LoRA Simple Add (Yogurt Nodes)

Sometimes you want two LoRAs to be one LoRA - same rank, same file size, no rank explosion - and you're willing to accept an approximate blend to get there. YogurtLoraSimpleAdd is that: a weighted tensor sum of two LoRA states. Feed it A and B, set how much of each you want, and it hands back a combined LoRA at the same rank as the inputs.

It's part of ComfyUI-YogurtNodes, yogurt7771's 150+ node all-in-one pack - obscure, MIT-licensed, auto-generated README - and it's the quick-and-dirty counterpart to the pack's exact-but-rank-growing YogurtLoraMerge. Knowing which one you're grabbing is the whole game.

How it works

Four inputs, all required:

• loraA - the base LoRA (from YogurtLoadLoraOnly).
• loraB - the secondary LoRA.
• alpha_a - weight for A, default 1, range -10 to 10.
• alpha_b - weight for B, default 1, range -10 to 10.

The output combined_lora is computed as alpha_a * A + alpha_b * B at the tensor level - every matching key blended, ranks staying put. Negative weights are allowed, so alpha_b = -1 subtracts B's contribution, which is a real technique for "this LoRA minus that one."

One output, combined_lora, ready to feed YogurtSaveLora (or further processing) exactly like any other LORA object in the pack.

Why it's not the same as a merge

This is the distinction worth internalizing. YogurtLoraMerge concatenates rank dimensions, which is slower to grow but exact - the output reproduces the summed model-side effect precisely. YogurtLoraSimpleAdd blends in place at the tensor level: the file stays small and the operation is instant, but you're linearly mixing two deltas, which is an approximation of what the pair would do applied together. For quick iteration on "does A plus a dash of B feel right?", Simple Add is the right tool. For something you're going to share or ship where correctness matters, use the merge.

Where you'll use it

• Prototyping a blend: A at 1.0, B at 0.3, save it, try it, tweak.
• Building a permanent hybrid of two styles that stays small on disk.
• Subtracting an unwanted trait with a negative weight.

Install

ComfyUI Manager, search ComfyUI-YogurtNodes, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes
cd ComfyUI-YogurtNodes
pip install -r requirements.txt

Restart, look under "Yogurt Nodes". No model downloads; inputs are LORA objects from YogurtLoadLoraOnly. Save the result to a new filename and the originals are safe to reuse.