Krea2 Prompt Generator 🐑

Krea2PromptGenerator takes a rough idea and has a large language model rewrite it into a rich, detailed prompt tuned for Krea 2. You type "a woman on a rooftop at dusk," it hands back a paragraph with lighting, composition, mood, and detail worked in - the kind of prompt Krea 2 actually rewards.

It's one of the model-specific prompt nodes in ComfyUI-MieNodes, MieMieeeee's utility pack. The key thing to understand up front: it doesn't run Krea 2, and it doesn't run an LLM by itself. It's a front-end that calls out to whatever LLM service you connect, using a system prompt shaped for Krea 2's preferences.

Why Krea 2 specifically wants this

Krea 2 is a 12B diffusion transformer that Krea trained from scratch and opened in mid-2026 - notable for being a genuinely undistilled base at a moment when nobody else was shipping one, which is why finetunes and LoRAs piled onto it within days (KB panel). Like the other 2026-era models, it's an LLM-encoded model: it responds to descriptive natural language, not a pile of comma-separated booru tags. Longer, well-structured prompts land better than terse ones. That's the gap this node fills - turning a one-liner into the sort of full description the model was built to read. The general prompt-engineering lesson applies: on these models, saying more (usefully) beats saying less.

How it works

You wire an LLM connector into it, give it your idea, and it returns the expanded prompt.

• llm_service_connector (required) - comes from a Set*LLMServiceConnector node (SiliconFlow, ZhiPu, DeepSeek, Gemini, MiniMax, or any OpenAI-compatible endpoint). This is the LLM that does the rewriting. No connector, no output.
• user_prompt (required, multiline) - your raw idea.
• seed (required) - controls variation. Same seed and inputs give a repeatable rewrite; change it to roll a different expansion of the same idea.

Optional knobs worth knowing:

• aspect_ratio (default 1:1) - a hint fed into the generation so the described composition suits the frame you're actually rendering.
• temperature (default 0.7) - creativity dial. Lower sticks closer to your input; higher wanders.
• max_tokens (default 4096) and timeout (30/60/120/300s) - headroom and patience for the LLM call.

The single output is krea2_prompt (STRING) - pipe it into your CLIP/text-encode node, or into a ShowAnything node first to read what the LLM produced before you commit a render to it.

Installing it

Install the pack via 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 itself. Nodes appear under the sheep-emoji 🐑 MieNodes menu.

The setup step people miss: the API key

This node is useless without an LLM behind it, and that means an API key. You configure it on the connector node - either type it into the connector's api_token, or (the tidier way) copy the pack's mie_llm_keys.json.example to mie_llm_keys.json in the plugin folder and fill in the service you use. Several providers, notably ZhiPu and SiliconFlow, have free-tier models, so you can try this without spending anything - check the provider's site for what's currently free and put that model id into the connector.

Common issues

Nothing comes out / it errors on run. Almost always the connector: no key, a wrong key, or the wrong config_key pointing at an empty entry in your JSON. Test the connection in isolation with CheckLLMServiceConnectivity before blaming this node.

The prompt is great but the image is wrong for the frame. Set aspect_ratio to match what you're actually rendering - the rewrite composes for the ratio you give it.

It's for Krea 2. Point the output at a different model family and results get muddier. If you're on Z-Image or LTX-2 or want a model-agnostic rewrite, use the matching ZImagePromptGenerator, LTX2PromptGenerator, or the general PromptGenerator instead.