Semantic Segment — Face / Clothes Parsing

SAM is great, but SAM needs a prompt - a point, a box, a text phrase - and it'll hand you three candidate masks and ask which one you meant. SemanticSegmentMEC takes a totally different route: you name the thing. Type skin,hair into classes_csv, hit queue, and out comes a binary mask covering exactly those semantic classes. It's face and clothes parsing via SegFormer, and for portrait work it's often the least-effort path to a mask in the whole ComfyUI-CustomNodePacks pack.

This is the node for the "isolate the subject's skin for a selective grade" or "mask the shirt so I can re-texture it" jobs. It's one of the few genuinely model-based nodes among the ones covered here, so it has real dependencies - but they're the auto-downloading kind.

How it works

The node runs one of two SegFormer segmentation models, both pulled from HuggingFace on first use:

• segformer_face (jonathandinu/face-parsing) - 19 classes of facial anatomy: skin, l_brow, r_brow, l_eye, r_eye, eye_g (glasses), ears, nose, mouth, u_lip, l_lip, neck, necklace, cloth, hair, hat.
• segformer_clothes (mattmdjaga/segformer_b2_clothes) - 18 classes of apparel and body: upper_clothes, skirt, pants, dress, belt, left_shoe, right_shoe, left_leg, right_leg, left_arm, right_arm, bag, scarf, plus face and hair.

Every pixel gets a class, you pick which classes to include, and the node merges them into one binary mask. Each run processes the full batch, so you get per-frame masks on video, too.

Inputs and outputs that matter

• image (IMAGE) - what to parse.
• model_name - segformer_face or segformer_clothes. The class list you can use depends on which you pick, and the tooltip carries the full class lists.
• classes_csv (STRING, default "skin,hair") - the classes to include, comma-separated. The default is a sensible portrait start.
• threshold (FLOAT, default 0.5) - confidence cutoff for class assignment.
• invert (BOOLEAN) - flip the output mask (keep everything except your classes).
• keep_model_loaded (BOOLEAN, default true) - keep the model in VRAM between runs; turn off to free memory if you're juggling other models.

Outputs:

• mask (MASK) - the combined binary mask.
• info (STRING) - JSON with the model, classes used, and per-class pixel counts. The pixel counts are genuinely handy for sanity-checking whether a class matched at all.

Installing it

This node needs the model stack, so the pack's full install story applies:

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git
pip install transformers pillow   # needed for SegFormer

then restart ComfyUI (or use Manager → "CustomNodePacks"). The SegFormer weights (~400 MB) auto-download from HuggingFace on first use, so no manual model placement - just a working internet connection the first time you run it. For the pack's other requirements (opencv-python, scipy, safetensors), comment out anything ComfyUI already bundles before installing, to avoid clobbering its torch/numpy.

Where it beats the alternatives - and where it doesn't

It beats SAM when your target is a semantic category, because there's no prompting to get wrong. It beats background-removal nodes when you want to keep the background but drop the clothes. Where it falls down is fine anatomy - the 19 face classes aren't going to separate individual teeth - and non-human subjects, where neither model knows what it's looking at. Use it as the quick pass, then refine edges with the pack's matting or ViTMatte nodes if you need compositing-grade alpha.