Load LoRA (full-pipe)

In a normal ComfyUI graph, applying a LoRA means breaking your model wire, inserting a LoraLoader, and dragging model and CLIP through it before they reach the sampler. In this pack's pipe world, you just place Load LoRA (full-pipe) inline: full_pipe in, pick a LoRA, and an updated full_pipe out with the LoRA merged into the model and CLIP slots. Nothing else in the chain changes - that's the entire point.

Mechanically it's a thin wrapper over ComfyUI's own LoRA loading: pick lora_name from your ComfyUI/models/loras/ folder, set strength_model and strength_clip, and the node clones the model and CLIP, loads the LoRA weights with load_lora_for_models, and writes the patched versions back into the pipe. Because the model is cloned, the original pipe on the input side is untouched - you can branch the same base pipe into multiple LoRA variants.

The inputs that matter

• lora_name - dropdown of your installed LoRAs. The obvious one.
• strength_model (default 1.0) - how hard the LoRA's weights hit the diffusion model. Range −20 to 20; most LoRAs live at 0.5–1.0, and negative strengths are a real trick for subtracting a style.
• strength_clip (default 1.0) - how hard it hits the text encoder. The community rule of thumb from the LoRA-training world still applies: if the LoRA changes style but the prompt stops being respected, ease the clip strength.

Output is the updated full_pipe. Since this is a pipe-to-pipe node, it slots anywhere in a mudknight workflow - before a detailer, before the base generation node, wherever the pipe flows.

Installing it

Part of mudknight-utils - ComfyUI Manager (search comfyui-mudknight-utils) or:

cd ComfyUI/custom_nodes
git clone https://github.com/mudknight/comfyui-mudknight-utils

then restart. No Impact Pack, no model files of its own - it reads whatever LoRAs you already have installed.

Where people get burned

Two things. First, the dropdown only shows LoRAs that were present when the graph loaded - drop a new LoRA into models/loras/ and you may need to refresh ComfyUI before it appears. Second, the classic: a LoRA trained for one architecture will happily load on another and just do nothing useful (or subtly wreck the image) because the tensors don't line up with the model. Architecture lock-in is baked into LoRAs - a Flux LoRA won't run on Anima or SDXL, and this node won't warn you. If your image looks subtly broken after adding a LoRA, check it's the right family for your model before touching the strength sliders.