Z-Engineer Prompt Enhancer (API)

This one genuinely calls an API - the name isn't a lie here, and it needs no key. The Z-Engineer Prompt Enhancer (API) is the old way to use Z-Image-Engineer: run the Qwen3-4B prompt model in an external OpenAI-compatible server (LM Studio, llama.cpp server, or Ollama) and let this node talk to it over HTTP. The README is upfront that it's the legacy path - the Z-Engineer Prompt Enhancer (Local) node does the same job in-process with the model you already loaded, no server required. So the honest question is: when do you want this one?

Its real niches: you already keep an external LLM server warm for other things, you want to use a different model than the one doing the encoding (say, a bigger or thinking model), or you've been running Z-Image-Engineer in LM Studio since December and this node just formalizes that workflow. If you're starting fresh and only doing Z-Image, use the Local node. If the setup is already there, this is a perfectly good string node.

How it works

It builds a chat request with the V6 system prompt (that long cinematography briefing is baked in as the system_prompt default), your seed as the user message, and your sampling parameters, then POSTs it to {api_url}/chat/completions. With model set to auto and discover_models on, it first queries /v1/models and picks the best Z-Image-Engineer match it can find. Then the same post-processing as the local node runs on the reply: strip reasoning, sanitize, re-append seed constraints, re-append keep_terms. Output is a prompt STRING, previewed on the node, ready to wire into CLIP Text Encode.

The inputs that matter

• api_url - defaults to http://localhost:1234/v1, which is LM Studio's port. Ollama is http://localhost:11434/v1. This is where people get burned, so check it before anything else.
• model - auto discovers from /v1/models; or type the exact model id the server reports to override its choice.
• error_mode - what happens when the server fails: return_input (passes your raw seed through untouched), return_error, or empty. return_input is the sane default for a queue that must not die mid-batch.
• keep_terms - comma-separated trigger words kept verbatim, re-appended if the model drops them. Same feature as the local node.
• timeout_seconds (120) and retries (1) are the resilience knobs. temperature 0.20, top_p 0.9, top_k 40, min_p 0.03, max_tokens 320 are the recommended V6 settings and the defaults.

Note there's no repetition_penalty here - the local node has it; the API path leaves repetition control to your server. That's a small reason the local node is the better default.

Install

The pack install is identical to the other nodes - Manager search "ComfyUI Z-Engineer," or git clone https://github.com/BennyDaBall930/ComfyUI-Z-Engineer into custom_nodes, pip install -r requirements.txt, restart. The pack's requirements.txt includes requests, which this node needs. The model is served by your server, not by ComfyUI - GGUF quants for LM Studio/Ollama come from BennyDaBall/Z-Image-Engineer-V6-GGUF, loaded into your LLM app's model folder.

Where people get burned

• Nothing comes back, and you're staring at your own seed. Check the port. LM Studio is 1234, Ollama is 11434, llama.cpp server exposes /v1. A dead server with return_input set will silently hand your seed back - useful, but confusing the first time.
• "model returned only reasoning_content." Your server loaded a thinking variant that spent the whole token budget reasoning. Raise max_tokens or load a non-think/no-think variant of the model.
• auto picks the wrong model. Discovery only sees what /v1/models reports; if your server's id isn't in the preferred list it falls back to the first available model. Type the exact id into model to fix it.
• Sanitization surprises. strip_reasoning and sanitize_output are on by default and will happily delete content the model put in <think> blocks. Leave them on for image prompts; turn sanitize_output off if you want to see the model's raw reply.

The bottom line: this is the comfortable path for people who already run a local LLM server, and a redundant one for everyone else. It works, it's well-behaved, and the same enhanced prompt lands in your CLIP Text Encode either way.