Load HF Lora

Every ComfyUI user has hit the wall: the workflow uses a LoRA that isn't in your models/loras folder, and the model file is only hosted on HuggingFace. Your options are browser-download, move into the right directory, refresh the list... or use Load HF Lora, which does the whole thing at queue time. You type a repo, a subfolder, a filename, and the node fetches the LoRA, loads it, and applies it to your model and CLIP - exactly like the stock LoraLoader, minus the file management.

How it works

It takes a model and clip like any LoRA loader, plus three strings that identify the file: repo_id (e.g. ostris/ikea-instructions-lora-sdxl), subfolder (usually empty, but some repos nest files), and filename. The download uses hf_hub_download, which is resumable and fast, into the pack's cache at ComfyUI/models/huggingface_cache/. Then it calls ComfyUI's own load_lora_for_models - the same function the built-in loader uses - so the mechanics after download are identical to what you're used to.

The strengths are the standard pair: strength_model and strength_clip (default 1.0, range -20 to 20). Leave both at 1.0 for most use; drop strength_model toward 0.5-0.8 when a LoRA is overpowering. If both are 0, the node short-circuits and returns the model and CLIP untouched - a handy bypass without unplugging wires. Outputs are the patched MODEL and CLIP.

One small implementation detail worth knowing: the node caches the loaded weights on its own instance, so re-running the same repo/file doesn't re-read from disk. If you switch files, it frees the old one. It's the kind of housekeeping that keeps long queue sessions from eating RAM.

What to watch

• It's a one-shot downloader. After the first run the file is cached, but it won't appear in the normal LoraLoader dropdown - this node is the only thing that knows about it. If you want the LoRA in your regular menu, grab the file from the cache and move it to models/loras.
• The repo must be public and unauthenticated. Gated repos will fail; this is a no-key node.
• Architecture lock-in applies. A Flux LoRA won't run on SDXL and vice versa - the loader will happily try and produce nothing or errors, because LoRAs are hard-bound to the backbone they were trained against.

Install

It's part of ComfyUI-GlifNodes, from the glif.app team. ComfyUI Manager → "ComfyUI-GlifNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/glifxyz/ComfyUI-GlifNodes

Restart ComfyUI. The pack's only pip requirement is diffusers>=0.28.0, and this node adds no model downloads at install - just whatever LoRA you point it at, on first run.

Reach for it when a shared workflow's LoRA lives on the Hub. It's one of those nodes you don't need often, and the day you do, you're very glad it exists.