🪐 📁 LoRA Loader (Folder)

Core's LoraLoader hands you one flat dropdown of every .safetensors in models/loras. That's fine at fifty files and genuinely annoying at four hundred, which is where most people end up once they've downloaded a year of character and style adapters. This node does the ordinary LoRA job - patch the model, patch the CLIP - but it filters the list down to one subfolder first, and it hands you back the name of the LoRA it just loaded as a STRING you can actually wire somewhere.

That last part is the real reason to reach for it. Wire the name into a filename_prefix, into a note, or into the next prompt in a queue batch, and your outputs stop being "some image with some LoRA in it."

How the loading actually works

The folder widget is a filter over the relative paths inside models/loras, and it has three modes:

• Celebrities - that exact subfolder only.
• PonyV6* - the trailing wildcard means that folder and everything nested under it.
• empty - no filter, i.e. the wall of names you were trying to escape.

The lora_name dropdown isn't baked at startup. The node registers its own local HTTP route and repopulates the list from the current folder value, which is why changing the folder updates the choices without restarting the server.

Loading itself is the standard path: read the file, then apply it through ComfyUI's own LoRA merge, which patches the UNet and the text encoder. That's why there are two strength values. If a LoRA is wrecking your prompt comprehension but you like its look, drop strength_clip to 0 and leave strength_model alone.

Inputs and outputs that matter

You set folder, then lora_name, then strength_model and strength_clip (both default 1.0, range -100 to 100). Note that 1.0 is frequently too strong - the old 0.5-0.8 rule of thumb was SDXL advice, and on newer bases people often run 1.0 or higher because dropping it costs likeness. Just don't assume 1.0 is "neutral."

output_format is the advanced widget that decides what the string output says: Filename (the default - krea2_Ana-De-Armas_v1.safetensors), Parsed Name (a cleaned-up human name), Filename without extension, Relative Path, Full Path, or Custom Regex. If you pick Custom Regex, custom_regex is searched against the relative path and whatever it matches is your string; if it doesn't match, you quietly get the basename without extension.

Outputs are MODEL and CLIP (the patched pair, same as core) plus lora_name. Selecting [ NONE ] passes model and clip straight through and outputs an empty string - a clean way to bypass the node without unwiring anything.

Install

ComfyUI Manager → search SaturnNodes → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/KOFiblto/ComfyUI-SaturnNodes

Restart ComfyUI. Dependencies are just Pillow, numpy and piexif from requirements.txt, and there's no install.py doing surprise pip work. No models to download. All sixteen nodes in the pack land under the 🪐 SaturnNodes/* categories, and this one is in Loaders.

Where people get burned

Silent no-op. The dropdown defaults to [ NONE ], so a fresh node does nothing at all and looks like it's working. Wire it, then actually pick something.

Empty dropdown. The filter is exact on folder name unless you add a wildcard. Pony won't match a folder called PonyV6; PonyV6* will. It's case-insensitive on the wildcard path, but the matching is on the relative path inside models/loras, so a nested characters/PonyV6 won't be found by typing PonyV6.

The LoRA loads and nothing changes. Architecture mismatch is the usual cause - an SDXL LoRA on a Flux base does nothing, and cross-family adapters (Pony on vanilla SDXL, Illustrious on Pony) are unreliable even where the code nominally loads. Also check for a required trigger word; plenty of LoRAs look broken until you type it.

Your lora_name string is a full path. That's output_format set to Full Path and a filename_prefix that now contains slashes. Switch it to Filename or Relative Path.