Krea2 Prompt Weight

If you've tried the old (word:1.4) emphasis syntax on Krea 2 and gotten nothing, you're not imagining it. Krea 2 reads your prompt through the Qwen3-VL text encoder, and the classic per-token weighting that grew up on CLIP models just doesn't route through an LLM encoder the same way. Krea2 Prompt Weight is kijai's fix: it makes (word:1.5) and (word:-1) actually do something on Krea 2 (aka K2), by scaling attention values rather than fiddling with token embeddings.

The node's own description spells out the mechanism and the one setting that will bite you: "Per-token prompt weighting for Krea2 (K2) via attention value scaling. Use (word:-1) to remove a concept, (word:1.5) to emphasize one -- works through the Qwen3-VL encoder where normal weighting doesn't. Outputs the patched model + conditioning; set the sampler CFG to 1.0."

Why this is needed

Krea 2 is one of the best open image models of 2026 - Krea's first fully in-house foundation model - and it deliberately uses Qwen3-VL with multilayer feature aggregation instead of a plain CLIP last-layer grab. That's great for prompt understanding and terrible for the muscle memory of anyone who spent years balancing prompts with parenthetical weights. The encoder doesn't expose weighting the old way. So kijai reaches into the attention instead: emphasize a concept by scaling up its attention values, remove one by scaling to negative. It's a genuinely different mechanism dressed in familiar syntax.

Inputs and outputs

You feed it your clip, your model, and your prompt text (with the weight syntax inline), plus a strength. It hands back two things:

• model - the patched model, which carries the attention-scaling behavior.
• conditioning - the encoded prompt, ready for your sampler.

You wire the model onward and the conditioning into your sampler's positive input. You use both outputs - that's the part people miss.

The one dial:

• strength (default 1, range 0–4) - global multiplier on the weighting effect. 1 respects your inline weights as written; push it up to make all your emphasis hit harder, down to soften it.

The setting that will get you: CFG 1.0

The description says it and it's not optional - set your sampler's CFG to 1.0. Krea 2's usable path is a CFG-1 model (Turbo, or Raw-plus-Turbo-LoRA), and this node is built for that regime. Run it at high CFG and you'll get bad results and blame the node. One related Krea 2 quirk worth carrying over from the wider community: on the Raw-plus-Turbo-LoRA path at CFG 1 you still want a negative prompt slot present even if it's empty, or output goes grainy - a ConditioningZeroOut won't substitute.

Installing it

Part of kijai's KJNodes pack. kijai, for what it's worth, had a hand in demonstrating Krea 2's zero-shot bounding-box prompting early on, so the K2-specific tooling here isn't a random guess.

• ComfyUI Manager - search KJNodes for ComfyUI, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r requirements.txt, and restart.

No extra downloads - but you do need a working Krea 2 setup, which means three files, not one: the model, the Qwen3-VL encoder (~8GB), and the Qwen-Image VAE (~250MB). People routinely grab only the model and then report everything as broken.

Common issues

It's in the experimental category, so treat it as a sharp tool. The failures are almost always one of: CFG not set to 1.0, only wiring one of the two outputs, or forgetting that Krea 2's baked-in safety filter can flatten faces and expressions no matter how you weight the prompt. The last one isn't the node's fault - it's a documented Krea 2 trait, and no amount of (smile:1.5) fully beats an alignment pass trained into the weights.