Clamped Add Opposite

Clamped Add Opposite is the version of add difference that refuses to overshoot. Where plain Add Difference computes A + alpha*(B − C) and lets the result wander wherever the math takes it, this node clamps every tensor so it stays between A and B, element by element. The name "opposite" comes from the older convention: B is the fine-tune, C is the anchor/base both A and B derive from, and the delta B − C is "opposite" in the sense that you're transplanting the change from one model onto another.

The README's own example workflow calls this "clipped add difference," and it's the poster child for the pack - a merge that keeps you inside the box both parents occupy, which is exactly the guardrail people want when they're afraid a merge will produce something degenerate.

How it works

result = clamp(A + alpha*(B − C), min(A, B), max(A, B)), per key, per element. The clamp is the whole point: if the delta pushes a value past the range either parent had, it's pulled back to the nearer parent's value instead of left to drift. With alpha = 1.0 (the default) you get the full clipped add. Lower alpha softens the whole thing.

Inputs from the schema:

• a (delta|param|weight) - the starting point, usually your base model in weight space.
• b (delta|param|weight) - the model you're stealing the change from.
• c (delta|param|weight) - the anchor both A and B are relative to; often the shared base.
• alpha (1.0) - strength, itself a MECHA_RECIPE, so you can feed it a Float for a flat value or Blocks Mecha Hyper for per-block weighting.
• merge_checkpointing - leave off unless you're rerunning a heavy graph and want the branch cached on CPU in fp16.

Output is one recipe (MECHA_RECIPE) → Mecha Merger → MODEL/CLIP/VAE.

When to reach for it

The classic setup: you have a base model, a fine-tune built on it, and a third model you want the fine-tune's behavior transplanted onto. C is the shared base, B is the fine-tune, A is the target. Clamping keeps the result expressible as "between A and B everywhere," which sounds restrictive and is - that restriction is what prevents the usual add-difference failure mode of drifting into noise outside either model's range.

If you don't want the clamp, use plain Add Difference Mecha Recipe. If you want the clamp but with a custom bound set rather than "between A and B," that's what Clamp Mecha Recipe plus a Recipe List is for.

Install & gotchas

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; only dependency is sd-mecha==1.1.7.

Where people burn themselves: forgetting that A and B both need to be weight-space recipes (the clamp requires A and B in the same space - the pack raises a type error if they aren't), and assuming the clamp means alpha can't overshoot. It can - the clamp bounds the result against A and B, so alpha = 2 still clamps, but you're forcing it against the wall everywhere and the output looks like a hard-cropped version of the blend. Keep alpha in the 0.5–1.0 range unless you know why you don't.