Eric Krea2 Prompt

Krea 2 is a natural-language model. Tag soup and JSON-style prompts underperform; long, specific, comma-joined descriptions of materials, light, and camera win. Eric Krea2 Prompt exists to make writing that kind of prompt painless: instead of composing a paragraph by hand, you fill in labeled fields and the node assembles a dense, comma-separated description - and it automatically quotes anything you want rendered as literal text in the image.

The one required input is subject, which is the honest design choice here. Everything else is optional structure: pose_action, appearance_clothing, props_materials, scene, composition, camera, lighting, color, mood, style_medium, and extra_details. Fill in the ones you care about, leave the rest empty, and the node stitches them into a single flowing prompt. text_to_render is the nice one - type the words you want baked into the picture (a street sign, a book title) and it wraps them in quotes automatically, which is exactly how Krea's own prompting guide says to do text.

The value here is twofold. First, consistency: the field order nudges you toward the prompt shape Krea 2 was trained on - subject, then action and appearance, then environment, then the visual treatment. You get a longer, denser prompt for free, and the ecosystem's findings are unambiguous that long specific prompts beat short ones on this model. Second, it pairs naturally with Magic Prompt: put this node's output into Magic Prompt's prompt input and the LLM expands it into a full paragraph. Or skip Magic Prompt entirely and wire prompt straight into Generate or Multi-Stage Ultra.

It's a single-output node (prompt), and it's about as opinionated as a text-composer gets - it won't write your prompt for you, it just arranges what you give it in the order that works. If you're the type who types prompts directly into Generate, this node is optional. If you've ever stared at a blank prompt box, or you're building a workflow where a non-expert should be able to fill in boxes and get a good image, it's quietly one of the most useful utilities in the pack.

A couple of field-by-field notes from Krea 2's known behavior: be specific with named things - "Teen Titans (2003)" gets the good version, "Teen Titans" gets the cheap one, so extra_details is the place to disambiguate series, years, and character versions. camera and lighting genuinely move the needle on this model because the training data is full of them; a sentence about a 85mm lens and golden hour beats a generic "high quality" tag. And the model has craters around violence and body-specific descriptions thanks to its alignment pass, so if a prompt "isn't working," that's often the model declining, not the phrasing - this node can't fix that.

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. It's pure text assembly, so it has no model dependencies - but the whole pack needs the shared setup (git diffusers with Krea2Pipeline, transformers>=4.57.0) before any of its sibling nodes work.

Common issues

• Prompt comes out as a wall of text - that's the point; Krea 2 likes dense comma-joined prose. It's a feature, not a formatting bug.
• Text you wanted in the image is missing - put it in text_to_render, not in extra_details, so it gets quoted.
• Same prompt, wildly different adherence between runs - that's the model's known alignment inconsistency, not the node. Disambiguate names and rephrase; don't fight the builder.