Krea 2 Raw

Krea 2 Raw is the unusual one in the family: an undistilled base model, CFG-guided, with no aesthetic finetune applied at all. The open-weights world treats it as the training base - the thing you build LoRAs and finetunes on rather than the thing you generate with directly. This node is the same Raw checkpoint, but hosted on Runware, and the interface reflects the difference. Where the Turbo node gives you a creativity slider and aspect presets, Raw hands you actual sampler controls: steps, CFG, and an acceleration level. If you've ever wanted to drive a modern base model rather than accept its default temperament, this is the node.

Why would you want that? Because Raw's relationship to the prompt is the honest one. Turbo was distilled to 8 steps with guidance disabled; Raw keeps classic CFG, which means your negative prompt actually does something, and you can trade steps for fidelity in a way a distilled model won't reward. The community's own tip for local Krea 2 is literally "train on Raw, apply on Turbo" - Raw is where the controllable, flexible behaviour lives.

The inputs that matter

• positivePrompt (required) and negativePrompt - both work here, because Raw is CFG-guided. Long, specific natural language beats short tags.
• width / height (required) - 64 to 2048, step 16, default 1024. You set dimensions directly instead of picking a preset.
• steps - default 20, up to 100. Note the open-weights reference recipe for Raw is 52 steps at CFG 3.5; the node defaults lean faster, so if you want the full-fidelity Raw character, crank steps up.
• CFGScale - default 4.5. Higher hugs the prompt harder; the reference recipe sits around 3.5.
• acceleration - none, low, medium, high: the optimization level, i.e. how aggressively Runware squeezes inference. Lower is more faithful, higher is cheaper/faster.
• lora - a RUNWARE_LORA socket. This is the "training base" payoff: wire a Runware LoRA builder in and you're doing the Krea 2 training workflow on the API, with the LoRA applied during generation.
• seed / numberResults - reproducibility, up to 20 variations.
• outputFormat / outputQuality - JPG/PNG/WEBP, quality 20–99.

Output: image (IMAGE), into Preview or Save.

How it fits a workflow

Two honest use cases. One: you want maximum control over a Krea 2 render - you have a negative prompt, you care about steps, you're not in a hurry. Two: you're testing or using a Krea 2 LoRA, and Raw is the base it was trained against, so it's the correct host. For quick bulk generation, the Turbo node is the better tool; Raw is for when you want to be in the driver's seat.

One thing to keep in mind: on the API you're getting the hosted Raw, which - like the whole hosted line - didn't get the alignment pass the downloadable weights received. If you've heard "Krea 2 open weights flatten expressions," that complaint is about the file on your disk, not necessarily this endpoint.

Installing it

Standard pack install - this is part of Runware/ComfyUI-Runware. ComfyUI Manager → search Runware → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

No model downloads, light deps (runware-sdk, pillow, soundfile). You need a Runware API key (Settings → Runware API key, RUNWARE_API_KEY, or runware auth login).

Gotchas

Raw is slower per image than Turbo - more steps and CFG both cost compute, and the title bar will show you exactly how much. Also, if your negative prompt seems ignored, double-check you didn't drop CFGScale to 1 or below; with guidance disabled the negative is ignored by design.