Load LoRA (Temporary)

The normal LoRA loader makes you shuffle files into models/loras before you can even see if the thing is any good. Load LoRA (Temporary) - class LoadTempLoRA - takes a URL instead and loads the LoRA straight into memory. It's the "try before you commit" node, and honestly it's the friendlier half of this pack: a LoRA is only 10–200 MB, so the whole in-memory trick costs you almost nothing, unlike a multi-gigabyte checkpoint.

How it works

It behaves exactly like the standard Load LoRA node, just with a URL field in place of the file dropdown. You feed it your MODEL and CLIP from the checkpoint, paste a direct link into ckpt_url, and it returns patched MODEL and CLIP outputs - same wires, same graph, nothing else changes. The other inputs are the ones you already know from the standard loader plus two pack additions:

• strength_model and strength_clip - both default to 1.0, range −20 to 20. Standard LoRA physics: 0.5–0.8 is a common starting point if 1.0 comes on too strong.
• ckpt_type - auto (default) reads the extension from the server's Content-Disposition header; safetensors or other forces it when the header is missing.
• download_split - parallel download threads, default 4.

Two things make it genuinely nicer than the checkpoint sibling. First, it caches: the node remembers the last LoRA it fetched, so re-running with the same URL doesn't re-download it every time - that alone makes it usable in a real workflow, not just one-off tests. Second, if you set both strengths to 0 it short-circuits and passes the inputs straight through without downloading anything, which is a neat way to toggle a LoRA off without deleting the URL.

The mechanism

Mechanically, it's the standard ComfyUI LoRA machinery (comfy.sd.load_lora_for_models) fed from a memory buffer. .pt and .ckpt LoRAs are loaded with the safe (weights_only) loader, so the risk profile is the same as any local LoRA - but remember you're pulling from an arbitrary URL, so only paste links you trust.

Common issues

Troubleshooting is mostly about the URL itself. It must be a direct file link - redirects are followed, so HuggingFace resolve URLs work, but anything that needs a browser session or an auth cookie will fail. And if ckpt_type is on auto but the server sends no Content-Disposition, the node can't sniff the format; set the type manually rather than chasing a weird load error.

How to install

Install: ComfyUI Manager (search "ComfyUI-TemporaryLoader"), or cd ComfyUI/custom_nodes && git clone https://github.com/pkpkTech/ComfyUI-TemporaryLoader, then pip install -r requirements.txt (just requests) and restart. It shows up alongside its siblings in the temporary_loaders category. If you're on a machine where disk space is the recurring argument, this is the node that wins it.