Prompt Enhance Advanced (LLM)

If you've got the plain Prompt Enhance node working, Advanced is the same thing with two extra inputs and one extra output - the kind of "just give me the rest of the controls" upgrade you install before you're sure you need it. It ships in the same pack, calls the same OpenAI-compatible API (DeepSeek by default), and it's still just an HTTP request plus a text string. No models, no local LLM, no heavy dependencies.

What it adds

system_prompt is the headline feature. The base node packs all its instructions into the single user message; Advanced sends an optional system message ahead of it. That's the standard way to steer a chat model's behavior across a whole conversation, and here it's your escape hatch from the author's baked-in template. Want the LLM to only ever answer in comma-separated Danbooru tags for your Illustrious checkpoint? Put that in system_prompt and leave the base template alone. Want a strict one-paragraph structure? Same place. It's the difference between using someone else's prompt philosophy and enforcing your own.

top_p is the other add. It's the nucleus-sampling knob that sits alongside temperature - default 1.0 means "use the whole probability distribution," and dropping it makes output more conservative. If you're tuning, move temperature first; top_p is the finer tool for when temperature alone gets you rambling text.

The genuinely useful extra is on the output side. Advanced returns two strings: enhanced_prompt and original_prompt - your input, passed through untouched. That's handy for two things. You can wire enhanced_prompt into your CLIP Text Encode while keeping the original available for a text-display node or a --prompt reference elsewhere in the graph. Or feed both into a comparison/Show Text setup so you can see exactly what the LLM did to your sentence. It's a small thing that saves you from adding a Reroute to hold onto the source string yourself.

The inputs you'll actually touch

Everything from the base node carries over: required prompt, api_endpoint, api_key, model; optional prompt_template, temperature, max_tokens, timeout. Then the two newcomers:

• system_prompt - multiline, empty by default. Leave it empty and this node behaves identically to the base one, minus the original-output convenience.
• top_p - 0–1.0 slider, default 1.0. Rarely needs touching.

The mechanism is otherwise identical: your prompt gets slotted into the template (again, {prompt} placeholder or it gets appended), the whole thing POSTs to the chat-completions endpoint, and the reply's first choice becomes enhanced_prompt. Adding top_p just means it's sent in the request payload like temperature already was. Same synchronous call, same 60-second default timeout, same red node if it fails.

Installing and running it

Identical to the base node - one pack, two classes:

cd ComfyUI/custom_nodes
git clone https://github.com/xiangyuT/ComfyUI-PromptEnhance
# restart ComfyUI, then find "Prompt Enhance Advanced (LLM)" under the prompt category

ComfyUI Manager's search also finds it by the pack title. The only dependency is requests, which ComfyUI already has, so there's nothing else to fetch.

Gotchas worth knowing

The base node's quirks all apply here, so don't re-learn them: your api_key sits in the workflow JSON in cleartext (don't share workflows with the key filled in), a missing key raises "API key is required," and the default template is in Chinese - replace prompt_template if you want English-only output. One extra trap specific to Advanced: because system_prompt is a separate field, it's easy to set a system prompt that contradicts the default template's instructions, and the template usually wins arguments because it's in the user message. If your system rules seem ignored, that's why - either keep them consistent with the template or replace the template too.

For most people the base node is enough. Advanced earns its keep if you're opinionated about how the LLM should phrase things, or if you're building a workflow where keeping the original prompt in the graph matters. It's the same simple tool with the training wheels off.