Random Multi-LoRA Model Loader

Here's the problem the Random Multi-LoRA Model Loader solves: you're cycling through 20 prompts, and prompt #7 needs a character LoRA plus a style LoRA at different strengths, while prompt #14 needs a completely different pair. Manually swapping LoRAs between runs is miserable. This node stores up to 20 prompts' worth of LoRA configs - 3 LoRA slots each, with per-slot strengths and trigger words - and applies the right stack automatically based on a prompt_index. It's the full-scale companion to the single Random LoRA Model Loader, and it's the node the pack's example workflow pairs with AUN PromptCycler.

How it works

A prompt_index (1–20) selects the active prompt's config. For that prompt, each of its three LoRA slots (pN_lora1/2/3) is applied in sequence to the incoming model using its per-slot pN_strength_model / pN_strength_clip values. Per-slot pN_enabled1/2/3 toggles let you turn individual slots on or off without deleting their config - which is the right workflow when you're A/B testing "is this third LoRA hurting or helping." apply_lora is the master kill switch (off = model passes through untouched). num_prompts (1–20) controls how many prompt configs are configured and shown.

Outputs mirror the single-slot version: MODEL, CLIP (with optional clip input), selected LoRAs (STRING, the names applied), index (INT, the active prompt), trigger words (STRING, all active slots' triggers), and trigger + prompt (STRING - triggers plus optional base_prompt). The last one is the trick that makes the whole thing work end to end: wire base_prompt from your cycler and trigger + prompt goes straight into CLIP Text Encode, so each prompt gets its LoRAs and its trigger words in one shot.

The interface, which is genuinely nice

This node is where AUN's compact-mode design earns its keep. Double-click to compact; in compact mode each prompt's LoRA rows collapse to a drag-and-drop list where you can drag one LoRA label onto another to swap their values (LoRA, strengths, and triggers together - the pack's own docs describe exactly this), and per-slot enable toggles sit right there. Right-click to hide/show clip strength or the trigger-words footer. It's the most thoughtfully built node in the pack, and for a node with ~200 parameters, that thoughtfulness matters.

The setup loop

Wire prompt_index from a PromptCycler's index output (or any INT source), configure each prompt's LoRA slots with their strengths and triggers, run the queue - each prompt change pulls its matching LoRA set. This is the workflow the README's flagship example is built around.

Install

Part of AUN ComfyUI Nodes (loz2754):

cd ComfyUI/custom_nodes
git clone https://github.com/loz2754/AUN-ComfyUI-Nodes

or ComfyUI Manager → "AUN ComfyUI Nodes" → install → restart. Pack deps (piexif, opencv-python-headless, imageio-ffmpeg, requests) via Manager; LoRAs come from your existing models/loras folder.

The honest caveats

It's a lot of widgets - configure it in compact mode or you'll drown. Strength values are per-slot floats; if a prompt looks wrong, check you didn't leave a stale strength_clip at a value that fought the model strength. And remember triggers only reach your prompt if you use the trigger + prompt output - a static prompt string bypasses them entirely. If you only need one LoRA per prompt, the single Random LoRA Model Loader is the lighter pick.