ARC Power LoRA Loader (Model Only)

The ARC Power LoRA Loader (Model Only) is the node you drop between your diffusion model loader and the sampler when you want to stack LoRAs and you do not care about CLIP - which, on the 2026 architectures this pack is built around, is most of the time.

It's one box with eight LoRA slots. Pick a LoRA in each, set a strength, wire a model in, get a patched model out. That's the whole job, and doing it well is harder than it sounds.

Why "model only" is the whole point

The regular ComfyUI LoraLoader patches the diffusion model and the text encoder - it demands a CLIP input and hands you a patched CLIP back. That's the right tool when you're on SDXL or SD 1.5, where a character LoRA rides on CLIP-encoded prompt features.

This node deliberately drops all of that. It only takes a MODEL in and only returns a MODEL out, because it's aimed at ARC-style workflows that have no CLIP to patch: Nano Banana runs over an API with no text encoder you own, and Flux 2 Klein conditions through Qwen3 - a full language model that LoRAs don't touch anyway. Dragging a CLIP wire through a graph that can't use it is just spaghetti for nothing. The author's README is blunt about the philosophy: no CLIP input, output, or patching, period.

How it works

Under the hood this is a thin wrapper over ComfyUI's own core LoraLoaderModelOnly class. For each slot from 1 to 8, in order, it takes the current model and - if that slot has a real LoRA selected at non-zero strength - applies the patch via load_lora_model_only, then hands the result to the next slot. Slots that are [None], empty, or sitting at strength 0.0 get skipped silently. Sequential application, top to bottom, one pass.

The inputs that actually matter

Of the seventeen inputs on the node, you'll touch four:

• model - your loaded diffusion model. Everything else is LoRA slots.
• lora_1 - a dropdown of every .safetensors in ComfyUI/models/loras, with [None] on top to disable the slot. Slots 2–8 are identical.
• strength_1 - model strength for that LoRA, from −100 to 100 in 0.01 steps. Default 1.0, and the usual 0.5–0.8 is often the sweet spot.
• model again, on the way out - the single output, the fully-patched model you feed your sampler. Yes, there's only one output. That's the point.

A strength of exactly 0.0 disables a slot even if a LoRA is sitting in it, and negatives are allowed, so you can weaken or invert a contribution without removing it.

Installing it

The README's method is copy-paste simple - drop the folder into your custom_nodes and restart:

cd ComfyUI/custom_nodes
git clone https://github.com/DHan315/ComfyUI-ARC-Power-LoRA

Then restart ComfyUI; the node appears under ARC > loaders. The repo ships a Comfy Registry publish workflow and a zero-dependency pyproject.toml, so it should also show up in ComfyUI Manager (search "ARC Power LoRA") once the registry listing goes live - until then, clone it. There are no extra Python dependencies and no model files to fetch; the only thing you need is LoRAs sitting in your loras folder for the dropdowns to have anything in them.

Workflow placement and the traps

The README suggests the obvious spot: Load Diffusion Model -> ARC Power LoRA Loader -> Configure Model Memory -> Configure Attention Backend -> Sampler - immediately after the model loader, before the ARC sampling path.

The failure modes are mostly the classic LoRA ones, not this node's fault. First, LoRAs are architecture-locked - an SDXL LoRA on a Klein model does nothing useful, and ComfyUI is famous for not yelling about it. Second, because the dropdown list is built from your loras folder at load time, a freshly-added LoRA won't appear until you refresh the node definitions or restart. And third, stacking order changes results: slot 1 hits the model first, so when output gets muddy, reorder rather than just re-strengthening.

One honest caveat: this pack is days old from an author with no community track record yet. The node is a clean, readable wrapper around core ComfyUI code, so it's low-risk - but if you're on an older SDXL flow that genuinely needs its text encoder patched, a model-only loader is the wrong shape for you. Grab it for the CLIP-less world, not the CLIP one.