see a cast only inside the face hole, that is LaMa’s inpaint tone (not a global profile change); try a smaller removal_area / lower boosts. With face_removal=disabled, final must match the input.
  • Tensor size mismatch (Final Overlay): match batch/H/W after decode

Tests

cd ComfyUI-GITS-Tracked-Identity
python -m pytest -q

Licenses

| Component | License | |-----------|---------| | This package | MIT (LICENSE) | | MediaPipe | Apache-2.0 | | OpenCV / YuNet | Apache-2.0 | | Big-LaMa weights (local) | Apache-2.0 |

No telemetry, no cloud inference. MediaPipe may log harmless portable_clearcut_uploader messages; they are not network model downloads from this package.

Limitations

  • Artwork is a camera-facing billboard (roll + optional yaw squash), not a full 3D mesh sticker
  • LaMa is an image inpainter; temporal cache/flow/blend improve continuity but are not a dedicated video-inpaint network
  • Extreme profiles, heavy occlusion, and tiny faces remain difficult
  • all_faces cost scales roughly with selected face count × LaMa