ModelMergeAdd

ModelMergeAdd is the bluntest tool in the merge family: it just adds two models' weights together - model1 + model2 - with no ratio, no blending, no safety margin. If ModelMergeSimple is the careful blend and ModelMergeSubtract is the scalpel, this is the hammer. It has exactly one job and you should use it exactly as rarely as the hammer.

The math is the giveaway: result = model1 + model2, every weight tensor added with full strength. There's no ratio because there's nothing to tune - it's pure addition. That's useful in exactly one class of situation: when you want to stack contributions rather than average them. The canonical example is the "merge all my LoRAs into one model" workflow people run with adapter weights - load several models that each carry a small additive change and sum them so the contributions compound. If you're adding a fine-tune's delta onto a base, ModelMergeAdd is the operation you're describing.

The community's own post-mortem on this is worth quoting, from the thread about stacking multiple LoRAs via merge nodes: pure addition over-amplifies - "If Lora 1 is [0,1,0,3] and Lora 2 is [1,1,3,2], pure addition gives [1,2,3,5], which is over amplifying the last value. This is why a difference merge is best." That's the honest engineering verdict: addition creates tensor peaks, and the more models you stack, the worse it gets.

How it works

Same patch machinery as the rest of the family - clone model1, apply model2's weights as a patch - but both patch strengths are set to 1.0, so every weight becomes model1 + model2 with no attenuation anywhere. No ratio input exists precisely because there's no partial state. It's instant, it's simple, and it's the most dangerous of the three arithmetic merge nodes because there's no knob to limit the damage.

The inputs that matter

• model1 (MODEL)
• model2 (MODEL)

One MODEL out. That's the entire surface area.

Where people get burned

The universal failure is expecting this to be a blend. It isn't - it's an addition, and a full-strength one. Give it two full checkpoints and you'll get a model that's overweighted everywhere, often with blown-out colors and incoherent outputs, because the sum of two neural nets is not a better neural net. The same family warning applies harder here: the KB's checkpoint notes that most community checkpoints are merges and that incompatible targets dilute each other - with addition, "dilute" becomes "double."

The right mental model: use ModelMergeAdd when the thing you're adding is small and known to be a delta (a fine-tune's difference over a base, adapter contributions). For anything else, ModelMergeSimple or ModelMergeSubtract is the safer choice. And if you're stacking multiple adapters, do it with a proper difference merge or a dedicated LoRA stacker - the peaks will thank you.

How you get it

Core ComfyUI, model/merging category, ships with the program. No installation - but also no reason to reach for it until you're sure addition is what you want.