CLIPSEG2

Full disclosure up front: WAS's own README doesn't document a node called CLIPSEG2 anywhere - it only lists "CLIPSeg Masking" and "CLIPSeg Masking Batch." CLIPSEG2 exists in the pack, and its name gives away most of what you need to know: it's a second implementation of the same idea, most likely a rewrite added after the original shipped, following the same pattern WAS used elsewhere in the suite (the README documents ASCII being deprecated in favor of STRING output while both stuck around for a while). Treat this as a variant of CLIPSeg Masking rather than a different tool with a different job.

The job itself, shared by both, is worth explaining properly because it's genuinely useful: CLIPSeg is a zero-shot segmentation model. Instead of training or fine-tuning a detector for a fixed list of classes, it jointly embeds your image and a short text description, then predicts which pixels the description actually refers to. Type "the red car" or "her hair" and it hands you back a relevance map for that region - no dataset, no class list, no manual painting required. That's a meaningfully different tool from something like MiDaS-based depth masking (which cuts by distance) or a hand-drawn mask (which needs you to actually draw it): CLIPSeg lets language do the pointing.

Expect the same basic contract as CLIPSeg Masking: an image and a text prompt in, likely a threshold control to decide how much of the model's soft prediction counts as "inside" the mask (standard for any CLIPSeg-based tool, since the raw output is a continuous relevance map, not a clean binary line), and a MASK out.

Because this one isn't documented, the more conservative move if you're not sure which CLIPSeg node to reach for is to start with the plain CLIPSeg Masking node - it's the one the README actually describes, so its behavior is at least on record. Come back to CLIPSEG2 if you specifically need whatever it does differently, or if you're already using it and it's working - this article isn't a reason to rip it out, just an honest flag that its exact differences from the original aren't written down anywhere public.

Installing it: ComfyUI Manager, search "WAS Node Suite," install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/WASasquatch/was-node-suite-comfyui, then pip install -r requirements.txt (portable: python_embeded\python.exe -s -m pip install -r requirements.txt; manual/venv: activate the venv first), restart ComfyUI. WAS's BLIP nodes are documented to fetch their model automatically on first use rather than at install time - expect any CLIPSeg-based node in this suite, this one included, to follow the same pattern and download its checkpoint from Hugging Face the first time you actually run it.

Where it bites: since this node has no public documentation of its own, the honest troubleshooting advice is limited - if it behaves unexpectedly, check whether CLIPSeg Masking (the documented one) does the same thing before concluding CLIPSEG2 specifically is at fault. More generally, WAS Node Suite has been unmaintained since December 2023 (its README says "Retired"), and the community's recurring complaint is the whole suite failing to import after a ComfyUI or dependency update - check your console for that before troubleshooting any single node, including this one.