Nox Prompt Enhancer

Nox Prompt Enhancer is the polish node: you feed it a base prompt, pick a style pack, and it layers on emphasis and quality tags. On the surface that's the most familiar idea in the pack - it's the descendant of every "prompt enhancer" from the SD 1.5 days. The twist is that the old bag of tricks it draws from has a shelf life, and this node is where that shows most clearly. It's still worth using, but you have to know which model family you're talking to.

How it works

Read the source and the mechanism is transparent. enhancement_mode picks a style pack - artistic, photorealistic, cinematic, fantasy, sci-fi, or portrait - each adding a bundle of style keywords to your base prompt. add_quality_tags (default on) appends quality descriptors. custom_style is free-text if you want your own pack. secondary_prompt gets appended, and negative_prompt passes through to the negative output.

Then comes emphasis_strength. Here's the thing to understand: when it's above 1.0, the node wraps your prompt in parentheses - (prompt) - scaled to the strength, capped at five levels. That is the classic SD 1.5/SDXL weighting syntax. It does not wrap in (word:1.3) style numeric weights; it just adds literal parens. Which is exactly right for SDXL-lineage checkpoints, where emphasis via parentheses still works, and exactly wrong for LLM-encoded models.

The model-family caveat (read this before you use it)

The KB's prompt-engineering essay is blunt here: on modern LLM-encoded models (Flux, Wan 2.x, Z-Image, Anima), (word:1.3) attention weights are silently discarded, "masterpiece, best quality" is inert because the vocabulary was never trained in, and the negative prompt does little at CFG 1. So this node's two headline features - parenthesis emphasis and quality tags - are both aimed at the SDXL lineage. If your checkpoint is Illustrious, Pony, NoobAI, or SDXL, Enhancer works as advertised. If it's Wan 2.2 or Flux, the quality tags and parens are mostly noise; the enhancement_mode style vocabulary still adds descriptive words, but you'd get more from writing plain prose and skipping the node.

Inputs and outputs that matter

Required: base_prompt, enhancement_mode, emphasis_strength, add_quality_tags. For SDXL work, start with emphasis_strength at 1.0 and only nudge it up if the subject is getting lost. Outputs: enhanced_prompt, negative_prompt, preset_status. It has the standard preset_action/preset_name save-load, useful here because your style pack is a real configuration you'll want to reuse.

Installing it

Pack install, once. ComfyUI Manager: search "Nox Prompter", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/jamesk9526/NoxPrompter-ComfyUI-CustomNode

Restart, find the NoxPrompter category. Pure stdlib, no dependencies.

Gotchas

Beyond the model-family issue, two small things. First, emphasis_strength between 0.0 and 1.0 does nothing visible - the node only adds parens above 1.0. Second, the README's stale ComfyNoxPrompter URL/folder name again - use the repo above. And don't stack this node with the Builder's model_emphasis and then dump both into an LLM-encoded model expecting magic; you're just paying for a longer prompt.