Keyword Preset Selector

If the prompt contains X, output Y. That's the whole node - and that's the point.

AUNKeywordPresetSelector is a keyword-to-value lookup table. Feed it a reference_phrase, and it scans up to 20 keyword/preset rows top to bottom, returning the first preset whose keyword appears in the phrase as a substring (case-insensitive by default). No scoring, no machine learning, no API. Just "does this text contain that word, and if so, give me the thing I attached to it."

You reach for it the moment your workflow's behavior should follow what the prompt says. Your prompt cycler (AUNPromptCycler or one of the Text Index Switches) emits a label or index for the active prompt - run that label into reference_phrase and this node hands you a preset value: a style phrase to append, a LoRA trigger string, a CFG value or scheduler name as text. Or feed it a captured filename, a wildcard token, anything. Anywhere you catch yourself thinking "if it mentions anime, use preset A, else B," this node is that sentence.

It's the exact shape of node the AUN pack loves: tiny, deterministic plumbing that turns text analysis into a decision. It looks boring on the canvas. It quietly deletes a pile of index switches from your graph.

How it works

The matching engine is shared with its FaceID sibling in the same pack: substring match, case-insensitive by default, first match wins scanning top to bottom. Each keywordN is a comma-separated list - any one keyword matching activates the row. Empty rows are skipped silently. If nothing matches, you get preset_default and a matched_index of 0, which is your "fell through" flag. It returns NaN from IS_CHANGED, so it re-evaluates on every run; a changed phrase always gets re-scanned.

The inputs that matter:

• reference_phrase: the text scanned for keywords (substring match, not exact equality).
• preset_default: what comes out when nothing matches. Empty string if you leave it - set it, or you'll be chasing empty outputs.
• visible_inputs (2–20): how many keyword/preset rows are active. Then it's just rows of keywordN + presetN. presetN is multiline, so a "preset" can be a full prompt paragraph, not just a token.

Three outputs: selected_value (STRING - the winning preset), matched_keyword (the exact keyword that fired, empty on fallback), matched_index (INT - 1-based row, or 0 when nothing matched). Wire selected_value into a CLIP Text Encode, a filename builder, or AUN Show Any Multi so you can see which preset won at a glance. matched_index feeds any of the pack's index-driven nodes if you want the choice to drive more than one thing.

A quick pattern:

AUNPromptCycler (label) ──▶ AUNKeywordPresetSelector (reference_phrase)
                                 └─ selected_value ──▶ CLIP Text Encode (append style)
                                 └─ matched_index  ──▶ AUN Show Any Multi

Install

cd ComfyUI/custom_nodes
git clone https://github.com/loz2754/AUN-ComfyUI-Nodes

...or ComfyUI Manager → search "AUN ComfyUI Nodes" → install → restart. Pure Python, no extra dependencies, no model downloads, no API key. The pack's requirements.txt (piexif, opencv-headless, imageio-ffmpeg) exists for the rest of the collection; this node doesn't touch any of it.

Gotchas

• Substring matching surprises. anime matches animesque, face matches interface. Since first match wins, put the longest, most specific keywords first - order is your only tiebreaker, so row order is load-bearing.
• Empty rows are skipped, not errors. A half-configured slot just never matches, and you get no warning. Check matched_index if a preset never seems to fire.
• The 0 fallback is your friend. Because selected_value can't tell you whether you fell through to preset_default, wire matched_index somewhere visible when you're debugging a row order.