Prompt Enhance (LLM)

You type a fox in the snow, the node hands a cloud LLM that sentence, and a few seconds later you get a dense, cinematic paragraph about a red fox mid-stride through blue dusk light with a frost-glazed forest behind it. That's the whole pitch of Prompt Enhance (LLM): outsource prompt-writing to an actual language model and feed the result to your text encoder.

The name slightly oversells nothing - it does exactly what it says, and it's refreshingly simple about it. It makes no local model call and downloads nothing; it's a thin HTTP wrapper around an OpenAI-compatible chat-completions API, DeepSeek by default. If you've ever watched an LLM describe a scene and thought "why am I writing prompts by hand," this is that thought turned into a node.

Why you'd reach for it

LLM-assisted prompting stopped being a novelty and became a normal workflow step - corpus chatter about prompt enhancers has grown ~20x since 2023, and it makes sense mechanically. Newer checkpoints use LLM text encoders that read your prompt as an instruction, so having an LLM write that instruction is two systems speaking the same language. For tag-based SDXL models it works too, as long as you tell the LLM to emit comma-separated tags. Either way, one sentence in, a structured paragraph out.

The tradeoff to swallow first: this needs an API key and internet. It's not local, it costs fractions of a cent per call with DeepSeek, and a commercial API will refuse NSFW prompts - which is exactly why some people prefer local wrappers like comfyui-ollama instead. If you're fine with a couple of pennies and safe content, this is the least-fiddly option in its lane.

How it works

The node takes your prompt, slots it into a prompt_template (look for the {prompt} placeholder - if your template doesn't contain one, the code just appends your text), and POSTs the whole thing to api_endpoint with your api_key as a Bearer token. It pulls choices[0].message.content out of the JSON reply and hands it to you as a string. One notable quirk: the default template is written in Chinese. It works - DeepSeek is a Chinese model and responds fine - but the phrasing it produces leans long, and the LLM may answer in whatever language it decides. Swap in your own English template via prompt_template if that bothers you.

The inputs that matter

Four of these are required:

• prompt - your raw idea. Multiline.
• api_endpoint - defaults to DeepSeek's /v1/chat/completions. Swap it for https://api.openai.com/v1/chat/completions or an Azure endpoint when you change providers.
• api_key - paste from your provider dashboard. This is the field that bites (see below).
• model - deepseek-chat by default; gpt-4 and friends work on OpenAI.

The optional trio is self-explanatory: temperature (0–2, default 0.7 - push toward 1 for wilder rewrites, down for strictness), max_tokens (default 2048 is plenty for a prompt), and timeout (default 60s).

The one output is enhanced_prompt - a STRING. Wire it straight into your CLIP Text Encode positive input and generate. That's the whole integration.

Installing it

Install from ComfyUI Manager by searching "ComfyUI-PromptEnhance", or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/xiangyuT/ComfyUI-PromptEnhance
# then restart ComfyUI

The README suggests pip install requests, but ComfyUI already ships requests, so there's no real dependency hunt and no model files to fetch. It's about as lightweight as a custom node gets.

Where people get burned

• The API key lives in the workflow. It's a plain widget input, so it's saved into your workflow JSON in cleartext. Fine locally; a disaster if you share a workflow on a public pastebin. Rotate the key if you've ever shared one with the key filled in.
• No key, no run. The node raises "API key is required" and turns red rather than silently returning garbage. Annoying, but honest.
• It's synchronous and non-streaming. The node blocks that queue item while the API thinks; a flaky endpoint past timeout gives you a TimeoutError in the node.
• The template is Chinese out of the box. Totally functional, mildly surprising on a first run. Replace prompt_template if you want English-only output and tighter phrasing.

For a $2 experiment with DeepSeek, this is the easiest "let an LLM write my prompts" node you'll find. Start with the base version; reach for the Advanced sibling only when you actually want the extra controls.