Multi LoRA Stack

Most ComfyUI LoRA loaders give you a single flat list: N LoRAs, one strength each, applied in order. That's fine for images, but the moment you're working with multi-model setups - Wan video with separate stacks for UNet and text encoder, or a pipeline that treats different LoRAs as interchangeable "slots" - a flat list is where your workflow starts to fall apart. Multi LoRA Stack (class MultiLoraStackerLM) gives you four named panels, A/B/C/D, each with its own list of LoRAs and per-entry strengths and on/off toggles.

It's the most visual node in the pack: a widget UI with four panels instead of four separate nodes stacked in the graph. You tick a LoRA on, set strength and clip strength inline, and the node does the rest.

How it works

Each panel serializes its LoRA list to a JSON string state (the loras_state_a through loras_state_d inputs - you mostly never touch these by hand; the UI writes them for you). On execution the node builds four independent LORA_STACK payloads, one per slot, and returns three outputs:

• multi_lora_stack - the structured payload with slots a/b/c/d, for anything that understands a MULTI_LORA_STACK (the pack's Prompt Manager, the recipe nodes)
• lora_stack - all four slots flattened into one plain LORA_STACK in A→B→C→D order, for a normal LoraLoaderModelOnly or similar
• trigger_words - the LoRA trigger words gathered from each active entry, comma-joined, ready to append to your prompt

The trigger-word trick is the part that needs a note: it pulls trigger words and absolute path resolution from ComfyUI-Lora-Manager when that pack is installed, and falls back to its own path resolver when it isn't. Install Lora-Manager and you also get the trigger words auto-read from each LoRA file, plus the hover preview. Without it, trigger words come back empty and you're typing them yourself.

Wiring it

The common pattern: connect the multi_lora_stack output to a Prompt Manager node's lora_stack_a (it accepts both stack types), and wire trigger_words into the same node's trigger input or directly into your prompt text. If your graph only takes one stack, use lora_stack. If you want to hand the four slots off separately - say, slot A goes to the image model and slot D goes to the video module - run the multi_lora_stack output through the pack's Multi LoRA Splitter, which peels it back into four separate LORA_STACK outputs.

Installing

It's part of the ComfyUI-Prompt-Manager pack:

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

Or search "Prompt Manager" in ComfyUI Manager and install there. Restart, then drag the node in. No models, no extra downloads. Install ComfyUI-Lora-Manager if you want the trigger-word and preview magic.

Gotchas

If your combined lora_stack output looks wrong, check whether a slot's LoRA is toggled off - the UI keeps disabled entries visible but the node skips them, so the count may not match what you see on the panel. And if a saved workflow references LoRAs you don't have, the pack highlights them as missing rather than silently failing, which is exactly the behavior you want when someone shares a stack-heavy workflow and you're missing half its LoRAs. Keep strengths sane: stacking the same effect across four slots is a great way to overbake a face into uncanny valley, no matter how good the node UI is.