FaceParsingModelLoader(FaceParsing)

This is the loader that gets you the actual brain of the pack. Everything downstream - FaceParse, FaceParsingResultsParser, the per-feature masks - depends on the model this node hands over, which is a segmentation network trained to label every pixel of a face: skin, nose, eyes, brows, lips, ears, hair, even hats and clothing collars. Think of it as the difference between "there's a face here" (which is what a bounding-box detector gives you) and "here's exactly where the left eyebrow ends and the skin begins." That per-pixel labeling is what lets the rest of this pack build precise masks for things like skin retouching or selective inpainting.

How it works

Under the hood this loads jonathandinu/face-parsing, a BiSeNet-style semantic segmentation model fine-tuned specifically for faces, from Hugging Face. On first run the node downloads its .json config and .safetensors weights straight from that repo into ComfyUI/models/face_parsing/. After that it's a local load, same as any checkpoint. The node itself doesn't do any parsing - it just instantiates the model on the device you pick and passes it forward as a FACE_PARSING_MODEL object. The actual parsing happens later, in the FaceParse node, which needs both this model and a matching processor to do its work.

The input that matters

There's exactly one setting here: device, an enum choice between cpu and cuda. Default is cpu, which is the safe choice but the slow one - face parsing on CPU works fine for a single image but will crawl if you're batching. If you've got a CUDA GPU (which, running on ComfyICU, you do), switch this to cuda and don't think about it again.

The output is a single FACE_PARSING_MODEL, which you wire straight into FaceParse's model input.

Installing it

This node ships as part of the comfyui_face_parsing pack, so you install the whole thing at once:

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

The model download happens automatically the first time you run this node, no manual step needed - but if your network can't reach Hugging Face, the README's fallback is to grab the .json and .safetensors files from the jonathandinu/face-parsing repo yourself and drop them into models/face_parsing/.

Common issues

The failure mode people actually hit is a network one: if the first-run download from Hugging Face stalls or fails (blocked region, flaky connection, corporate proxy), the node will error out rather than silently retry. The fix is the manual download path from the README - get the files onto disk yourself and the node picks them up on the next launch, no different from any other model that ComfyUI auto-fetches.

Beyond that, this is a lightweight loader with very little to go wrong on its own - most of the real trouble in this pack shows up downstream, in FaceParse or in the auxiliary nodes that depend on a working OpenCV install. If this node loads fine but things break later, the model itself usually isn't the culprit.