Lora Stacker (LoraManager)

The Stacker does the same picking-and-syntax dance as the Lora Loader, but it doesn't touch your model. Instead it packages your chosen LoRAs into a LORA_STACK - the standard stack format that Efficiency, Impact, and other packs' loaders already understand - and hands that off. Think of it as "describe the LoRA set now, apply it later, somewhere else."

That indirection is the point. A loader bakes the LoRAs into a specific MODEL right there. A stack is portable: you can merge two of them, combine one with a randomizer's output, or feed one stack into several loaders. If your workflow already has an Efficient Loader or an Impact Pack loader with a lora_stack input, this node lets you build that stack with the manager's autocomplete and its trigger-word tracking instead of the clunky per-LoRA widgets.

How it works

You type or paste <lora:name:strength> entries into the autocomplete text box, same as the loader. The node bundles them - names and strengths - into a LORA_STACK object and also gives you the combined trigger words and a readable list of what's active. Nothing is applied to a model until that stack reaches a loader downstream.

Inputs and outputs that matter

• text (the autocomplete LoRA box, required) - the only input. Put your <lora:...> list here; / autocomplete pulls from whatever the manager has scanned.

Three outputs: LORA_STACK (wire it into any stack-consuming loader, or into Lora Stack Combiner to merge with another), trigger_words (STRING, into your prompt encode) and active_loras (STRING, a human-readable summary of what's in the stack).

How to install it

Via ComfyUI 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. Light Python deps, no model downloads. Add your CivitAI API key in the LoRA Manager settings page if you want the built-in downloads and metadata fetching.

Common issues & troubleshooting

The stack "does nothing." A stack is inert on its own - it's just a description. Make sure it's actually plugged into a loader that has a lora_stack input. If the downstream loader ignores it, check that the loader accepts the standard LORA_STACK type (most do; a few packs use their own).

Trigger words are missing. Same story across this pack: the words only come through if the manager has them for each LoRA. Set or fetch them in the UI.

Base-model mismatch. The stacker will happily package a Flux LoRA and an SDXL LoRA together; the loader will apply whichever matches your checkpoint and silently skip the rest. LoRAs are architecture-bound, so keep a stack to one base family unless you know what you're doing.