Lodestone LoRA Multi-Merge

The moment you're juggling more than three LoRAs, the sensible move stops being "stack them all in the graph" and becomes "bake the combination into a single file once." Lodestone LoRA Multi-Merge is the one node in the pack's Lodestone family that lets you go past three: it merges anywhere from 2 to 8 equal-strength LoRAs, expanding each into its full per-layer weight change and combining them with Frobenius-norm-aware math before writing one merged full-difference .safetensors into your loras folder.

Mechanically it's the same honest approach as the fixed 2- and 3-input variants - no averaging of compressed low-rank factors. Each LoRA's real delta per layer is materialized (alpha normalized first), merged by the method you pick, and saved as canonical .diff tensors. The multi version just throws in a lora_count selector and seven slots, then applies the chosen method with every included LoRA treated as one equal-prior contribution.

The inputs that matter

• lora_count (default 2, range 2–8) - the one control that trips people up. It must equal the number of LoRA slots you actually filled. Set lora_count to 5 and leave lora_6 unselected and the node refuses to run with an "unselected input" error rather than quietly merging the wrong set.
• lora_1 through lora_8 - lora_1 is mandatory and anchors the output layer set and metadata; lora_2 is always included too, and lora_3–lora_8 default to None.
• calc_mode (default sum) - with four or more LoRAs, blind sum is usually wrong: stacked deltas drift far from any single LoRA's intended strength. mean fixes the count scaling; slotnorm equalizes each input's Frobenius magnitude before averaging so a strong LoRA can't dominate; normmatch keeps the summed direction but rescales to a typical magnitude; slotnorm-normmatch does both. For big merges, treat the default as a starting point and expect to reach for the norm-aware modes.
• mismatch_mode (default skip) - with many independently trained LoRAs, missing layers are common. skip keeps LoRA 1's version of the layer, zeros treats absent inputs as zero updates, error aborts.
• output_filename - defaults to lodestone_merged_multi_lora; the .safetensors lands in models/loras.

The only output is output_filename. This node exists to produce a file - run the workflow once, then the result sits in your loras folder like any other .safetensors, ready for the pack's own LoRA tooling or for LoRAMergeToModel if you'd rather bake it into a checkpoint.

Where the batch merge earns its keep

Memory is the pleasant surprise. It streams one logical layer at a time and writes each completed delta incrementally, so merging six LoRAs costs roughly the memory of one layer times contributors, not six full LoRAs in RAM. Each CUDA out-of-memory layer is retried on CPU, and the destination file is replaced only after a fully successful write - no half-written outputs when you walk away mid-merge.

Two real caveats. First, the "equal-strength" framing is baked in: there are no per-input strength sliders, so if your inputs are wildly different in loudness, that's exactly what slotnorm (or slotnorm-normmatch) is for. Second, this stays in canonical plain-LoRA territory - DoRA-scale vectors and quantized/low-bit factors raise rather than merge, and every input needs matching layer keys where the merge matters.

Installing it

ComfyUI Manager: search Model Utility Toolkit. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-ModelUtils

Restart ComfyUI. No API keys, no model downloads - it reads LoRAs already sitting in your models/loras folder. Look for it under ModelUtils → LoRA → Merge → Lodestone, or search "Lodestone" and pick the Multi entry (the 2 and 3 variants are the same machinery with fixed slots). These mergers are recent additions to a pack that's still gathering community mileage, so sanity-check any surprising merge against a simple stack before you treat the baked file as gospel.