VRGDG LoRA From Path Model Only

VRGDG LoRA From Path Model Only applies a LoRA to a model directly from a filesystem path and hands back the patched model. If you've ever trained a LoRA and had to copy the .safetensors into ComfyUI/models/loras/, restart, and only then test it - this node exists to skip all of that. Point it at the file wherever it lives, and it patches.

The "Model Only" in the name is the scope. It's a lean version of the usual LoRA-apply node: it takes a MODEL in, applies one LoRA from a full path, and returns a MODEL out. No CLIP, no conditioning, no strength on multiple components. If you just want the model patched, this is the minimal surface you need.

Where it shines: the workflow-relevant case the tooltip calls out - "hidden workflows that need to preview a freshly trained LoRA before it is copied into ComfyUI's loras folder." The pack trains LoRAs (musubi/Krea 2 workflows) into arbitrary project folders, and this node lets you generate a test image with a brand-new LoRA without touching your model directory. That's a genuinely nice loop for iterating on training runs: train, sample with this node, tweak, re-train.

The inputs

• model - the model to patch. Wire in a checkpoint/diffusion model loader.
• lora_path - the full path to a .safetensors LoRA file. Path-based, so it doesn't need to be registered in ComfyUI's loras folder.
• strength_model - how strongly to apply it. 1.0 is normal, 0.5 is lighter, 0.0 passes the model through unchanged. The range goes negative, which is occasionally useful for subtraction, but you'll live at 0.5–1.0.

The output

• model - the patched model, ready to feed a sampler.

Installing it

Part of the VRGameDevGirl pack: Manager → search vrgamedev, or:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git

Restart, hard-refresh. No extra dependencies beyond the pack.

Common issues

• The path doesn't load. This node does not scan folders - it takes a literal full path. Make sure it's the absolute path to the file, including the .safetensors extension, and that the path has no trailing quote.
• Wrong architecture errors. A LoRA trained for one architecture won't apply to another (a Krea 2 native LoRA won't patch an LTX model, and so on). That's the model's fault, not the node's.
• Strength 1.0 too strong. 1.0 is the LoRA author's intended weight, but some LoRAs are trained hot. Drop to 0.7–0.8; the node makes iterating cheap, so use it.

It's a small node with a specific niche, but for anyone training through this pack it's quietly one of the most useful in the set - train, test from path, iterate, only promote the winner to your loras folder.