🚀 强力lora加载器 Magic Power LoRA Loader

If you have more than ten LoRAs, the stock LoRA loader has already betrayed you. Every model is one line in a dropdown, and there's no way to remember which one was the "good detail" LoRA versus the "hands" LoRA without clicking through all of them. Magic Power LoRA Loader is the fix: a visual gallery where each LoRA shows its preview image, lives in a folder you defined, has tags, and gets enabled, disabled, weighted, and reordered by dragging. It's the LoRA manager that should have been core.

The comparison the pack is chasing is the difference between a file list and a library. Folder organization, drag-to-reorder, preview images auto-detected from a magicloradate subfolder or the LoRA's own directory, tags for search, batch enable/disable, and one-click toggles for whole folders. The UI is where you'll spend your time; the node itself is downstream of it.

How it works

The lora_stack input is a string (JSON-ish list) that the front-end builds as you click LoRAs in the gallery - you're not meant to edit it by hand. When the node runs, it applies each LoRA through ComfyUI's official comfy.sd.load_lora_for_models path and caches the loaded weights per (path, weight) so repeated runs don't re-read files off disk. That's the standard-loading path, and it's the one that now auto-handles INT8/FP8/INT4-quantized models: an older version had a dedicated INT8 mode, and the V1.3.9 changelog deleted it because ComfyUI's core model_patcher.py learned to do it natively. If you see an int8_mode field in an old workflow, it's ignored.

Beyond that default path there are modes you flip in the settings popup - Klein mode for the Nunchaku FLUX.2 Klein loader, SDNQ mode for SDNQ-quantized models, and an adaptive mode that tries to detect the model type automatically. The newest one (V1.4.1) is Anima 2.9B mode, which remaps legacy 28-layer Anima LoRAs onto the 40-layer Anima 2.9B in memory - no disk writes, no copies, and non-Anima LoRAs pass through untouched.

The chain feature

This is the one that trips people up, so read the changelog's explanation: a node can output a lora串 (LoRA chain) instead of loading anything itself. A loader that outputs a chain doesn't apply its LoRAs - it just passes its list downstream. Only the final loader in the chain, the one that isn't outputting a chain, loads everything it received. Chain nodes don't even need model/clip connected; the chain-end does. It's built for SDNQ-style "unload the model once, apply everything, load once" workflows, and it survives bypass correctly. If your LoRAs mysteriously aren't applying, check whether you're mid-chain.

The inputs and outputs that matter

• lora_stack (STRING) - the list the gallery builds; leave it alone
• model / clip - connect the checkpoint, get them back LoRA'd
• lora串接受 (MAGIC_LORA_CHAIN) - the chain input for chaining

Outputs are model, clip, lora_preview (an IMAGE list - this is what shows in the gallery), tags_output (STRING, the LoRA trigger-word/tag set), and lora串输出 for chaining.

Installing

It's part of the Magic-Assistant pack:

cd ComfyUI/custom_nodes/
git clone https://github.com/shigjfg/ComfyUI-Magic-Assistant.git
cd ComfyUI-Magic-Assistant && pip install -r requirements.txt

Restart and find it under "✨ Magic Assistant". No models to download - the gallery reads your existing models/loras/. If previews are missing, put a same-named .png (or use the magicloradate subfolder) next to each LoRA; that's the convention the node looks for. Weight controls default to compact arrow buttons with a clickable value; the settings offer a slider mode if that's more your style.