Krea 2 Medium Turbo · Text to Image

The one you'll actually use most

About $0.015 an image - roughly a quarter of Large, half of Medium - makes Krea 2 Medium Turbo the tier for thinking out loud. You want to see eight versions of a composition before you commit, you want a quick style test, you're generating placeholder art for a layout at 2am. At this price the "just run it again" reflex, which is a bad habit everywhere else on this pack, is basically fine.

It's Krea 2 underneath - Krea's own 12B diffusion transformer from June 2026, Qwen3-VL reading the prompt, Qwen-Image's VAE chosen deliberately because Krea wanted a model good at non-photographic work. Turbo means what it means locally: fewer steps, faster answers, slightly less polish. The KB's local verdict on the Krea line is worth carrying over here: the reason Krea 2 took the open ecosystem off Ideogram 4 wasn't raw rendering quality, it was what you could build with it. On a hosted endpoint you're renting exactly that knowledge - the character and style recall, the pop-culture literacy - without any of the ecosystem work.

Inputs

• prompt - up to 5,000 characters, and Turbo still rewards the long-form natural-language prompt Krea 2 is known for. Same disambiguation trick as the rest of the line: name the year or the variant, or the model will happily generate the wrong thing it knows more about.
• aspect_ratio - 1:1 by default, eight shapes including 2.35:1 anamorphic. That's two more shapes than the Medium tier gets, which is a slightly odd bit of tiering but a nice bonus if you're framing for widescreen.
• reference_image_urls - optional, up to ten style references, an Autogrow group so you add sockets by connecting images. Style steering, not casting: these set the look rather than reproducing a face. Sockets are named reference_image_0 upward in the wiring, and since the group label isn't drawn on the node, count them yourself if your prompt refers to order.

Outputs are images (a normal IMAGE batch - Save Image, a detailer, an upscaler, whatever) and task id, which names the exact SpicyAPI request for the console or for support.

Note what isn't here: no seed. Nothing on this node reproduces a previous result, so if you're testing two prompt variants, test them side by side on a fresh run rather than trying to hold everything else constant.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI, add your key under Settings > SpicyAPI (a key from spicyapi.ai/console/keys), then restart once more so the node list is rebuilt from your account instead of the bundled snapshot. Servers can set SPICY_API_KEY. No pip install, no model file, no GPU. ComfyUI Manager: search SpicyAPI by pack title. Tested on ComfyUI 0.39 with the V3 node API; Python 3.10+.

Things to know before you lean on it

Cheap is not free. A template workflow that fires twenty Turbo images per queue press is thirty cents - fine, until it runs while you're asleep. Set Max cost per run in the settings panel and it'll refuse anything quoted above your limit before a charge lands.

Style references are uploads, and uploads have a ceiling. Images are re-encoded to PNG, fall back to JPEG if the PNG exceeds 10 MiB, and then fail outright with a "scale it down first" message. Ten references is also ten uploads the first time each one runs; they're cached by content hash per account for twelve hours afterwards.

Turbo is the sketch tier, so don't judge the model by it. If a composition is right but the fine detail doesn't hold up, that's what Large is for. Run the good idea again on the expensive tier rather than re-prompting Turbo repeatedly and wondering why it never resolves.

Cancelling stops your wait, not the job. The task keeps running on SpicyAPI, is charged if it succeeds, and its output stays in your console for up to 14 days. Hitting cancel and immediately queueing again is a two-charge move.

Refusals are the endpoint's, not the node's. Krea's hosted tiers didn't receive the alignment pass that made the open weights drop whole categories of prompt, so this is the less-filtered version of the model - but an API node still can't bypass whatever the provider's moderation decides.