Krea2 System Prompt

Tiny node, one job - and it fixes the single most confusing thing about the Krea 2 reference-image workflow. If you've plugged a picture into Text Encode (Krea2) and found that the model describes your reference beautifully but flatly refuses to change it - you say "give her blonde hair," the output keeps the reference's brown hair - this is the node you were missing.

Here's the why. Krea 2's text encoder is Qwen3-VL, and LLM/VLM encoders don't read your prompt as a bag of tags; they wrap it in a chat template and read it like an instruction. That chat turn has a system message, and by default Text Encode (Krea2) fills it with Krea's trained descriptor - an instruction that basically says "describe this image." Great for faithfulness, useless for editing: the model dutifully describes the reference and treats your prompt as a separate thing sitting next to it. This node swaps that system message for an instruct-style one that tells the model to describe the reference and then explain how your instruction should alter it and generate accordingly - the Qwen-Image-Edit way of thinking, where the picture and the words combine.

How it works

It's a text source, nothing more. It holds a STRING preloaded with the instruct prompt and outputs it, so you can wire it straight into the system_prompt input on Text Encode (Krea2). The encoder does the actual chat-template scaffolding - you just supply the instruction text. The preloaded default is:

Describe the key features of the reference image (color, shape, size, texture, objects, background), then explain how the user's instruction should combine with or alter it, and generate a new image meeting the instruction while staying consistent with the reference where appropriate:

You're free to rewrite it. It's a plain editable text box; the shipped wording is just a sensible starting point that someone already tuned for you.

Inputs and output

There's really only one knob:

• text - the multiline system instruction, preloaded with the text above. Edit it, or leave it.
• Output system_prompt (STRING) - drag this into the system_prompt input on your Text Encode (Krea2) node. That's the whole hookup.

It does nothing on its own. Drop it, wire its output into the encoder, done.

The honest caveat

This is experimental and out-of-distribution, and it's worth saying plainly. Krea 2 was trained on that fixed descriptor, so feeding it a different system instruction pushes the model off the path it knows - results can drift, and prompt adherence or quality may take a hit. That's the same recurring theme with Krea 2 in general: the open weights went through an alignment pass the hosted version never got, and every workaround for its quirks tends to cost you a little something. So treat this as a tool to reach for when the default descriptor is too faithful and you genuinely need the prompt to reshape the reference - and A/B it against just leaving system_prompt unconnected. If the reference still dominates, get more explicit in the instruction text about what should change versus what should stay.

Install

It ships in the same pack as Text Encode (Krea2), so if you've got one you've got both. If not - ComfyUI Manager, search Text Encode (Krea2) and install, or manually:

cd ComfyUI/custom_nodes && git clone https://github.com/ethanfel/ComfyUI-Krea2TextEncoder

Restart ComfyUI afterward. No models of its own to download - it's just text.