Multi-LoRA Loader v02 (Filtered)

If your model speaks CLIP - SD1.5, SDXL, Pony, Illustrious, that whole world - and you want to stack LoRAs with real filtering, this is the loader to grab. Multi-LoRA Loader v02 (Filtered) is the CLIP-aware generalist of the AAA Metadata System pack: one node, a model plus a clip in, up to eight LoRAs applied with separate model and CLIP strengths, and a search/filter panel that's the most complete in the family.

The "Filtered" in the display name is doing real work. Where the platform-specific loaders assume your LoRAs live in tidy subfolders, v02 assumes nothing - it lists everything and lets you carve the list down with filters. It's the closest thing to "Multi-LoRA Loader Model Only, but with CLIP," and if you run SDXL or Flux-class CLIP models it's arguably the one you should just use.

How it works

It scans your full LoRA library and gives you the master search controls:

• search_directory and search_filename - comma-separated substring matches on paths and filenames.
• search_category - style, character, pose, clothing, effect, and two dozen more.
• search_trigger_word - match against trigger words in the pack's local LoRA database.
• filter_architecture - SD1.5, SDXL, Flux, Pony, Illustrious, Noobai, SD3.5, HiDream, Stable Cascade and others, guessed from filenames and paths.
• min_rating - 0–5, from the database's quality ratings.

Then eight slots, each with lora_N_enable, lora_N_name, lora_N_strength and lora_N_clip_strength. The CLIP strength matters for models where the text encoder is trained alongside - on SDXL/Pony a clip strength under 1.0 is often what keeps a style LoRA from overpowering the prompt.

Trigger words work like the rest of the family: query_civitai pulls them from the Civitai API when the DB has none, force_civitai_fetch bypasses the cache, and trigger_position/trigger_separator control injection into your prompt.

Outputs

• model and clip - both patched, into your conditioning and sampler.
• prompt_with_triggers - your prompt with trigger words injected.
• prompt - your original prompt, passed through.
• filter_info / filtered_loras_list - what your filters matched, in text form.
• all_trigger_words / loaded_loras_info - all triggers found and per-LoRA load details.

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" via ComfyUI Manager. No model downloads. License: CC BY-NC non-commercial; commercial use needs a paid license from the author.

Common issues

• Filters don't change the dropdown - expected, and the refresh_lists tooltip says so: "Filters are applied during execution - check filter_info output to see results." The dropdown is fixed at node creation; toggle refresh_lists to rescan the folder.
• filter_architecture feels unreliable - it's a name/path heuristic; badly named files land in "Unknown" and only match Any. Treat it as a convenience, not a guarantee.
• LoRA effects too strong even at low model strength - drop clip_strength too; on CLIP models the text encoder is often where style leaks in.

There are slicker LoRA managers out there with prettier UIs, but v02 is the solid workhorse: filters that don't require you to reorganize your folder tree, eight slots, and trigger-word help if you want it. For daily SDXL/Pony stacking, it earns its place in a default graph.