Nodes/ComfyUI/Krea 2 Image
ComfyUI Node Runs on cloud

Krea 2 Image

Krea 2 in one node — no 15GB download, and the style-chaining the open release kept

By Comfy-Org·Created 4 years ago·Updated about 13 hours ago· 130,663
Krea 2 Image
    • IMAGE
    prompt
    model
    seed0

    Krea 2 is open-weights now - that much was a big deal when it landed in June 2026, and it's genuinely one of the most aesthetic image models around. But "open weights" doesn't mean "free to run." The full model is a big download and a bigger VRAM bill, and there's a specific part of Krea's product - the moodboard and the chained style references - that stayed behind the hosted API even after the weights went out.

    The Krea 2 Image node is the shortcut for both problems. One node, no weights on your disk, and the style-reference pipeline that didn't ship with the open checkpoint.

    What it actually is

    Another core partner node: built into ComfyUI, no install, cloud-backed. Sign in with your ComfyUI account, buy credits, and the node calls Krea's API through Comfy's proxy. It's synchronous - submit the request, poll until the job finishes, get an IMAGE tensor back into your local graph. The model runs on Krea's servers, so your GPU's only job is being a ComfyUI host.

    The node landed right around Krea 2's release (mid-2026), which is typical for these partner nodes - the Comfy team wires them in as the models launch.

    Picking a model

    The model dropdown is where Krea's three variants live:

    • Krea 2 Medium - expressive illustrations. The stylized, painterly end of Krea's range.
    • Krea 2 Medium Turbo - the same medium line, distilled and cheaper per call. Good for iterating.
    • Krea 2 Large - expressive photorealism. This is the one for convincing, film-like output.

    Each variant carries sub-options you set under the same widget: aspect_ratio (1:1, 16:9, 4:3, 2.35:1 and friends), resolution (1K), and creativity (raw / low / medium / high). Raw sticks closest to your prompt text; high lets the model riff. Start at medium and move down if you feel it's ignoring you.

    The one tooltip worth memorizing: Medium is the illustration call, Large is the photoreal call. It's an aesthetic split, not a quality ranking.

    The part local weights can't give you

    Two optional inputs hide under the model widget and they're the real reason to use the hosted node:

    • moodboard_id - a UUID you copy from the Krea website. It steers the whole image toward that moodboard's look, with a moodboard_strength dial (default 0.35) to blend it in.
    • style_reference - a chain of up to 10 image references wired in from the Krea 2 Style Reference nodes. This is the style-consistency system Krea kept proprietary; it's only available through the API, which is exactly what this node reaches.

    That's the headline: you can run Krea 2 locally all day and you still won't have the style-chaining. This node does.

    Output, and the fine print

    One IMAGE output, wired to a Save Image node like anything else. The seed (0–2,147,483,647) gives you reproducibility within a run.

    Gotchas: the moodboard ID must be a real UUID copied from the Krea site - the node validates the format and rejects garbage. Max 10 style references, and each one adds to the per-call cost, so a moodboard-plus-five-refs run is visibly pricier than a bare text generation. Check the price badge before you queue something elaborate.

    If you want Krea's taste without committing 20GB of disk and a GPU upgrade to it - or you need the style references - this is the node. If you're running Krea 2 locally already, you're mostly paying for the convenience of not loading it.

    Categorypartner/image/Krea

    Inputs (3)

    NameTypeDefaultDescription
    promptSTRINGText prompt for the image.
    modelCOMBOKrea 2 Medium is best for expressive illustrations; Krea 2 Large is best for expressive photorealism.
    seedINT00–2147483647Random seed for reproducibility.

    Outputs (1)

    NameTypeDescription
    IMAGEIMAGE