krea2 / comfy / base / createImage

Krea 2 was the open-model event of mid-2026: a 12B diffusion transformer Krea trained from scratch, shipped as a genuinely undistilled base checkpoint at a moment when nobody else was shipping one, and adopted within days. But there's a wrinkle the model card glosses over - the open weights went through an alignment pass that the API version never did. People who downloaded Krea 2 found flattened facial expressions, dropped prompt clauses about bodies and violence, refusal craters in a supposedly-open model. The hosted Krea 2 that Civitai's orchestrator calls has none of that. So this node, CivitaiImageGenComfyKrea2BaseCreateImage, isn't just "Krea 2 in the cloud." It's the version of Krea 2 the community actually wanted, with the download and the VRAM removed.

Mechanically it's the same story as every node in the pack: your prompt goes to Civitai's orchestration API, the job runs on their fleet, and the result comes back as an IMAGE tensor, billed in Buzz. This is the "base" variant of the comfy ecosystem - the full sampler, 20 steps at cfg 4.0 by default - as opposed to the turbo variant in the same menu, which is 8 steps at cfg 1.0. If you want Krea 2's full range, this is the node.

The inputs that matter

prompt is the only required input. The ones you'll actually set:

• negative_prompt - works, and unlike some newer models it's genuinely useful with Krea 2.
• sampler / scheduler - the full ComfyUI sampler menu (euler, euler_ancestral, heun, dpm_2, lms, and two dozen more) with the scheduler list (normal, karras, exponential, sgm_uniform…). Leave them empty and the platform picks; set them if you have a sampler you trust for a specific style.
• steps (20) / cfg_scale (4) - the base-model defaults. Bump steps toward 30–40 for polish; treat cfg 4 as a starting point, not a law.
• width / height - 1024×1024 default, up to 2048. Krea 2 handles the big end fine.
• quantity - up to 12 per call. This is the batch knob; 12 images in one billed job is a lot of Krea for the money.
• diffusion_model / loras - CIVITAI_AIR and CIVITAI_LORAS sockets. Wire a Civitai Model Selector to swap checkpoints or a Civitai LoRA Selector to stack LoRAs. There's no dedicated checkpoint input for Krea - the base model is built in.
• seed, output_format (jpeg/png/webP), image_metadata - standard.

Outputs: images (IMAGE), errors, workflow_id, raw_json. The node shows the workflow id and per-job cost inline - with quantity up to 12, watch that number.

Install and auth

One install covers the whole pack. ComfyUI Manager → search Civitai Comfy Nodes → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt   # just `requests`

Then authenticate with a Civitai account that has Buzz: export CIVITAI_API_TOKEN=..., wire a Civitai Auth node, or sign in via the Civitai sidebar.

Where people get burned

The main gotcha is picking the wrong node. There are three Krea 2 creation paths in this pack: this comfy/base one, the comfy/turbo one, and the fal one with a totally different input shape (style references, aspect ratio, creativity). If you're looking for the classic text-to-image Krea experience with LoRA support, this is it. Also: this node calls the API version of Krea 2 - the less-aligned, less-flattened one - which is a feature for most people and a compliance consideration if you're building for a client. And remember the pack is early preview; Civitai warns nodes "may change without notice," so pin your installed version if this workflow matters.