Z-Image Prompt Generator 🐑

ZImagePromptGenerator takes a short idea and has an LLM rewrite it into a fuller prompt tuned for Z-Image, Alibaba's small-but-mighty image model. It's a prompt front-end, not a generator: it calls whatever LLM you connect and shapes the output for how Z-Image likes to be prompted.

It ships in ComfyUI-MieNodes, MieMieeeee's utility pack, in the prompt-generator group.

Why Z-Image benefits from this

Z-Image is the story of 2026 in local image gen: a 6B model that put near-Flux quality on a 12GB card, uncensored and Apache 2.0, and got the community calling it "SDXL 2.0" (KB panel). It's the default for local photorealism now. But the same panel names its real weakness plainly - weaker complex-prompt adherence than Flux 2. It won't reliably nail a dense multi-clause instruction the way a bigger model does.

That's precisely where a well-structured prompt earns its keep. A clearer, better-organized description - subject first, then setting, lighting, style, cleanly ordered instead of a run-on pile - gives a model with softer adherence its best shot at getting your intent. The prompt-engineering KB is the theory; this node is the shortcut, letting an LLM do the structuring for you.

How it works

Three inputs, no fuss:

• llm_service_connector (required) - from a Set*LLMServiceConnector node. The LLM that rewrites.
• prompt (required, multiline) - your raw idea.
• seed (required) - repeatability and variation. Lock it to reproduce a rewrite; change it to roll a new one.

The single output is zimage_prompt (STRING). Wire it into your Z-Image text-encode, or preview it with a ShowAnything node first.

Installing it

Install the pack. ComfyUI Manager → search ComfyUI-MieNodes (ComfyUI_MieNodes) → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/MieMieeeee/ComfyUI-MieNodes

then restart. No model download for the pack. Nodes appear under the sheep-emoji 🐑 MieNodes menu.

The API key step

This node needs an LLM behind it, so you need an API key on the connector - typed into api_token, or in mie_llm_keys.json (copy the .example in the plugin folder). ZhiPu and SiliconFlow offer free-tier models if you want to test without paying; grab a current free model id from their site.

Common issues

Run errors point at the connector. No key, wrong key, or a config_key aimed at an empty entry. Prove the connection with CheckLLMServiceConnectivity before suspecting this node.

Don't over-stuff the rewrite. Z-Image's soft complex-prompt adherence cuts both ways - a bloated, contradictory paragraph confuses it more than a lean one. If a heavy rewrite muddies the result, lower the LLM's ambition (a simpler input, or the general PromptGenerator in simple mode) and keep the prompt tight and ordered.

Mind the resolution ceiling. The KB notes Z-Image has a roughly 2MP native ceiling. A gorgeous prompt won't buy you a clean 4K render straight from the model - plan to upscale for large output rather than pushing native size past what it handles.