Krea2 风格语义条件

Krea's hosted style-reference feature - drag a reference image in, and the model borrows its look - is deliberately not part of the open Krea 2 weights. The company said so on the record. This node is a community attempt to rebuild that capability locally. It doesn't call any API, needs no key, and doesn't pretend to be an official adapter. It just feeds your reference image into Krea 2's own text encoder - which is Qwen3-VL, a vision-language model that can actually look at pictures. That's the entire trick, and it's why the node can be this small.

What it actually is

This is the text-to-image style-reference node. You hand it one image and a prompt, and it returns a standard CONDITIONING you plug straight into a KSampler's positive input. Think of it as "borrow the palette, material, brushwork and mood from this picture, but draw what I'm describing." It's the instant alternative to a style LoRA: no training set, no trigger word, per-image style guidance that costs a few extra seconds of encoding.

Two things it explicitly is not. It's not Krea's cloud style adapter, so don't expect a faithful recreation of the reference - read the README as "transfers visual language, not the subject." And it doesn't touch the pipeline beyond conditioning: no diffusion-weight edits, no latent token injection, no sampler changes. The source is small enough to read in one sitting, which is the best sign a node like this can give.

How it works

The mechanism is one line, plus housekeeping. The node builds a text prompt and calls clip.tokenize(prompt, images=[style_image]), then clip.encode_from_tokens_scheduled(tokens). Because Krea 2's encoder is Qwen3-VL - the VLM Krea picked precisely so the stack could later do image editing - passing the image as an argument makes the model see both your words and the reference in one conditioning pass.

What you don't see is the text prepended to your prompt. The node writes an English style instruction ("Use the attached reference image as a style guide. Transfer its visual language, color palette, material texture, brushwork, lighting, composition rhythm, and overall mood. Do not copy the depicted subject."), then appends Target prompt: .... That's where style_strength lives - it picks between three phrasings (轻微 / 平衡 / 强烈, i.e. subtle, balanced, strong). This is the trap: it's a wording dial, not a mathematical strength. Raising it doesn't scale a weight like CFG; it just tells the model in plainer terms to lean on the reference.

The inputs that matter

• CLIP - must be the krea2 text encoder (the loader set to type krea2), not a checkpoint's bundled CLIP. The node calls tokenize(..., images=[...]), which stock encoders can't do.
• 风格参考图 (style_image) - the reference. Pick something with clear style and minimal subject clutter; a busy subject tends to leak into the output.
• 正面提示词 (prompt) - describe the target subject and scene. The author's advice: don't re-describe the reference's subject, or the model will happily borrow it.
• 语义风格强度 (style_strength) - 轻微/平衡/强烈. Start at 轻微 or 平衡.
• 视觉编码分辨率 (vision_resolution) - default 384, range 128–1024. It's a total-pixel budget the reference is area-downscaled to before encoding. 384 is stable; 512 keeps more detail at the cost of speed and VRAM.
• 自定义风格指令 (custom_instruction) - optional. Fill it in and it replaces the preset wording (your prompt is still appended), useful when the built-in phrasing isn't pushing the direction you want.

Output is one 条件 (CONDITIONING) → KSampler's positive. The README's Turbo starting point: 8 steps, CFG 1.0, er_sde + simple, denoise 1.0.

Install

Easiest via ComfyUI Manager - search "Krea2 Style Refrence" and install - or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/DocWorkBox/ComfyUI_Krea2_Style_Refrence.git

Then restart ComfyUI. No requirements.txt, no Python deps (dependencies = [] in pyproject) - it uses only Comfy core and torch. The work is in the model files, which the pack doesn't bundle. You need all three, not one:

diffusion_models/krea2_turbo_fp8.safetensors
text_encoders/qwen3vl_4b_fp8_scaled.safetensors
vae/qwen_image_vae.safetensors

People routinely grab the diffusion model and skip the other two, then report the pack as broken. Missing the Qwen3-VL text encoder in particular means no image path at all.

Common issues

• Style too weak - before raising vision_resolution, try a more explicit custom_instruction. The author's own ordering: instruction first, resolution second.
• Reference subject creeps in - lower style_strength and make the prompt's subject unambiguous. The built-in instruction already says "do not copy the subject," but wording only goes so far.
• It's still Krea 2 - the open weights went through an alignment pass the hosted model didn't, so flattened expressions and dropped body/violence clauses still bite here. That's a model property, not a node bug.
• Expect a style guide, not a copy - you'll get the visual language, not the reference's subject. If you need the subject kept too, that's the pack's fusion node and an img2img denoise dial.