krea2 / comfy / raw / createImage

Krea 2 Raw is the checkpoint every LoRA trainer in the ecosystem now starts from: the undistilled Krea 2 base, no aesthetic finetune, meant to be trained on and tuned. It's also a 12B diffusion transformer that wants around 25GB in BF16 - or a decent chunk of VRAM in FP8. This node runs the exact same Raw model on Civitai's fleet, so you get the raw base without the download, the quantization decisions, or the text-encoder-and-VAE triathlon that trips people up locally.

Who's this for? The person who wants Krea 2's character and style knowledge in their workflow but doesn't have the GPU headroom, and doesn't want to debug why their local Krea 2 setup renders mush. It's text-to-image, not editing - feed it a prompt, get an image.

How it works

Standard civitai-comfy-nodes machinery: the node serializes your prompt (and any model/LoRA choices), submits a Krea 2 Raw job to the Civitai Orchestration API, long-polls until it succeeds, and hands back an IMAGE tensor. "Raw" in the discriminator means the hosted recipe targets the undistilled base, so you get the model's native look rather than the turbo-distilled one. Defaults reflect that: 20 steps at CFG 4.0, with the full sampler/scheduler menu if you want to tune it.

One honest caveat from the KB: the open weights you can download went through alignment the hosted model didn't. Civitai's orchestration is running Krea's hosted flavor here, so prompt adherence and expression should be better than the open checkpoint - but it's still a hosted model with hosted filters and hosted pricing.

The inputs that matter

• prompt - required, and Krea 2 rewards long, specific natural language. Characters and styles it knows render well without LoRAs.
• diffusion_model (CIVITAI_AIR) - optional. Empty means stock Krea 2 Raw. Wire a Civitai Model Selector to run a specific Krea 2 finetune from Civitai instead - this is the node's superpower and the reason the socket exists.
• loras (CIVITAI_LORAS) - optional; stack Krea 2 LoRAs via a Civitai LoRA Selector. This is the "Raw plus Turbo LoRA at 0.6" trick people converged on locally, doable here too.
• steps / cfg_scale - 20 / 4.0 defaults. If you drop a turbo LoRA in, you'll want to drop steps toward 8 and CFG to 1.
• sampler / scheduler / seed / quantity / width / height - standard.

Outputs

images → wire into Preview/Save Image. errors (JSON), workflow_id, and raw_json are diagnostics; raw_json is the full workflow result for cost and debugging.

Installing and authenticating

Install is the pack-wide recipe - there's no separate Krea 2 package:

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

or ComfyUI Manager → search "Civitai Comfy Nodes" → Install → restart. Dependencies: requests, python-socketio[client].

Authenticate via a Civitai Auth node, CIVITAI_API_TOKEN, a stored key/OAuth in ~/.civitai/, or browser login. Headless servers want the env var:

export CIVITAI_API_TOKEN=your_token

Where people get burned

• It's not free. You're renting Krea 2 Raw inference by the job, billed in Buzz. Fine for exploration, expensive as a bulk pipeline - budget accordingly.
• Raw needs steps. It's the undistilled base; at 4 steps it'll look unfinished. Leave it near the 20-step default unless you've added a turbo LoRA.
• The model you get is Civitai's hosted Krea 2, not a byte-for-byte copy of the HF checkpoint. That's mostly an upgrade, but seeds and results won't match your local run - don't expect reproducibility across the boundary.
• Your prompt and any reference images leave your machine. Standard API-node deal; don't send anything you wouldn't post on civitai.com.