Color Clip ADE20k

If you've ever stared at an ADE20k semantic segmentation map and wondered which RGB triple belongs to "wall," this node is for you. Color Clip ADE20k is the class-aware sibling of the pack's Color Clip: instead of typing a color, you pick a class name from the 151-class ADE20k list - person, wall, sky, tree, car - and it clips that class out of the image, exactly like the plain version clips a color.

Same family, same workflow. The difference is entirely in how you choose the target.

What it does

ADE20k is the scene-understanding dataset behind a lot of segmentation models, and it ships with a fixed palette - each semantic class has a canonical color. The pack's version (via the gray2color library) builds a dictionary mapping every ADE20k class name to its palette color. When you select class_name, the node looks up the color and runs the same clip logic as Color Clip against it.

Two practical details separate it from the basic node:

• A small tolerance. The palette colors are close together - the author notes the max Euclidean distance between any two dictionary colors is ~7, with typical differences like (4,5,3) - so the node clips with a per-channel leeway of 2. That matters, because exact matching on a JPG'd segmentation map would leave speckles.
• It's only useful for ADE20k-style images. The color it clips is the ADE20k palette color for that class. If your image isn't an ADE20k semantic map, this is the wrong tool - use plain Color Clip instead.

Inputs: image, target and complement (the same TO_BLACK / TO_WHITE / NOTHING enums), and class_name - a dropdown of the 151 classes with "person" as the default. Output is an IMAGE (a binary-ish matte if you set target to black and complement to white).

The catch: it needs gray2color

This is one of the pack's 📓 nodes - it depends on the external gray2color library. In practice ComfyUI Manager installs the pack's requirements.txt (which pins gray2color) automatically. If you cloned manually and skipped the requirements, the whole pack prints a "not loaded" note at startup and this node won't appear in your node list. The fix:

cd ComfyUI/custom_nodes/comfyui_bmad_nodes
pip install -r requirements.txt

then restart ComfyUI. As long as gray2color is importable you'll see the node.

Install

Full install, for the record:

cd ComfyUI/custom_nodes
git clone https://github.com/bmad4ever/comfyui_bmad_nodes
cd comfyui_bmad_nodes
pip install -r requirements.txt

or grab "comfyui_bmad_nodes" via ComfyUI Manager.

Where this pays off: pull a segmentation map out of a SAM or segmentation model, pick the class you care about, and you've got a clean subject mask without touching a color picker - one dropdown instead of RGB spelunking.