Eric Krea2 Magic Prompt

Krea 2 wants long, specific, natural-language prompts - and most of us don't want to type them. Eric Krea2 Magic Prompt is the pack's answer: feed it a short idea and it expands it into a rich paragraph using a local LLM. No API key, no cloud call, nothing leaves your machine. It's the "I know what I want, I just don't want to write it out" button.

The prompt expansion is opinionated in the right way. The default system_prompt instructs the model to write the kind of description Krea 2 was actually trained on: one paragraph for subject and scene (who/what, pose, appearance, materials, backdrop), one for visual treatment (composition, framing, camera, lighting, palette, mood), no meta phrases like "in this image," text-to-render words kept in quotes, all under ~350 words. That last constraint matters because of Krea 2's fixed 512-token budget on the conditioning side - a sprawling expansion isn't necessarily better.

Inputs that matter:

• backend - auto (default) tries local Qwen3-VL first, then LM Studio, then Ollama. Or pin one: qwen3_vl (runs a local VLM from models/VLM), lm_studio, ollama, or endpoint (any OpenAI-compatible server).
• endpoint - the base URL. Defaults to http://localhost:1234/v1 (the LM Studio default). Ollama's :11434 is auto-handled, /v1 stripped.
• model - server model id, blank for auto-detect; for qwen3_vl it's a local folder path, blank to auto-detect.
• temperature (0.7), max_tokens (512) - expansion flavor and length.
• system_prompt - the instructions, editable. And system_prompt_path, which is the one worth knowing about: point it at Krea's own expansion.txt (the exact prompt template Krea uses) and you get the highest-fidelity expander, per the README.
• enabled - off = pass the prompt through untouched. Handy for A/B.
• device - qwen3_vl tier only; cuda:1 exists so a second GPU can hold the VLM while your Krea 2 pipeline keeps the main one. The tooltip's phrasing - "keeps the 6000 free for the Krea2 pipeline" - tells you exactly the rig this pack was tuned on.

Outputs: prompt (the expansion, on to Generate or Multi-Stage Ultra) and status (a string so you can see which backend actually answered).

The auto backend order is the pragmatic bit: if you've got a spare VLM, it uses it; if not, it leans on whatever OpenAI-compatible server you already run. People with an LM Studio or Ollama install get this for free - wire the prompt output of Eric Krea2 Prompt into this node's input and you have a structured-to-expanded prompt chain with zero external dependencies. That's the setup the example workflow uses, and it's a good one.

How to install

Part of EricRollei/Krea2_ComfyUI_Advanced. ComfyUI Manager → search "Krea2_ComfyUI_Advanced", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Krea2_ComfyUI_Advanced

Restart. The node itself talks to a local server you supply - install the pack, and the only extra you may want is Krea's expansion.txt for system_prompt_path.

Common issues

• "No backend available" / falls through to passthrough - auto needs at least one of the three tiers running. Start LM Studio or Ollama first, or set backend explicitly.
• Local Qwen3-VL eats your VRAM - use device: cuda:1 if you have a second GPU, or just let auto skip to LM Studio/Ollama.
• Expansions ignore your subject - the system prompt is editable; if your local model drifts, tighten it or lower temperature. The status output tells you which backend produced the text so you can pin it.
• Long output truncated - raise max_tokens, and remember Krea 2's fixed token budget is a real ceiling on how much conditioning is useful anyway.