BBoxDetectorLoader(FaceParsing)

Before this pack can parse a face, it has to find one - and finding faces is a completely different job from segmenting them. That's what this node is for: it loads a YOLO object detector that draws a bounding box around each face in your image, which downstream nodes (FaceBBoxDetect, BBoxDetect) then use to crop tight before the actual segmentation model runs. Cropping first matters a lot - face parsing works far better on a zoomed-in face than on a tiny face buried in a full-body shot, for the same reason ADetailer-style face fixes always crop first: the detail budget is limited, and you want it spent on the face, not the background.

How it works

This is a thin loader around Ultralytics YOLO. The one option it exposes, model_name, currently has a single choice: bbox/face_yolov8m.pt. That's the medium YOLOv8 face-detection checkpoint from Bing-su's adetailer project - yes, the same weights that power ADetailer's face-fixing in Automatic1111/Forge, repurposed here as the detection front-end for face parsing. It's a well-worn, reliable detector; this pack doesn't reinvent it, it just borrows it, which is a sensible move given how much mileage that model has gotten across the ecosystem.

The input that matters

Just model_name - right now there's only the one choice (bbox/face_yolov8m.pt), so there's nothing to actually decide. The node exists as a slot for future detector options more than a real configuration point today.

The output is a single BBOX_DETECTOR object, which you feed into FaceBBoxDetect or BBoxDetect to actually run detection against an image.

Installing it

Part of the comfyui_face_parsing pack:

• ComfyUI Manager - search "comfyui_face_parsing", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Ryuukeisyou/comfyui_face_parsing.git, restart ComfyUI.

The face_yolov8m.pt weights download automatically on first use into ComfyUI/models/ultralytics/bbox/. If your connection to Hugging Face is unreliable, the README's manual fallback is to grab face_yolov8m.pt from the Bingsu/adetailer repo yourself and drop it in that folder.

Common issues

If you already run ADetailer or Impact Pack's UltralyticsDetectorProvider, you may already have face_yolov8m.pt sitting on disk somewhere - but this pack looks specifically in models/ultralytics/bbox/, so a copy living under a different custom node's folder won't be picked up automatically; you'll need it in the right place (or symlinked there) for this loader to see it.

Otherwise this node is about as low-drama as it gets - it's a detector load, not a heavy inference step, so the main thing to watch for is the same first-run download hiccup that affects every model this pack pulls in: if it can't reach Hugging Face, fetch the file by hand and place it yourself.