Lora Loader By Path

The full sibling of LoraLoaderOnlyModelByPath - same idea, but it also updates CLIP. If your LoRA collection doesn't live inside models/loras (an external drive, a training output directory, a shared folder you don't want to duplicate into ComfyUI's tree) this node lets you load straight from a path instead of copying files around or symlinking.

How it works

You give it model and clip, a lora_path pointing at a .safetensors file anywhere on disk, and two separate strength dials - strength_model and strength_clip - and it applies the LoRA to both. That split matters: some LoRAs, especially anything trained on an SD1.5 or SDXL base, genuinely modify the text encoder, and if you only load them model-only you lose part of what the LoRA actually does. On newer architectures where text-encoder training is largely off, strength_clip will do less work, but it's still there for the LoRAs that need it.

The inputs and outputs that matter

• model, clip - both required, both get modified.
• lora_path - a raw filesystem path, not a dropdown. The default (c:\loras\my_lora.safetensors) is a giveaway that this expects a literal path string - flip the slashes on Linux/macOS, and know there's no autocomplete to catch a typo before you queue.
• strength_model, strength_clip - each independently adjustable from -20 to 20 in 0.01 steps. Realistic values sit close to 1 for both; the wide range exists for people doing deliberately unusual blends, not as a hint that extreme values are normal.
• Outputs: MODEL and CLIP, both updated - wire them onward the same way you would the stock LoRA loader's outputs.

How to install it

Through ComfyUI Manager: search "SP-Nodes," install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/bananasss00/ComfyUI-SP-Nodes

Restart afterward. Nothing else to download - it just reads the LoRA file you point it at.

Common issues

Same path gotchas as the model-only version: a wrong or mistyped lora_path fails at queue time with a generic "not found" rather than anything more specific, so get the path right before you spend a queue slot on it. If you notice the LoRA's effect looks weaker than expected compared to loading it through the standard dropdown-based loader, check that both strengths are actually set the way you think - it's easy to nudge strength_clip while meaning to adjust strength_model, since they're two separate widgets stacked right next to each other. And if you're loading a LoRA that genuinely has no CLIP component (increasingly common on Flux-era models), you won't see any difference from using the model-only loader instead - in that case LoraLoaderOnlyModelByPath is the simpler, equally correct choice.