PM SCE (Mergekit)

SCE stands for Select, Calculate, Erase - a three-stage merge method that decides, per parameter matrix, how much each LoRA actually matters, instead of trusting fixed weights. It's the pack's take on the question "which LoRA owns this matrix?" When you merge a character LoRA and a style LoRA, some layers are dominated by one and some by the other; SCE tries to let each matrix be claimed by whichever LoRA is genuinely changing it.

You reach for it when your LoRAs contribute to different parts of the network and a flat weighted sum flattens that structure. It needs 2+ LoRAs plus a base model.

How it works

Three stages, per matrix:

1. Select - variance-based masking. Parameters with low variance across the different LoRAs' task vectors are candidates to be zeroed; select_topk sets the fraction of highest-variance elements you keep.
2. Calculate - matrix-level weights are computed for each LoRA's contribution, so each matrix gets its own weighting rather than a global one.
3. Erase - TIES-style sign consensus removes conflicting changes.

The result is a merge that weights each matrix adaptively, favoring high-variance, consistently-signed changes.

The inputs that matter

• select_topk - fraction of the highest-variance elements to retain (default 0.1, range 0–1). Lower = more selective and weaker; the tooltip's guidance is that 1.0 keeps everything (strongest). For a first run, the middle of the range is a safer bet than the default if the merge comes out faint.

Output is a MergeMethod config for PM LoRA Merger.

Installing

Ships in the LoRA Power-Merger pack. ComfyUI Manager (search "LoRA Power-Merger") or:

cd ComfyUI/custom_nodes
git clone https://github.com/larsupb/LoRA-Merger-ComfyUI
cd LoRA-Merger-ComfyUI
pip install -r requirements.txt

Restart ComfyUI. Dependencies: PyTorch, lxml, mergekit - SCE runs on the mergekit path.

Common issues

SCE's signature failure mode is a merge that's too weak, because aggressive selection zeroes most of every matrix. If the output barely moves the image, raise select_topk before you try anything else. It's also a method that quietly assumes your LoRAs genuinely differ per-matrix - merging two near-identical LoRAs with SCE just produces noise, and plain PM Linear or PM TIES would serve you better. Ranks must match at the merger; reconcile with PM LoRA Stack Decompose (rSVD) first. And pair it with the Parameter Sweep Sampler sweeping select_topk if you want to find the sweet spot visually.