🧬 LoRA Prompter

The most common LoRA failure is a typo

"Trigger word missing" is item two on every LoRA troubleshooting checklist, right under "you loaded an SDXL LoRA onto a Flux checkpoint." Trigger words are deliberately weird - ohwx, ch9ractername, styl3name - trained as rare tokens precisely so they don't collide with normal vocabulary. Which also means you will never remember them, and you will mistype them, and the image will come out looking like the LoRA did nothing at all.

🧬 LoRA Prompter is a fix for that specific, small, extremely common annoyance. It is not a conditioning node, it doesn't touch your model, and it makes no API calls.

How it works (the honest version)

The backend is three lines. Given prompt and triggers, it returns prompt. That's the whole execute - you can check it yourself:

grep -n "NodeOutput(prompt)" ComfyUI/custom_nodes/FiL_Design_ImageMind/nodes/node_lora_prompter.py

All the work happens in the pack's Vue panel, which is why the node feels smarter than its Python. It parses whatever is in triggers - split on commas, newlines or semicolons, deduplicated case-insensitively - and renders one clickable chip per word. Click a chip and the word is inserted at the front of your prompt (trigger, your existing prompt). Click it again and it's excised cleanly, with the spare commas tidied up rather than left as , ,. There are add-all and clear-all actions when you've got a long trigger list.

The nice trick: if the triggers field itself is empty, the panel walks the graph, finds whatever node is wired into the triggers input, and reads that node's own triggers or text widget value directly. So the chips populate off the link even when the value hasn't been resolved into this node.

Inputs and output

Two optional strings, one output. prompt is your base text - multiline, and the chips edit it in place. triggers takes a trigger-word string from 🧬 LoRA Loader's trigger output, or any comma-separated list you type. Both are optional, so the node happily runs empty.

The single output is prompt (STRING). Wire it into your CLIP Text Encode's text field. That's it - no second downstream consumer to think about.

Reach for it when you're running the pack's LoRA Loader, which cycles adapters on each queue and emits the trigger words for whichever one is loaded. The chips give you a menu of what's actually in the adapter you're about to generate with, instead of a wall of text you keep re-typing. If you're not cycling LoRAs, it's still a decent place to park trigger words you use constantly - just be aware it's an editor with a text output, not automation.

Installing it

ComfyUI Manager → search FiL_Design_ImageMind → install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/FiL-Design-Ai/FiL_Design_ImageMind.git
pip install -r FiL_Design_ImageMind/requirements.txt

On a portable install, use its bundled interpreter (python_embeded\python.exe -m pip install -r ...) or run install_requirements.bat in the pack folder. Restart ComfyUI and look under 🎨 FiL Design/🧰 Tools.

No model downloads, no GPU work, no API key - everything the README says about provider keys and cloud vision models is for the pack's LLM nodes, and none of it applies here. Two requirements that do matter: ComfyUI 0.3.60+ (the pack is built on the V3 node API, so an older core won't register any of it), and the pack's bundled frontend, which ships pre-built in frontend/dist so you never need Node.js.

Where it bites

No chips showing up. Either triggers is empty and nothing is wired into it, or the upstream node has no triggers/text widget with content. Type cat, dog into the field to confirm the panel is alive, then chase the wire.

The prompt and triggers widgets aren't where you expect. The pack hides the native widgets and replaces them with the panel - that's the design, not a bug. Both fields can still be wired as inputs.

Stale triggers after a cycle. This is the one that'll get you. When the LoRA Loader steps to a different adapter, the chip list changes, but the text already in your prompt does not - the previous adapter's trigger word is still sitting there doing nothing (or worse, nudging the model toward a concept it half-recognises). Glance at the chips after a cycle and clear what no longer lights up. Nothing auto-removes triggers for you.

If you wire into prompt, that's what downstream sees. The panel's toggles write into the node's own prompt value; a connected link wins over anything typed in the panel. Compose in the panel, or drive it from the graph - not both at once.