Diffusers Load LoRA Only

In normal ComfyUI, loading a LoRA means applying it - the loader node slaps the weights onto the model and clip immediately. This pack splits that in two, and DiffusersLoadLoraOnly is the first half: it reads a LoRA file off disk and hands you the raw state dictionary, without touching any model. Applying happens later, through DiffusersMergeLoraToPipeline.

Why bother? Because it's the only way to do surgery on a LoRA before it ever reaches the transformer. The canonical workflow is: load the LoRA here, inspect or modify it with DiffusersLoraStatViewer / DiffusersLoraLayersOperation, optionally save the edited result with DiffusersSaveLora, and only then merge it into the pipeline. Keep the load and the apply as separate steps and you can test a whole family of edits against one cached load.

The one input that matters

Just lora_path - and this is where people stumble. It's a plain text field, not a dropdown. You type or paste an absolute path to the LoRA file, so there's no file browser and no models/loras resolution. If the path is wrong you get a FileNotFoundError and a dead end; if it's right, the node calls ComfyUI's own load_torch_file with safe_load=True (so safetensors it is).

There's a per-node-instance cache: load the same path twice in one session and the second run returns the already-loaded dict instead of re-reading disk. Different path, and it reloads. Cheap either way - these files are tens of MB.

Output

A single lora output of the pack's custom LORA type - essentially the state dict. It wires into:

• DiffusersLoraStatViewer to understand the layer structure,
• DiffusersLoraLayersOperation to zero out or scale specific transformer blocks,
• DiffusersSaveLora to persist an edit,
• DiffusersMergeLoraToPipeline to actually apply it.

Note that the README mentions "MergeLoraToModel" - that node doesn't exist. The real one is DiffusersMergeLoraToPipeline. README rot, not a new node.

Install

ComfyUI Manager → search "Comfyui-DiffusersUtils", or:

cd ComfyUI/custom_nodes
git clone https://github.com/lrzjason/Comfyui-DiffusersUtils

Restart ComfyUI. The README's pip install -r requirements.txt step will error out - there is no requirements.txt in the repo; the pack leans on ComfyUI's bundled diffusers/safetensors. (If you only use the LoRA nodes you don't need the GLM git-installs at all.)

Common issues

• File not found - check you're passing an absolute path, including the .safetensors extension, and that the file lives somewhere the ComfyUI user can actually read.
• A LoRA that loads but does nothing downstream - you've loaded it but never merged it. Follow it with DiffusersMergeLoraToPipeline; loading alone changes nothing.

It's a boring node, on purpose. The point isn't what it does - it's that it lets the rest of the pack edit a LoRA before it ever gets applied.