Add Cosine B

Add Cosine B is the little sibling of Add Cosine A with one genuine difference: it blends a magnitude-aware similarity score in with the plain cosine similarity before deciding how much of model b to add to model a. Cosine similarity is scale-blind - it only cares about direction - which is great until you're merging two checkpoints trained at different scales, where a direction match can hide a big size mismatch. B folds both signals together: (cosine_similarity + magnitude_similarity) / 2, then the same lerp-based blending as its sibling.

The practical difference

Both nodes answer "how much should b contribute, per weight?" with "a lot, if a and b agree there." The difference is what counts as agreement. A is pure direction; B is direction plus magnitude, so it's a touch more conservative when the two models disagree in scale but happen to point the same way. In practice, for same-base merges you'll struggle to tell them apart. Where B earns its keep is when your parents come from different training runs with different effective scales - exactly the case where the scale-blindness of A can let a badly-scaled key through. If you're only going to wire up one of them, B is the safer default; A is the purist option.

Inputs & wiring

• a (weight) and b (weight) - model recipes in weight space.
• alpha - default 1.0. The similarity threshold that gates how much of b is added.

Output: one recipe (MECHA_RECIPE) in weight space. Chain it into a Weighted Sum, Add Difference, or straight to Mecha Merger. The node computes per-key, streams through lazy evaluation, and carries the usual merge_checkpointing toggle to cache the branch on CPU in fp16 when the recipe is stable - worth flipping on for big merges you iterate on.

The mechanism is straightforward sd-mecha code: normalize, compute the two similarity scores, average them, then k = 1 − clamp(sim − alpha, 0, 1) and lerp(a, b, k). Same family as the masked-add tricks in this pack (Train Difference Mask, Add Opposite), but cheaper - one similarity computation per key instead of a full mask construction.

When it's worth it

Reach for it when a plain weighted sum gives you a muddy blend and you want the merge to preserve each parent's character where they agree. It won't rescue incompatible models - near-zero similarity just means "don't add here," and a merge of two fundamentally different architectures stays broken. But for the common case of "two fine-tunes of the same base that feel slightly off in combination," this is a cheap, principled improvement over the naive formula.

Install

ComfyUI Manager → search mecha → "Mecha Merge Node Pack", or:

cd ComfyUI/custom_nodes
git clone https://github.com/ljleb/comfy-mecha.git
pip install -r comfy-mecha/requirements.txt

Restart after. Dependency is sd-mecha==1.1.7. Feed it weight-space recipes - deltas will confuse the similarity math and silently weaken the result.