Train Difference Mask

This is a node from the training world that wandered into merging, and it's a good one to know even if you never train. Train Difference Mask computes a per-tensor mask that answers the question: where do B and C actually differ, without B also differing from A? The output is a param recipe - a set of 0-to-1 weighting values - that you can multiply into other merge operations to focus their effect on specific regions of the network.

The formula, in plain words

Give it three models: a, b, and c. It computes |b − a| / (|b − a| + |b − c|) per weight, scales by 1.8 and alpha, and clamps the noise. Read that ratio: the numerator is "how much B changed relative to A"; the denominator adds "how much B also changed relative to C." So the mask is high where B's change is unique to the A-to-B direction and low where B is just a generic drift. In the training community this exact mask is used to tell a fine-tune "emphasize the weights that actually matter for this task and leave the rest alone" - which, applied to merging, translates to "apply my delta mainly where it's doing something real."

Inputs:

• a, b, c - three recipe inputs (models from Model Mecha Recipe, or deltas). All take delta|param|weight.
• alpha - default 1.0. Scales the whole mask.

Output: one recipe (MECHA_RECIPE) in param space. That's the part to get right: this isn't a model, it's a weighting layer. You use it by feeding it into a param slot of another merge node - e.g. as the alpha input of a Scale or a masked operation - so the operation applies per-key weights instead of one flat scalar.

Why you'd reach for it

The honest pitch: it's a surgical tool for merges that keep polluting unrelated weights. Say you're adding a style delta and it's bleeding into the parts of the model you wanted untouched. Compute the train-difference mask between the involved models and route the delta through it - the add now happens hardest where the delta is genuinely informative, and near-zero elsewhere. The downside is that a per-key mask adds a full extra computation over the model, which is why the pack's lazy streaming and the merge_checkpointing toggle (cache this branch on CPU in fp16 once it's stable) matter here more than on most nodes.

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 after. Dependency is sd-mecha==1.1.7.

The usual trap: treating the param output as if it were a model and wiring it into Mecha Merger directly. It's a mask - it multiplies other things, it doesn't stand alone. And don't expect it to fix a fundamentally bad merge; it makes the placement of a change smarter, it doesn't make the change itself good.