Load LoRA Only

The stock LoraLoader in ComfyUI does two jobs at once: it reads the file and immediately patches your checkpoint with it. That's fine when you just want to drop a style on top of a model. It's annoying when you want to inspect the LoRA, zero out a few overbaked layers, merge it with another LoRA, and then decide how hard to apply it. Load LoRA Only is the first half of that split: it reads a .safetensors file from your loras folder and hands you the raw state dict, with nothing applied to anything yet.

That raw dict is the currency this whole pack trades in. The LORA output isn't a patched model - it's the tensor weights plus their metadata, passed around so the other nodes in the LoraUtils family can chew on it. You can wire it straight into LoRA Stat Viewer, LoRA Layers Operation, Lora Add, or skip all the surgery and send it directly to Merge LoRA to Model, which is the node that finally patches your checkpoint and CLIP.

Here's the part the name undersells: it's not just a loader. Two things happen while you're in here.

Strength is applied at load time. The strength_model slider (default 1.0, range −5 to 5) multiplies the tensor weights the moment the file is read, so the dict you pass downstream is already scaled. Negative values work and give you the classic "inverse LoRA" trick - flip a LoRA's sign to subtract its effect instead of adding it. One honest quirk: it scales every tensor in the file, including the alpha scalar if your LoRA stores one (kohya-style files usually do). So cranking the slider way up doesn't follow quite the same curve you'd get from applying a big strength in LoraLoader. Keep the load-time strength modest and do fine control later at Merge LoRA to Model, and you'll never notice.

It can merge with another loaded LoRA. There's an optional lora input. If you feed it one, the node adds the newly loaded file's tensors on top of it, key by key - matching keys get summed, keys only one side has are kept as-is. That's a clean way to layer two LoRAs before any scaling decisions, though for a proper weighted merge with different ranks you'd reach for Lora Add instead.

The lora_name dropdown lists every LoRA in your models/loras folder, exactly like the built-in loader. The node caches the last file it read, so re-running the graph doesn't re-read the same .safetensors off disk - a small thing that matters when you're iterating on a heavy file.

Installing this pack is a two-minute job, and it has no special requirements beyond what ComfyUI already ships (pure PyTorch and safetensors, no model downloads). Either grab it through ComfyUI Manager by searching for Comfyui-LoraUtils, or:

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

Then restart ComfyUI and the nodes show up under the LoraUtils category. It's a small, semi-neglected pack by lrzjason (you may know the same person as xiaozhijason on Civitai, who publishes Qwen Image fine-tunes like Rebalance) - one of those useful-but-niche utilities that has basically no community footprint, so don't expect a big user base to copy workflows from.

One last reminder, straight from the LoRA playbook: a LoRA is architecture-bound. This node will happily load an SDXL file and hand it to you - it won't stop you from merging it into a Flux checkpoint later and getting nothing but noise. Check the model page for what base your LoRA was trained on before you start editing.