Multi-LoRA Loader (Model Only)

Somewhere between "one LoRA per loader node" and "a whole platform-specific loader family" there's this node: the generalist. Multi-LoRA Loader (Model Only) takes a MODEL and no CLIP, applies up to eight LoRAs to it, and returns a model plus a trigger-word-aware prompt. No CLIP in, no CLIP out - that's the whole point.

Why would you want that? Because plenty of modern models don't route their text conditioning through a CLIP object in a way you'd patch with the standard LoraLoader. Video models like Wan, and most diffusion models where you're only nudging the UNet/DiT weights, are exactly this case. This loader is the "if you don't know which one you need, or you have an odd model, this one" option of the pack - and it's also the most filter-heavy of them all.

How it works

Unlike the platform-specific siblings, which hardcode a directory filter, this one gives you the full search toolbox:

• search_directory and search_filename - comma-separated substring matches on paths and filenames (prefix a term with a dash to exclude, the classic pattern).
• search_category - filter by a category list (style, character, pose, clothing, effect…).
• search_trigger_word - match against trigger words from the pack's local LoRA database.
• filter_architecture - SD1.5, SDXL, Flux, Pony, Illustrious, Noobai, SD3.5, HiDream, Stable Cascade, and more. It sniffs architecture hints from filenames and paths and lets you cut the list down hard.
• min_rating - 0–5, from the DB's quality ratings.

Then the eight slots (lora_1_enable / lora_1_name / lora_1_strength, up to slot 8). Just model strength per slot - there's no CLIP strength here because there's no CLIP.

The trigger-word machinery is the same as the rest of the family: query_civitai fetches trigger words from the Civitai API when the local DB has none, force_civitai_fetch bypasses the cache, and trigger_position/trigger_separator control how injected words land in your prompt.

Outputs

• model - your patched model, straight into the sampler.
• prompt_with_triggers - prompt plus injected trigger words; feed to your encoder.
• prompt - your prompt back unchanged (useful if you only want the loader for the model patch).
• filter_info and filtered_loras_list - the reports that tell you what your filters actually matched.
• all_trigger_words and loaded_loras_info - every trigger found and a per-LoRA load summary.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/AAA_Metadata_System.git
cd AAA_Metadata_System
pip install -r requirements.txt

Restart ComfyUI (or install "AAA Metadata System" from ComfyUI Manager). No model downloads. CC BY-NC licensed for non-commercial use; commercial use needs a license from the author.

Common issues

• The dropdown lists everything, filters change nothing visible - expected. Filters apply at execution time, and the refresh_lists tooltip says it plainly: "Filters are applied during execution - check filter_info output to see results." Toggle refresh_lists to rescan the filesystem.
• filter_architecture hides nothing - architecture is guessed from names/paths; badly-named LoRAs fall through to "Unknown" and match Any only. Don't rely on it as a hard wall.
• No triggers injected - DB entry missing and query_civitai off. Enable it, or type the trigger yourself.

It's the least glamorous node in the family - no cute platform branding, no special handling - which is exactly why it's the one to reach for when your model doesn't fit the pre-made buckets. If your LoRAs live in a flat pile and you'd rather filter than reorganize folders, start here.