Model Math

Model Math is the pack's take on model merging: it takes up to four checkpoints and evaluates an expression over their weights, producing a merged model. Where core ComfyUI's ModelMergeSimple gives you one blend weight for the whole model, this node lets the expression change per layer - and per tensor inside a layer. a*(1-w)+b*w merges models a and b by weight w, which is a bog-standard merge; the interesting stuff happens when the expression starts referencing layer indices.

How it works

The node iterates over the union of every weight (state_dict) key across all input models. For each layer it binds the weights to a, b, c, d (models missing that layer get zeros), the floats w, x, y, z, and a few special variables:

• L / layer - the layer index (0, 1, 2, ...)
• LC / layer_count - total layers
• K / key - the layer's state_dict key string

Your expression is evaluated per layer, the result is compared to the base model's weight, and only the differences become patches applied to a clone of a. That's how merges are supposed to work under the hood - ComfyUI stores model edits as patch diffs, and this node hands you the merge math directly. Want to blend only the attention layers? Filter on key containing "attn". Want a merge weight that ramps with depth? Use L.

The inputs that matter

• a - the main model (base), required.
• b, c, d - optional additional models.
• Model - the expression applied to weights, default a*(1-w)+b*w.
• length_mismatch - tile / error / pad, default error. (For models, the tooltip notes this usually ends up broadcasting.)
• w, x, y, z - floats.

Output is a single MODEL.

Installing it

Part of More Math (mcDandy/more_math). ComfyUI Manager - search "More Math" - or:

cd ComfyUI/custom_nodes
git clone https://github.com/mcDandy/more_math
cd more_math
pip install -r requirements.txt

Restart. Dependencies are antlr4-python3-runtime and torch - no downloads. Needs a current ComfyUI.

Where people get burned

This is the slow one. Evaluating an expression against every layer of a multi-gigabyte checkpoint takes real time - this is not a lightweight node, and you'll see the run grind on big models. It's CPU-side tensor work, so RAM and patience matter more than VRAM.

Second, merging across mismatched architectures produces garbage. The zero-filling for missing layers means a 1.5 model merged with an SDXL model will have huge invented zeros in places - technically no crash, practically nonsense. Merge models that share an architecture.

And remember what the KB has to say about merges generally: they're fast and need no training, but they trade away coherence, and over-merged models all start to look alike. a*(1-w)+b*w at sane w values is fine; chasing extreme mixes is how you end up with a model that can't do faces. Also, the classic form is deprecated in favor of an autogrow variant with the same "Model Math" name - both ship, and the classic described here is perfectly usable.