HuggingFace LoRA Loader (Model Only)

HuggingFace LoRA Loader (Model Only) applies a LoRA to a diffusion model - the model only, skipping the CLIP side entirely - with the LoRA weights coming from the Hugging Face cache or a local path. It's the focused sibling of the pack's full HfLoraLoader (which patches model and CLIP); this one is for when the LoRA only touches the UNet/diffusion transformer, which is true of most modern LoRAs. It also inherits the pack's defining HF rule: no auto-downloading. The LoRA file must already be cached, or you give it a local path.

Why model-only matters in practice: plenty of LoRAs - especially for newer architectures like FLUX, where the CLIP is often a frozen T5 that nobody finetunes - don't contain CLIP weights at all. The full ComfyUI LoraLoader expects a CLIP socket and applies CLIP-strength math that does nothing for such LoRAs. This node drops the pretense: you hand it a MODEL, it patches just that, and there's no empty CLIP socket cluttering the graph. If you've ever fought the "this LoRA has no CLIP weights" edge case, you know exactly why this variant exists.

The inputs

• model - the diffusion model the LoRA applies to (a MODEL).
• repo_id - the HF repo ID, absolute path, or ./relative path from the ComfyUI root.
• strength_model - how strongly to apply, default 1.0. Can go negative (invert the LoRA's effect) and ranges −100 to 100.
• subfolder / filename (optional) - locate the LoRA file inside the repo; auto-detects if filename is blank.

One output: MODEL, patched.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Duanyll/duanyll_nodepack
cd duanyll_nodepack && pip install -r requirements.txt

or ComfyUI Manager → "Duanyll Nodepack" → install → restart.

The gotchas

Same cache rule as the other HF loaders - if the repo isn't in the HF cache you'll get the "Run hf download first" error, so make sure you've cached the LoRA file ahead of time. And the usual ComfyUI caveat applies: this is a model-only LoRA application, so a LoRA that does contain CLIP weights will have those silently ignored - for old SD1.5/SDXL style LoRAs that modify the text encoder, use the full HfLoraLoader instead. The model output can be chained, so stacking several cache-sourced LoRAs onto one FLUX model is trivial. It's a niche convenience in an already-niche pack, but for the HF-cache LoRA workflow it's exactly the node you'd want: no CLIP socket, no forced pairing, just the patch.