RuYi multi-Lora-loader (model only)

The sibling node in RuYi-Nodes is easy to misread as the same thing minus a wire, but the difference is the point: RuYi multi-Lora loader (model only) applies LoRAs to the diffusion model and leaves the text encoder strictly alone. Same picker, same trigger-word logic, same MIT-licensed pack - just no CLIP input anywhere. If you mostly run Flux or Krea 2, this is the one you actually want.

Why model-only is a real thing, not a stripped-down version

A stock LoraLoader has two strengths because SDXL-era LoRAs patched both the UNet and the CLIP text encoder. That era is gone for a lot of current workflows. Flux LoRAs are trained to patch the diffusion model, and Krea 2 goes further - its text encoder is a separate Qwen3-VL stack that you load with CLIPLoader (type: krea2), entirely independent of the model. Load a Krea 2 LoRA through the full loader and you're setting strength_model and strength_clip to the same value and pretending the CLIP half matters. This node just drops the half that doesn't.

So you feed it MODEL in, you get MODEL back out, with trigger_words as the only other output. The README shows the intended shape for Krea2 / Flux-style pipelines:

MODEL / UNET ──> RuYi multi-Lora-loader (model only) ──> MODEL

Keep the text encoder / CLIP path on its own branch, because a model-only LoRA loader has nothing to attach it to.

How it works

Identical mechanism to the full loader, minus the CLIP wiring. A stack_config JSON list (built for you by the node's own UI) holds one entry per LoRA - name, enable toggle, strength, and whether it contributes trigger words. The node chains ComfyUI's core LoraLoaderModelOnly node for each active LoRA, applies them in list order, and skips zero-strength entries. The picker, folder/base-model filters, in-node previews, and the per-node "Show N LoRAs" viewport (default 3, 0 = unlimited) are all shared with its sibling, as is the LoRA-Manager .metadata.json sidecar reading.

The trigger_words output behaves exactly the same too: each enabled LoRA's Output trigger toggle decides whether its trained words join the deduplicated, comma-joined string. No sidecar, no trigger words - just an empty string, which is fine, and loading still works.

Install

It ships in the same pack, so one install covers both loaders and the text-preview node. ComfyUI Manager: search RuYi-Nodes. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/RuYi-Xiao/RuYi-Nodes.git

Restart ComfyUI. Zero additional Python dependencies - the pack only uses what a stock ComfyUI already has.

The honest verdict

This is a convenience wrapper over stock LoraLoaderModelOnly, and it doesn't pretend to be more. What you're paying for is the management layer: one node instead of a LoRA chain, a searchable picker with previews and base-model filters, and trigger-word auto-collection if you run ComfyUI-Lora-Manager for sidecars. It's a young pack with no community reputation to lean on yet, so the selling point is purely practical.

Watch out for: trying to wire a CLIP in - there's no port for it, and the fix is to keep your text encoder branch separate. And remember LoRAs are architecture-locked: a Flux LoRA won't do anything useful on Krea 2 or vice versa, no matter which loader you use. It's new, it's niche, but if you stack model-only LoRAs regularly it's a genuinely tidy way to do it.