Lora File Loader (Soze)

Stock ComfyUI's LoRA loader only knows about files it finds by scanning your models/loras folder at startup. This one skips that scan entirely - you give it a filesystem path, and it loads whatever's there. That matters more than it sounds: freshly trained LoRAs sitting in a training-output directory, files on a shared drive, anything you don't want to symlink or copy into ComfyUI's own folder structure - all of that becomes usable without touching your models directory or restarting ComfyUI to pick up a new file.

How it works

It applies the LoRA the same way any LoRA loader does - patching the diffusion model's and CLIP's weights by the given strengths - the only real difference is where it reads the file from. The status output's own tooltip tells you it reports "LoRA name, strengths, and cache hit/miss," which means the node keeps the last-loaded file around rather than re-reading it off disk every single run. That's why re-running with the same lora_filepath tends to be fast the second time.

The inputs and outputs that matter

• model / clip - the diffusion model and CLIP the LoRA gets applied to, same as any LoRA loader.
• lora_filepath - the file's absolute path. It's read literally, so don't type ~/loras/thing.safetensors on Linux or Mac expecting shell-style expansion - most file-loading code doesn't do that, and it'll fail as "not found" even though the path looks right to you.
• strength_model / strength_clip - separate weights for the diffusion model and CLIP, -100 to 100. Default is 1.0 for both, but per the LoRA ecosystem generally, 1.0 is often too strong - 0.5 to 0.8 is the more common comfortable range. If your results look right stylistically but the prompt seems to be fighting the LoRA, try backing off strength_clip a bit further than strength_model before you touch anything else.

Outputs are MODEL, CLIP (both modified, ready to feed straight into your sampler chain), and status.

Installing it

ComfyUI Manager: search ComfyUI_Soze. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/SozeInc/ComfyUI_Soze.git, pip install -r ComfyUI_Soze/requirements.txt, restart. No models bundled - you supply the LoRA file yourself, from wherever it lives.

This pack overall reads as a personal QoL toolkit rather than a widely-adopted "essential" pack - it's a small, obscure project (it started by folding in the abandoned Comfy_KepListStuff helpers), so don't expect community threads dissecting its edge cases. What it does, it does plainly, and the README is honestly your best source of truth.

Common issues

The path has to be exact and absolute - no ~, no relative paths from ComfyUI's own directory. Beyond that, the usual LoRA landmine applies regardless of loader: the LoRA has to match the checkpoint's architecture. An Illustrious LoRA won't do anything sensible on a Flux model and vice versa - if results look garbled or the LoRA seems to have zero effect, check compatibility before assuming this node is misbehaving.