AutopromptProcessor (CRT)

A lot of modern workflows pipe a prompt out of an LLM or an "autoprompt" helper and straight into the sampler. Which is great, until the model spits out phrasing your checkpoint has never seen, or a word that trips over your LoRA's trigger, or the same token appearing in three different forms. AutopromptProcessor sits between that text source and your text encoder and cleans house: it applies a list of word replacements and optionally prefixes a trigger word, then hands you a tidy STRING to encode.

The pitch is boring and that's the point. It's prompt hygiene as a reusable node instead of a regex you wrote once in a text editor and lost.

How it works

Three inputs, one pass of work:

• autoprompt (STRING) - the raw text to process.
• replacements - one rule per line in old_word|new_word format. Lines starting with # are treated as comments and skipped. Replacement is whole-word and case-insensitive, so dark|moody changes dark but not darkness, and catches Dark too. That whole-word behavior is deliberate - you don't want old|new mangling every word that merely contains the target.
• custom_prefix - text prepended to the result. This is the field designed for LoRA trigger words: add_detail, masterpiece in front of whatever came in.

Output is processed_text (STRING), ready for a CLIP Text Encode.

Order of operations: replacements run first, then the prefix is joined on. Empty prefix is skipped entirely, and if the processed text ends up empty the prefix alone comes out.

A practical example

Say your autoprompt helper writes "a photo of a woman with long brown hair, cinematic lighting". Your model is fine with that, but your style LoRA responds to a specific trigger and your checkpoint hates the word "photo". Set replacements to:

photo|portrait
woman|girl

and custom_prefix to lora_trigger_token,. Result: lora_trigger_token, a portrait of a girl with long brown hair, cinematic lighting. Took three seconds and now every run is consistent.

Installing it

CRT-Nodes again - ComfyUI Manager → CRT-Nodes, or clone + pip install -r requirements.txt, restart, and it lives under CRT/Text. No model downloads, no extra dependencies; it's pure string work.

Where people get burned

The replacement format is old|new with the pipe as separator, and there's no escaping - if your old or new text contains a | it'll split wrong. Keep rules to single tokens and it's fine. Also note it processes the text as one blob: no comma-splitting, no weighting. It replaces and prefixes, that's all. If you need per-segment rewriting or dynamic scheduling, you want the pack's Dynamic Prompt Scheduler instead - this node is the simple hammer.