Z-Engineer Prompt Enhancer (Local)

Here's the thing that makes ComfyUI-Z-Engineer more than "a GGUF loader with a longer name." Z-Image is a prose model - feed it a sentence and it renders a scene, not a tag soup. But prose quality is exactly where most people's "cute girl, neon, rain, 8k" seeds fall apart, because the encoder is a 4B LLM and like any LLM it does a lot better when the prompt is well-written. The Z-Engineer Prompt Enhancer (Local) is a prompt-writing model that runs on the same Qwen3-4B weights you already loaded as your text encoder. No external server, no API key, no second copy of the model in VRAM.

Drop a raw seed into input_prompt, queue, and it returns a polished 180–250 word natural-language prompt, previewed right on the node. That's the whole loop: it writes the prompt, and the same model encodes it.

How it works

You feed it the clip output from one of the Z-Engineer CLIP loaders (GGUF or Safetensors/Shards). The node wraps your seed in the Qwen chat template, calls the model's generate() with the sampling knobs, decodes the result, then runs a post-processing pass that's honestly the most thoughtful part: it strips <think> reasoning blocks, chatml tags, negative-prompt sections and camera-brand names, re-appends seed constraints (counts, colors, quoted text) the model dropped, and re-appends anything in keep_terms. Because the model is already resident as your text encoder, ComfyUI's model management treats it as one model - no VRAM double-billing, and it unloads normally.

The inputs that matter

Most of the required fields are sampling knobs with sensible defaults. You'll actually touch:

• clip - must come from a Z-Engineer loader (or any Z-Image Qwen3-4B CLIP). Feed it a stock CLIP and it dies with "this CLIP does not support text generation."
• input_prompt - your raw seed. Multiline, and each box is vertically resizable via the grip in its corner (double-click the grip to reset).
• keep_terms - comma-separated LoRA trigger words or phrases that must survive verbatim (m4rty style, neon glow). The model is told to weave them in unchanged, and anything it still drops gets deterministically re-appended. This is the one input I'd genuinely call essential if you run LoRAs.
• enforce_seed_terms / strip_reasoning / sanitize_output - the safety nets. Leave them on.

The defaults are the README's recommended V6 settings: temperature 0.20, top_p 0.9, top_k 40, min_p 0.03, repetition_penalty 1.05, max_tokens 320. They're well-chosen; don't fight them until you've seen output.

Output is a single prompt STRING, and there's a batch_mode + batch_separator (default \n---\n) if you want to process several seeds in one call.

The workflow

[Z-Engineer CLIP Loader (GGUF)] ──clip──┬──> [Z-Engineer Prompt Enhancer (Local)] ──prompt──> [CLIP Text Encode] ──> ...
                                        └──────────────────────clip──────────────────────────────^

The same clip feeds both the enhancer and CLIP Text Encode. One model, two jobs.

Install

Pack install is the usual: Manager search "ComfyUI Z-Engineer," or git clone https://github.com/BennyDaBall930/ComfyUI-Z-Engineer into custom_nodes, pip install -r requirements.txt, restart. Then load the model with either Z-Engineer loader - Q4_K_M GGUF for low VRAM (~3–4 GB), sharded FP16 for full fidelity (~9 GB). Requires ComfyUI v0.3.75+ with native Z-Image support.

Where people get burned

• The "does not support text generation" error means the CLIP didn't come from a Z-Engineer loader. Re-wire it.
• Samey output. Z-Image Turbo is already notorious for near-identical results across seeds, and a prompt enhancer can compound it. Change the seed, nudge temperature toward 0.3–0.4, and check keep_terms if a trigger isn't showing up.
• Nothing happens on an empty seed. An empty input_prompt returns empty by design. Type something first.
• The model has favorite phrases - earlier Z-Image-Engineer versions were heavy on "weathered wood" and "ethereal" people. If every result starts smelling the same, that's the seed/temperature dial, not a broken node.