Fashion Cloth Mask 2

Fashion Cloth Mask 2 is the pack's second mask node, and it's a study in confusing defaults. Same job as its sibling - segment a photo, output a MASK - but it uses a different, coarser model (Segformer B2 Clothes, mattmdjaga/segformer_b2_clothes) with just 13 classes: Face, Hair, Hat, Upper_clothes, Skirt, Pants, Dress, Belt, shoe, leg, arm, Bag, Scarf. And here's the kicker: every toggle defaults to ON.

That changes the output completely. With all boxes checked, the mask is the person silhouette - white where the person is, black background. That's a perfectly good "person mask" for other jobs, but if you wire it straight into CatVTONNode, you'll tell the diffusion model to redraw the entire person, not just the garment. That's the trap.

How to actually use it

To build a try-on mask for a t-shirt: leave Upper_clothes ON and turn everything else OFF. What stays checked stays white (the replace region); everything you uncheck turns black, and the background is always black. So the two mask nodes approach the same end state from opposite directions:

• FashionClothMask - starts with an empty (black) mask; you check what to replace.
• FashionClothMask2 - starts with the whole person; you uncheck what to keep.

Which one you reach for comes down to the job. If you're working in broad garment groups - swap the top, swap the pants - the 13 toggles of FashionClothMask2 get you there with less fiddling. If you need fine detail classes (zipper, ruffle, pocket) or you're isolating one specific accessory, FashionClothMask's 46 labels win. And if you wanted a person silhouette for something unrelated to try-on - a cutout, a background job, an inpainting boundary - this node gives you one for free, which its sibling can't do.

Installing it

Same pack, same install - Manager → comfyui-try-on, or:

cd ComfyUI/custom_nodes
git clone https://github.com/shadowcz007/comfyui-try-on

restart, then the model goes to:

ComfyUI/models/segformer/segformer_b2_clothes

Notice the underscore this time - FashionClothMask's folder has a dash (segformer-b3-fashion). The two nodes hardcode their own literal paths, so getting the folder names crossed is the classic first-run error. Both nodes need transformers and torchvision (no requirements.txt ships with this pack, so Manager won't install them), and both reload their model from disk on every run - the first mask on a fresh boot is the slow one.

Honestly, this node is the less-used sibling - its 4 impressions on this site reflect that. But if your workflow is "swap the whole top" rather than "swap the pocket", its bigger-picture defaults are genuinely faster to drive than the 46-toggle version.