Krea 2 Large · Text to Image

What you're buying

Krea 2 is the model that took the open image ecosystem off Ideogram 4 in under a week in June 2026, mostly on licence terms and trainability rather than outright render quality. The three "Krea 2" nodes in this pack don't run the open checkpoints - they're SpicyAPI's hosted tiers of Krea's own image line, and Large is the top of it: the tier you use for a finished frame rather than a draft. It's Krea 2 on SpicyAPI's hosted tiers, roughly $0.06 an image, and what separates it from the cheaper tiers below it is the per-image quality you're buying plus the frame shapes on offer - eight, including 2.35:1 anamorphic - and the reference input.

Worth having straight in your head before you spend anything: what Krea ships as open weights is not what Krea sells through an API. Their head of research said the open checkpoint went through an alignment pass the hosted version didn't, and the visible symptom is a model that quietly ignores parts of prompts near "unsafe" content - craters, not deformities. Whatever you think of that, it cuts in favour of a hosted endpoint: you're getting the version that wasn't lobotomised. Against it, the endpoint still runs the provider's own moderation, so you can't push past a refusal from inside the graph. That's the standing rule for every API node, not a quirk of this one.

The inputs

Three fields. That's the entire node.

• prompt - up to 5,000 characters, and Krea 2 wants them used. Natural language, described the way you'd describe a picture to a person, wins over keyword soup. The trick worth stealing from the local Krea community is disambiguation: the model has deep pop-culture knowledge and will hand you the wrong version of a thing if you don't narrow it. Name the series, the year, the variant. Specificity is prompt adherence here.
• aspect_ratio - 1:1 by default, across eight shapes that include 2.35:1 anamorphic. If you're building frames for a sequence, that plus the reference input is the reason to pay Large prices rather than Medium ones.
• reference_image_urls - optional, and it's an Autogrow group: you get one socket, and a new one appears each time you connect something. These steer the style of the result rather than acting as a literal cast, so think mood board, not casting sheet. Sockets are named reference_image_0, reference_image_1 and so on in the wiring; the group label isn't drawn on the node, so count your connections.

Outputs are images - an ordinary IMAGE batch, ready for Save Image, a local upscaler, or a VAE round-trip - and task id, a string identifying the exact request for the console or for support. Keep the task id when a result looks wrong; it's the fastest way to get an answer.

Install

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

Restart ComfyUI, then Settings > SpicyAPI and paste a key from spicyapi.ai/console/keys; restart once more so the node list is built from your account. On a server, set SPICY_API_KEY in the environment instead - the key lives in the ComfyUI user folder, never in the workflow or in a saved image's metadata, which is the one security question people ask about API nodes and the answer is genuinely good here. ComfyUI Manager users can search the pack title, SpicyAPI. No dependencies, no checkpoints, current ComfyUI (V3 node API, tested on 0.39), Python 3.10+.

Honest notes

Six cents an image is cheap until you're iterating. There's no seed on this node, so nothing is reproducible - you can't lock the batch you liked and vary one word. At $0.06 a roll, the sensible shape is a handful of Large frames per project and the rest of your exploration on the cheaper tiers.

Know what Krea is bad at. The line chose Qwen-Image's VAE over Flux's deliberately, for non-realistic images, and the CEO's own comment was that for photorealism he'd use the Flux VAE. Stylized, illustrative and character-heavy work is where this line earns its keep; if your goal is a photograph, spend your money elsewhere.

Reference images have a 10 MiB upload ceiling - the node re-encodes to PNG and falls back to high-quality JPEG, and only then fails with a "scale it down first" error. Big scanner files get rejected.

Cancelling stops the wait, not the job. The request keeps running on SpicyAPI, is charged if it succeeds, and stays in your console for up to 14 days. Set Max cost per run in the settings panel if you don't want surprises; anything quoted above it stops before a charge.