LoRA Suggester

You have a thousand LoRAs and you use nine, because the rest are named v4_final_fix_02 and you have no idea what they do.

EricLoRASuggester goes further: it decides which LoRAs belong in this prompt and hands your loader a stack.

The catch comes first, because it's the whole story: the node calls no AI. No API, no key, no latency. The thinking happened earlier, in the rest of the pack - a local vision model wrote a card for every LoRA (what it does, its role, triggers, good weights, NSFW level) and you correct the wrong ones in a browser editor. No cards, no suggestions: it can't pick a LoRA you never cataloged. Cards get written once.

How it actually decides

Every carded LoRA is scored for the role it plays, and that list is the genuinely useful part: style (the base look), treatment (linework, halftone, grain, colour grade - layered on a style), transform, detailer, environment, lighting, subject, layout, utility (speed LoRAs), slider.

Scoring is plain text matching, not vibes: prompt words against the card's facets, weighted per role, plus similarity against the card's name and triggers, squashed so the threshold means something. Then your own signals multiply in - stars, LoRA Manager favourites, card confidence, a Civitai popularity nudge weaker than one star. Your taste wins.

Winners get per-stage weights (S1 composition / S2 refine / S3 detail): detailers sit at zero early and full late, environments the reverse. Two picks both claiming composition or faces, and the weaker drops.

Sliders are never scored on text, because they're dials: the prompt is read for its position on axes ("wasp waist" lands near the negative end of a waist axis) and the slider is set to match. slider_min_strength decides how loudly it must be asked for before a dial moves; leave that at 2.

The inputs you'll actually touch

• platform - the gate. Only LoRAs on that platform's shelf (or matching its Civitai base model) count; default krea2. Pick an unshelved platform and the report is full of nothing - which looks like a broken node, but isn't.
• mode - suggest by default: nothing loads, you just read the report. It's the best design call in here, so live with it for a week before switching to apply.
• max_loras (4) and the per-role caps (style_max through slider_max). Set a cap to 0 and that role never loads - the cheapest way to stop a detailer sneaking into a clean illustration. apply_threshold (0.35) is how adventurous you are; weight_budget (2.5) caps total Stage-1 weight, scaling the accents down to fit.
• nsfw is a cap, not a switch: NSFW LoRAs load only when the prompt itself asks, and off still allows suggestive styles. variety + seed let close runners-up trade places as the seed moves.
• exclude / pin - one entry per line, with folder:, role: and axis: prefixes. Pins skip the threshold, role caps and NSFW gate, which is how you stop arguing with it about your favourite LoRA.

Outputs: suggestions (picks with stage weights; empty in suggest mode) goes to the Eric Krea2 Multi-LoRA Stack's suggestions input. report tells you what won, why, the runner-up per role, and what an exclude entry killed - the thing that makes this usable. triggers and prompt_with_triggers cover the trigger step (skip the latter if a rewriter runs after). lora_text emits <lora:name:weight> for loaders outside the Eric_Krea2 family.

Installing it

You're installing a catalog pipeline that contains a node. Manager finds it under the pack title, or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_LoRA_Catalog
# ComfyUI portable: ..\..\python_embeded\python.exe -m pip install -r Eric_LoRA_Catalog\requirements.txt

The Python deps are trivial (requests, pyyaml, Pillow), so the real prerequisites are elsewhere: Python 3.12+ (the pack declines to load on older ones rather than half-working), and an OpenAI-compatible local server - LM Studio or Ollama - with a vision model for the cards, e.g. Qwen3-VL 8B at Q4 for 6–8 GB of VRAM. A Qwen3 text encoder file in your workflows is not a chat server and won't work here. Then sidebar → Settings: LoRA folders, server, models, optional Civitai key, Save, Test LLM, Refresh catalog. Data goes to ComfyUI/user/lora_catalog/, so updating the pack won't eat your card edits.

Where people get burned

Test LLM passes and every card times out. The README flags it because everyone hits it: the test sends one tiny image, while a real card carries two previews, 7,000 characters of evidence and needs ~8,192 tokens of context. Either ComfyUI is holding the GPU so the server spills to CPU (unload its models, or refresh with ComfyUI closed), or the context is too small - set OLLAMA_CONTEXT_LENGTH=16384 and restart Ollama, or load at --context-length 16384 in LM Studio.

Also: the first card pass is thousands of vision-model captions, so run it overnight. And the node is new - it moved here from EricRollei's video_prompter on 2026-10-11, same node id and input order so saved workflows load unchanged, but there's essentially no field experience with it yet. License is CC BY-NC 4.0: free locally, commercial use needs a license.