Load External LoRA Model Only

Every ComfyUI user has been here: someone posts a killer LoRA, hands you a direct download link, and you go through the ritual - save the file, rename it, drop it in models/loras, refresh the node list. Load External LoRA Model Only kills that whole dance. Paste the URL straight into lora_name and the node downloads the file into your loras folder and applies it, all in one run.

The catch is in the name, so read it carefully: model only. Unlike the built-in LoRA loader, this one returns just a MODEL - no CLIP output, no strength_clip. That makes it a natural fit for Flux, which is exactly what the author says it was tested with. Flux Dev and its relatives mostly use model-only LoRAs anyway (the README frames it that way), so for a lot of modern workflows you're not giving up anything. If you genuinely need a text-encoder-touching LoRA, this is the wrong tool and you want the standard loader.

How it works

The mechanism is simple and honestly a little refreshing. Give it a URL and it parses the filename off the end of the path, then downloads into the first entry of ComfyUI's loras folder via a plain requests.get with a browser-ish user agent:

lora_name: "https://example.com/models/my_style.safetensors"
# downloads to  ComfyUI/models/loras/my_style.safetensors

If that filename already exists, it skips the download entirely - a deliberate touch so re-running a workflow doesn't re-fetch a 200MB file. Pass a bare filename instead of a URL and it behaves exactly like any LoRA loader, looking the file up in the loras folder. An empty lora_name short-circuits and returns your model untouched, which is a nice safety valve. It caches the loaded weights per node instance, and strength_model accepts negatives, which inverts the LoRA's effect (subtracting instead of adding the learned delta).

The inputs and output

• model - the diffusion model to patch. Comes from your checkpoint, UNET, or dual-checkpoint loader.
• lora_name - the one that matters. Either a local filename from models/loras or a full http(s):// URL to download.
• strength_model - default 1.0, range −100 to 100. Negative values invert the effect.

Output: a single MODEL, ready to feed your sampler.

Installing

It's part of Ruucm's ComfyUI Nodes, so one clone covers it:

cd ComfyUI/custom_nodes
git clone https://github.com/ruucm/ruucm-comfy

Restart ComfyUI and you'll find it under "loaders". Note the pack's requirements.txt also installs insightface and onnxruntime-gpu - heavy, and only needed by the pack's other nodes. If you only want this loader, you can install the pack anyway; just be ready for a slow pip step.

The traps

• The URL has to be a direct file link. Paste a Civitai model page URL and you'll save an HTML page, not a .safetensors - then ComfyUI fails to parse it as a LoRA with a confusing error. Use the actual direct-download endpoint. If a URL has no filename in the path, the node invents one with a UUID and .safetensors extension, which is usually a sign you gave it the wrong kind of link.
• No checksum, no versioning. The skip-if-exists check means if the author updates the file at the same URL, you keep using the stale cached copy until you delete it manually.
• It doesn't verify what it downloads. A 404 or an HTML error page gets saved to disk and then surfaces as a load error one step later. Keep the console open the first time.

For a small utility node it does exactly one job well: turn "here's a direct LoRA link" into a working load in one paste. If you live on Flux LoRAs, this is the one from the pack you'll actually use every day. Just don't expect CLIP patching - it's in the name.