Replicate (krea/krea-2-medium)

This is the sibling node of Replicate (krea/krea-2-large) with exactly one difference: it tells Replicate to run krea/krea-2-medium instead of the large tier. Same pack, same generated wrapper, same inputs, same single IMAGE output - the packed schemas for the two are byte-for-byte identical. If you've read the large article, you've read this one's mechanism; here's the short version and the one fact worth knowing before you pick a tier.

That fact comes straight from Krea: their head of research identified the open-weights release as the medium version of the hosted product. So if you've been running Krea 2 Raw or Turbo locally, this node hits the closest hosted sibling of what you already know - the same model family, but without the alignment pass the open checkpoint went through, and with Krea's in-house prompt expansion applied. Add the moodboard and style-reference inputs, which are hosted-only features Krea never released for local weights, and you can see why someone with no VRAM budget (or no patience for a 12B download plus the Qwen3-VL encoder and Qwen-Image VAE) would route here instead.

Large vs. medium: don't benchmark one against the other and expect a match - they're different hosted tiers, and "large" is the heavier sibling. Which one you want depends on whether the extra spend buys you anything you can see in your own subject matter; try medium first, since it's the tier whose behavior the community has actually been able to compare against the open weights.

What you set

• prompt - long, specific natural language; quote any words you want rendered as text.
• creativity - raw (literal), low, medium, high (creative liberty). Default medium.
• style_reference_images + style_reference_strength (0–1, default 0.5) - up to 10 style-transfer reference images.
• moodboard_id + moodboard_strength (default 0.35) - a Krea moodboard UUID from krea.ai/moodboards, grabbed from the share URL's ?share=<uuid>.
• seed - no default on the node, so set it yourself for reproducible runs.

Output is one output IMAGE tensor - straight into Save Image or anything else downstream.

Installing

ComfyUI Manager → search "Dados Nodes", or:

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

Restart, then set your REPLICATE_API_TOKEN environment variable (with a billing-enabled Replicate account). No model files to download - the weights live on Replicate's side, and every generation is pay-per-image with your prompt and reference images leaving your machine. If a run fails, the model's logs print to the ComfyUI console while the node polls - that's the first place to look. And since this wrapper calls the model unpinned (no version hash), Krea can update it underneath you; outputs can drift between days, which is the price of always running the current hosted model.

The honest take: if you want Krea 2's hosted behavior in your graph without running 12B locally, medium is the sensible default - it's the tier the open community actually knows, minus the alignment damage and plus the features Krea only sells hosted.