LoRA Tester v03

Testing LoRAs is usually a grind: swap the file, re-run, compare, swap again. LoRA Tester v03 turns that into a search-and-select problem. It scans your LoRA folders, filters by filename, architecture, category, and trigger words, picks one (or a random one), applies it to your model and CLIP, and can even insert the trigger words into your prompt for you. It's the highest-leverage node in the AAA Metadata System pack's LoRA family, and it's the one I'd reach for when I have forty style LoRAs and no idea which ones are actually good.

The killer feature is the filtering. Instead of scrolling a dropdown, you type comma-separated search terms - and -term to exclude. So -pony,-inpaint in file search terms skips everything tagged for those variants. Combined with seed (0 = pick a random matching LoRA), it becomes a lottery machine for your own library: re-run and see what different LoRAs do to the same prompt.

How it works

The node maintains a local catalog of your LoRAs in lora_tester_db.json, recording each file's architecture, category, and trigger words. On each run it filters that catalog by your criteria, picks the LoRA at seed (or a random one), and in direct_load mode applies it to the MODEL and CLIP you feed in. In info_only mode it skips loading entirely and just hands you the info - handy for building selection UIs. If a LoRA has no trigger words recorded, query_civitai will hit the Civitai API to fetch them (that's a network call, so keep it off unless you need it). It also gathers "associated images" - same-basename previews sitting next to your LoRA files - into a gallery.

The prompt_placement setting decides whether trigger words get spliced into the beginning or end of your input_prompt, or not at all. If you're testing whether a LoRA does anything, wire the combined prompt and the images together and compare.

Inputs and outputs that matter

• dir_search_terms / file_search_terms / additional_path - how you find the LoRAs. Comma-separated, -term to exclude, and additional_path for folders outside the standard ones.
• architecture and category - filters against the catalog. "Any" if you don't care.
• mode - direct_load (needs model and clip wired in) or info_only.
• model_strength / clip_strength - the multipliers, default 0.8 / 1.0.
• seed - 0 = random, 1 = first in the filtered list.
• query_civitai / force_civitai_fetch - the network options. Leave off unless a LoRA is missing triggers.

Outputs: model, clip (the modified ones in direct_load), lora_name, lora_path, trigger_words_all, trigger_words_selected, combined_prompt, lora_list, total_loras, selected_loras, associated_images, gallery_images, and the two strength values used.

Installing it

Part of AAA Metadata System by Eric Hiss (GitHub: EricRollei). Via ComfyUI Manager (search "AAA Metadata System") or:

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. requests (for the Civitai calls) is in the core requirements.

Common issues

If nothing ever matches, the catalog is empty or the tags are wrong - run it once with empty filters so it scans and indexes, then check lora_list / total_loras. If direct_load mode errors, you forgot to wire model and clip in; that mode genuinely needs them. Random selection (seed 0) can feel broken when you "get the same one twice" - that's the RNG, not a bug; bump the seed. And remember Civitai queries are real network requests: on a fresh folder with query_civitai on, the first run can be slow while it fetches trigger words for everything.