🎨 Easy LoRA Merger

This is the flagship - the display name, "Easy LoRA Merger," is the pack's name, and it's the node the author built the whole thing around. It merges two or three LoRAs into one and applies the result to a model in a single node. Instead of the usual "stack three LoRA Loaders at different weights and hope" approach, you get a merged adapter that behaves like one LoRA.

The origin story explains why it exists. The author (Terpentinas) built it after failing to merge a Musubi-tuner LoRA with an AI-Toolkit one for Flux Klein 4B - every existing merger choked on the trainer mismatch. So the whole pack is built around one hard problem: merging LoRAs that don't share a trainer, a rank, or even an architecture.

How it works

At the base, it's the same weighted-combination math as any merger - method defaults to linear, and for LoRAs the weights are genuinely independent (they're tiny deltas, not absolute weights, so 1.0 + 1.0 is fine). The difference is everything layered on top to make mismatched LoRAs play nice:

• balancing_mode - the equalization brain. safe (default) hard-matches energy on shared layers only; creative is looser; intensity uses peak-energy detection for cross-concept merges; impact adds sparsity correction for sparse-vs-dense cases like Anima. This is why the weights don't behave like a simple 70/30 mix.
• magnitude_scaling - scales LoRAs B and C to match A's energy using RMS or a top-X% percentile, so one loud LoRA doesn't drown the others.
• energy_preservation - on by default; keep it that way. It stops the merge from pumping up overall magnitude.
• active_threshold - only merges the "active" non-zero regions of each tensor.

The honest caveat, straight from the author's beta post: cross-trainer merges aren't 1:1. You may need 2–4x weight on one side to get the balance you want. That's the price of the compatibility.

Inputs that matter

• model / clip - required; the merge is applied straight onto them. Wire the outputs into a sampler to see the result.
• lora_a / lora_b / lora_c - dropdowns, or lora_data_a/b/c chained from another merge.
• weight_a / weight_b / weight_c - start all at 1.0, then treat them as "loudness," not ratio. Tune in 0.1 steps.
• method - linear first. Try magnitude or feature_mix later; the DARE/TIES family is mostly for when you know what you're doing.
• save_trigger - off = preview through the live model/clip outputs; on = write the merged .safetensors to your loras folder.

Outputs: lora (the merged adapter - chain it into the Easy LoRA Baker to permanently bake it), model and clip (the applied result), output_path, and forensic_report.

Installing it

Manager (search "Easy LoRA Merger") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Terpentinas/EasyLoRAMerger

Restart ComfyUI. No model downloads.

Gotchas

Community feedback on the beta: one tester got a genuinely nice stable blend of two character LoRAs but found it fiddly compared to stacking LoRAs in rgthree's Power Lora Loader - "body horror material" before it clicked. That's real: this node is more powerful than a simple mixer and less predictable. If you just want a 70/30 of two same-trainer LoRAs, a plain weighted stack is simpler. If your LoRAs come from different trainers or you want them as one permanent file, this is the tool. And if you merge a 9B and a 4B Klein LoRA together, the results are... an experience, per the author. Don't say you weren't warned.