SegToMask PipelineLoader

What it does

SegToMask PipelineLoader is the least interesting node in this pack, and that's exactly what a loader should be. It takes no inputs at all. Drop it in the graph, hit run, and out come MODEL and PROCESSOR - the two objects that every SegToMask node in the pack wants to eat. Put one loader down, wire both outputs to as many consumers as you like, and the expensive part happens exactly once instead of on every node.

How it works

You're used to loaders with dropdowns - "which checkpoint, which VAE." This one has none, because the model choice is baked into the node. It boots a fixed HuggingFace-style segmentation pipeline: a checkpoint plus its image processor, loaded together. The practical consequence is that the first run is slow. ComfyUI has to pull the model from the HuggingFace Hub before anything renders, so your first queue after installing is a "looks frozen, isn't" moment.

Now the awkward bit, because you should hear it before you invest an afternoon. The pack's README is empty, and the GitHub repo 404s as of this writing; comfy.icu records roughly two installs in three years. I can't confirm the exact checkpoint name, and I won't pretend to. The shape tells the story though: a MODEL + PROCESSOR output pair with no model picker is the fingerprint of a transformers pipeline, and the classes box on its sibling nodes points at a class-label segmentation model - most plausibly a SegFormer ADE20k-style checkpoint with a fixed vocabulary of everyday scene objects. Treat "which model, exactly" as unverifiable until the repo comes back.

Install

Install is the standard two-lane road. In ComfyUI Manager, search for segment_to_mask_comfyui (or "SegToMask") and install; or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/ginlov/segment_to_mask_comfyui

Then restart ComfyUI. One caveat that's easy to misread as your own screw-up: if the clone dies with an auth error asking for a username, that's the repo being gone, not your git setup.

Troubleshooting

Troubleshooting is short, because the node has almost nothing to go wrong except its dependency:

• First-queue hang - the hardcoded checkpoint is downloading from the HF Hub. Needs internet and HuggingFace access; a firewall or an offline box will stall here forever.
• Pipeline load failure / OSError - it can't reach whatever model it ships with. There's no input to point it at a different model, so you're stuck with the baked-in choice.
• Dependency conflicts - it's a ~3-year-old pack, and on a fresh ComfyUI the transformers version it wants can clash with something newer. That's the classic custom-node dependency hell; ComfyUI Manager will usually surface it.

If any of that blocks you, step back and skip the loader entirely: the maintained alternatives (CLIPSeg, or Grounding DINO + SAM via Impact Pack) will get you a mask with a model you can actually choose. This loader is the pack's weak link - the other two nodes are the reason anyone found it at all.