[SK] LoRA (Model Only)

This is a LoRA loader with the CLIP half amputated on purpose. It takes a MODEL in, applies one LoRA to it at a given strength, and hands you back a MODEL - no clip input, no CLIP output, nothing to wire to your text encoder. That's the entire pitch: for the growing pile of workflows where CLIP isn't even in the same part of the graph (a UNETLoader-style split checkpoint, a video model where the text encoder is loaded separately, or a LoRA that was only ever trained against the diffusion weights), this node is the one that doesn't ask you to pretend it has something to do with text encoding.

It's one of eight loaders in SK-ComfyUI-FolderingLoader, a small, obscure pack by a GitHub author going by SKFRMSEHF. Every node in it is a rebuild of a stock ComfyUI loader - checkpoint, VAE, UNet, LoRA - with one twist: the file picker renders as a folder tree instead of ComfyUI's flat alphabetical dropdown. Worth knowing going in: the repo ships with an empty README and has essentially zero footprint anywhere online, so you're mostly working from the node itself rather than any documentation. That's not a red flag on its own - plenty of perfectly fine utility packs are one person's itch scratched in public - but don't expect a wiki to bail you out if something's unclear.

How it works

Same mechanism as any LoRA application: it reads the low-rank adapter weights from your chosen file and merges them into the diffusion model's layers, scaled by strength_model. The difference from ComfyUI's regular LoRA loader is just what it doesn't touch - the text encoder is left completely alone, so whatever CLIP conditioning you built elsewhere in the graph passes through unaffected. If the LoRA you're loading was trained to also nudge the text encoder, this node quietly skips that half; you'd want the full SK_LoraLoader (with both model and clip inputs) for that instead.

The inputs that matter

• model - the diffusion model to patch. Comes from a checkpoint or UNet loader upstream.
• lora - pick the file from the folder-tree combo. It only lists whatever's sitting in your loras/ directory (or wherever your model paths point).
• strength_model - how hard the LoRA hits, default 1.0, range -100 to 100. In practice you're almost always somewhere between 0.5 and 1.2; the wide range exists so you can go extreme (or negative, to subtract a style) but most LoRAs look best well inside it.

Output is a single MODEL - wire it straight into your sampler's model input, or chain another LoRA loader after it if you're stacking.

Installing it

There's no install section in the README to quote, so this is the standard path: open ComfyUI Manager, search for SK-ComfyUI-FolderingLoader (or the class name SK_LoraLoaderModelOnly), install, and restart. Manually it's cd ComfyUI/custom_nodes && git clone https://github.com/SKFRMSEHF/comfyui_SK_Loader then a restart. There are no bundled models and nothing in the node's schema suggests any exotic Python dependency - it's a loader, not a pipeline, so it's riding on whatever ComfyUI already ships with.

Common issues

• Empty dropdown. If the lora combo is blank, the pack hasn't found any files under your configured LoRA path. Drop files into ComfyUI/models/loras/ (or wherever extra_model_paths.yaml points) and restart - these tree pickers build their list at startup, not live.
• "Why isn't my LoRA doing anything to the prompt?" Because it can't - this node never sees CLIP. If the LoRA you're using is supposed to shift word associations or add a trigger-word-driven concept, that's the text-encoder half, and you need SK_LoraLoader or the stock LoraLoader, not this one.
• No community trail to lean on. This pack has essentially no discussion anywhere, so if something behaves oddly there's no thread to search - check the GitHub repo's issues directly, or fall back to ComfyUI's own LoraLoaderModelOnly, which is functionally identical and far better trodden.