LoRA Loader (Finding LoRA)

If you've got more than a few hundred LoRAs, the stock LoRA loader stops being a tool and becomes a chore: a dropdown that stretches off your monitor, no way to remember which trigger word goes with which file, and manual rewiring every time you want to stack a second LoRA. LoRA Loader (Finding LoRA) is one developer's fix for exactly that - a model-only loader with bookmarks, stored trigger words, fuzzy search, and a button that chains another loader into your model line for you.

It's a small, opinionated pack, and honestly that's its strength. The author (Peter Neill, aka shootthesound) built it because he has over a thousand LoRAs and wanted the loader he wished ComfyUI shipped with. It's new and low-traffic, so you won't find a big community footprint behind it yet - judge it on the feature list.

How it works

At the core it's a drop-in replacement for the stock LoraLoaderModelOnly: MODEL in, MODEL out, no CLIP side. That makes it a natural fit for pipelines where the LoRA chain is model-only - the README calls out Klein 9B, Flux 2, Wan, and Z-image. LoRAs are patches on the model (10–200 MB adapters that tweak cross-attention layers rather than replace the checkpoint, as our concepts doc puts it), and this node just applies them, one or many.

The quality-of-life is all around the edges:

• Bookmark bar (📚) and a bookmark button (📖) - one click saves the current LoRA, and bookmarks persist globally in <ComfyUI>/user/finding-lora/bookmarks.json, synced live across every node you have open.
• Fuzzy search picker - click the LoRA bar and you get a proper modal over a giant dropdown. Type a few characters and it fzf-matches across your whole loras/ folder, with substring hits ranked above scattered matches.
• Trigger words - when you bookmark, you can attach a trigger phrase. It displays on a read-only row, and clicking it copies it to your clipboard.
• Chain another LoRA Loader - a button that spawns a fresh copy of the node beside the current one and re-routes the MODEL chain through it automatically. No dragging, no rewiring.

There are also stack rows on the node face for up to ten extra LoRAs, each with its own strength, applied on top of the main one.

The inputs and outputs that matter

• model (MODEL) - the diffusion model the LoRA patches. Wire it from your checkpoint/UNET loader.
• lora_name - the LoRA to load, chosen through the picker rather than typed.
• strength_model (FLOAT, default 1.0) - how hard the LoRA hits. Range −100 to 100, step 0.01. If it looks overcooked, 0.7 is a common starting point.

Outputs: model (the patched model, feeds your sampler) and trigger (STRING) - the trigger phrase saved against the bookmarked LoRA, or an empty string if none. Wire that into your prompt-encoding chain with a string-concat node and your trigger words auto-prepend. That's the feature that saves your memory: no more opening CivitAI to find out what ch9ractername stood for.

Installing it

It has no Python dependencies beyond what ComfyUI already ships - no requirements.txt, nothing heavy. Easiest route:

1. In ComfyUI Manager, search for "Finding LoRA" and install.
2. Restart ComfyUI.
3. Add LoRA Loader (Finding LoRA) from the loaders category.

Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/shootthesound/comfyui-lora-FindingLora
# restart ComfyUI

Gotchas worth knowing

• It's model-only. If you're on an SD1.5/SDXL workflow where the LoRA patches CLIP too, this loader skips that side - reach for the stock LoraLoader there and this one where the model chain is all you need.
• Bookmarks live in a plain JSON file. It's deliberately editable for bulk import, but manual edits need a restart to take effect, and a malformed file gets ignored rather than crashing your session.
• Custom nodes run arbitrary code with no sandbox, so if you're cautious (and per the ecosystem doc, the only real "scan" is reading the code), this one's easy to eyeball - the whole backend is a single nodes.py.

If you live in LoRA-stacking workflows, this is the loader that makes the most-used node in your graph stop feeling like the worst one.