FaceTag Crop (YOLO) — ComfyUI custom node

A ComfyUI custom node that uses an Ultralytics YOLO model to find a specific bbox in an image (or a batch of video frames) and crops the image centered on that detection, with temporal smoothing, normalization, and graceful fallbacks.

It outputs the cropped image, the aligned mask, and a preview with the mask drawn on top.

Install

Clone into your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/adbrasi/facetaggg.git

Requires ultralytics, opencv-python, numpy, and torch (torch ships with ComfyUI). Install ultralytics/opencv if you don't have them:

pip install ultralytics opencv-python

Put your YOLO models under ComfyUI/models/ultralytics (subfolders like bbox/ are searched recursively). The node lists them in a dropdown.

Node: FaceTag Crop (YOLO)

Category: image/facetag

Inputs

| Input | Description | | --- | --- | | image | A single image or a batch of frames (video). | | model_name | YOLO model from models/ultralytics (recursive). | | confidence | Detection confidence threshold. | | padding | Crop size as a multiple of the detected box. | | output_width / output_height | Output resolution. | | smoothing | Temporal smoothing (zero-phase EMA) across the batch. | | zoom_mode | fixed (one zoom for the whole batch) or smooth. | | zoom_percentile | Percentile used to pick the fixed zoom level. | | mask_mode | follow_detection (steady camera, mask tracks the box) or locked_center (static centered mask, the video moves/zooms to keep the subject on it). | | max_gap | Max missing-frame gap bridged by interpolation. | | mask_feather | Gaussian blur radius on the mask edges (px). | | detector_imgsz | Inference image size for the detector. |

Outputs

• image — crop centered on the detection, at the chosen output size.
• mask — the detection as a mask, aligned to the crop.
• preview — the crop with the mask drawn as a translucent green overlay.

Robustness

• Missing detections in some frames are interpolated; gaps larger than max_gap hold the nearest detection instead of sweeping across the frame.
• If no detection is found in the whole batch, it falls back to a full-frame crop and mask (with a console warning) instead of erroring.
• The YOLO model is cached between runs; uses CUDA + FP16 when available.