Mask Operations

Mask Operations takes two MODEL_MASKs and performs set logic on them. Union, intersection, difference, xor - the same operations you'd use on image masks in Photoshop, but applied to tensors of booleans where every boolean says "merge this parameter" or "leave it alone." It's a combinator, and it's how you build compound protection schemes out of simple masks.

A mask in this pack is a per-parameter whitelist for a merge. Build a Magnitude Masker for one pair of models, another for a different concern, and you've got two opinions about what a merge should be allowed to touch. This node reconciles them:

• union - selected by either mask. The loosest combination: a parameter merges if either protection scheme says it's fine.
• intersect - selected by both. The strictest: a parameter only merges if both schemes agree it should. Use this when you want both of two conditions to hold.
• difference - mask_a minus mask_b: everything in A that isn't also in B. "Protect these, except where the other mask says don't bother."
• xor - the symmetric difference: in one mask or the other, but not both.

The pack's README is charmingly honest about how speculative all this is - it muses about "potentially binning the parameters of a model up with some fancy set thresholding, and then merging a different model in to each slice." Nobody has published a recipe for that, but it's the kind of thing these nodes make possible: split a model into four non-overlapping random slices with Quad Masker, then run a different merge into each slice using the same base model. Mask Operations is the glue.

The inputs

• mask_a, mask_b - two MODEL_MASKs. They needn't cover identical layer sets; the operations handle missing keys sensibly (difference and union keep one-sided keys, intersect requires both).
• operation - union (default), intersect, difference, xor.

The output

A single MODEL_MASK, ready to feed any merger's model_mask input, or onward into Mask Edit for further surgery.

Install and gotchas

ComfyUI Manager (search "DareMerge") or:

cd ComfyUI/custom_nodes
git clone https://github.com/54rt1n/ComfyUI-DareMerge

Restart ComfyUI. Dependencies: matplotlib, numpy, torch, pillow; no downloads.

Two practical notes. First, masks are memory-heavy - boolean tensors at full model scale - and each operation builds new tensors, so a chain of several mask ops can add up; the pack chunks the heavy math to cope, and you'll see peak VRAM printed to the console. Second, the mental model that keeps you out of trouble: the mask is a whitelist for change, so union is the permissive direction and intersect is the restrictive one. If a masked merge comes out looking like pure model A, your mask likely shrank the mergeable set to nearly nothing - that's the mask doing its job, not a broken node.