Krea 2 Medium Turbo

Krea 2 is the 12B diffusion transformer that took the open ecosystem away from Ideogram 4 in a week, and this node is the hosted version of it - specifically the medium tier with the turbo distillation, running on Runware's cloud. The part people don't expect: the hosted Krea 2 is not the same model you can download. Krea's own head of research confirmed the open weights went through an alignment pass the API version never got, and the API adds in-house prompt expansion on top. Net effect for you: the version behind this node follows your prompts better than the weights on your disk. That's a genuinely useful thing to know before you blame the model for a flat expression or a dropped prompt clause.

Turbo is the fast, guidance-distilled variant of Krea 2 - the kind of thing you reach for when you want Krea 2's look without babysitting a 50-step run. On Runware that speed is the point: it's a billed API, so the fewer steps, the cheaper and quicker each image is. It's the "I need a good image now, not a thesis" choice.

The inputs that matter

• positivePrompt (required) - Krea 2 wants long, specific natural language. Describe the image as you would to a person; quote words you want rendered as text. There's no negative-prompt slot on this variant - Krea's guidance-based models lean on settings.creativity instead.
• settings.creativity - raw, low, medium (default), or high. This is your prompt-adherence dial: lower follows the prompt closely, higher lets the model improvise. This is the single most useful knob on the node.
• referenceImages - an IMAGE socket for style/reference images, the hosted equivalent of Krea's style-reference features.
• size - the hosted model exposes aspect presets rather than freeform dimensions: 1K (1:1), 1K (4:3), 1K (16:9), portrait options, even 1K (2.35:1) for ultrawide. All 1K; there's no bigger bucket here.
• seed / numberResults - reproducibility and up to 20 variations per run.
• outputFormat / outputQuality - JPG/PNG/WEBP, quality 20–99.
• advanced_json - the escape hatch for settings.moodboards, Krea's moodboard conditioning, which isn't exposed as widgets.

Output is a native image (IMAGE), ready to wire into Preview or Save.

How it fits a workflow

If you're generating many images fast - concept exploration, moodboards, "which of these eight directions works" - this is the node to reach for. It's Krea 2's aesthetic with turbo latency and API pricing. For prompt-sensitive final renders, consider that the Krea 2 open weights are the training base and behave differently; the hosted turbo is the consumer-facing tier and is tuned accordingly.

Installing it

Part of the Runware/ComfyUI-Runware pack. 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 (it's all remote), light deps (runware-sdk, pillow, soundfile). You need a Runware API key: Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Gotchas

The usual API things: it's paid per image, and the cost shows on the node title bar - the "turbo" keeps it low, but numberResults multiplies it. And one mental-model fix: because this is the hosted Krea 2, don't judge the open-weights model by what this node produces. They differ on purpose, and in this direction the hosted one usually wins on prompt adherence.