🪐 ✨ LoRA Loader (Pretty)

Civitai hands you files named krea2_Ana-De-Armas_v1.safetensors. ComfyUI shows you exactly that. This node shows you Ana De Armas V1 and loads the right file behind the scenes. Same wiring as the plain folder loader, same two model outputs, plus a name string - it just costs you less squinting, and it adds a trick the plain one doesn't have.

How the pretty name is derived

There's no database lookup and no network call. The parser is dumb and predictable, which is worth knowing because it decides whether your library looks elegant or absurd:

1. Take the filename, drop the extension, split on _.
2. If there are two or more parts, take the second part - i.e. the chunk after the creator prefix (krea2_).
3. Split camelCase, turn hyphens into spaces, capitalize each word, but keep a whitelist uppercase (NSFW, LORA, V1–V4, FP16, HM).
4. If the third part looks like a version (v1, v2.5), append it: Ana De Armas V1.

So it's built for the creator_Name_v1.safetensors convention that Civitai, and most trainers, produce. Files that follow it look great. Files named my cool lora final FINAL.safetensors will come out weird, because the parser is guessing from positions.

When the folder filter reaches into nested folders, the picker prefixes the subfolder it found the file in (Celebrities - Ana De Armas V1), and the string output strips that prefix back off, so downstream you still get a clean name.

Two things it does that the plain loader doesn't

[ RANDOM ]. There's a special entry at the top of the picker alongside [ NONE ]. Pick it and the node rolls a random LoRA from the current folder at execution time. Combined with the usual queue-it-twenty-times habit, that's a cheap "which of these characters is funniest" sweep - no extra nodes.

A zero-strength fast path. If both strength_model and strength_clip are 0, it skips reading the weight file entirely and just returns the formatted name. Useful when you want the name string for a filename or a prompt without paying the load.

Inputs and outputs

model and clip in, folder filter, lora_name (the pretty dropdown), both strengths (default 1.0, range -100 to 100), and the advanced output_format / custom_regex pair. Parsed Name is the format that strips the trailing version back off. Out comes MODEL, CLIP, and lora_name as a STRING.

Install

Same pack, same install - ComfyUI Manager → SaturnNodes, or:

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

Restart. Requirements are Pillow, numpy and piexif; nothing heavy, no model downloads, no install script. The node shows up in 🪐 SaturnNodes/Loaders, and it's a separate node from the plain folder loader - both are registered, so you can mix them.

Where people get burned

Two LoRAs, one entry. The pretty list is a name-to-path dictionary. If two files in the same filtered folder parse to the same pretty name, one of them wins and the other is simply not selectable. If a LoRA seems to have vanished from the dropdown, rename it with a distinguishing version suffix or split the folder.

The name in the output isn't the file you picked. With output_format on Parsed Name, the string is the human name, not the filename, so pasting it into a tag-based loader later won't necessarily resolve. Use Filename if the string is going somewhere that parses names literally.

Nothing loads, no error. You're on [ NONE ]. Also worth checking the folder filter matches the actual path inside models/loras - the pretty display doesn't change how the filter matches, so PonyV6 still needs PonyV6* to catch subfolders.

Nothing appears to change in the image. Same architecture trap as any LoRA loader: adapters are base-model bound, so an SDXL LoRA on Flux is a very expensive no-op. Reorder or lower the weight before you blame the node.