💫 Isulion Prompt Enhancer

Let me save you some time: the Isulion Prompt Enhancer does not rewrite your prompt, does not call an LLM, and does not sprinkle ten adjectives over everything. It appends one word or phrase. That's the whole node. And once you know that, it's actually useful in a way the flashier "enhancers" aren't.

It's one of the three enhancement nodes in the Isulion/ComfyUI_Isulion pack, and it sits in a nice spot: you've built a prompt from generator fragments, and you want a single quality/mood/lighting word on the end. That's a genuinely common SDXL habit - a focused boost term does more than a sentence of vague praise, because CLIP's attention budget is finite.

How it works

You give it a base_prompt string and it tacks one enhancement term onto the end with a comma. The term is picked from a small curated table by two axes: focus (what you're enhancing) and enhancement_level (how hard). So focus detail + level moderate gives you something like "highly detailed"; focus lighting + level dramatic gives "volumetric lighting" or "god rays". The seed input picks which synonym from the bucket you get.

The inputs that matter

• base_prompt - required. Whatever text you're enhancing. Can be wired from another node or typed.
• focus (optional) - detail, mood, composition, lighting, or color. This is the lever you'll actually use. Detail and lighting are the safe picks; color leans on words like "vibrant" and "saturated."
• enhancement_level (optional) - subtle, moderate (default), dramatic. Subtle appends "detailed" or "balanced"; dramatic appends "hyper-realistic" or "god rays."
• seed (optional) - which synonym you draw.

Output

One STRING: base_prompt, <enhancement>. Wire it into a CLIP Text Encode and you're done.

Installing it

One pack, all Isulion nodes:

1. ComfyUI Manager → Custom Nodes Manager → search Isulion → Install → Restart.
2. Or clone it manually:

cd "your_ComfyUI_install_dir/custom_nodes"
git clone https://github.com/Isulion/ComfyUI_Isulion

No requirements.txt, no models, no API calls. It's a random picker over a text table.

Common issues

Expectations are the main issue, so: don't feed it a full paragraph and expect a rewrite. It's an append-a-word node, and on long prompts one extra term is barely visible. The other trap is mixing levels and models - dramatic detail can drop "photo-realistic" or "hyper-realistic" into a prompt for an anime checkpoint, and that fights your model. Pick a level that matches the base style. On the model question: this is squarely aimed at CLIP-era checkpoints where "highly detailed" as a weighted tag still does something; on an LLM-encoded model like Flux the appended phrase is harmless but don't expect it to change much. And because the whole thing is seed-driven, the same prompt with seed 0 gives you a different synonym every run - which is usually exactly what you want for variations.