Lora Loader (LoraManager)

If you've ever stared at a workflow full of stacked LoraLoaderModelOnly nodes trying to remember which one needed which trigger word, this is the node that fixes that. It's the workflow-side half of ComfyUI-Lora-Manager - a whole model-management app that lives at http://localhost:8188/loras - and its one job is to take LoRAs you picked in that browser UI and drop them into your graph with the strengths and trigger words already filled in.

The trick is the "LoRA syntax" idea. Instead of a dropdown per LoRA, this node has a single text box. You browse your collection in the manager, right-click a LoRA, hit "Copy LoRA syntax," and paste something like <lora:my_style:0.8> into the box. The node reads that, applies the LoRA at that strength, and - the part people actually install this for - hands you the LoRA's trigger words on a separate output so they flow straight into your prompt. The text field has / autocomplete for everything the manager has scanned, so most of the time you're typing, not pasting.

This is openly inspired by rgthree's Power Lora Loader, which the README credits directly. If you know that node ("I swear by the rgthree Power Lora Loader," as the standard recommendation goes), this is the same idea wired into a fuller management app - the manager can fetch trigger words and preview images off CivitAI and keep them locally, which is the only copy you've got once a model gets pulled from the site.

How it works

You wire your base MODEL in, put one or more LoRAs in the text box, and the node patches the model (and CLIP) with each LoRA at the strength written in the syntax. Stack several by putting several <lora:...> entries in the box. Because the strengths and trigger words travel with the syntax you copied from the manager, you're not re-typing the values the model page told you to use.

Inputs and outputs that matter

• model (MODEL, required) - your checkpoint's model. The output MODEL is the patched one; feed it to your sampler.
• text (the autocomplete LoRA box, required) - where the <lora:name:strength> entries go. This is the whole interface.

Four outputs: MODEL and CLIP (the patched pair, into your sampler and text encode), trigger_words (a STRING of the activation words for what you loaded - send it into a Prompt (LoraManager) node or your CLIP encode), and loaded_loras (a STRING listing what actually got applied, handy for sanity-checking).

How to install it

Via ComfyUI Manager: open Manager > Custom Node Manager, search lora-manager, Install, restart. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/willmiao/ComfyUI-Lora-Manager.git
cd ComfyUI-Lora-Manager
pip install -r requirements.txt

then restart. No model downloads, no GPU deps - it's a management tool, so requirements.txt is light. If you want the built-in CivitAI downloads, paste your CivitAI API key into the LoRA Manager settings page and save.

Common issues & troubleshooting

trigger_words comes out empty. The node can only hand you words the manager actually has. If it never fetched metadata for that LoRA (no CivitAI match, no local .metadata.json), there's nothing to pass. Open the LoRA in the manager UI, set the trigger words there, and they'll flow.

A LoRA does nothing. Same checklist as always: LoRAs are architecture-bound, so an SDXL LoRA is inert on Flux and vice versa - match the base model. After that, check the trigger word is actually in your prompt and the strength isn't cranked so high it fries the image.

The LoRA isn't in autocomplete. The manager hasn't indexed it. Make sure it's in a folder the manager knows about (check the folder paths in settings) and rescan.