一体节点 · 内部触发词

What it does, in one sentence

It reads your LoRA list and emits the trigger words of the enabled ones as a single comma-separated string.

That's it. 一体节点 · 内部触发词 has one required input, config, and one output, 触发词, of type STRING. No model, no CLIP, no parsing of your prompt. It doesn't know what a token is. It's a formatter that turns six LoRA cards with a phrase each into ohwx person, film grain style, ch9ractername.

Why such a node exists

Because in vanilla ComfyUI a trigger word is folklore. The LoRA file carries its weights; the phrase it was trained with lives in a Civitai description, an _readme you never open, or your memory. Miss it and the LoRA does almost nothing - you conclude the LoRA is bad and go download another one. This is the single most common LoRA complaint in the wild, and it's a bookkeeping failure, not a training failure.

This pack solves it by giving every entry in its LoRA list a trigger field, saved into the workflow alongside the file and the strengths. The triggers stage then collects those fields and hands you one string you can paste anywhere. In the full console flow you never see this node at all - the console injects it, takes its output, and appends that string to the positive prompt as upstream "extra" text before the prompt stage merges everything.

The exact rules

It reads the same version-1 LoRA snapshot the LoRA stage does and walks entries in order. An entry contributes if it's enabled and has a non-empty trigger after trimming. Duplicates collapse - two LoRAs that both want film grain give you one phrase, first one wins. The result is joined with ", ".

Two consequences worth knowing. First, disabling a LoRA removes its trigger, so a character LoRA switched off in the console quietly drops ohwx person from your prompt - usually what you want, occasionally a surprise when two characters share a base likeness. Second, the order follows the LoRA list, which is also the order the LoRAs patch in, so the string reads the way your stack is built.

One nice interaction: because the prompt stage deduplicates normalized terms when merging, typing the trigger phrase yourself and having it emitted here doesn't double it up. You get it once.

When you'd reach for it standalone

It's registered under 创界/临时测试/内部, so you can add it from the node menu and feed a CLIPTextEncode alongside your own text. Handy if you keep a config snapshot and want the phrase to follow your LoRA list automatically instead of being retyped in three places. Build the config by hand like this:

{"version": 1, "entries": [
  {"file": "characters/ohwx_v3.safetensors", "enabled": true,
   "strength_model": 0.8, "strength_clip": 0.8,
   "trigger": "ohwx person", "name": "ohwx"}
]}

Trailing junk is refused - version must be exactly 1, enabled must be a real boolean, and file must be a relative path inside models/loras. A malformed config gives 节点配置版本无效（需要 version: 1） or a filename error, not a silent empty string.

And keep the expectations calibrated: a trigger word is a rare token the LoRA was trained against. If a LoRA has no meaningful trigger - many style LoRAs don't - leaving the field blank is correct, and this node will simply skip that entry. Nothing here will find triggers for you; it only moves the ones you've recorded.

Install

Part of the pack, zero extra dependencies (no torch, no models, no GGUF).

cd ComfyUI/custom_nodes
git clone https://github.com/lefango/ComfyUI-ChuangJie.git

or install ComfyUI-ChuangJie through ComfyUI Manager and restart.

Note that this node has no panel of its own - no custom UI is registered for it, so it shows as a plain config text box, and that's expected. The triggers come from the LoRA cards inside the console, not from this node's widget. If the output is empty, either no LoRA in the list is enabled, or none of them has a trigger recorded. Fix it in the console's LoRA editor, not here.