EvoLink Krea 2 Turbo (Official)

Fair question. Krea 2 Turbo shipped open weights in 2026 - a 12B diffusion transformer you can genuinely run yourself, no API required. So why would you put a Krea 2 node in your graph that phones home? Because the thing Krea hosts and the thing it released aren't the same model. The open weights went through an alignment pass the hosted version never got - Krea admitted this on the record. The API version is the one Krea actually sells, filter and all. If you want that Krea 2, the cloud is the only door, and this node is the doorbell.

That, plus the usual reasons: no GPU, no 20GB download, and Turbo is fast enough that the round-trip barely registers.

What it does

EvoLink Krea 2 Turbo (Official) is text-to-image only - this node has no image inputs, so don't come here for Krea's strong img2img. You give it a prompt and it returns a photorealistic-style render. It's pitched as fast and it is; EvoLink's default 300-second timeout rarely gets stressed.

The knobs:

• prompt (required) - the scene description. Krea 2's photorealism is the reason to be here, so write for that.
• size - 1:1 by default, with the full ratio menu including 1:2 / 2:1.
• quality - 1K or 2K. Default is 1K, which is fine for drafts.
• seed - 0 is random; set a value to reproduce a result. Handy for the "same scene, small change" iteration loop.
• nsfw_check - off by default. Turning it on runs content detection on the request. Don't read that as "off = unmoderated"; the provider still enforces its own policy, and moderation-blocked tasks just fail without charging.

Outputs are the standard EvoLink set: image (downloaded to a tensor, wire it into Save Image), task_id, status_info (model, elapsed time, credits used), result_urls (24-hour expiry), and response_json.

Install and key

It's the shared EvoLink routine. ComfyUI Manager → search EvoLink → install → full restart (close the console window, don't just refresh the tab). Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/deeplearning-goethe/comfyui-evolink

Only dependency is requests; no model files, no VRAM. Get a key at evolink.ai/dashboard/keys (sk- prefix), paste it into api_key once - it persists to a local config, so the box can stay empty after, and it never rides into a shared workflow JSON. Chinese prompts work, and error messages come back in Chinese, which is either a feature or a thing to know depending on your 中文.

When to reach for it, and when not to

Reach for it when you want the hosted Krea 2 (the unaligned one), when your prompt is photorealism-heavy, or when you need a fast look without spending an afternoon on a 12B local setup. Skip it if you're going to do heavy iteration - metered per call, and drafts add up - or if you want Krea's editing abilities, which this node doesn't expose. For that you'd want the open weights locally, or a different EvoLink node entirely. And the usual API-node reminder: your prompt leaves the machine. For a model with open weights that you could run yourself, that's a real cost to weigh, not a default.