H3 Rewrite Clean

A chat model doesn't hand you a clean prompt. It hands you chat - a preamble, code fences, maybe a "Here's your enhanced prompt:" flourish - and all of it bleeds straight into your conditioning if nothing strips it. H3 Rewrite Clean is the counterpart to H3 Rewrite Prompt: take the raw text straight out of Generate Text, get back the production prompt.

How it works

One input, generated_text (straight from Generate Text), and four outputs:

• prompt - the trimmed production prompt, chat tokens, fences and preamble gone.
• integrated_multimodal_description - the first of the rewriter's three split fields.
• overall_soundscape - the second.
• non_diegetic_music - the third.

The rewriter LoRA was trained to answer in three parts, and this node splits them out so you can use them separately - feed the whole thing to H3 Encode, or grab just the soundscape for an audio-focused pass. Text it doesn't recognise passes through unchanged, which is the sane failure mode: you never lose the rewriter's answer because the cleaner got confused about its format.

Why you need it at all

The failure mode it prevents is the classic one from prompt-enhancer history: an LLM's conversational scaffolding becomes literal prompt tokens, which is worse than no enhancer at all. H3's text encoder is a language model reading an instruction, so a stray assistant: or a markdown fence isn't just ugly - it's conditioning the model. If you've ever watched a "prompt enhancement" workflow produce gibberish for no visible reason, this is the step that was missing.

Install

Part of the NynxzH3 pack - ComfyUI Manager → "Nynxz H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzH3

Restart. Pure Python, no dependencies, ComfyUI 0.30.0+. It's just string handling - nothing to download, no model files.

The one-liner

Wire generated_text from Generate Text, send prompt to H3 Encode, and keep the three split fields for when you want to touch the soundscape separately. It's a small node, but it's the difference between a rewriter that works and a rewriter that poisons its own output.