CLIPSeg Model Loader

This node's whole job is upstream plumbing: it loads the CLIPSeg model that WAS's masking nodes - CLIPSeg Masking and CLIPSeg Masking Batch - actually run. On its own it doesn't do anything visible; it exists so those two masking nodes don't each have to load the model independently, and so you can load it once and feed it into more than one masking node in the same graph.

What CLIPSeg actually does, briefly

CLIPSeg is a real, published open segmentation model: give it an image and a plain-language phrase - "the red car," "the person's hair," "the background" - and it returns a rough mask of where that phrase applies in the image. That's the feature WAS is exposing here: text-prompted masking, no manual painting or bounding box required, using natural language instead of clicking. This loader is the first step; the actual masking happens in the CLIPSeg Masking nodes downstream, which take this loaded model plus your image and your text prompt.

Installing it

Bundled with the full WAS Node Suite pack:

• 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 from inside that folder, then restart ComfyUI.

Like WAS's other model-loading nodes, expect this one to fetch the CLIPSeg weights automatically the first time you run it rather than requiring a manual download - the same pattern the pack uses for BLIP, which the README documents explicitly as downloading on first use.

Common issues

The first run is the one to be patient with: if it's fetching the model over the network, that's a one-time delay, and if you're offline or behind a restrictive firewall at that moment, expect it to fail there rather than partway through your generation. Re-running once you're back online usually resolves it.

On the masking quality itself - text-prompted segmentation like CLIPSeg is genuinely handy for "roughly where is X" but isn't pixel-precise the way a dedicated detection model can be. If the resulting mask is close but rough around the edges, that's normal for this approach; WAS's own mask-cleanup nodes (Mask Erode Region, Mask Dilate Region, Mask Smooth Region) are the natural next step to tighten it up rather than expecting the CLIPSeg pass alone to be surgical.

And the pack-wide caveat: WAS Node Suite has been unmaintained since December 2023, and "Import Failed" errors after a ComfyUI update are a real, recurring issue tied to the pack's older pinned dependencies colliding with newer ones. Reinstalling requirements.txt inside your ComfyUI venv is the usual fix if the whole suite stops loading.