[SK] LoRA

This is the LoRA loader you reach for by default: it takes both your MODEL and CLIP, applies one LoRA to each with independently tunable strength, and hands both back modified. Most LoRAs are trained to touch the text encoder as well as the diffusion weights - a character or style LoRA that only nudges the UNet but leaves CLIP alone tends to underperform, because part of what it learned lives in how the model associates words with the concept. If you need the model-only variant (say, a video-model workflow where CLIP is loaded and wired separately), this pack ships that too as SK_LoraLoaderModelOnly; this is the full one.

It's part of SK-ComfyUI-FolderingLoader, a small pack by GitHub author SKFRMSEHF that rebuilds several of ComfyUI's stock loaders around one idea: file pickers that render as a folder tree instead of one long alphabetical dropdown. The pack ships no README and has no visible discussion anywhere - treat it as a solo convenience tool rather than something with a support community, but the outputs are plain ComfyUI MODEL/CLIP types, so it mixes freely with core nodes and anything from other packs.

How it works

A LoRA is a small file of low-rank weight adjustments layered on top of a checkpoint - not a full model, just a patch, which is why they run 10–200MB instead of gigabytes. Loading and applying one means reading those adjustment matrices and merging them into both the diffusion model's cross-attention layers and the text encoder's weights, each scaled by its own strength value. Chain multiple LoRA loaders in sequence and you're stacking style, character, and concept adjustments together - the standard way to combine effects, though interactions between LoRAs get unpredictable the more you pile on.

The inputs that matter

• model / clip - the pair to modify, straight from your checkpoint loader.
• lora - pick the file from the folder-tree combo, listing whatever's in your loras/ directory.
• strength_model - how hard it hits the diffusion weights, default 1.0, range -100 to 100. Most LoRAs actually look best somewhere in 0.5–0.8; the full-strength default is frequently too much.
• strength_clip - same idea for the text encoder, tunable independently of strength_model. If a LoRA is oversaturating your output at strength 1 on both, dropping just the model strength while keeping CLIP closer to 1 (or vice versa) is a common fine-tuning move.

Outputs are the modified MODEL and CLIP - wire them onward the same way you would the checkpoint's originals.

Installing it

No install section shipped with the pack, so it's the usual two paths: ComfyUI Manager → search SK-ComfyUI-FolderingLoader or the class name → install → restart, or manually cd ComfyUI/custom_nodes && git clone https://github.com/SKFRMSEHF/comfyui_SK_Loader and restart. No bundled models - you bring your own LoRA files, same as with any LoRA loader.

Common issues

• Base model mismatch. A LoRA trained for one architecture won't work on another - an Illustrious LoRA on a Flux checkpoint, for instance, does nothing useful. Even within a nominally shared codebase, SDXL/Pony/Illustrious LoRAs aren't reliably cross-compatible despite all showing up in the same picker.
• Missing trigger word. Plenty of LoRAs need a specific word in your prompt to activate their concept - check the model's source page. ComfyUI won't warn you if a LoRA is loaded but effectively inert; it just quietly does nothing.
• Strength too high by default. 1.0 is the widget default, and it's very often too strong - oversaturated, overcooked-looking output is usually a strength problem before it's anything else.
• Empty combo. The folder tree only lists what's already in your LoRA directory at ComfyUI startup - add the file and restart, it isn't a live scanner.
• No community trail for this specific pack. If something behaves oddly, there's no thread to search since the pack has no online footprint; the stock LoraLoader is a functional drop-in replacement while you sort it out.