LoraListUrlLoader

The usual LoRA dance is: go to Civitai, click Download, drag the file into ComfyUI/models/loras, then hunt for it in a LoraLoader dropdown. LoraListUrlLoader (displayed as "Lora URL List Loader") skips the middle two steps. You paste a direct download URL straight into the node, it pulls the file into your loras folder on the first run, caches it, and applies it like any other LoRA. It was built for one company's internal pipeline, but the trick generalizes nicely: share a workflow that references LoRAs by URL and whoever runs it gets the files fetched automatically.

Mechanically it's nothing exotic - underneath it's just ComfyUI's standard load_lora_for_models machinery, which is why anything a normal LoraLoader handles works here too. A LoRA is a small adapter file that patches a checkpoint's attention layers rather than replacing the model, so this node still needs a compatible base model fed into it and outputs a patched MODEL and CLIP. You can also type a plain local filename instead of a URL and it loads it directly, which means the node quietly doubles as a three-at-once LoRA stacker.

The inputs that matter

You get three slots, each a lora_N_url string plus its own model/clip strength:

• model and clip - wire in your loaded checkpoint (and its CLIP) like you would for any LoRA loader.
• lora_1_url / lora_2_url / lora_3_url - a direct file URL, or a local filename. Leave empty to skip that slot.
• model_strength_N / clip_strength_N - floats from -10 to 10, default 1.0. That default is usually too hot; 0.5–0.8 is the more common landing zone, and you'll almost always want model and clip strength equal.

Outputs are MODEL and CLIP, wired exactly like a LoraLoader's - MODEL into your sampler, CLIP into CLIPTextEncode. Stacking order is left to right, and stacked LoRAs interact in ways you can't fully predict, so expect to tune strengths when you combine a style and a character LoRA.

Installing it

ComfyUI Manager is the easy path: search for ComfyUI_Lora_List_With_Url_Loader, install, restart. Or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/liushuchun/ComfyUI_Lora_List_With_Url_Loader
pip install -r requirements.txt
# restart ComfyUI

One real trap in that README step: the requirements file lists torch, which ComfyUI already installed. Running pip install -r requirements.txt in your venv can happily swap your carefully-built torch/CUDA combo for the latest wheel and break everything. The code actually only needs requests (ships with ComfyUI) - the wget it imports is never even called. Install nothing, or at most pip install wget, and move on.

Where people get burned

• URLs without a real filename. The saved file is just the last path segment of the URL, and ComfyUI's torch loader picks its path by file extension. A Civitai /api/download/models/12345 link saves a file literally named 12345 with no extension, and it can fail to load. Use URLs that end in a real .safetensors filename - HuggingFace resolve/main/name.safetensors links are ideal.
• The README overpromises. It advertises "force refresh" and SHA-256 validation for free renaming. Neither is in the shipped code: the node only downloads when the file is missing, and there's no hash check anywhere. To grab an updated version of a LoRA you must delete the old file first. Same story with gated models - the downloader sends no API token, so anything that requires Civitai auth just 401s. Public files only.
• First run looks hung. Downloads happen when you hit Queue, streamed to disk, and a big LoRA over a slow link can sit there for a while. Give it a beat before assuming it's dead.

It's a niche tool with rough edges, and if you're on one machine manually downloading anyway, ComfyUI's built-in LoraLoader beats it. But for sharing reproducible workflows - or any setup where LoRAs should arrive on demand - pasting a URL and letting the node fetch and cache is genuinely handy.