ImageCropWithBBox(FaceParsing)

About as simple as nodes get: give it an image and a single bounding box, it hands back the cropped region. There's no smoothing, no padding, no logic - this is the plumbing node that connects a detector's output to a parsing model's input. If you've used FaceBBoxDetect or BBoxDetect to find a face and BBoxListItemSelect to pull one box out of the list, this is what actually turns that box into pixels you can feed FaceParse.

The inputs that matter

Just two, both required:

• bbox (BBOX) - a single bounding box, usually from BBoxListItemSelect (pulling one entry out of a BBOX_LIST) or from a detector node that already returns a single box, like FaceBBoxDetect wired through a list-select, or anywhere else a single BBOX gets produced in this pack.
• image (IMAGE) - the full image you're cropping from.

The output is a single IMAGE - the cropped region defined by that box.

Where it fits

This is the single-box counterpart to ImageCropWithBBoxList, which does the same job but for a whole list of boxes at once (useful when there are multiple faces and you want all the crops in one shot). Reach for this node instead when you're processing one face at a time - pull a box with BBoxListItemSelect, crop it here, run FaceParse on the crop, then put it back with ImageInsertWithBBox.

Installing it

Ships with the whole 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.

No models or extra dependencies - it's pure image slicing.

Common issues

There's very little to go wrong with this node in isolation. If you get an unexpected crop size or an error, the box itself is almost always the culprit, not this node - check what upstream detector or padding step produced the BBOX in the first place. If a detected face box ran off the edge of the image (which can happen after generous dilation padding on a face near the frame border), the crop will come back clipped or oddly sized; that's the padding step to revisit, not this one.

One thing that surprises people coming from other cropping nodes: because this takes a single bbox object rather than four separate coordinate widgets, you can't just type numbers into the node to nudge a crop by hand - you have to adjust whatever's feeding the box (usually the dilation setting on whichever detector produced it) and re-run. That's a deliberate trade-off in how this pack structures data: boxes stay opaque and pass cleanly between detection, cropping, resizing, and insertion nodes instead of every step reinventing its own four number inputs, at the cost of losing quick manual tweaking at this specific step.